Classification of Breast Cancer Histopathological
Classification of Breast Cancer Histopathological
Research Article
Classification of Breast Cancer Histopathological Images Using
DenseNet and Transfer Learning
Received 9 May 2022; Revised 19 June 2022; Accepted 30 July 2022; Published 10 October 2022
Copyright © 2022 Musa Adamu Wakili et al. *is is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
Breast cancer is one of the most common invading cancers in women. Analyzing breast cancer is nontrivial and may lead to
disagreements among experts. Although deep learning methods achieved an excellent performance in classification tasks in-
cluding breast cancer histopathological images, the existing state-of-the-art methods are computationally expensive and may
overfit due to extracting features from in-distribution images. In this paper, our contribution is mainly twofold. First, we perform
a short survey on deep-learning-based models for classifying histopathological images to investigate the most popular and
optimized training-testing ratios. Our findings reveal that the most popular training-testing ratio for histopathological image
classification is 70%: 30%, whereas the best performance (e.g., accuracy) is achieved by using the training-testing ratio of 80%: 20%
on an identical dataset. Second, we propose a method named DenTnet to classify breast cancer histopathological images chiefly.
DenTnet utilizes the principle of transfer learning to solve the problem of extracting features from the same distribution using
DenseNet as a backbone model. *e proposed DenTnet method is shown to be superior in comparison to a number of leading
deep learning methods in terms of detection accuracy (up to 99.28% on BreaKHis dataset deeming training-testing ratio of 80%:
20%) with good generalization ability and computational speed. *e limitation of existing methods including the requirement of
high computation and utilization of the same feature distribution is mitigated by dint of the DenTnet.
histopathological images is one of the important tasks in optimized training-testing ratios to classify histopathological
CAD (Computer-Aided Detection/Diagnosis) systems, and images considering existing state-of-the-art deep learning
deep learning models play a remarkable role by detecting, models? *ere exist many surveys enriched to sufficient
classifying, and segmenting prime breast cancer histo- contemporary methods and materials with systematic deep
pathological images. Many researchers worldwide have discussion of automatic classification of breast cancer his-
invested appreciable efforts in developing robust computer- topathological images [68–72]. Nevertheless, to the best of
aided tools for the classification of breast cancer histo- our knowledge, the direct or indirect indication of this
pathological images using deep learning. At present, in this question was not reported in any of the previous studies.
research arena, the most popular deep learning models Henceforth, we perform a succinct survey to investigate this
proposed in the literature are based on CNNs [6–66]. question. Our findings include that the most popular
A pretrained CNN model, for example, DenseNet [67], training-testing ratio for histopathological image classifi-
utilizes dense connection between layers, reduces the cation is 70%: 30%, whereas the best performance (accuracy)
number of parameters, strengthens propagation, and en- is achieved by using the training-testing ratio of 80%: 20% on
courages feature reutilization. *is improved parameter the identical dataset.
efficiency makes the network faster and easier to train. In summary, the main contributions of this context are
Nevertheless, a DenseNet [67] has an excessive connection, as follows:
as all its layers have a direct connection to each other. *ose
(i) Determine the most popular and/or optimized
lavish connections have been shown to decrease the com-
training-testing ratios for classifying histopatho-
putational and parameter efficiency of the network. In ad-
logical images using the existing state-of-the-art
dition, features extracted by a neural network model stay in
deep learning models.
the same distribution. *erefore, the model might overfit as
the features cannot be guaranteed to be sufficient enough. (ii) Propose a novel approach named DenTnet that
Besides, a CNN-training task demands a large number of amalgamates both DenseNet [67] and transfer
training samples; otherwise, it leads to overfitting and re- learning technique to classify breast cancer histo-
duces generalization ability. However, it is arduous to secure pathological images. DenTnet is anticipated to
labeled breast cancer histopathological images, which se- achieve high accuracy and fasten the learning
verely limits the classification ability of CNN [27]. process due to its utilization of dense connections
On the other hand, the use of transfer learning can from its backbone architecture (i.e., DenseNet [67]).
expand prior knowledge about data by including informa- (iii) Determine the generalization ability of DenTnet and
tion from a different domain to target future data [68]. the superiority measure considering nonparametric
Consequently, it is a good idea to extract data from a related statistical tests.
domain and then transfer those extracted data to the target
*e rest of the paper is organized as follows: Section 2
domain. *is way, resources can be saved and the efficiency
hints some preliminaries; Section 3 surveys briefly the
of the model can be improved during training. A great
existing deep models for histopathological image classifi-
number of breast cancer diagnosis methods based on
cation and reports our findings; Section 4 depicts the ar-
transfer learning have been proposed and implemented by
chitecture of our proposed DenTnet and its implementation
distinct researchers (e.g., [57–66]) to achieve state-of-the-art
details; Section 5 demonstrates the experimental results and
performance (e.g., ACC, AUC, PRS, RES, and F1S) on
comparison on BreaKHis dataset [33]; Section 6 evaluates
different datasets. Yet, the limitations of such performance
the generalization ability of DenTnet; Section 7 discusses
indices, algorithmic assumptions, and computational
nonparametric statistical tests, their reported results, and
complexities are indicating a further development of smart
reasons for superiority along with few hints of further study;
algorithms.
and Section 8 concludes the paper.
In this paper, we aim to propose a novel neural-network-
based approach called DenTnet (see Figure 1) for classifying
breast cancer histopathological images by taking the benefits 2. Preliminaries
of both DenseNet [67] and transfer learning [68]. To address
the cross-domain learning problems, we employ the prin- Breast cancer is one of the oldest known kinds of cancer first
ciple of transfer learning for transferring information from a found in Egypt [73]. It is caused by the uncontrolled growth
related domain to the target domain. Our proposed DenTnet and division of cells in the breast, whereby a mass of tissue
is anticipated to increase the accuracy of breast cancer called a tumor is created. Nowadays, it is one of the most
histopathological images classification and accelerate the terrifying cancers in women worldwide. For example, in
learning process. *e DenTnet demonstrates better per- 2020, there were 2.3 million women diagnosed with breast
formance over its alternative CNN and/or transfer-learning- cancer and 685000 deaths globally [74]. Early detection of
based methods (e.g., see Table 1) on the same dataset as well breast cancer can save many lives. Breast cancer can be
as training-testing ratio. diagnosed in view of histology and radiology images. *e
To find the best performance scores of deep learning radiology images analysis can help to identify the areas,
models for classifying histopathological images, contrasting where the abnormality is located. However, they cannot be
training-testing ratios were applied for divergent models on used to determine whether the area is cancerous [75]. On the
the same dataset. What would be the most popular and/or other hand, a biopsy is an examination of tissue removed
Computational Intelligence and Neuroscience 3
3 Color
Channels
per image
Input Volume
Input Images
(224×224×3)
Transition Layer 1
Output size of
112 × 112
56 × 56
Output size
Dense Block 1 of 28 × 28 Dense Block 2
Input size of 56 × 56
Input size of 56 × 56
2 × 2 Avg Pool,Stride 2
3 × 3 Mx Pool, stride 2
Input size of 28 × 28
Input size of 224 × 224
1 × 1 conv
Channel-wise concatenation Channel-wise concatenation
2 × 2 Avg Pool,Stride 2
Input size of 14 × 14
Input size of 28 × 28
Input size of 14 × 14
Input size of 7 × 7
1 × 1 conv
1 × 1 conv
Benign
ImageNet Fine-tune
Weights Network Weights
Softmax Activation
7× 7 Glob.Avg Pool
Malignant
Table 1: Comparison of results of various methods using training-testing ratio of 80%: 20% on BreaKHis [33]. *e best result is shown in
bold.
Year Method PRS RES F1S AUC ACC (%)
Togacar et al. [26] — — — — 97.56
2020 Parvin et al. [31] — — — — 91.25
Man et al. [36] — — — — 91.44
Boumaraf et al. [63] — — — — 92.15
2021
Soumik et al. [60] — — — — 98.97
Liu et al. [172] — — — — 96.97
Zerouaoui and Idri [56] — — — — 93.85
2022
Chattopadhyay et al. [174] — — — — 96.10
DenTnet [ours] 0.9700 0.9896 0.9948 0.99 99.28
4 Computational Intelligence and Neuroscience
from a living body to discover the presence, cause, or extent (a) (b)
of a disease (e.g., cancer). Biopsy is the only reliable way to
make sure if an area is cancerous [76]. Upon completion of
the biopsy, the diagnosis will be based on the qualification of
the histopathologists who determine cancerous regions and
malignancy degree [7, 75]. If the histopathologists are not
well trained, the histopathology or biopsy report can lead to
an incorrect diagnosis. Besides, there might be a lack of (c) (d)
specialists, which may cause keeping the tissue samples for
up to a few months. In addition, diagnoses made by un-
specialized histopathologists are sometimes difficult to
replicate. As if that were not enough of a problem, at times,
even expert histopathologists tend to disagree with each
other. Despite notable progress being reached by diagnostic Figure 2: A sample breast cancer histopathological image [79] with
imaging technologies, the final breast cancer grading and four magnification levels of (a) 40x, (b) 100x, (c) 200x, and (d) 400x.
staging are still done by pathologists using visual inspection
of histological samples under microscopes.
datasets has been used in the literature as incorporated in
As analyzing breast cancer is nontrivial and would get
Table 2:
down to disagreements among experts, computerized and
interdisciplinary systems can improve the accuracy of di- (i) BreaKHis [33] ⇒ It is considered as the most
agnostic results by reducing the processing time. *e CAD popular and clinical valued public breast cancer
can help to assist doctors in reading and interpreting medical histopathological dataset. It consists of 7909
images by locating and identifying possible abnormalities in breast cancer histopathology images, 2480 benign
the image [69]. It is proclaimed that the utilization of CAD to and 5429 malignant samples, from 82 patients
automatically classify histopathological images does not only with different magnification factors (e.g., 40x,
improve the diagnostic efficiency with low cost but also 100x, 200x, and 400x) [33].
provide doctors with more objective and accurate diagnosis (ii) Bioimaging2015 [122] ⇒ *e Bioimaging2015
results [77]. Consequently, there is an adamant demand for [122] dataset contained 249 microscopy training
the CAD [78]. *ere exist several comprehensive surveys for images and 36 microscopy testing images in total,
CAD based methods in the literature. For example, Zebari equally distributed among the four classes.
et al. [71] provided a common description and analysis of
(iii) ICIAR2018 [78] ⇒ *is dataset, available as part
existing CAD systems that are utilized in both machine
of the BACH grand challenge [78], was an ex-
learning and deep learning methods as well as their current
tended version of the Bioimaging2015 dataset
state based on mammogram image modalities and classi-
[8, 122]. It contained 100 images in each of four
fication methods. However, the existing breast cancer di-
categories (i.e., normal, benign, in situ carcinoma,
agnosis models take issue with complexity, cost, human-
and invasive carcinoma) [8].
dependency, and inaccuracy [73]. Furthermore, the limi-
tation of datasets is another practical problem in this arena (iv) BACH [78] ⇒ *e database of BACH holds
of research. In addition, every deep learning model demands images obtained from ICIAR2018 Grand Chal-
a metric to judge its performance. Explicitly, performance lenge [78]. It consists of 400 images with equal
evaluation metrics are the part and parcel of every deep distribution of normal (100), benign (100), in situ
learning model as they indicate progress indices. carcinoma (100), and invasive carcinoma (100).
In the two following subsections, we discuss the com- *e high-resolution images are digitized with the
monly used datasets for classifying histopathological images same conditions and magnification factor of 200x.
and the performance evaluation metrics of various deep In this dataset, images have a fixed size of 2048 ×
learning models. 1536 pixels [175].
(v) TMA [99] ⇒ *e TMA (Tissue MicroArray)
database from Stanford University is a public
2.1. Brief Description of Datasets. Accessing relevant images resource with an access to 205161 images. All the
and datasets is one of the key challenges for image analysis whole-slide images have been scanned by a 20x
researchers. Datasets and benchmarks enable validating and magnification factor for the tissue and 40x for the
comparing methods for developing smarter algorithms. cells [176].
Recently, several datasets of breast cancer histopathology (vi) Camelyon [97] ⇒ *e Camelyon (cancer me-
images have been released for this purpose. Figure 2 shows a tastases in lymph nodes) was established based on
sample breast cancer histopathological image from BreaK- a research challenge dataset competition in 2016.
His [33] dataset of a patient who suffered from papillary *e Camelyon organizers trained CNNs on
carcinoma (malignant) with four magnification levels: (a) smaller datasets for classifying breast cancer in
40x, (b) 100x (c) 200x, and (d) 400x [79]. *e following list of lymph nodes and prostate cancer biopsies. *e
Table 2: A succinct survey of deep-learning-based histopathological image classification methods. NA indicates either “not available” or “no answer” from the associated authors.
Training Testing Performance
Year Ref Aim Technique Dataset Sample Result
(%) (%) AUC ACC
ACC of 97.90%, 16.50%, 16.50%, and 25.30% obtained for
Chan and To predict tumor Employed binarization, fractal
BreaKHis [33] 7909 50 50 40x, 100x, 200x, and 400x magnification factors, NA 39.05%
Tuszynski [80] malignancy in breast cancer dimension, SVM
respectively
To classify histopathological ACC of 90.0%, 88.4%, 84.6%, and 86.1% obtained for 40x,
Spanhol et al. [33] Employed CNN based on AlexNet [81] BreaKHis [33] 7909 70 30 NA 87.28%
images 100x, 200x, and 400x magnification factors, respectively
For single-task CNN, ACC of 83.08%, 83.17%, 84.63%, and
2016
82.10%, obtained for 40x, 100x, 200x, and 400x
Bayram-oglu et al. To classify breast cancer Employed single-task CNN and
BreaKHis [33] 7909 70 30 magnification factors, respectively; accordingly, for NA 82.69%
[38] histopathology images multitask CNN
multitask CNN, ACC of 81.87%, 83.39%, 82.56%, and
80.69%
Overall 92%, 84.20%, 91.50%, and 0.91 obtained for
Abbas [77] To diagnose breast masses Applied SURF [82], LBPV [83] DDSM [84], MIAS [85] 600 40 60 0.91 91.50%
sensitivity, specificity, ACC, and AUC, respectively
ACC of 94.42%, 89.49%, 87.25%, and 85.62% obtained for
To classify histopathology Employed a model of CNN, Fisher BreaKHis [33],
Song et al. [21] 8283 70 30 40x, 100x, 200x, and 400x magnification factors, NA 89.19%
images vector [86], SVM IICBU2008 [87]
respectively
Computational Intelligence and Neuroscience
images gradient boosted trees classifier class classification ACC was reported to be 93.8%
Applied regional deep learning Distinguished between benign and malignant lesions with
Almasni et al. [119] To detect breast masses DDSM [84] 600 80 20 0.96 97.00%
technique an overall ACC of 97%
Table 2: Continued.
Training Testing Performance
Year Ref Aim Technique Dataset Sample Result
(%) (%) AUC ACC
Multimodel method got better predictions than single
To use deep learning for BreaKHis [33], ICIAR2018 classifiers and other algorithms with ACC of 98.13%,
Usage of VGG19 [98], MobileNet [120],
Kassani et al. [8] binary classification of [78], PCam [121], 8594 87 13 95.00%, 94.64% and 83.10% obtained for BreaKHis [33], NA 92.72%
and DenseNet [67]
breast histology images Bioimaging2015 [122] ICIAR2018 [78], PCam [121], and Bioimaging2015 [122],
respectively
To classify breast cancer From BreaKHis [33], ACC of 97.90%, 97.50%, 97.30%, and
BreaKHis [33],
Alom et al. [10] from histopathological Inception recurrent residual CNN 8158 70 30 97.40%, obtained for 40x, 100x, 200x, and 400x 0.98 97.53%
Bioimaging2015 [122]
images magnification factors, respectively
ACC of 95.00%, 96.60%, 93.500%, and 94.20% obtained for
Nahid and Kong To classify histopathological Employed RGB histogram [123] with
Computational Intelligence and Neuroscience
BreaKHis [33] 7909 85 15 40x, 100x, 200x, and 400x magnification factors, NA 94.68%
[12] breast images CNN
respectively
To use CNN for breast *e achieved accuracy between 98.87% and 99.34% for the
Employed CNN, Squeeze-and-
Jiang et al. [16] cancer histopathological BreaKHis [33] 7909 70 30 binary classification as well as between 90.66% and 93.81% 0.99 95.67%
Excitation [124] based ResNet [75]
images for the multiclass classification
ACC of 86.59%, 84.98%, 83.47%, and 82.79% obtained for
Sudharshan et al. To use instance learning for Employed CNN-based multiple
BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, NA 84.46%
[18] image sorting instance learning algorithm
respectively
ACC of 88.37%, 90.29%, 90.54%, and 86.11% obtained for
Gupta and Bhavsar To segment breast cancer Employed ResNet [75] for multilayer
2019 BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, NA 88.82%
[25] images feature extraction
respectively
ACC of 95.10%, 96.30%, 96.90%, and 93.80% obtained for
To extract visual features Combined weak classifiers into a BreaKHis [33],
Vo et al. [125] 8194 87 13 40x, 100x, 200x, and 400x magnification factors, NA 95.56%
from training images stronger classifier Bioimaging2015 [122]
respectively
ACC of 91.48%, 92.20%, 93.01%, and 92.58% obtained for
To organize breast cancer Employed a CNN network to complete
Qi et al. [32] BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, NA 92.32%
images the classification task
respectively
DenseNet161 pretrained and ResNet50 achieved ACC of
To detect and classify DenseNet [67], ResNet [75] (e.g.,
Talo [41] KimiaPath24 [126] 25241 80 20 97.89% and 98.87% on grayscale and color images, NA 98.38%
diseases in images DenseNet161, ResNet50)
respectively
To detect invading Convolutional autoencoder-based 361 samples of the breast ACC was taken into account. *e overall ACC achieved
Li et al. [127] 361 90 10 NA 76.00%
component in cancer images contrast pattern mining framework cancer was 76.00%, whereas 77.70% was presented for F1S
To detect breast cancer from DDSM [84], CBIS-DDSM *e deep CNN presented an ACC of 73.6%, whereas the
Ragab et al. [44] AlexNet [81] and SVM 2781 70 30 0.88 73.60%
images [128] SVM demonstrated an ACC of 87.2%
A modification of Inception module From deep learning networks, an overall ACC of 89.00%
Romero et al. [45] To study cancer images HASHI [129] 151465 63 37 0.96 89.00%
[88] was demonstrated along with F1S of 90.00%
To diagnose breast cancer A modification of ResNet [75] and ACC with 95% for 4 types of cancer classes and ACC with
Minh et al. [46] BACH [78] 400 70 20 0.97 96.25%
images InceptionV3 [88] 97.5% for two combined groups of cancer
7
8
Table 2: Continued.
Training Testing Performance
Year Ref Aim Technique Dataset Sample Result
(%) (%) AUC ACC
To visualize a health system Employed region covariance [131], FABCD [133], BreaKHis ACC of 91.27% and 92.00% at the patient and image level,
Stanitsas et al. [130] 7949 70 15 0.98 91.64%
for clinicians SVM, multiple instance learning [132] [33] respectively
ACC of 97.99%, 97.84%, 98.51%, and 95.88% obtained for
To analyze breast cancer Employed a ResNet [75] architecture
Togacar et al. [26] BreaKHis [33] 7909 80 20 40x, 100x, 200x, and 400x magnification factors, NA 97.56%
images rapidly with attention modules
respectively
ACC of 93.58%, 91.04%, 93.38%, and 91.00% obtained for
To study breast cancer Employed self-training and self-paced
Asare et al. [134] BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, NA 92.25%
images learning
respectively
ACC of 90.69%, 91.12%, 95.36%, and 90.24% obtained for
To diagnose breast cancer
Gour et al. [28] Employed a modification of ResNet [75] BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, 0.91 92.52%
tumors images
respectively
ACC of 95.13%, 95.21%, 94.09%, and 91.42% obtained for
To grade pathological Employed a modification of Xception BreaKHis [33], VLAD
Li et al. [29] 8583 60 40 40x, 100x, 200x, and 400x magnification factors, NA 93.96%
images network [135] [136], LSC [137]
respectively
ACC of 90.12%, 88.89%, 91.57%, and 90.25% obtained for
To allocate breast cancer Deep neural-network-based manifold
Feng et al. [138] BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, NA 90.53%
images preserving autoencoder [139]
respectively
ACC of 89.00%, 92.00%, 94.00% and 90.00% obtained for
Parvin and Mehedi To study CNN models for LeNet [140], AlexNet [81], VGGNet
BreaKHis [33] 7909 80 20 40x, 100x, 200x, and 400x magnification factors, 0.85 91.25%
Hasan [31] cancer images [98], ResNet [75], Inception V3 [88]
respectively
ACC of 95.40%, 94.70%, 97.60%, and 95.50% obtained for
Carvalho et al. To classify histological Entropies of Shannon [142], Renyi
BreaKHis [33] 4960 70 30 40x, 100x, 200x, and 400x magnification factors, 0.99 95.80%
[141] breast images [112], Tsallis [143]
respectively
2020 ACC of 96.00%, 96.16%, 98.01%, and 95.97% obtained for
To analyze breast cancer Employed global covariance pooling
Li et al. [144] BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, NA 94.93%
images module [145]
respectively
ACC of 97.72%, 96.19%, 86.66%, and 85.18% obtained for
Usage of generative adversarial
Man et al. [36] To classify cancer images BreaKHis [33] 7909 80 20 40x, 100x, 200x, and 400x magnification factors, NA 91.44%
networks, DenseNet [67]
respectively
For BreaKHis [33], ACC of 95.94%, 96.22%, 98.15%, and
To classify human breast Employed a framework based on a
BreaKHis [33] and 94.41% obtained for 40x, 100x, 200x, and 400x
Kumar et al. [37] cancer and canine variant of VGGNet [98] (e.g., 8261 70 30 0.95 96.93%
CMTHis [37] magnification factors, respectively; the same for CMTHis
mammary tumors VGGNet16) and SVM
[37], ACC of 94.54%, 97.22%, 92.07%, and 82.84% obtained
Kaushal and Singla To detect cancerous cells in Employed a CNN model of self-training Total 50 images of various ACC was taken into account. Estimation of the standard
50 90 10 NA 93.10%
[40] images. and self-paced learning patients error of mean was approximately 0.81
Variants of VGGNet [98] (e.g., fully
To use deep learning for Breast cancer images: 675 *e ensemble of fine-tuned VGG16 and VGG19 models
trained VGG16, fine-tuned VGG16,
Hameed et al. [43] classification of breast for training and 170 for 845 80 20 offered sensitivity of 97.73% for carcinoma class and overall NA 95.29%
fully trained VGG19, and fine-tuned
cancer images testing accuracy of 95.29%. It also offered an F1 score of 95.29%
VGG19 models)
In INbreast [92] mean ACC of 89%, 93%, and 95% for
To detect breast lesions in
Alantari et al. [48] Adopted three deep CNN models INbreast [92], DDSM [84] 1010 70 20 CNN, ResNet50, and Inception-ResNet V2, respectively; 0.96 94.08%
digital X-ray mammograms
95%, 96%, and 98% for DDSM [146]
ResNet [75], DenseNet [67], VGGNet CBIS-DDSM [128], Overall ACC of 90.91% and 87.93% obtained from CBIS-
Zhang et al. [49] To classify breast mass 3168 70 30 0.96 89.42%
[98] INbreast [92] DDSM [128] and INbreast [92], respectively
With CBIS-DDSM [128] and INbreast [92] databases, the
To classify breast cancer Modification of AlexNet [22] and CBIS-DDSM [128], MIAS
Hassan et al. [59] 600 75 17 modified GoogLeNet achieved ACC of 98.46% and 92.5%, 0.97 96.98%
masses GoogLeNet [88] [85], INbreast [92], etc
respectively
Computational Intelligence and Neuroscience
Table 2: Continued.
Training Testing Performance
Year Ref Aim Technique Dataset Sample Result
(%) (%) AUC ACC
To use high-resolution info Multiview attention-guided multiple BreaKHis [33], BACH Overall ACC of 94.87%, 91.32%, and 90.45% obtained from
Li et al. [147] 12329 70 30 0.99 92.21%
of images instance detection network [78], PUIH [148] BreaKHis [33], BACH [78], and PUIH [148], respectively
ACC of 92.71%, 94.52%, 94.03%, and 93.54% obtained for
To divide breast cancer Employed a model of CNN and
Wang et al. [27] BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, NA 93.70%
images CapsNet [149]
respectively
ACC of 96%, 95.10%, and 87% obtained for polynomial
Albashish et al. [30] To analyze VGG16 [98] Employed a variation of VGGNet [98] BreaKHis [33] 7909 90 10 SVM, Radial Basis SVM, and k-nearest neighbors, NA 92.70%
respectively
ACC of 93.35%, 93.86%, 93.73%, and 94.00% obtained for
To classify breast cancer Employed SURF [82], DSIFT [151],
Computational Intelligence and Neuroscience
Kundale et al. [150] BreaKHis [33] 7909 70 30 40x, 100x, 200x, and 400x magnification factors, NA 93.74%
from histology images linear coding [152]
respectively
To classify breast cancer For BreaKHis [33], ACC of 99.03%, 99.53%, 98.08%, and
Employed several deep learning BreaKHis [33], ICIAR2018
Attallah et al. [153] from histopathological 7909 70 30 97.56% got for 40x, 100x, 200x, and 400x magnification NA 98.43%
techniques including autoencoder [139] [78]
images factors, respectively; for ICIAR2018 [78], ACC was 97.93%
Stochastic [155], Nesterov [156], ACC was taken into account. *e overall ACC of 97.00%,
To classify breast cancer
Burçak et al. [154] Adaptive [157], RMSprop [158], BreaKHis [33] 7909 70 30 97.00%, 96.00%, and 96.00% obtained for 40x, 100x, 200x, NA 96.50%
histopathological images
AdaDelta [159], Adam [160] and 400x magnification factors, respectively
To label breast cancer Images from four different data samples achieved an
Hirra et al. [161] Patch-based deep belief network [162] HASHI [129] 584 52 30 NA 86.00%
images accuracy of 86%
2021
*e overall ACC for the subimage classification was 97.29%
To extract eminent breast
Elmannai et al. [42] A combination of two deep CNNs BACH [78] 400 60 20 and for the carcinoma cases the sensitivity achieved was NA 97.29%
cancer image features
99.58%
Baker and Abu To segment breast cancer Clustering and global thresholding *e maximum ACC obtained from classifiers and neural
BACH [78] 400 70 30 NA 63.66%
Qutaish [163] images methods network using BACH [78] to detect breast cancer
ACC of 99.50%, 98.90%, 98.96% and 98.51% obtained for
To classify breast cancer
Soumik et al. [60] Employed Inception V3 [88] BreaKHis [33] 7909 80 20 40x, 100x, 200x, and 400x magnification factors, NA 98.97%
images
respectively
AUC values of 0.698, 0.639, and 0.654 obtained for max-
To analyze gigapixel Employed CNN with a compressing Camelyon16 [164],
Brancati et al. [50] 892 68 32 pooling, average pooling, and combined attention maps, 0.66 NA
histopathological images path and a learning path TUPAC16 [165]
respectively
Mahmoud et al. To classify breast cancer Mammography images Maximum ACC of 97.80% was claimed by using the given
Employed transfer learning 7500 80 20 NA 94.45%
[61] images [166] dataset [166]. Sensitivity and specificity were estimated
To classify breast cancer Overall ACC of 98.33% obtained from ICIAR2018 [78].
Munien et al. [62] Employed EfficientNet [167] ICIAR2018 [78] 400 85 15 NA 98.33%
images Sensitivity was also taken into account
To analyze breast cancer Employed ResNet [75] on ImageNet ACC of 94.49%, 93.27%, 91.29%, 89.56% obtained for 40x,
Boumaraf et al. [63] BreaKHis [33] 7909 80 20 NA 92.15%
images [168] images 100x, 200x, and 400x magnification factors, respectively
Overall ACC, PRS, F1S, and AUC of 98.96%, 97.35%,
Saber et al. [64] To detect breast cancer Employed transfer learning technique MIAS [85] 322 80 20 0.995 98.96%
97.66%, and 0.995, respectively, got from MIAS [85]
9
10
Table 2: Continued.
Training Testing Performance
Year Ref Aim Technique Dataset Sample Result
(%) (%) AUC ACC
Ameh Joseph et al. To classify breast cancer Employed handcrafted features and ACC of 97.87% for 40x, 97.60% for 100x, 96.10% for 200x,
BreaKHis [33] 7909 90 10 NA 97.08%
[169] images dense layer and 96.84% for 400x demonstrated from BreaKHis [33]
Employed probabilistic transition rules ACC, PRS, RES, F1S, and GMN of 89.13%, 86.23%, 81.47%,
Reshma et al. [52] To detect breast cancer BreaKHis [33] 7909 90 10 NA 89.13%
with CNN 85.38%, and 85.17% demonstrated from BreaKHis [33]
To detect nuclei on breast PRS, RES, and F1S of 91.28%, 87.68%, and 89.44%
Huang et al. [53] Employed mask-region-based CNN H&E images of patients 537 80 20 NA 95.00%
cancer demonstrated from the used dataset
To learn efficient Employed magnification prior Maximum mean ACC of 97.04% and 97.81% were got from
Chhipa et al. [170] BreaKHis [33] 7909 70 30 NA 97.42%
representations contrastive similarity patient and image levels, respectively using BreaKHis [33]
Average ACC, PRS, RES, and F1S of 97.75%, 95.19%,
To classify breast cancer Employed channel attention module
Zou et al. [171] BreaKHis [33], BACH [78] 8309 90 10 97.30%, and 96.30% obtained from BreaKHis [33], NA 91.37%
images with nondimensionality reduction
respectively. ACC of 85% got from BACH [78]
Average ACC, PRS, RES, F1S, and RTM of 96.97%, 96.47%,
To classify breast cancer Employed autoencoder and Siamese
Liu et al. [172] BreaKHis [33] 7909 80 20 99.15%, 97.82%, and 335 seconds obtained from BreaKHis NA 96.97%
images framework
[33], respectively
Maximum ACC of 98.67% obtained from MIAS [85].
Jayandhi et al. [54] To diagnose breast cancer Employed VGG [98] and SVM MIAS [85] 322 80 20 NA 98.67%
Sensitivity and specificity were also calculated
2022
Average ACC, PRS, RES, F1S, and AUC of 95.58%, 95%,
Sharma and Kumar To classify breast cancer
Employed Xception [135] and SVM BreaKHis [33] 2000 75 25 95%, 95%, and 0.98 obtained from BreaKHis [33], 0.98 95.58%
[55] images
respectively
Zerouaoui and Idri To classify breast cancer Employed multilayer perceptron, ACC of 92.61%, 92%, 93.93%, and 91.73% on four
BreaKHis [33] and others NA 80 20 NA 93.85%
[56] images DenseNet201 [67] magnification factors of BreaKHis [33]
To classify breast cancer 323 colored lymphoma A total of 27 misclassifications for 323 samples were
Soltane et al. [65] Employed ResNet [75] 323 50 50 NA 91.6%
images images claimed. PRS, RES, F1S, and Kappa score were estimated
Random forest algorithm achieved better result for
To analyze breast cancer Employed random forest, k-nearest
Naik et al. [173] 699 whole-slide images 699 80 20 classifying benign and malignant images from 190 testing NA 98.2%
images neighbors, SVM
samples
Average ACC, PRS, RES, and F1S of 96.10%, 96.03%,
Chattopadhyay To classify breast cancer Employed dense residual dual-shuffle
BreaKHis [33] 7909 80 20 96.08%, and 96.02%, respectively, obtained from four NA 96.10%
et al. [174] images attention network
different magnification levels of BreaKHis [33]
Best results for ACC, PRS, RES, F1S, and AUC of 89.5%,
Alruwaili and Employed the principle of transfer
To detect breast cancer MIAS [85] 322 80 20 89.5%, 90%, and 89.5% obtained from MIAS [85], NA 89.5%
Gouda [66] learning, ResNet [75]
respectively
Computational Intelligence and Neuroscience
Computational Intelligence and Neuroscience 11
training dataset consists of 270 whole-slide im- real-world type cohort which contains a larger
ages; among them 160 are normal slides and 110 degree of stain and scanning variances [137]. It
slides contain metastases [97]. consists of 374 images with resolution of 1388 ×
(vii) PCam [121] ⇒ It is a modified version of the 1040 [29].
Patch Camelyon (PCam) dataset, which consists (xviii) KimiaPath24 [126] ⇒ *e official KimiaPath24
of 327680 microscopy images with 96 × 96-pixel [126] dataset consists of a total of 23916 images
sized patches extracted from the whole-slide for training and 1325 images for testing. It is
images with a binary label hinting the presence of publicly available. It shows various body parts
metastatic tissue [8]. with texture patterns [41].
(viii) HASHI [129] ⇒ Each image in the dataset of
HASHI (high-throughput adaptive sampling for
2.2. Performance Evaluation Metrics. Performance evaluation
whole-slide histopathology image analysis) [129]
of any deep learning model is an important task. An algorithm
has the size of 3002 × 2384 [161].
may give very pleasing results when evaluated using a metric
(ix) MIAS [85] ⇒ *e Mammographic Image Anal- (e.g., ACC), but it may give poor results when evaluated against
ysis Society (MIAS) database of digital mam- other metrics (e.g., F1S) [177]. Usually, we use classification
mograms [85] contains 322 mammogram images, accuracy to measure the performance of deep learning algo-
each of which has a size of 1024 × 1024 pixels with rithms. But it is not enough to determine the model perfectly.
PGM format [59]. For truly judge any deep learning algorithm, we can use
(x) INbreast [92] ⇒ *e INbreast database has a total nonidentical types of evaluation metrics including classifica-
of 410 images collected from 115 cases (i.e., pa- tion ACC, AUC, PRS, RES, F1S, RTM, and GMN.
tients) indicating benign, malignant, and normal (i) ACC ⇒ It is normally defined in terms of error or
cases having sizes of 2560 × 3328 or 3328 × 4084 inaccuracy [178]. It can be calculated using the
pixels. It contains 36 benign and 76 malignant following equation:
masses [92].
(xi) DDSM [84] ⇒ *e DDSM [84] dataset was col- (100)tn + tp
ACC � , (1)
lected by the expert team at the University of South tp + tn + fp + fn
Florida [84]. It contains 2620 scanned film mam-
mography studies. Explicitly, it involves 2620 breast where tn is true negative, tp is true positive, fp is
cases (i.e., patients) categorized in 43 different vol- false positive, and fn is false negative. Sometimes,
umes with average size of 3000 × 4800 pixels [48]. ACC and the percent correct classification (PCC)
(xii) CBIS-DDSM [128] ⇒ *e CBIS-DDSM [128] is can be used interchangeably.
an updated version of the DDSM providing easily (ii) PRS ⇒ Its best value is 100 and the worst value is
accessible data and improved region-of-interest just 0. It can be formulated using the following
segmentation [128, 146]. *e CBIS-DDSM equation:
dataset comprises 2781 mammograms in the
PNG format [49]. (100)tp
PRS � . (2)
(xiii) CMTHis [37] ⇒ *e CMTHis (Canine Mam- tp + fp
mary Tumor Histopathological Image) [37]
dataset comprises 352 images acquired from 44 (iii) RES ⇒ It should ideally be 100 (the highest) for a
clinical cases of canine mammary tumors. good classifier. It can be calculated using the fol-
(xiv) FABCD [133] ⇒ *e FABCD (Fully Annotated lowing equation:
Breast Cancer Database) [133] consists of 21 (100)tp
annotated images of carcinomas and 19 images of RES � . (3)
tp + fn
benign tissue taken from 21 patients [130].
(xv) IICBU2008 [87] ⇒ *e IICBU2008 (Image In- (iv) AUC ⇒ It is one of the most widely used metrics
formatics and Computational Biology Unit) for evaluation [177–179]. *e AUC of a classifier
malignant lymphoma dataset contains 374 H&E equals the probability that the classifier ranks a
stained microscopy images captured using bright randomly chosen positive sample higher than a
field microscopy [21]. randomly chosen negative sample. *e AUC varies
(xvi) VLAD [136] ⇒ *e VLAD (Vector of Locally in value from 0 to 1. If the predictions of a model
Aggregated Descriptors) dataset [136] consists of are 100% wrong, then its AUC � 0.00; but if its
300 annotated images with resolution of 1280 × predictions are 100% correct, then its AUC � 1.00.
960 [29]. (v) F1S ⇒ It is the harmonic mean between precision
(xvii) LSC [137] ⇒ *e LSC (Lymphoma Subtype and recall. It is also called the F-score or F-mea-
Classification) [137] dataset has been prepared by sure. It is used in deep learning [177]. It conveys
pathologists from different laboratories to create a the balance between the precision and the recall. It
12 Computational Intelligence and Neuroscience
also tells us how many instances it classifies cor- 3.2. Key Techniques and Challenges. *e CNNs can be
rectly. Its highest possible value is 1, which indi- regarded as a variant of the standard neural networks. In-
cates perfect precision and recall. Its lowest stead of using fully connected hidden layers, the CNNs
possible value is 0, when either the precision or the introduce the structure of a special network, which com-
recall is zero. It can be formulated as prises so-called alternating convolution and pooling layers.
2 tp *ey were first introduced for overcoming known problems
F1S � � , (4) of fully connected deep neural networks when handling high
1/PRS + 1/RES tp + fp + fn /2 dimensionality structured inputs, such as images or speech.
From Table 2, it is noticeable that CNNs have become state-
where PRS is the number of correct positive results of-the-art solutions for breast cancer histology images
divided by the number of positive results predicted classification. However, there are still challenges even when
with the classifier and RES is the number of correct using the CNN-based approaches to classify pathological
positive results divided by the number of all rel- breast cancer images [16], as given below:
evant samples.
(i) Risk of overfitting ⇒ *e number of parameters of
(vi) RTM ⇒ Estimating the RTM complexity of al-
CNN increases rapidly depending on how large the
gorithms is mandatory for many applications (e.g.,
network is, which may lead to poor learning.
embedded real-time systems [180]). *e optimi-
zation of the RTM complexity of algorithms in (ii) Being cost-intensive ⇒ To get a huge number of
applications is highly expected. *e total RTM can labeled breast cancer images is very expensive.
prove to be one of the most important determi- (iii) Huge training data ⇒ CNNs need to be trained
native performance factors in many software-in- using a lot of images, which might not be easy to
tensive systems. find considering that collecting real-world data is a
(vii) GMN ⇒ It indicates the central tendency or typical tedious and expensive process.
value of a set of numbers by considering the (iv) Performance degradation ⇒ Various hyper-
product of their values instead of using their sum. parameters have a significant influence on the per-
It can be used to attain a more accurate measure of formance of the CNN model. *e model’s
returns than the mean or arithmetic mean or parameters need to be tuned properly to achieve a
average. *e GMN for any set of numbers desirable result [75], which usually is not an easy task.
x1 , x2 , x3 , . . . , xm can be defined as (v) Employment difficulty ⇒ In the process of training
m
1/m
���� CNN model, it is usually inevitable to rearrange the
⎝ x ⎞
GMN � ⎛ ⎠ � [m]x1 x2 x3 · · · xm . (5) learning rate parameters to get a better performance.
i
i�1 *is makes it arduous for the algorithm to use in real-
life applications by nonexpert users [181].
Many methods had been proposed in the literature
(viii) MCC ⇒ *e Matthews correlation coefficient
considering the aforementioned challenges. In 2012, Alex-
(MCC) is used as a measure of the quality of binary
Net [81] architecture was introduced for ImageNet Chal-
classifications, introduced by biochemist Brian
lenge having error rate of 16%. Later various variations of
W. Matthews in 1975.
AlexNet [81] with denser network were introduced. Both
(ix) κ⇒ *e metric of Cohen’s kappa (κ) can be used to AlexNet [81] and VGGNet [98] were the pioneering works
evaluate binary classifications. that demonstrated the potential of deep neural networks
[182]. AlexNet was designed by Alex Krizhevsky [81]. It
3. A Succinct Survey of State of the Art contained 8 layers; the first 5 were convolutional layers,
some of them followed by max-pooling layers, and the last 3
*is section deals with a summary of existing studies were fully connected layers [81]. It was the first large-scale
apposite for the classification of breast cancer histo- CNN architecture that did well on ImageNet [183] classi-
pathological images followed by a short discussion and fication. AlexNet [81] was the winner of the ILSVRC [183]
our findings. classification, the benchmark in 2012. Nevertheless, it was
not very deep. SqueezeNet [184] was proposed to create a
3.1. Summary of Previous Studies. Table 2 provides a short smaller neural network with fewer parameters that could be
summary of previous studies carried out to classify breast easily fit into computer memory and transmitted over a
cancer from images. Experimental results of miscella- computer network. It achieved AlexNet [81] level accuracy
neous deep models in the literature on publicly available on ImageNet with 50x fewer parameters. It was compressed
datasets demonstrated various degrees of accurate cancer to less than 0.5 MB (510x smaller than AlexNet [81]) with
prediction scores. However, as AUC and ACC are the model compression techniques. *e VGG [98] is a deep
most important metrics for breast cancer histopatho- CNN used to classify images. *e VGG19 is a variant of
logical images classification [49], the experimental results VGG which consists of 19 layers (i.e., 16 convolution layers
in Table 2 take them into account as the performance and 3 fully connected layers, in addition to 5 max-pooling
indices. layers and 1 SoftMax layer) [98]. *ere exist many variants of
Computational Intelligence and Neuroscience 13
VGG [98] (e.g., VGG11, VGG16, VGG19, etc.). VGG19 has artificial data is a tentative solution of this issue, but the
19.6 billion FLOPs (floating point operations per second). determination of the best hyperparameters is extremely
VGG [98] is easy to implement but slow to train. Nowadays, difficult. Besides efficient deep learning models, the datasets
many deep-learning-based methods are implemented on themselves have some limitations, for example, overinter-
influential backbone networks; among them, both DenseNet pretation, which cannot be diagnosed using typical evalu-
[67] and ResNet [75] are very popular. Due to the longer ation methods based on the ACC of the model. Deep
path between the input layer and the output layer, the in- learning models trained on popular datasets (e.g., BreaKHis
formation vanishes before reaching its destination. Dense- [33]) may suffer from overinterpretation. In overinterpre-
Net [67] was developed to minimize this effect. *e key base tation, deep learning models make confident predictions
element of ResNet [75] is the residual block. DenseNet [67] based on details that do not make any sense to humans (e.g.,
concentrates on making the deep learning networks move promiscuous patterns and image borders). When deep
even deeper as well as simultaneously making them well learning models are trained on datasets, they can make
organized to train by applying shorter connections among apparently authentic predictions based on both meaningful
layers. In short, ResNet [75] adopts summation, whereas and meaningless subtle signals. *is effect, eventually, can
DenseNet [67] deals with concatenation. Yet, the dense reduce the overall classification performance of deep models.
concatenation of DenseNet [67] creates a challenge of de- Most probably, this is one of the reasons why any state-of-
manding high GPU (Graphics Processing Unit) memory the-art deep learning model in the literature for classifying
and more training time [182]. On the other hand, the breast cancer histopathological images (see Table 2) could
identity shortcut that balances training in ResNet [75] curbs not show an ACC of 100%.
its representation dimensions [182]. Compendiously, there In addition, the training-testing ratio can regulate the
is a dilemma in the alternative between ResNet [75] and performance of a deep model for image classification. We
DenseNet [67] for many applications in terms of perfor- wish to determine the most popular and/or optimized
mance and GPU resources [182]. training-testing ratios for classifying histopathological im-
ages using Table 2. To this end, we have calculated the usage
3.3. Our Findings. Although various deep learning models in frequency of the training-testing ratio (i.e., percentage of the
Table 2 often achieved pretty good scores of AUC and ACC, number of papers that used the same ratio) by considering
the models demand a large amount of data but breast cancer data in Table 2 and the following equation:
diagnosis always suffers from a lack of data. To adopt
Figure 3 demonstrates the frequency of usage of 4.1. Architecture of Our Proposed DenTnet. *e architecture
training-testing ratio considering data in Table 2. From of our proposed DenTnet is shown in Figure 1, which
Figure 3, it is noticeable that the most popular training- consists of four different blocks, namely, the input volume,
testing ratio for histopathological image classification is training from scratch, transfer learning, and fusion and
70%: 30%. *e second-best used training-testing ratio is recognition.
80%: 20%, followed by 90%: 10%, 75%: 25%, 50%: 50%,
and so on. Figure 4 presents the GMN of ACC for the
most frequently used training-testing ratios considering 4.1.1. Input Volume. *e input is a 3D RGB (three-di-
data in Table 2. It shows a different history; in terms of mensional red, green, and blue) image with a size of
ACC, the rate of 80%: 20% became the best option for the 224 × 224, that is, 224 × 224 × 3.
training-testing ratio to classify histopathological im-
ages. Explicitly, the GMN of ACC formed like a Gaussian 4.1.2. Training from Scratch. Initially, features are extracted
shaped curve and the ratio of 80%: 20% owned its highest from the input images by feeding the input to the con-
peak. To cut a long story short, by considering ACC, the volutional layer. *e convolution (conv) layers contain a set
training-testing ratio of 80%: 20% became the finest and of filters (or kernels) parameters, which are learned
the optimal choice for classifying histopathological throughout the training. *e size of the filters is usually
images. smaller than the actual image, where each filter convolves
with the image and creates an activation map. *ereafter, the
4. Methods and Materials pooling layer progressively decreases the spatial size of the
representation for reducing the number of parameters in the
*is section explains in detail our proposed DenTnet model network. Instead of differentiable functions such as sigmoid
and its implementation. Figure 5 demonstrates a general and tanh, the network utilizes the ReLU as an activation
flowchart of our methodology to classify breast cancer function. Finally, the extracted features or the output of the
histopathological images automatically. last layer from the training from scratch block is then
14 Computational Intelligence and Neuroscience
35 Splitting Data
30
25 Training and Validation Sets
20
15 Model Fitting Deep Model
10 Process
Training
5 No
0 Optimization / Parameter Tuning
Training
40-60
50-50
52-30
60-20
60-40
63-37
65-35
68-32
70-20
70-30
75-17
75-25
80-20
85-15
87-13
90-10
98-02
Successful
?
Training-testing ratios
Figure 3: Determination of the most popular training-testing Yes
ratios using data from Table 2.
70
60 Figure 5: Flowchart of our methodology to classify breast cancer
50 histopathological images.
40
30 ImageNet [168] dataset. Since transfer learning involves
20 transferring knowledge from one domain to another, we
10
have utilized the ImageNet weight as the models developed
0
50-50 70-30 75-25 80-20 90-10 98-02 in the ImageNet [168] classification competition are mea-
Training-testing ratios (%) sured against each other for performance. Henceforth, the
ImageNet weight provides a measure of how good a model is
Figure 4: GMN of ACC for the most popular training-testing
ratios deeming data from Table 2.
for classification. Besides, the ImageNet weight has already
showed a markedly high accuracy [185]. *e extracted
amalgamated with the features extracted from the transfer features are then used by the network before being passed to
learning approach. Figure 1 includes the design of the the fusion and recognition block, where the features are
DenseNet [67] architecture used to extract the feature using amalgamated with the extracted features from the learning-
the learning-from-scratch approach. from-scratch block for recognition.
4.1.3. Transfer Learning. In transfer learning, given that a 4.1.4. Fusion and Recognition. *e extracted features based
domain D consists of feature space X and a marginal on the ImageNet weights are then amalgamated with the
probability distribution P(X), where X = x1 , x2 , . . . , xn ∈∈X, features extracted by the block of training from scratch. A
and a task T consists of a label space Y and an objective global average pooling is performed. Dropout technique
predictive function f: X ⟶ Y, the corresponding label helps to prevent a model from overfitting. It is used with
f(x) of a new instance x is predicted by function f, where dense fully connected layers. *e fully connected layer
the new tasks denoted by T = Y, f(x) are learned from the compiles the data extracted by previous layers to form the
training data consisting of pairs xi and yi , where xi ∈ X and final output. *e last step passes the features through the
yi ∈ Y. When utilizing the learning-from-scratch approach, fully connected layer, which then uses SoftMax to classify the
the extracted features stay in the same distribution. To solve class of the input images.
this problem, we amalgamated both learning-from-scratch
and the transfer learning approach. *e learned parameters
are further fine-tuned by retraining the extracted features. 4.2. Implementation Details
*is is anticipated to expand the prior knowledge of the 4.2.1. Data Preparation. We have adopted data augmen-
network about the data, which might improve the efficiency tation, stain normalization, and image normalization
of the model during training, thereby accelerating the strategies to optimize the training process. Hereby, we have
learning speed and also increasing the accuracy of the model. explained each of them briefly.
As shown in Figure 1, there is a connection between the
blocks of the input volume and transfer learning. *e
transfer learning approach extracted features from the 4.2.2. Data Augmentation. Due to the limited size of the
ImageNet [168] weights. *e weight is the parameters (in- input samples, the training of our DenTnet was prone to
cluding trainable and nontrainable) learned from the overfitting, which caused low detection rate. One solution to
Computational Intelligence and Neuroscience 15
alleviate this issue was the data augmentation, which gen- model during training. To evaluate the images classification,
erated more training data from the existing training set. we have computed the recognition rate at the image level
Dissimilar data augmentation techniques (e.g., horizontal over the two different classes: (i) correctly classified images
flipping, rotating, and zooming) were applied to datasets for and (ii) the total number of images in the test set.
creating more training samples.
Table 3: List of hyperparameter values for the proposed deep learning model.
Hyperparameters
Model
Beta_1 Beta_2 Learning rate Epoch Batch size Epsilon Decay AMSGrad
DenTnet 0.60 0.90 0.001 50 32 None 0.0 True
0.9925
Loss
0.10
0.9900 0.09
0.9875 0.08
0.9850 0.07
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Epoch Epoch
Figure 6: (a) hints ACC and (b) shows loss charts of DenTnet during training.
0.8
Benign 216 0
True positive rate
True label
300
0.6
200 0.4
Malignant 5 476
0.2
100
0.0
Benign Malignant 0.0 0.2 0.4 0.6 0.8 1.0
0 False positive rate
Predicted label Roc curve (AUC=0.99)
ACC = 0.9928: Misclass = 0.0072
(a) (b)
1.0
0.9
Precision
0.8
0.7
0.6
0.5
0.0 0.2 0.4 0.6 0.8 1.0
Recall
(c)
Figure 7: (a) hints confusion matrix for benign and malignant classification, (b) shows ROC curve, and (c) demonstrates precision-recall
curve.
Computational Intelligence and Neuroscience 17
Table 1 compares the results obtained by several 6.2. Experimental Results Comparison. Using four datasets in
methods. *e methods of Togacar et al. [26], Parvin et al. the experiment, DenTnet has been compared with six widely
[31], Man et al. [36], Soumik et al. [60], Liu et al. [172], used and well-known deep learning models, namely,
Zerouaoui and Idri [56], and Chattopadhyay et al. [174] were AlexNet [81], ResNet [75], VGG16 [98], VGG19 [98], In-
centered on mainly CNN models, but they were tested ception V3 [88], and SqueezeNet [184]. To evaluate and
against the same training-testing ratio of 80%: 20% on the analyze the performance of DenTnet, four different cases are
BreaKHis dataset [33]. However, Boumaraf et al. [63] considered. *e first case is the evaluation of different deep
suggested a transfer-learning-based method deeming the learning methods, which are trained and tested on BreaKHis
residual CNN ResNet-18 as a backbone model with block- [33] dataset. *e second case studies the performance of the
wise fine-tuning strategy and obtained a mean ACC of deep-learning-based classification methods that are trained
92.15% applying a training-testing ratio of 80%: 20% on and tested on Malaria [191] dataset. *e third case is to train
BreaKHis dataset [33]. From Table 1, it is notable that and test the deep learning models on SkinCancer [193]
DenTnet [ours] achieved the best ACC on the same ground. dataset. *e final one is to understand and analyze the
performance of the deep learning models on CovidXray
[192] dataset. *e overall results are tabulated in Tables 5–9.
6. Generalization Ability Evaluation of Besides, the RTM in seconds of various datasets using the
Proposed DenTnet deep learning models is shown in Table 10.
According to the results in terms of GMN of ACC, RES,
What would be the performance of the proposed DenTnet
F1S, and AUC as shown in Tables 5–9, respectively, the
compared with other types of cancer or disease datasets? To
proposed DenTnet architecture provides the best scores as
evaluate the generalization ability of DenTnet, this section
compared to AlexNet [81], ResNet [75], VGG16 [98],
presents the experimental result obtained not only from the
VGG19 [98], Inception V3 [88], and SqueezeNet [184]. On
dataset of BreaKHis [33] but also from additional datasets of
the other hand, DenTnet gets the third best result.
Malaria [191], CovidXray [192], and SkinCancer [193].
Moreover, in most of the cases, AlexNet [81] obtains the
lowest results.
6.1. Datasets Irrelevant to Breast Cancer. *e three following
datasets are not related to breast cancer. Herewith, their
6.3. Performance Evaluation. *e deepening of deep models
primary aim is to evaluate the generalization ability of our
makes their parameters rise rapidly, which may lead to
proposed method DenTnet:
overfitting of the model. To take the edge off the overfitting
(i) Malaria [191] ⇒ *is dataset contains a total of problem, predominantly a large number of dataset images
27558 infected and uninfected images for malaria. are required as the training set. Considering a small dataset,
(ii) SkinCancer [193] ⇒ *is dataset contains balanced it is possible to reduce the risk of overfitting of the model by
images from benign skin moles and malignant skin reducing the parameters and augmenting the dataset. Ac-
moles. *e data consist of two folders, each con- cordingly, DenTnet used fewer parameters along with the
taining 1800 pictures (224 × 244) from the two dense connections in the construction of the model, instead
types of mole. of the direct connections among the hidden layers of the
network. As DenTnet used fewer parameters, it attenuated
(iii) CovidXray [192] ⇒ Corona (COVID-19) virus af-
the vanishing gradient descent and strengthened the feature
fects the respiratory system of healthy individual.
propagation. Consequently, the proposed DenTnet out-
*e chest X-ray is one of the key imaging methods
performed its alternative state-of-the-art methods. Yet, its
to identify the coronavirus. *is dataset contains
runtime was a bit longer in Malaria [191] and SkinCancer
chest X-ray of healthy versus pneumonia (Corona)
[193] datasets as compared to ResNet [75]. *e main reason
infected patients along with few other categories
why the DenTnet model may require more time is that it uses
including SARS (Severe Acute Respiratory Syn-
many small convolutions in the network, which can run
drome), Streptococcus, and ARDS (Acute Respira-
slower on GPU than compact large convolutions with the
tory Distress Syndrome) with a goal of predicting
same number of GFLOPS. Still, DenTnet includes fewer
and understanding the infection.
parameters compatibility when compared to ResNet [75].
Figure 8 specifies some sample images from Malaria Henceforth, it is more efficient in solving the problem of
[191], SkinCancer [193], and CovidXray [192] datasets. overfitting. In general, all of the used algorithms suffered
18 Computational Intelligence and Neuroscience
Figure 8: (a), (b), and (c) specify images of Malaria [191], SkinCancer [193], and CovidXray [192] datasets, respectively.
from some degree of overfitting problem on all datasets. We 7. Nonparametric Statistical Analysis
minimized such problems by reducing the batch size and
adjusting the learning rate and the dropout rate. In some Figure 9 depicts performance evaluation of various algo-
cases, the proposed DenTnet predicted fewer positive rithms deeming the numerical values of the ineffectualness
samples as compared to ResNet [75]. *is is due to the lack metrics and RTM from Table 11. It is noted that, for a better
of its conservative designation of the positive class. *us, the visualization purpose, the RTM scores in Figure 9 use log-
GMN PRS of the proposed DenTnet was about 2% lower normal distribution [195] with a mean of 10 and standard
than that of ResNet [75]. deviation of 1. However, from this graph, it is extremely hard
As VGG16 [98] is easier to implement, many deep to rank each algorithm. However, statistically, it is possible
learning image classification problems benefit from the to show that one algorithm is better than its alternatives.
technique by using the network either as a sole model or as a Friedman test [196] and its derivatives (e.g., Iman-Daven-
backbone architecture to classify images. While VGG19 [98] port test [197]) are normally referred to as examples of the
is better than the VGG16 [98] model, they are both very slow most well-known nonparametric tests for multiple com-
to train—for example, a ResNet with 34 layers only requires parisons. *e mathematical equations of Friedman [196],
18% of operations as a VGG with 19 layers (around half the Friedman’s aligned rank [198], and Quade [199] tests can be
layers of the ResNet) will require [194]. Regarding AlexNet found in the works of Quade [199] and Westfall and Young
[81], the model struggled to scan all features as it is not very [200]. Friedman test [196] takes measures in preparation for
deep, resulting in poor performance. *e SqueezeNet [184] ranking of a set of algorithms with performance in
model achieved approximately the same performance as the descending order. But it can solely inform us about the
AlexNet [81] model. VGG19 [98] and Inception V3 [88] appearance of differences among all samples of results under
showed almost the same level of effectiveness. Although the comparison. Henceforth, its alternatives (e.g., Friedman’s
ResNet [75] model has proven to be a powerful tool for aligned rank test [198] and Quade test [199]) can give us
image classification and is usually fast, it has been shown to further information. Consequently, we have performed the
take a long time to train. Concisely, using all benefits of tests of Friedman [196], Friedman’s aligned rank [198], and
DenseNet [67] with optimization, DenTnet obtained the Quade [199] for average rankings based on the features of
highest GMN ACC of 0.9463, RES of 0.9649, F1S of 0.9531, our experimental study. On rejecting null-hypotheses, we
and AUC of 0.9465 from all four datasets. *is implies that have continued to use post hoc procedures to find the special
DenTnet has the best generalization ability compared to its pairs of algorithms that give idiosyncrasies. In the case of
alternative methods. 1 × N comparisons, the post hoc procedures make up for
Often, it is important to measure that certain deep Bonferroni-Dunn’s [201], Holm’s [202], Hochberg’s [203],
learning models are more efficient and practical as compared Hommel’s [204, 205], Holland and Copenhaver’s [206],
to their alternatives. Seemingly, it is difficult to measure such Rom’s [207], Finner’s [208], and David Li’s [209] proce-
superiority from the obtained experimental results in dures, whereas the post hoc procedures of Nemenyi [210],
Tables 5–10. Nonetheless, nonparametric statistical test can Shaffer [211], and Bergmann-Hommel [212] are involved in
make a clear picture of this issue. N × N comparisons. *e details can be found in the works of
20 Computational Intelligence and Neuroscience
Performance Evaluation Deeming Failures and GMN of RTM statistic, Friedman’s aligned statistic, Iman-Davenport sta-
0.14 tistic, and Quade statistic are 0.000422, 0.000762847204,
0.12 0.000002458229, and 0.000820133186, respectively.
Numerical values
Table 11: Summary of performance failure and RTM scores of miscellaneous deep learning algorithms.
GMN scores of performance failure
Models GMN of RTM
ACC PRS RES F1S AUC
AlexNet [81] 0.0951 0.0979 0.0634 0.0811 0.1036 2762.8
ResNet [75] 0.0578 0.0623 0.0403 0.0506 0.0559 3368.5
VGG16 [98] 0.0855 0.0666 0.0633 0.0658 0.0846 3567.2
VGG19 [98] 0.0672 0.0652 0.0543 0.0603 0.0740 4059.4
Inception V3 [88] 0.0704 0.0554 0.0499 0.0678 0.0761 3597.3
SqueezeNet [184] 0.1142 0.0964 0.0941 0.0969 0.1297 3795.3
DenTnet [ours] 0.0537 0.0650 0.0351 0.0469 0.0535 3196.3
Table 12: Average ranking of each algorithm using nonparametric statistical tests. *e best results are shown in bold.
Multiple comparison tests
Algorithms
Friedman ranking [196] Friedman’s aligned ranking [198] Quade ranking [199]
AlexNet [81] 5.3333 26.0000 4.6189
ResNet [75] 2.1667 09.0000 2.2857
VGG16 [98] 4.6667 27.8333 4.6191
VGG19 [98] 4.0000 21.6667 4.3333
Inception V3 [88] 3.6667 22.1667 4.0952
SqueezeNet [184] 6.6667 36.6667 6.6667
DenTnet [ours] 1.5000 07.1667 1.3809
Various statistics 24.500000 23.102557 5.274194
p value 0.000422 0.000763 0.000820
Average rankings : e tallest bar indicates the best model Figure 11 depicts the Nemenyi [210] post hoc critical
0.8
distance diagrams at three distinct levels of significance α
values. If the distance between algorithms is less than the
0.7
critical distance, then there is no statistically significant
Numerical Values
0.6
difference between them. *e diagrams in Figures 11(a) and
0.5
11(b) associated with α � 0.10 with the critical distance of
0.4
3.3588 and with α � 0.05 with the critical distance of
0.3 3.6768, respectively, are identical, whereas the diagram in
0.2 Figure 11(c) related to α � 0.01 with the critical distance of
0.1 4.3054 is different. Any two algorithms are considered as
0 significantly different if their performance variation is
Friedman [196] F.al.rank [198] Quade [199] greater than the critical distance. To this end, from Fig-
Nonparametric Tests
ure 11, it is noticeable that, at α � 0.01, both SqueezeNet
DenTnet [ours] VGG16 [98] [184] versus DenTnet [ours] and SqueezeNet [184] versus
ResNet [75] AlexNet [81] ResNet [75] are remarkably different, while other pairs are
Inception V3 [88] SqueezeNet [184] not remarkably divergent as their performance differences
VGG19 [98] are less than 4.3054. As compared to ResNet [75], DenTnet
Figure 10: Plotting of average rankings data from Table 12, where [ours] differs from SqueezeNet [184] by a greater distance.
each value x is plotted as 1/x to visualize the highest ranking with On the other hand, SqueezeNet [184] versus DenTnet
the tallest bar. [ours] and AlexNet [81] versus DenTnet [ours] are sig-
nificantly different at both α � 0.10 and α � 0.05, whereas
SqueezeNet [184] versus ResNet [75] is significantly dis-
Shaffer’s [211] procedure rejects those hypotheses that have similar at those α values. Straightforwardly, DenTnet [ours]
an unadjusted p value ≤ 0.004762; and (iv) Bergmann’s is outstandingly unalike both SqueezeNet [184] and
[212] procedure rejects those hypotheses of AlexNet [81] AlexNet [81], but ResNet [75] is not outstandingly unalike
versus DenTnet [ours], ResNet [75] versus SqueezeNet [184], AlexNet [81]. *is implies that the method of DenTnet
and SqueezeNet [184] versus DenTnet [ours]. [ours] outperforms that of ResNet [75], which also agrees
with the finding in Figure 10.
7.4. Critical Distance Diagram from Nemenyi [210] Test.
Nemenyi [210] test is very conservative with a low power,
and hence it is not a recommended choice in practice [215]. 7.5. Reasons of Superiority. In this study, DenseNet [67] was
Nevertheless, it has a unique advantage of having an as- a great choice as it was very compact and deep. It used less
sociated plot to demonstrate the results of fair comparison. training parameters and reduced the risk of model
22 Computational Intelligence and Neuroscience
Table 13: Results achieved on post hoc comparisons for adjusted p values, with α � 0.05 and α � 0.10.
α � 0.05 α � 0.10
Index Algorithms p values
Holm [202] Shaffer [211] Holm [202] Shaffer [211]
1 VGG19 [98] versus Inception V3 [88] 0.789268 0.050000 0.050000 0.100000 0.100000
2 ResNet [75] versus DenTnet [ours] 0.592980 0.025000 0.025000 0.050000 0.050000
3 VGG16 [98] versus VGG19 [98] 0.592980 0.016667 0.016667 0.033333 0.033333
4 AlexNet [81] versus VGG16 [98] 0.592980 0.012500 0.016667 0.025000 0.033333
5 VGG16 [98] versus Inception V3 [88] 0.422678 0.010000 0.016667 0.020000 0.033333
6 AlexNet [81] versus SqueezeNet [184] 0.285049 0.008333 0.008333 0.016667 0.016667
7 AlexNet [81] versus VGG19 [98] 0.285049 0.007143 0.007143 0.014286 0.014286
8 ResNet [75] versus Inception V3 [88] 0.229102 0.006250 0.006250 0.012500 0.012500
9 AlexNet [81] versus Inception V3 [88] 0.181449 0.005556 0.005556 0.011111 0.011111
10 ResNet [75] versus VGG19 [98] 0.141579 0.005000 0.005000 0.010000 0.010000
11 VGG16 [98] versus SqueezeNet [184] 0.108809 0.004545 0.004545 0.009091 0.009091
12 Inception V3 [88] versus DenTnet [ours] 0.082352 0.004167 0.004167 0.008333 0.008333
13 VGG19 [98] versus DenTnet [ours] 0.045021 0.003846 0.003846 0.007692 0.007692
14 ResNet [75] versus VGG16 [98] 0.045021 0.003571 0.003571 0.007143 0.007143
15 VGG19 [98] versus SqueezeNet [184] 0.032509 0.003333 0.003333 0.006667 0.006667
16 Inception V3 [88] versus SqueezeNet [184] 0.016157 0.003125 0.003333 0.006250 0.006667
17 VGG16 [98] versus DenTnet [ours] 0.011118 0.002941 0.003333 0.005882 0.006667
18 AlexNet [81] versus ResNet [75] 0.011118 0.002778 0.003333 0.005556 0.006667
19 AlexNet [81] versus DenTnet [ours] 0.002116 0.002632 0.003333 0.005263 0.006667
20 ResNet [75] versus SqueezeNet [184] 0.000309 0.002500 0.003333 0.005000 0.006667
21 SqueezeNet [184] versus DenTnet [ours] 0.000034 0.002381 0.002381 0.004762 0.004762
Table 14: Adjusted p values for various tests considering DenTnet [ours] as control method.
overfitting and improved the learning rate. In the dense of data for large number of columns. *e dense block of
block of DenTnet, the outputs from the previous layers were DenTnet contained convolution and nonlinear layers, which
concatenated instead of using the summation. *is type of applied several optimization techniques (e.g., dropout and
concatenation helped to markedly speed up the processing BN). DenTnet scaled to hundreds of layers, while exhibiting
Computational Intelligence and Neuroscience 23
Table 15: Adjusted p values of tests for multiple comparisons among all methods.
N × N post hoc procedures and p values
Index Hypothesis
Unadjusted Nemenyi [210] Holm [202] Shaffer [211] Bergmann [212]
1 SqueezeNet [184] versus DenTnet [ours] 0.000034 0.000721 0.000721 0.000721 0.000721
2 ResNet [75] versus SqueezeNet [184] 0.000309 0.006479 0.006171 0.004628 0.004628
3 AlexNet [81] versus DenTnet [ours] 0.002116 0.044428 0.040197 0.031734 0.031734
4 AlexNet [81] versus ResNet [75] 0.011118 0.233469 0.200116 0.166763 0.111176
5 VGG16 [98] versus DenTnet [ours] 0.011118 0.233469 0.200116 0.166763 0.122293
6 Inception V3 [88] versus SqueezeNet [184] 0.016157 0.339296 0.258511 0.242354 0.177726
7 VGG19 [98] versus SqueezeNet [184] 0.032509 0.682698 0.487642 0.487642 0.292585
8 ResNet [75] versus VGG16 [98] 0.045021 0.945439 0.630292 0.495230 0.315146
9 VGG19 [98] versus DenTnet [ours] 0.045021 0.945439 0.630292 0.495230 0.405188
10 Inception V3 [88] versus DenTnet [ours] 0.082352 1.729397 0.988227 0.905874 0.494113
11 VGG16 [98] versus SqueezeNet [184] 0.108809 2.284998 1.196904 1.196904 0.652857
12 ResNet [75] versus VGG19 [98] 0.141579 2.973156 1.415789 1.415789 0.652857
13 AlexNet [81] versus Inception V3 [88] 0.181449 3.810433 1.633043 1.633043 1.270144
14 ResNet [75] versus Inception V3 [88] 0.229102 4.811140 1.832815 1.633043 1.270144
15 AlexNet [81] versus VGG19 [98] 0.285049 5.986038 1.995346 1.995346 1.270144
16 AlexNet [81] versus SqueezeNet [184] 0.285049 5.986038 1.995346 1.995346 1.425247
17 VGG16 [98] versus Inception V3 [88] 0.422678 8.876240 2.113390 2.113390 1.690712
18 AlexNet [81] versus VGG16 [98] 0.592980 12.452582 2.371920 2.371920 1.778940
19 VGG16 [98] versus VGG19 [98] 0.592980 12.452582 2.371920 2.371920 1.778940
20 ResNet [75] versus DenTnet [ours] 0.592980 12.452582 2.371920 2.371920 1.778940
21 VGG19 [98] versus Inception V3 [88] 0.789268 16.574629 2.371920 2.371920 1.778940
(a)
Critical Distance = 3.6768 deeming α=0.05
7 6 5 4 3 2 1 Average ranking
(b)
Critical Distance = 4.3054 deeming α=0.01
7 6 5 4 3 2 1 Average ranking
(c)
Figure 11: Nemenyi [210] post hoc critical distance diagrams for three α values using data in Table 11.
no optimization difficulties. Overall, this model was applied positively affected the performance of the model due to
to a very large number of preprocessed augmented images expansion in the size of training data, which is the foremost
from BreaKHis [33], Malaria [191], SkinCancer [193], and requirement of a deep network for its proper working. Our
CovidXray [192] datasets. To the best of our knowledge, no DenTnet was well trained through various parameters’
other studies in the literature had such an edge. Additionally, tuning. For example, in the case of BreaKHis [33], unlike
the use of data augmentation approach in this study other existing models, our model was trained on all the
24 Computational Intelligence and Neuroscience
magnifications combined (40x, 100x, 200x, and 400x) to attenuated. To test the generalizability of DenTnet, we
avoid any loss of generality. conducted experiments on three additional datasets
In sum and substance, based on the aforementioned (Malaria, SkinCancer, and CovidXray) with varying diffi-
experimental and nonparametric statistical test results, it is, culties. Experimental results on all four datasets demon-
therefore, possible to conclude that the proposed DenTnet strated that DenTnet achieved a better performance in terms
[ours] outperformed AlexNet [81], ResNet [75], VGG16 of accuracy and computational speed than a large number of
[98], VGG19 [98], Inception V3 [88], and SqueezeNet [184] effective state-of-the-art classification methods (AlexNet,
in terms of computational speed. Significantly, the accuracy ResNet, VGG16, VGG19, InceptionV3, and SqueezeNet).
achieved by the proposed DenTnet [ours] surpassed those of *ese findings contributed to our understanding of how a
existing state-of-the-art models in classifying images of the lightweight model could be used to improve the accuracy
BreaKHis [33], Malaria [191], SkinCancer [193], and the and accelerate the learning process of images, including
CovidXray [192] dataset. histopathology image classification on using the wild state-
of-the-art datasets. Future work shall investigate the efficacy
of DenTnet on datasets with multiclass labels.
7.6. Limitation of Proposed Model and Methodology.
Despite these promising results, questions remain as to Abbreviations
whether the proposed DenTnet model could be utilized to
classify categorical images. Moreover, DenTnet was tested BreaKHis: Breast cancer histopathological image
with one breast cancer dataset (i.e., BreaKHis [33]) only. classification
Although the generalization ability of DenTnet with three BACH: Breast cancer histology images
non-breast-cancer-related datasets was studied in Section 6, CNN: Convolutional neural network
it is unknown whether DenTnet can generalize to other DenseNet: Dense convolutional Network
state-of-the-art breast cancer datasets. Future work should, ResNet: Resi du al network
therefore, investigate the efficacy and generalizability of VGG: Visual geometry group at the University
DenTnet with datasets along with multiclass labels, as well as of Oxfor d
other publicly available breast cancer datasets (e.g., the most DDSM: Digital da tabase for screening
recently introduced MITNET dataset [216]). mammography
*e classification effect of breast cancer histopatholog- CBIS − DDSM: Curate d breast imaging subset of DDSM
ical images of any deep learning methodology is related to SVM: Support vector machine
the features and many studies predominantly focused on H&E: Haematoxylin an d eosin
how to develop good feature descriptors and better extract BN: Batch normalization
features. Different from traditional handcrafted feature- ReLU: Rectifie d linear unit
based models, DenTnet can automatically extract more ROC: Receiver operating characteristic
abstract features. Nevertheless, it is worth noting that al- AUC: Area un de r the ROC curve
though the proposed DenTnet has addressed the cross- ACC: Accuracy
domain problem by utilizing the transfer learning approach, GMN: Geometric mean
features extracted in the methodology are solely deep-net- PRS: Precision score
work-based features, which are extracted by feeding images RES: Recall score
directly to the model. However, feeding deep models directly F1S: F1 score
with images would not generalize as the models consider RTM: Runtime.
color distribution of an image. It is understood that local
information can be captured from color images using Local Data Availability
Binary Pattern (LBP) [217]. *erefore, future work can use
multiple types of features by combining the features *e four following publicly available datasets were used in
extracted by the proposed method with LBP features to this study: BreaKHis [33] (https://www.kaggle.com/datasets/
address this issue. ambarish/breakhis), Malaria [191] (https://www.kaggle.
com/iarunava/cell-images-for-detecting-malaria), CovidX-
ray [192] (https://github.com/ieee8023/covid-chestxray-
8. Conclusion dataset), and SkinCancer [193] (https://www.kaggle.com/
We presented that, for classifying breast cancer histopath- fanconic/skin-cancer-malignant-vs-benign).
ological images, the most popular training-testing ratio was
70%: 30%, while the best performance was indicated by the Conflicts of Interest
training-testing ratio of 80%: 20%. We proposed a novel
*e authors have no conflicts of interest to declare.
approach named DenTnet to classify histopathology images
using training-testing ratio of 80%: 20%. DenTnet achieved a
very high classification accuracy on the BreaKHis dataset.
References
Several impediments of existing state-of-the-art methods [1] Who, Breast Cancer Now Most Common Form of Cancer:
including the requirement of high computation and the WHO Taking Action, World Health Organization, Geneva,
utilization of the identical feature distribution were Switzerland, 2021.
Computational Intelligence and Neuroscience 25
[2] A. A. Ewees, L. Abualigah, D. Yousri et al., “Improved Slime [16] Y. Jiang, L. Chen, H. Zhang, and X. Xiao, “Breast cancer
Mould Algorithm Based on Firefly Algorithm for Feature histopathological image classification using convolutional
Selection: A Case Study on QSAR Model,” Engineering with neural networks with small SE-ResNet module,” PLoS One,
Computers, vol. 38, pp. 1–15, 2021. vol. 14, no. 3, Article ID 0214587, 2019.
[3] A. A. Ewees, Z. Y. Algamal, L. Abualigah et al., “A cox [17] Y. Benhammou, S. Tabik, B. Achchab, and F. Herrera, “A
proportional-hazards model based on an improved aquila first study exploring the performance of the state-of-the art
optimizer with whale optimization algorithm operators,” CNN model in the problem of breast cancer,” in Proceedings
Mathematics, vol. 10, no. 8, p. 1273, 2022. of the International Conference on Learning and Optimiza-
[4] M. Abd Elaziz, A. A. Ewees, D. Yousri, L. M. Abualigah, and tion Algorithms: Heory and Applications (LOPAL), no. 1–47,
M. A. A. Al-qaness, “Modified marine predators algorithm p. 47, Rabat, Morocco, June 2018.
for feature selection: case study metabolomics,” Knowledge [18] P. J. Sudharshan, C. Petitjean, F. A. Spanhol, L. E. Oliveira,
and Information Systems, vol. 64, no. 1, pp. 261–287, 2022. L. Heutte, and P. Honeine, “Multiple instance learning for
[5] L. Abualigah, K. H. Almotairi, M. A. Al-qaness et al., “Ef- histopathological breast cancer image classification,” Expert
ficient text document clustering approach using multi-search Systems with Applications, vol. 117, pp. 103–111, 2019.
Arithmetic Optimization Algorithm,” Knowledge-Based [19] S. Sharma and R. Mehra, “Breast cancer histology images
Systems, vol. 248, Article ID 108833, 2022. classification: training from scratch or transfer learning?”
[6] M. Jannesari, M. Habibzadeh, H. Aboulkheyr et al., “Breast ICT Express, vol. 4, pp. 247–254, 2018.
cancer histopathological image classification: a deep learning [20] S. Cascianelli, R. Bello-Cerezo, F. Bianconi et al., “Dimen-
approach,” in Proceedings of the International Conference on sionality reduction strategies for CNN-based classification of
Bioinformatics and Biomedicine, pp. 2405–2412, BIBM, histopathological images,” Intelligent Interactive Multimedia
Madrid, Spain, June 2018. Systems and Services, pp. 21–30, Springer International
[7] D. Bardou, K. Zhang, and S. M. Ahmad, “Classification of Publishing: Cham, New York, NY, USA, 2018.
breast cancer based on histology images using convolutional [21] Y. Song, H. Chang, H. Huang, and W. Cai, “Supervised intra-
neural networks,” IEEE Access, vol. 6, Article ID 24680, 2018. embedding of Fisher vectors for histopathology image
[8] S. H. Kassani, P. H. Kassani, M. J. Wesolowski, classification,” in Proceedings of the International Conference
K. A. Schneider, and R. Deters, “Classification of histo- on Medical Image Computing and Computer Assisted In-
pathological biopsy images using ensemble of deep learning tervention (MICCAI), pp. 99–106, Quebec City, Canada,
networks,” in Proceedings of the Annual International
March 2017.
Conference on Computer Science and Software Engineering [22] B. Wei, Z. Han, X. He, and Y. Yin, “Deep learning model
(CASCON), pp. 92–99, Markham, Ontario, July 2019.
based breast cancer histopathological image classification,”
[9] K. Kumar and A. C. S. Rao, “Breast cancer classification of
in Proceedings of the International Conference on Cloud
image using convolutional neural network,” in Proceedings of
Computing and Big Data Analysis (ICCCBDA), pp. 348–353,
the International Conference on Recent Advances in Infor-
Chengdu, April 2017.
mation Technology (RAIT), pp. 1–6, Dhanbad, India, October
[23] K. Das, S. P. K. Karri, A. G. Roy, J. Chatterjee, and D. Sheet,
2018.
“Classifying histopathology whole-slides using fusion of
[10] M. Z. Alom, C. Yakopcic, M. S. Nasrin, T. M. Taha, and
decisions from deep convolutional network on a collection of
V. K. Asari, “Breast cancer classification from histopatho-
logical images with inception recurrent residual convolu- random multi-views at multi-magnification,” in Proceedings
tional neural network,” Journal of Digital Imaging, vol. 32, of the International Symposium on Biomedical Imaging
no. 4, pp. 605–617, 2019. (ISBI), pp. 1024–1027, Melbourne, Australia, March 2017.
[11] K. Das, S. Conjeti, A. G. Roy, J. Chatterjee, and D. Sheet, [24] Y. Song, J. J. Zou, H. Chang, and W. Cai, “Adapting Fisher
“Multiple instance learning of deep convolutional neural vectors for histopathology image classification,” in Pro-
networks for breast histopathology whole slide classifica- ceedings of the International Symposium on Biomedical
tion,” in Proceedings of the Int. Symposium on Biomedical Imaging (ISBI), pp. 600–603, Melbourne, Australia, April
Imaging (ISBI), pp. 578–581, New York, NY, USA, July 2018. 2017.
[12] A. A. Nahid and Y. Kong, “Histopathological breast-image [25] V. Gupta and A. Bhavsar, “Partially-Independent framework
classification using concatenated R–G–B histogram infor- for breast cancer histopathological image classification,” in
mation,” Annals of Data Science, vol. 6, no. 3, pp. 513–529, Proceedings of the Conference on Computer Vision and
2019. Pattern Recognition Workshops, pp. 1123–1130, Long Beach,
[13] B. Du, Q. Qi, H. Zheng, Y. Huang, and X. Ding, “Breast CA, USA, June 2019.
cancer histopathological image classification via deep active [26] M. Togacar, K. B. Ozkurt, B. Ergen, and Z. Comert,
learning and confidence boosting. Artificial neural networks “BreastNet: a novel convolutional neural network model
and machine learning (ICANN),” in Proceedings of the 27th through histopathological images for the diagnosis of breast
international conference on artificial neural networks, cancer,” Physica A: Statistical Mechanics and Its Applications,
vol. 11140, pp. 109–116, Greece, May 2018. vol. 545, Article ID 123592, 2020.
[14] Z. Gandomkar, P. C. Brennan, and C. Mello-*oms, [27] P. Wang, J. Wang, Y. Li, P. Li, L. Li, and M. Jiang, “Automatic
“MuDeRN: multi-category classification of breast histo- classification of breast cancer histopathological images based
pathological image using deep residual networks,” Artificial on deep feature fusion and enhanced routing,” Biomedical
Intelligence in Medicine, vol. 88, pp. 14–24, 2018. Signal Processing and Control, vol. 65, Article ID 102341,
[15] V. Gupta and A. Bhavsar, “Sequential modeling of deep 2021.
features for breast cancer histopathological image classifi- [28] M. Gour, S. Jain, and T. Sunil Kumar, “Residual learning
cation,” in Proceedings of the Conference on Computer Vision based CNN for breast cancer histopathological image clas-
and Pattern Recognition Workshops, pp. 2254–2261, Salt sification,” International Journal of Imaging Systems and
Lake City, UT, USA, June 2018. Technology, vol. 30, no. 3, pp. 621–635, 2020.
26 Computational Intelligence and Neuroscience
[29] L. Li, X. Pan, H. Yang et al., “Multi-task deep learning for [44] D. A. Ragab, M. Sharkas, S. Marshall, and J. Ren, “Breast
fine-grained classification and grading in breast cancer cancer detection using deep convolutional neural networks
histopathological images,” Multimedia Tools and Applica- and support vector machines,” PeerJ, vol. 7, 2019.
tions, vol. 79, no. 21-22, pp. 14509–14528, 2020. [45] F. P. Romero, A. Tang, and S. Kadoury, “Multi-level batch
[30] D. Albashish, R. Al-Sayyed, A. Abdullah, M. H. Ryalat, and normalization in deep networks for invasive ductal carci-
N. Ahmad Almansour, “Deep CNN Model Based on VGG16 noma cell discrimination in histopathology images,” in
for Breast Cancer Classification,” in Proceedings of the In- Proceedings of the International Symposium on Biomedical
ternational Conference on Information Technology (ICIT), Imaging (ISBI), pp. 1092–1095, Venice, Italy, July 2019.
pp. 805–810, Amman, Jordan, July 2021. [46] H. L. Minh, M. M. Van, and T. V. Lang, “Deep feature fusion
[31] F. Parvin and M. A. Mehedi Hasan, “A comparative study of for breast cancer diagnosis on histopathology images,” in
different types of convolutional neural networks for breast Proceedings of the International Conference on Knowledge
cancer histopathological image classification,” in Proceedings and Systems Engineering (KSE), pp. 1–6, Da Nang, Vietnam,
of the 2020 IEEE Region 10 Symposium (TENSYMP), September 2019.
pp. 945–948, Dhaka, Bangladesh, June 2020. [47] Y. Liu, K. Gadepalli, M. Norouzi et al., “Detecting cancer
[32] Q. Qi, Y. Li, J. Wang et al., “Label-efficient breast cancer metastases on gigapixel pathology images,” 2017, https://
histopathological image classification,” IEEE Journal of arxiv.org/abs/1703.02442.
Biomedical and Health Informatics, vol. 23, no. 5, [48] M. A. Alantari, S. M. Han, and T. S. Kim, “Evaluation of deep
pp. 2108–2116, 2019. learning detection and classification towards computer-
[33] F. A. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte, aided diagnosis of breast lesions in digital X-ray mammo-
“Breast cancer histopathological image classification using grams,” Computer Methods and Programs in Biomedicine,
Convolutional Neural Networks,” in Proceedings of the In- vol. 196, Article ID 105584, 2020.
ternational Joint Conference on Neural Networks (IJCNN), [49] H. Zhang, R. Wu, T. Yuan et al., “DE-Ada∗: a novel model
pp. 2560–2567, Vancouver, BC, 2016. for breast mass classification using cross-modal pathological
[34] F. A. Spanhol, L. S. Oliveira, P. R. Cavalin, C. Petitjean, and semantic mining and organic integration of multi-feature
L. Heutte, A. B. Banff, Deep features for breast cancer fusions,” Information Sciences, vol. 539, pp. 461–486, 2020.
histopathological image classification,” in Proceedings of the [50] N. Brancati, G. De Pietro, D. Riccio, and M. Frucci, “Gig-
International Conference on Systems, Man, and Cybernetics apixel histopathological image analysis using attention-
(SMC), pp. 1868–1873, Canada, CA, USA, 2017. based neural networks,” IEEE Access, vol. 9, Article ID 87552,
[35] Z. Han, B. Wei, Y. Zheng, Y. Yin, K. Li, and S. Li, “Breast
2021.
cancer multi-classification from histopathological images
[51] E. Mahraban Nejad, L. S. Affendey, R. B. Latip, I. B. Ishak,
with structured deep learning model,” Scientific Reports,
and R. Banaeeyan, “Transferred semantic scores for scalable
vol. 7, no. 1, p. 4172, 2017.
retrieval of histopathological breast cancer images,” Inter-
[36] R. Man, P. Yang, and B. Xu, “Classification of breast cancer
national Journal of Multimedia Information Retrieval, vol. 7,
histopathological images using discriminative patches
no. 4, pp. 241–249, 2018.
screened by generative adversarial networks,” IEEE Access,
[52] V. K. Reshma, N. Arya, S. S. Ahmad et al., “Detection of
vol. 8, Article ID 155362, 2020.
breast cancer using histopathological image classification
[37] A. Kumar, S. K. Singh, S. Saxena et al., “Deep feature learning
for histopathological image classification of canine mam- dataset with deep learning techniques,” BioMed Research
mary tumors and human breast cancer,” Information Sci- International, vol. 202213 pages, Article ID 8363850, 2022.
ences, vol. 508, pp. 405–421, 2020. [53] H. Huang, X. Feng, J. Jiang, P. Chen, and S. Zhou, “Mask
[38] N. Bayramoglu, J. Kannala, and J. Heikkilä, “Deep learning RCNN algorithm for nuclei detection on breast cancer
for magnification independent breast cancer histopathology histopathological images,” International Journal of Imaging
image classification,” in Proceedings of the International Systems and Technology, vol. 32, no. 1, pp. 209–217, 2022.
Conference on Pattern Recognition (ICPR), pp. 2440–2445, [54] G. Jayandhi, J. Leena Jasmine, and S. Mary Joans, “Mam-
Cancun, Mexico, December 2016. mogram learning system for breast cancer diagnosis using
[39] J. Sun and A. Binder, “Comparison of deep learning ar- deep learning SVM,” Computer Systems Science and Engi-
chitectures for H&E histopathology images,” in Proceedings neering, vol. 40, no. 2, pp. 491–503, 2022.
of the IEEE Conference on Big Data and Analytics (ICBDA), [55] S. Sharma and S. Kumar, “*e Xception model: a potential
pp. 43–48, Kuching, Malaysia, November 2017. feature extractor in breast cancer histology images classifi-
[40] C. Kaushal and A. Singla, “Automated segmentation tech- cation,” ICT Express, vol. 8, no. 1, pp. 101–108, 2022.
nique with self-driven post-processing for histopathological [56] H. Zerouaoui and A. Idri, “Deep hybrid architectures for
breast cancer images,” CAAI Trans. Intell. Technol.vol. 5, binary classification of medical breast cancer images,” Bio-
no. 4, pp. 294–300, 2020. medical Signal Processing and Control, vol. 71, Article ID
[41] M. Talo, “Convolutional neural networks for multi-class 103226, 2022.
histopathology image classification,” 2019, https://arxiv.org/ [57] W. Zhi, H. W. F. Yeung, Z. Chen, S. M. Zandavi, Z. Lu, and
ftp/arxiv/papers/1903/1903.10035.pdf. Y. Y. Chung, “Using transfer learning with convolutional
[42] H. Elmannai, M. Hamdi, and A. AlGarni, “Deep learning neural networks to diagnose breast cancer from histopath-
models combining for breast cancer histopathology image ological images,” International Conference on Neural In-
classification,” International Journal of Computational In- formation Processing (ICONIP), China, vol. 10637,
telligence Systems, vol. 14, no. 1, pp. 1003–1013, 2021. pp. 669–676, 2017.
[43] Z. Hameed, S. Zahia, B. Garcia-Zapirain, J. Javier Aguirre, [58] J. Chang, J. Yu, T. Han, H. Chang, and E. Park, “A method
and A. Marı́a Vanegas, “Breast cancer histopathology image for classifying medical images using transfer learning: a pilot
classification using an ensemble of deep learning models,” study on histopathology of breast cancer,” in Proceedings of
Sensors, vol. 20, no. 16, p. 4373, 2020. the International Conference on E-Health Networking,
Computational Intelligence and Neuroscience 27
Applications and Services (Healthcom), pp. 1–4, Dalian, cancer classification from ultrasound images,” Computers &
China, June 2017. Electrical Engineering, vol. 70, pp. 871–882, 2018.
[59] S. A. Hassan, M. S. Sayed, M. I. Abdalla, and M. A. Rashwan, [74] WHO, “Breast Cancer,” 2021, https://www.who.int/news-
“Breast cancer masses classification using deep convolutional room/fact-sheets/detail/breast-cancer.
neural networks and transfer learning,” Multimedia Tools [75] K. He, X. Zhang, S. Ren, and J. Sun, “Identity Mappings in
and Applications, vol. 79, no. 41-42, pp. 30735–30768, 2020. Deep Residual Networks,” in Proceedings of the European
[60] M. F. I. Soumik, A. Z. B. Aziz, and M. A. Hossain, “Improved Conference on Computer Vision (ECCV), vol. 9908,
transfer learning based deep learning model for breast cancer pp. 630–645, Netherlands, Europe, 2016.
histopathological image classification,” in Proceedings of the [76] N. B. C. Foundation, Biopsy, *e National Breast Cancer
2021 International Conference on Automation, Control and Foundation, Texas, TX, USA, 2018.
Mechatronics for Industry 4.0 (ACMI), pp. 1–4, Rajshahi, [77] Q. Abbas, “DeepCAD: a computer-aided diagnosis system
Bangladesh, June 2021. for mammographic masses using deep invariant features,”
[61] H. Mahmoud, A. H Alharbi, and D. S Khafga, “Breast cancer Computers, vol. 5, no. 4, p. 28, 2016.
classification using deep convolution neural network with [78] G. Aresta, T. Araujo, S. Kwok et al., “BACH: grand challenge
transfer learning,” Intelligent Automation & Soft Computing, on breast cancer histology images,” Medical Image Analysis,
vol. 29, no. 3, pp. 803–814, 2021. vol. 56, pp. 122–139, 2019.
[62] C. Munien and S. Viriri, “Classification of hematoxylin and [79] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,”
eosin-stained breast cancer histology microscopy images Nature, vol. 521, no. 7553, pp. 436–444, 2015.
using transfer learning with EfficientNets,” Computational [80] A. Chan and J. A. Tuszynski, “Automatic prediction of tu-
Intelligence and Neuroscience, vol. 2021, pp. 1–17, Article ID mour malignancy in breast cancer with fractal dimension,”
5580914, 2021. Royal Society Open Science, vol. 3, no. 12, Article ID 160558,
[63] S. Boumaraf, X. Liu, Z. Zheng, X. Ma, and C. Ferkous, “A 2016.
new transfer learning based approach to magnification de- [81] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet
pendent and independent classification of breast cancer in classification with deep convolutional neural networks.
Advances in Neural Information Processing Systems 25,” in
histopathological images,” Biomedical Signal Processing and
Proceedings of the 26th Annual Conference on Neural In-
Control, vol. 63, Article ID 102192, 2021.
formation Processing Systems 2012, pp. 1106–1114, Lake
[64] A. Saber, M. Sakr, O. M. Abo-Seida, A. Keshk, and H. Chen,
Tahoe, Nevada, USA, October 2012.
“A novel deep-learning model for automatic detection and
[82] H. Bay, T. Tuytelaars, and L. V. Gool, “SURF: Speeded Up
classification of breast cancer using the transfer-learning
Robust Features,” in Proceedings of the European Conference
technique,” IEEE Access, vol. 9, pp. 71194–71209, 2021.
on Computer Vision ECCV, vol. 3951, pp. 404–417, Graz,
[65] S. Soltane, S. Alshreef, and S. MSerag Eldin, “Classification
Austria, July 2006.
and diagnosis of lymphoma’s histopathological images using
[83] Z. Guo, L. Zhang, and D. Zhang, “Rotation invariant texture
transfer learning,” Computer Systems Science and Engi-
classification using LBP variance (LBPV) with global
neering, vol. 40, no. 2, pp. 629–644, 2022. matching,” Pattern Recognition, vol. 43, no. 3, pp. 706–719,
[66] M. Alruwaili and W. Gouda, “Automated breast cancer
2010.
detection models based on transfer learning,” Sensors, [84] S. F. University, “Digital Database for Screening Mam-
vol. 22, no. 3, p. 876, 2022. mography,” 2006, http://www.eng.usf.edu/cvprg/
[67] G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, mammography/database.html.
“Densely Connected Convolutional Networks,” in Pro- [85] J. Suckling and J. Parker, “Mammographic Image Analysis
ceedings of the Conference on Computer Vision and Pattern Society (MIAS) Database v1.21 [Dataset],” 2015, https://
Recognition (CVPR), pp. 2261–2269, Honolulu, HI, USA, www.repository.cam.ac.uk/handle/1810/250394.
July 2017. [86] C. Szegedy, W. Liu, Y. Jia et al., “Going deeper with con-
[68] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE volutions,” in Proceedings of the Conference on Computer
Transactions on Knowledge and Data Engineering, vol. 22, Vision and Pattern Recognition (CVPR), pp. 1–9, Boston,
no. 10, pp. 1345–1359, 2010. MA, USA, July 2015.
[69] S. Z. Ramadan, “Methods used in computer-aided diagnosis [87] L. Shamir, N. Orlov, D. Mark Eckley, T. J. Macura, and
for breast cancer detection using mammograms: a review,” I. G. Goldberg, “IICBU 2008: a proposed benchmark suite
Journal of Healthcare Engineering, vol. 2020, pp. 1–21, Article for biological image analysis,” Medical, & Biological Engi-
ID 9162464, 2020. neering & Computing, vol. 46, no. 9, pp. 943–947, 2008.
[70] Y. Benhammou, B. Achchab, F. Herrera, and S. Tabik, [88] C. Szegedy, W. Liu, Y. Jia et al., “Going deeper with con-
“BreakHis based breast cancer automatic diagnosis using volutions,” in Proceedings of the Conference on Computer
deep learning: taxonomy, survey and insights,” Neuro- Vision and Pattern Recognition (CVPR), pp. 1–9, Boston,
computing, vol. 375, pp. 9–24, 2020. MA, USA, July 2015.
[71] D. A. Zebari, D. A. Ibrahim, D. Q. Zeebaree et al., “Sys- [89] M. A. Kahya, W. Al-Hayani, and Z. Y. Algamal, “Classifi-
tematic review of computing approaches for breast cancer cation of breast cancer histopathology images based on
detection based computer aided diagnosis using mammo- adaptive sparse support vector machine,” Journal of Applied
gram images,” Applied Artificial Intelligence, vol. 35, no. 15, Mathematics and Bioinformatics, vol. 7, pp. 49–69, 2017.
pp. 2157–2203, 2021. [90] V. Gupta and A. Bhavsar, “An integrated multi-scale model
[72] K. R. Weiss, T. M. Khoshgoftaar, and D. Wang, “A survey of for breast cancer histopathological image classification with
transfer learning,” J. Big Data, vol. 3, no. 1, p. 9, 2016. joint colour-texture features,” in Proceedings of the Inter-
[73] M. A. Mohammed, B. Al-Khateeb, A. N. Rashid, national Conference on Computer Analysis of Images and
D. A. Ibrahim, M. K. Abd Ghani, and S. A. Mostafa, “Neural Patterns (CAIP), vol. 10425, pp. 354–366, Ystad, Sweden,
network and multi-fractal dimension features for breast August 2017.
28 Computational Intelligence and Neuroscience
[91] N. Dhungel, G. Carneiro, and A. P. Bradley, “A deep learning Conference on Intelligent Tutoring Systems (ITS), Virtual
approach for the analysis of masses in mammograms with Event, vol. 12677, pp. 343–349, 2021.
minimal user intervention,” Medical Image Analysis, vol. 37, [108] D. S. Morillo, J. Gonzalez, M. G. Rojo, and J. Ortega,
pp. 114–128, 2017. “Classification of breast cancer histopathological images
[92] I. C. Moreira, I. Amaral, I. Domingues, A. Cardoso, using KAZE features,” Int. Work-Conference on Bio-
M. J. Cardoso, and J. S. Cardoso, “INbreast: toward a full- informatics and Biomedical Engineering (IWBBIO), Spain,
field digital mammographic database,” Academic Radiology, vol. 10814, pp. 276–286, 2018.
vol. 19, no. 2, pp. 236–248, 2012. [109] P. F. Alcantarilla, A. Bartoli, and A. J. Davison, “KAZE
[93] Y. Jia, E. Shelhamer, J. Donahue et al., “Caffe: Convolutional Features,” in Proceedings of the European Conference on
Architecture for Fast Feature Embedding,” in Proceedings of Computer Vision (ECCV), vol. 7577, pp. 214–227, Florence,
the ACM Int. Conf. on Multimedia (MM), pp. 675–678, Italy, June 2012.
Florida, FL, USA, 2014. [110] S. Chattoraj and K. Vishwakarma, “Classification of histo-
[94] S. Kaymak, A. Helwan, and D. Uzun, “Breast cancer image pathological breast cancer images using iterative VMD aided
classification using artificial neural networks. Procedia Zernike moments & textural signatures,” 2018, https://arxiv.
Computer Science,” in Proceedings of the International org/abs/1801.04880.
Conference on Heory and Application of Soft Computing, [111] T. *eodoridis, K. Loumponias, N. Vretos, and P. Daras,
Computing with Words and Perception (ICSCCW), vol. 120, “Zernike pooling: generalizing average pooling using zernike
pp. 126–131, Budapest, Hungary, June 2017. moments,” IEEE Access, vol. 9, pp. 121128–121136, 2021.
[95] R. Hecht-Nielsen, “*eory of the backpropagation neural [112] N. Lassance and F. Vrins, “Minimum Rényi entropy port-
network,” Neural Networks, vol. 1, pp. 445–448, 1988. folios,” Annals of Operations Research, vol. 299, no. 1-2,
[96] C. Bishop, “Improving the generalization properties of radial pp. 23–46, 2021.
basis function neural networks,” Neural Computation, vol. 3, [113] M. Rahimi and M. Mohammadi Anjedani, “A local view on
no. 4, pp. 579–588, 1991. the Hudetz correction of the Yager entropy of dynamical
[97] Camelyon16, “Challenge on Cancer Metastases Detection in systems,” International Journal of General Systems, vol. 48,
Lymph Node,” 2016, https://camelyon16.grand-challenge. no. 3, pp. 321–333, 2019.
org. [114] Y. Zheng, Z. Jiang, H. Zhang et al., “Size-scalable content-
[98] K. Simonyan and A. Zisserman, “Very deep convolutional based histopathological image retrieval from database that
networks for large-scale image recognition,” in Proceedings consists of WSIs,” IEEE Journal of Biomedical and Health
of the 3rd International Conference on Learning Represen- Informatics, vol. 22, no. 4, pp. 1278–1287, 2018.
tations (ICLR), San Diego, CA, USA, July 2015. [115] R. W. Hamming, “Error detecting and error correcting
[99] S. J. Stanford, “Tissue Microarray Database,” 2021, https:// codes,” Bell System Technical Journal, vol. 29, no. 2,
tma.im/cgi-bin/home.pl. pp. 147–160, 1950.
[100] A. A. Nahid, A. Mikaelian, and Y. Kong, “Histopathological [116] R. Mukkamala, P. S. Neeraja, S. Pamidi, T. Babu, and
breast-image classification with restricted Boltzmann ma- T. Singh, “Deep PCANet framework for the binary cate-
chine along with backpropagation,” Biomedical Research, gorization of breast histopathology images,” in Proceedings
vol. 29, pp. 2068–2077, 2018. of the International Conference on Advances in Computing,
[101] W. Lu, C. S. Leung, and J. Sum, “Analysis on noisy Boltz- Communications and Informatics (ICACCI), pp. 105–110,
mann machines and noisy restricted Boltzmann machines,” India, September 2018.
IEEE Access, vol. 9, pp. 112955–112965, 2021. [117] T. H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma,
[102] H. Tamura, S. Mori, and T. Yamawaki, “Textural features “PCANet: a simple deep learning baseline for image clas-
corresponding to visual perception,” IEEE Trans. Syst. Man sification?” IEEE Transactions on Image Processing, vol. 24,
Cybern.vol. 8, no. 6, pp. 460–473, 1978. no. 12, pp. 5017–5032, 2015.
[103] J. A. Badejo, E. Adetiba, A. Akinrinmade, and M. B. Akanle, [118] A. Rakhlin, A. Shvets, V. Iglovikov, and A. A. Kalinin, “Deep
“Medical image classification with hand-designed or ma- convolutional neural networks for breast cancer histology
chine-designed texture descriptors: a performance evalua- image analysis,” in Proceedings of the International Con-
tion,” International Work-Conference on Bioinformatics and ference on Image Analysis and Recognition (ICIAR),
Biomedical Engineering (IWBBIO), Granada, Spain, vol. 10882, pp. 737–744, Portugal, July 2018.
vol. 10814, pp. 266–275, 2018. [119] M. A. Almasni, M. A. Alantari, J. M. Park et al., “Simulta-
[104] P. Alirezazadeh, B. Hejrati, A. Monsef-Esfahani, and neous detection and classification of breast masses in digital
A. Fathi, “Representation learning-based unsupervised do- mammograms via a deep learning YOLO-based CAD sys-
main adaptation for classification of breast cancer histopa- tem,” Computer Methods and Programs in Biomedicine,
thology images,” Biocybernetics and Biomedical Engineering, vol. 157, pp. 85–94, 2018.
vol. 38, no. 3, pp. 671–683, 2018. [120] A. G. Howard, M. Zhu, B. Chen et al., “MobileNets: Efficient
[105] J. Spencer and K. S. John, “Random sparse bit strings at the Convolutional Neural Networks for Mobile Vision Appli-
threshold of adjacency,” in Proceedings of the Annual sym- cations,” 2017, https://arxiv.org/abs/1704.04861.
posium on theoretical aspects of computer science (STACS), [121] B. S. Veeling, J. Linmans, J. Winkens, T. Cohen, and
vol. 1373, pp. 94–104, Paris, France, October 1998. M. Welling, “Rotation equivariant CNNs for digital pa-
[106] C. Y. Maa and M. A. Shanblatt, “Linear and quadratic thology,” Int. Conf. on Medical Image Computing and
programming neural network analysis,” IEEE Transactions Computer Assisted Intervention (MICCAI), Spain, vol. 11071,
on Neural Networks, vol. 3, no. 4, pp. 580–594, 1992. pp. 210–218, 2018.
[107] F. Giannakas, C. Troussas, A. Krouska, C. Sgouropoulou, [122] A. Pego and P. Aguiar, “Bioimaging Challenge 2015 Breast
and I. Voyiatzis, “XGBoost and deep neural network com- Histology Dataset,” 2015, http://www.bioimaging2015.ineb.
parison: the case of teams’ performance,” International up.pt/dataset.html.
Computational Intelligence and Neuroscience 29
[123] R. Lenz and P. L. Carmona, “Transform Coding of RGB- images,” IEEE/ACM Transactions on Computational Biology
Histograms,” in Proceedings of the International Conference and Bioinformatics, vol. 17, no. 1, pp. 91–101, 2020.
on Computer Vision Heory and Applications (VISAPP), [139] E. J. Zaferani, M. Teshnehlab, and M. Vali, “Automatic
pp. 117–124, Lisboa, Portugal, November 2009. personality traits perception using asymmetric auto-en-
[124] J. Hu, L. Shen, and G. Sun, “Squeeze-and-Excitation net- coder,” IEEE Access, vol. 9, pp. 68595–68608, 2021.
works,” in Proceedings of the Conference on Computer Vision [140] Y. LeCun, B. E. Boser, J. S. Denker et al., “Backpropagation
and Pattern Recognition (CVPR), pp. 7132–7141, Salt Lake applied to handwritten zip code recognition,” Neural
City, UT, USA, May 2018. Computation, vol. 1, no. 4, pp. 541–551, 1989.
[125] D. M. Vo, N. Q. Nguyen, and S. W. Lee, “Classification of [141] R. H. Carvalho, A. S. Martins, L. A. Neves, and M. Z. do
breast cancer histology images using incremental boosting Nascimento, “Analysis of features for breast cancer recog-
convolution networks,” Information Sciences, vol. 482, nition in different magnifications of histopathological im-
pp. 123–138, 2019. ages,” in Proceedings of the International Conference on
[126] M. Babaie, S. Kalra, A. Sriram et al., “Classification and Systems, Signals and Image Processing (IWSSIP), pp. 39–44,
Retrieval of Digital Pathology Scans,” in Proceedings of the A Brazil, August 2020.
New Dataset. Conference on Computer Vision and Pattern [142] M. H. Sharif and C. Djeraba, “An entropy approach for
Recognition Workshops, pp. 760–768, CVPR Workshops, abnormal activities detection in video streams,” Pattern
Honolulu, HI, USA, July 2017. Recognition, vol. 45, no. 7, pp. 2543–2561, 2012.
[127] X. Li, M. Radulovic, K. Kanjer, and K. N. Plataniotis, [143] A. Toomaj and H. A. Atabay, “Some new findings on the
“Discriminative pattern mining for breast cancer histopa- cumulative residual Tsallis entropy,” Journal of Computa-
thology image classification via fully convolutional tional and Applied Mathematics, vol. 400, Article ID 113669,
autoencoder,” IEEE Access, vol. 7, pp. 36433–36445, 2019. 2022.
[128] R. S. Lee, F. Gimenez, A. Hoogi, K. K. Miyake, M. Gorovoy, [144] J. Li, J. Zhang, Q. Sun et al., “Breast cancer histopathological
and D. L. Rubin, “A curated mammography data set for use image classification based on deep second-order pooling
in computer-aided detection and diagnosis research,” Sci- network,” in Proceedings of the Int. Joint Conf. On Neural
entific Data, vol. 4, no. 1, Article ID 170177, 2017. Networks (IJCNN), pp. 1–7, London, U K, November 2020.
[129] C. Roa, “Data from: High-*roughput Adaptive Sampling [145] P. Li, J. Xie, Q. Wang, and Z. Gao, “Towards Faster Training
for Whole-Slide Histopathology Image Analysis (HASHI) of Global Covariance Pooling Networks by Iterative Matrix
via Convolutional Neural Networks: Application to Invasive Square Root Normalization,” in Proceedings of the Confer-
Breast Cancer Detection,” 2018, https://datadryad.org/stash/
ence on computer vision and pattern recognition (CVPR),
dataset/doi:10.5061/dryad.1g2nt41.
pp. 947–955, Salt Lake City, UT, USA, September 2018.
[130] P. Stanitsas, A. Cherian, V. Morellas, R. Tejpaul,
[146] L. Shen, L. R. Margolies, J. H. Rothstein, E. Fluder,
N. Papanikolopoulos, and A. Truskinovsky, “Image de-
R. McBride, and W. Sieh, “Deep learning to improve breast
scriptors for weakly annotated histopathological breast
cancer detection on screening mammography,” Scientific
cancer data,” Frontiers in Digital Health, vol. 2, Article ID
Reports, vol. 9, no. 1, Article ID 12495, 2019.
572671, 2020.
[147] G. Li, C. Li, G. Wu, D. Ji, and H. Zhang, “Multi-view at-
[131] M. A. Asan and A. Ozsoy, “cuRCD: region covariance de-
tention-guided multiple instance detection network for in-
scriptor CUDA implementation,” Multimedia Tools and
terpretable breast cancer histopathological image diagnosis,”
Applications, vol. 80, no. 13, pp. 19737–19751, 2021.
[132] T. G. Dietterich, R. H. Lathrop, and T. Lozano-Pérez, IEEE Access, vol. 9, pp. 79671–79684, 2021.
“Solving the multiple instance problem with axis-parallel [148] R. Yan, F. Ren, Z. Wang et al., “Breast cancer histopatho-
rectangles,” Artificial Intelligence, vol. 89, no. 1-2, pp. 31–71, logical image classification using a hybrid deep neural
1997. network,” Methods, vol. 173, pp. 52–60, 2020.
[133] A. H. Fischer, K. A. Jacobson, J. Rose, and R. Zeller, “He- [149] S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing
matoxylin and eosin staining of tissue and cell sections,” CSH between capsules Advances in Neural Information Pro-
Protocols, vol. 2008, 2008. cessing Systems 30,” in Proceedings of the Annual Conference
[134] S. K. Asare, F. You, and O. T. Nartey, “A semisupervised on Neural Information Processing Systems, pp. 3856–3866,
learning scheme with self-paced learning for classifying Long Beach, CA, USA, April 2017.
breast cancer histopathological images,” Computational [150] J. Kundale and S. Dhage, “Classification of breast cancer
Intelligence and Neuroscience, vol. 2020, pp. 1–16, Article ID using histology images: handcrafted and pre-trained features
8826568, 2020. based approach,” IOP Conference Series: Materials Science
[135] F. Chollet, “Xception: Deep Learning with Depthwise Sep- and Engineering, vol. 1074, no. 1, Article ID 012008, 2021.
arable Convolutions,” in Proceedings of the Conference on [151] A. R. H. Khayeat, X. Sun, and P. L. Rosin, “Improved DSIFT
Computer Vision and Pattern Recognition (CVPR), descriptor based copy-rotate-move forgery detection. Image
pp. 1800–1807, Honolulu, HI, USA, May 2017. and video technology - 7th pacific-rim symposium (PSIVT),
[136] K. Dimitropoulos, P. Barmpoutis, C. Zioga, A. Kamas, auckland,” New Zealand, vol. 9431, pp. 642–655, 2015.
K. Patsiaoura, and N. Grammalidis, “Grading of invasive [152] J. Wang, J. Yang, K. Yu, F. Lv, T. S. Huang, and Y. Gong,
breast carcinoma through Grassmannian VLAD encoding,” “Locality-constrained linear coding for image classification,”
PLoS One, vol. 12, no. 9, Article ID e0185110, 2017. in Proceedings of the Conference on Computer Vision and
[137] A. Janowczyk and A. Madabhushi, “Grading of invasive Pattern Recognition (CVPR), pp. 3360–3367, San Francisco,
breast carcinoma through Grassmannian VLAD encoding,” USA, January 2010.
Journal of Pathology Informatics, vol. 7, Article ID 27563488, [153] O. Attallah, F. Anwar, N. M. Ghanem, and M. A. Ismail,
2016. “Histo-CADx: duo cascaded fusion stages for breast cancer
[138] Y. Feng, L. Zhang, and J. Mo, “Deep manifold preserving diagnosis from histopathological images,” PeerJ Computer
autoencoder for classifying breast cancer histopathological Science, vol. 7, p. e493, 2021.
30 Computational Intelligence and Neuroscience
[154] K. C. Burçak, Ö. K. Baykan, and H. Uguz, “A new deep Pattern Recognition (CVPR), pp. 248–255, Miami, Florida,
convolutional neural network model for classifying breast USA, April 2009.
cancer histopathological images and the hyperparameter [169] A. Ameh Joseph, M. Abdullahi, S. B. Junaidu, H. Hassan
optimisation of the proposed model,” He Journal of Ibrahim, and H. Chiroma, “Improved multi-classification of
Supercomputing, vol. 77, no. 1, pp. 973–989, 2021. breast cancer histopathological images using handcrafted
[155] H. Iimori, G. T. F. de Abreu, O. Taghizadeh, R. A. Stoica, features and deep neural network (dense layer),” Intelligent
T. Hara, and K. Ishibashi, “A stochastic gradient descent Systems with Applications, vol. 14, Article ID 200066, 2022.
approach for hybrid mmWave beamforming with blockage [170] P. C. Chhipa, R. Upadhyay, G. G. Pihlgren, R. Saini,
and CSI-error robustness,” IEEE Access, vol. 9, S. Uchida, and M. Liwicki, “Magnification prior: a self-su-
pp. 74471–74487, 2021. pervised method for learning representations on breast
[156] A. Botev, G. Lever, and D. Barber, “Nesterov’s accelerated cancer histopathological images,” 2022, https://arxiv.org/
gradient and momentum as approximations to regularised abs/2203.07707.
update descent,” in Proceedings of the International Joint [171] Y. Zou, J. Zhang, S. Huang, and B. Liu, “Breast cancer
Conference on Neural Networks (IJCNN), pp. 1899–1903, histopathological image classification using attention high-
Anchorage, AK, USA, August 2017. order deep network,” International Journal of Imaging
[157] A. Byerly and T. Kalganova, “Homogeneous vector capsules Systems and Technology, vol. 32, no. 1, pp. 266–279, 2022.
enable adaptive gradient descent in convolutional neural [172] M. Liu, Y. He, M. Wu, and C. Zeng, “Breast histopathological
networks,” IEEE Access, vol. 9, pp. 48519–48530, 2021. image classification method based on autoencoder and si-
[158] N. Shi, D. Li, M. Hong, and R. Sun, “RMSprop converges amese framework,” Information, vol. 13, no. 3, p. 107, 2022.
with proper hyper-parameter,” in Proceedings of the Inter- [173] D. A. Naik, R. M. Mohana, G. Ramu, Y. S. Lalitha,
national Conference on Learning Representations (ICLR), M. SureshKumar, and K. V. Raghavender, “Analyzing his-
Virtual Event, Austria, February 2021. topathological images by using machine learning tech-
[159] Z. Qu, S. Yuan, R. Chi, L. Chang, and L. Zhao, “Genetic niques,” Applied Nanoscience, pp. 1–7, 2022.
optimization method of pantograph and catenary compre- [174] S. Chattopadhyay, A. Dey, P. K. Singh, and R. Sarkar,
hensive monitor status prediction model based on adadelta “DRDA-Net: dense residual dual-shuffle attention network
deep neural network,” IEEE Access, vol. 7, pp. 23210–23221, for breast cancer classification using histopathological im-
2019. ages,” Computers in Biology and Medicine, vol. 145, Article
[160] I. K. M. Jais, A. R. Ismail, and S. Q. Nisa, “Adam optimi-
ID 105437, 2022.
zation algorithm for wide and deep neural network,”
[175] F. Shahidi, S. Mohd Daud, H. Abas, N. A. Ahmad, and
Knowledge Engineering and Data Science, vol. 2, no. 1,
N. Maarop, “Breast cancer classification using deep learning
pp. 41–46, 2019.
approaches and histopathology image: a comparison study,”
[161] I. Hirra, M. Ahmad, A. Hussain et al., “Breast cancer clas-
IEEE Access, vol. 8, pp. 187531–187552, 2020.
sification from histopathological images using patch-based
[176] C. Kampf, I. Olsson, U. Ryberg, E. Sjostedt, and F. Ponten,
deep learning modeling,” IEEE Access, vol. 9, pp. 24273–
“Production of tissue microarrays, immunohistochemistry
24287, 2021.
staining and digitalization within the human protein atlas,”
[162] G. E. Hinton, S. Osindero, and Y. W. Teh, “A fast learning
algorithm for deep belief nets,” Neural Computation, vol. 18, Journal of Visualized Experiments, vol. 31, no. 63, p. 3620,
no. 7, pp. 1527–1554, 2006. 2012.
[163] Q. B. Baker and A. Abu Qutaish, “Evaluation of histo- [177] H. A. Shehu, M. H. Sharif, M. H. U. Sharif et al., “Deep
pathological images segmentation techniques for breast sentiment analysis: a case study on stemmed Turkish twitter
cancer detection,” in Proceedings of the International Con- data,” IEEE Access, vol. 9, pp. 56836–56854, 2021.
ference on Information and Communication Systems (ICICS), [178] M. H. Sharif, “An eigenvalue approach to detect flows and
pp. 134–139, Valencia, Spain, May 2021. events in crowd videos,” Journal of Circuits, Systems, and
[164] B. Ehteshami Bejnordi, M. Veta, P. Johannes van Diest et al., Computers, vol. 26, no. 07, Article ID 1750110, 2017.
“Diagnostic assessment of deep learning algorithms for [179] H. A. Shehu, M. H. Sharif, and R. A. Ramadan, “Distributed
detection of lymph node metastases in women with breast mutual exclusion algorithms for intersection traffic prob-
cancer,” JAMA, vol. 318, no. 22, pp. 2199–2210, 2017. lems,” IEEE Access, vol. 8, pp. 138277–138296, 2020.
[165] M. Veta, Y. J. Heng, N. Stathonikos et al., “Predicting breast [180] P. Jungklass and M. Berekovic, “Static allocation of basic
tumor proliferation from whole-slide images: the TUPAC16 blocks based on runtime and memory requirements in
challenge,” Medical Image Analysis, vol. 54, pp. 111–121, embedded real-time systems with hierarchical memory
2019. layout,” in Proceedings of the Second Workshop on Next
[166] M. Z. D. Nascimento, A. S. Martins, L. A. Neves, R. P. Ramos, Generation Real-Time Embedded Systems, vol. 87, no. 1–3,
E. L. Flôres, and G. A. Carrijo, “Classification of masses in pp. 3–14, Budapest, Hungary, February 2021.
mammographic image using wavelet domain features and [181] I. Loshchilov and F. Hutter, “SGDR: stochastic gradient
polynomial classifier,” Expert Systems with Applications, descent with warm restarts,” in Proceedings of the Interna-
vol. 40, no. 15, pp. 6213–6221, 2013. tional Conference on Learning Representations (ICLR),
[167] M. Tan and Q. V. Le, “EfficientNet: rethinking model scaling pp. 1–16, Toulon, France, September 2017.
for convolutional neural networks,” in Proceedings of the [182] C. Zhang, P. Benz, D. M. Argaw et al., “ResNet or DenseNet?
36th International Conference on Machine Learning (ICML), Introducing Dense Shortcuts to ResNet,” in Proceedings of
vol. 97, pp. 6105–6114, Long Beach, California, USA, De- the Winter Conference on Applications of Computer Vision
cember 2019. (WACV), pp. 3549–3558, HI, USA, December 2021.
[168] J. Deng, W. Dong, R. Socher, L. J. Li, K. Li, and L. F. Fei, [183] I. L. S. V. R. C. ImageNet, “Large Scale Visual Recognition
“ImageNet: A Large-Scale Hierarchical Image Database,” in Challenge (ILSVRC),” 2010, https://image-net.org/
Proceedings of the Conference on Computer Vision and challenges/LSVRC.
Computational Intelligence and Neuroscience 31
[184] F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, [201] O. J. Dunn, “Multiple comparisons among means,” Journal
W. J. Dally, and K. Keutzer, “SqueezeNet: AlexNet-level of the American Statistical Association, vol. 56, no. 293,
accuracy with 50x fewer parameters and <1MB model size,” pp. 52–64, 1961.
CoRR, 2016, https://arxiv.org/abs/1602.07360?context�cs. [202] S. Holm, “A simple sequentially rejective multiple test
[185] S. Kornblith, J. Shlens, and Q. V. Le, “Do Better ImageNet procedure,” Scandinavian Journal of Statistics, vol. 6,
Models Transfer Better?” in Proceedings of the Conference on pp. 65–70, 1979.
Computer Vision and Pattern Recognition (CVPR), [203] Y. Hochberg, “A sharper Bonferroni procedure for multiple
pp. 2661–2671, Canada CA, USA, March 2019. tests of significance,” Biometrika, vol. 75, no. 4, pp. 800–802,
[186] M. N. Gurcan, L. Boucheron, A. Can et al., “Histopatho- 1988.
logical image analysis: a review,” IEEE Rev. Biomed. [204] G. Hommel, “A stagewise rejective multiple test procedure
Eng.vol. 2, pp. 147–171, 2009. based on a modified Bonferroni test,” Biometrika, vol. 75,
[187] M. Macenko, M. Niethammer, J. S. Marron et al., “A method no. 2, pp. 383–386, 1988.
for normalizing histology slides for quantitative analysis,” in [205] G. Hommel and G. Bernhard, “A rapid algorithm and a
Proceedings of the International Symposium on Biomedical computer program for multiple test procedures using logical
Imaging: From Nano to Macro, pp. 1107–1110, Boston, MA, structures of hypotheses,” Computer Methods and Programs
USA, June 2009. in Biomedicine, vol. 43, no. 3-4, pp. 213–216, 1994.
[188] Keras, Keras API, 2021. [206] B. S. Holland and M. D. Copenhaver, “An improved se-
[189] H. A. Shehu, R. A. Ramadan, and M. H. Sharif, “Artificial quentially rejective Bonferroni test procedure,” Biometrics,
vol. 43, no. 2, pp. 417–423, 1987.
intelligence tools and their capabilities,” PLOMS AI, p. 1,
[207] D. M. Rom, “A sequentially rejective test procedure based on
2021.
a modified Bonferroni inequality,” Biometrika, vol. 77, no. 3,
[190] D. P. Kingma, J. Ba, and Adam, “A Method for Stochastic
pp. 663–665, 1990.
Optimization,” in Proceedings of the International Confer-
[208] H. Finner, “On a monotonicity problem in step-down
ence on Learning Representations (ICLR), Y. Bengio and
multiple test procedures,” Journal of the American Statistical
Y. LeCun, Eds., San Diego, CA, USA, May 2015. Association, vol. 88, no. 423, pp. 920–923, 1993.
[191] N. I. H. Malaria, “Datasets of National Institutes of Health [209] J. David Li, “A two-step rejection procedure for testing
(NIH),” 2021, https://www.kaggle.com/iarunava/cell- multiple hypotheses,” Journal of Statistical Planning and
images-for-detecting-malaria. Inference, vol. 138, no. 6, pp. 1521–1527, 2008.
[192] Kaggle, “CoronaHack - chest X-ray-dataset,” 2021, https:// [210] P. Nemenyi, Distribution-free Multiple Comparisons, PhD
github.com/ieee8023/covid-chestxray-dataset. thesis, Princeton University, New Jersey, NY, USA, 1963.
[193] Kaggle and Skin Cancer, “Malignant vs. Benign,” 2021, [211] J. P. Shaffer, “Modified sequentially rejective multiple test
https://www.kaggle.com/fanconic/skin-cancer-malignant- procedures,” Journal of the American Statistical Association,
vs-benign. vol. 81, no. 395, pp. 826–831, 1986.
[194] H. A. Shehu, W. Browne, and H. Eisenbarth, “An Adversarial [212] G. Bergmann and G. Hommel, “Improvements of general
Attacks Resistance-Based Approach to Emotion Recognition multiple test proceduresfor redundant systems of hypoth-
from Images Using Facial Landmarks,” in Proceedings of the eses,” in Multiple Hypotheses Testing, P. Bauer and
IEEE International Conference on Robot and Human In- G. Hommel, Eds., pp. 100–115, Springer, New York, NY,
teractive Communication (RO-MAN), pp. 1307–1314, USA, 1988.
Naples, Italy, August 2020. [213] S. Garcı́a and F. Herrera, “An extension on ”Statistical
[195] M. H. Sharif and C. Djeraba, “A simple method for eccentric comparisons of classifiers over multiple data sets” for all
event espial using mahalanobis metric. Progress in pattern pairwise comparisons,” Journal of Machine Learning Re-
recognition, image analysis, computer vision, and applica- search, vol. 9, pp. 2677–2694, 2008.
tions,” in Proceedings of the 14th iberoamerican conference on [214] G. University, “Soft Computing and Intelligent Information
pattern recognition (CIARP), vol. 5856, pp. 417–424, Gua- Systems,” 2020, https://sci2s.ugr.es/sicidm.
dalajara, Mexico, April 2009. [215] B. Calvo and G. Santafé, “Scmamp: statistical comparison of
[196] M. Friedman, “*e use of ranks to avoid the assumption of multiple algorithms in multiple problems,” Rice Journal,
normality implicit in the analysis of variance,” Journal of the vol. 8, no. 1, pp. 248–256, 2016.
American Statistical Association, vol. 32, no. 200, pp. 675– [216] S. Cayir, G. Solmaz, H. Kusetogullari et al., “MITNET: a
701, 1937. novel dataset and a two-stage deep learning approach for
[197] R. L. Iman and J. M. Davenport, “Approximations of the mitosis recognition in whole slide images of breast cancer
critical region of the fbietkan statistic,” Communications in tissue,” Neural Computing & Applications, pp. 1–15, 2022.
Statistics - Heory and Methods, vol. 9, no. 6, pp. 571–595, [217] Q. U. Ain, H. Al-Sahaf, B. Xue, and M. Zhang, “Generating
1980. knowledge-guided discriminative features using genetic
[198] J. L. Hodges and E. L. Lehmann, “Rank methods for com- programming for melanoma detection,” IEEE Trans. Emerg.
bination of independent experiments in analysis of vari- Top. Comput. Intell.vol. 5, no. 4, pp. 554–569, 2021.
ance,” He Annals of Mathematical Statistics, vol. 33, no. 2,
pp. 482–497, 1962.
[199] D. Quade, “Using weighted rankings in the analysis of
complete blocks with additive block effects,” Journal of the
American Statistical Association, vol. 74, no. 367, pp. 680–
683, 1979.
[200] P. Westfall and S. Young, Resampling-based Multiple Testing:
Examples and Methods for P-Value Adjustment, John Wiley
& Sons, New Jersey, NY, USA, 2004.