Received 15 March 2024, accepted 9 May 2024, date of publication 16 May 2024, date of current version 24 May 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3401729

FastLeakyResNet-CIR: A Novel Deep Learning

Framework for Breast Cancer Detection and
Classification
RUIMING ZENG1 , BOAN QU2 , WEI LIU 2 , JIANGHAO LI1 , HONGSHEN LI1 ,
PINGPING BING 1 , SHUANGNI DUAN3 , AND LEMEI ZHU1
1 Hunan Provincial Key Laboratory of the Research and Development of Novel Pharmaceutical Preparations, Changsha Medical University, Changsha 410219,
China
2 College of Mechanical and Electrical Engineering, Beijing University of Chemical Technology, Beijing 100029, China
3 School of Public Health, Xiangnan University, Chenzhou 423000, China

Corresponding authors: Pingping Bing ([email protected]), Shuangni Duan ([email protected]), and Lemei Zhu
([email protected])
This work was supported in part by the National Key Research and Development Program of China under Grant 2018YFB2000800 and
Grant 2022YFB3303600; in part by Hunan Provincial Education Department under Grant 20A056, Grant 22C0669, and Grant 22C0565;
and in part by Hunan Provincial Health Commission Foundation under Grant D202302088596.

ABSTRACT Breast cancer is a type of disease that primarily affects the breast tissue, and it is crucial
to achieve early diagnosis for successful treatment and recovery. In recent years, the residual network
(ResNet) has gained significant attention in the detection of breast cancer using medical images. In this paper,
we propose an efficient and robust deep learning framework called FastLeakyResNet-CIR, an improved
ResNet architecture, for breast cancer detection and classification. The FastLeakyResNet-CIR achieves an
impressive accuracy of 98.94% when evaluated on a dataset of 7909 microscopic images of breast tumor
tissue from BreakHis dataset, which outperforms the state-of-the-art methods, e.g. ResNet18, ResNet50,
InceptionV3 and VGG16. The experiment results further highlight the potential of FastLeakyResNet-CIR
for accurate and rapid diagnosis of breast cancer, thus facilitating effective medical treatment for patients.

INDEX TERMS Breast cancer, convolutional neural network, FastLeakyResNet-CIR, medical image
classification, residual network.

I. INTRODUCTION most popular methods because of the relative high-accuracy,

Breast cancer poses a significant risk to women’s lives and low-cost, and high detectability, which is able to provide
overall health, making it a pressing concern. Globally, breast an effective imaging tool for breast cancer identifica-
cancer stands out as one of the most commonly diagnosed tion and classification [7]. However, the performance of
forms of cancer among women [1], [2], [3], [4], [5]. Egypt, mammography may be weak in some cases, especially in
in particular, has witnessed a rapid rise in cancer case, with patients with dense breast tissue [8]. Thermography is an
breast cancer being prominently affected. Detecting breast emerging breast cancer screening technology by analyzing
cancer at an early stage plays a pivotal role in ensuring a thermogram obtained by a thermal infrared camera. It is
effective treatment and reducing mortality rate. characterized by radiation-free, low-cost, and non-invasive.
Several diagnosis techniques have currently been Compared with mammography, breast thermography has
employed for early detection of breast cancer [6], such its own advantages including its ability to applicable to
as mammography, thermography, ultrasound, and magnetic individuals with dense breast tissue and effectiveness for
resonance imaging (MRI). Mammography is one of the all age women [9]. Therefore, thermography is regarded as
a safe, quick and effective early diagnosis technique for
The associate editor coordinating the review of this manuscript and patients with breast cancer. Ultrasound is a primary tool
approving it for publication was Humaira Nisar . for breast cancer screening, which provides a potentially
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 70825
 R. Zeng et al.: FastLeakyResNet-CIR: A Novel Deep Learning Framework

viable solution for early breast cancer detection because outperforms several architectures, e.g. ResNet 50, SeResNet
it is easy to carry, and low cost than mammography [10]. 50 and Inception. Mohiyuddin et al. [24] put forward a
It plays a crucial role in differentiating between solid-cystic modified network of YOLOv5 to detect and classify breast
breast lesions and characterizing solid lesions. Thanks tumors, and it performs better than YOLOv3 and faster
to technological advances, high-frequency ultrasonography regions with CNN (RCNN) with lowering the false positive
probes have enabled the visualization of even small breast ratio (FPR) and false negative ratio (FNR) and boosting
lesions, especially for patients with dense breast [11]. the Matthews correlation coefficient (MCC) value. Asadi
Therefore, it is the ideal imaging tool to guide subsequent and Memon [25] employed a cascade network with U-Net
procedures. Magnetic resonance imaging is primarily utilized for segmentation and a ResNet for efficient breast cancer
as a supplemental means to breast cancer screening along detection, which has a classification accuracy of 98.61% and
with mammography or ultrasound [12]. Since 2000, the F1 score of 98.41%.
breast MRI, with excellent sensitivity and specificity, has To improve the accuracy and interpretability of breast
become an important tool in high risk screening and diagnosis cancer detection and classification, we propose a novel deep
for earlier-stage breast cancer by measuring the size of the learning framework called fast leaky residual network with
cancer, and looking for other tumors in the breast. Recent class imbalance reduction, termed as FastLeakyResNet-CIR,
research has found that the MRI is capable of locating which has higher accuracy and faster convergence rate. The
some small breast lesions that may be sometimes missed by main contribution of this study is as follows.
mammography [13], thus, the MRI has been extensively used (1) The novel scheme FastLeakyResNet-CIR is presented
for patients with newly diagnosed breast cancer. for predicting breast cancer.
Nowadays, the methods based on deep learning (DL) (2) We strengthen the robustness of the model by
have been extensively applied to medical image analysis introducing the improved loss function and improved Adam
and breast cancer diagnosis [14], [15], [16]. Convolutional algorithm.
neural network (CNN) is one of the most popular approaches (3) The proposed method performs well in a performance
in DL [17], [18], which achieves an end-to-end mapping comparison against the state-of-the-art breast cancer predic-
learning from input image to output result and becomes an tion methods.
efficient classification model [19]. Subsequently, a variety of The rest of the paper is organized as follows. Section II
DL models are emerging, such as RNN, AlexNet, GoogleNet, provides the theoretical basis. Section III introduces a
ResNet, VGG, and U-Net. Tan et al. [20] utilized CNN detailed description of the proposed breast cancer detection
acting on mammogram imaging system to produce three scheme. Section IV presents the experiments and results.
classifications of normal, benign and malignant, which assists Section 5 gives the conclusions.
specialist to diagnosis and classify the breast cancer quickly.
Mambous et al. [21] developed a deep neural network (DNN) II. METHODOLOGY
model that is related to pre-trained Inception V3 model Figure 1 shows the details of the proposed breast cancer
for sick and healthy breast classification, where they pay detection scheme that consists of three main steps. First,
attention to the physical model camera sensitivity of breast. the data preprocessing aims to reduce training time, enhance
Ting et al. [22] presented a convolutional neural network image, alleviate overfitting and learn feature, which includes
improvement for breast cancer classification (CNNI-BCC), several techniques such as resizing and balance, Gamma
which can classify breast cancer medical images into healthy, transform, data augmentation and shuffling and splitting.
benign and malignant patients in the context of no prior Second, the FastLeakyResNet-CIR is designed to predict
information about cancerous lesion. It improves the breast breast cancer. Third, the improved loss function and improved
cancer lesion classification and helps experts for breast cancer Adam algorithm are introduced to strengthen the robustness
identification. Barbosa et al. [23] introduced the deep-wavelet of the model.
neural network (DWNN) for feature extraction in image
representation, combining it with intelligent algorithms for A. DATA AND PROCESSING
pattern recognition. They demonstrated a sensitivity of 95% 1) RESIZING AND BALANCE
and specificity of 79% using the DWNN for detection of In this study, we utilize the BreaKHis dataset, Breast Cancer
breast lesions in thermographic images. While promising, Histopathological Database, to validate the proposed scheme.
there are some limitations about the DWNN approach. The dataset was previously used by Spanhol et al. [26], which
It relies on handcrafted feature extraction, which may not contains 7909 breast cancer histopathology images acquired
capture discriminative features as effectively as end-to-end on 82 patients with two states benign and malignant. Since
feature learning methods like CNNs. Also, their evaluation the number of normal training sets is much smaller than the
was on a small proprietary dataset, so further validation on number of abnormal training sets in the dataset, we replicate
larger benchmark datasets would be needed. Naseem et al. the normal training sets itself to be the same as the number of
[9] proposed a computer-aided detection approach based on the two training sets. To cut down the data size, all images are
CNN using thermal images, it is faster, reliable and robust, resized to 224*224 before training, which will significantly
and obtains a higher accuracy (92%) and F1-score (92%) that reduce the training time of the model.
70826 VOLUME 12, 2024
R. Zeng et al.: FastLeakyResNet-CIR: A Novel Deep Learning Framework

FIGURE 2. (a) The image before Gamma transform and (b) the image
after Gamma transform.

scaling, and flipping the original images, we effectively

expand the dataset and provide the model with a richer set of
training examples. This approach allows the model to learn
robust features that are invariant to these transformations,
improving its ability to generalize well to unseen data. The
increased diversity through data augmentation is beneficial
to alleviate overfitting and enhance the model’s performance
across different scenarios and conditions.

4) SHUFFLING AND SPLITTING

Shuffling is able to prevent model oscillation during the
training, which is beneficial for the robustness of the model.
Moreover, it can also hinder overfitting and help the model
learn more correct features. Therefore, it’s necessary to
shuffle the dataset. Then, we divide the dataset into training
set, validation set and test set in the ratio of 8:1:1.

B. NETWORK ARCHITECTURE AND IMPLEMENTATION

The FastLeakyResNet-CIR, an improved ResNet, is designed
for breast cancer detection, constructed by four convolutional
blocks, adaptive average pooling layer, linear layer and
output layer. Each convolutional block is comprised of five
convolutional layers with the sizes of 3 × 3 × 64, 3 ×
3 × 128, 3 × 3 × 256 and 3 × 3 × 512, respectively,
which gradually extract more sophisticated features. The
FIGURE 1. Illustration of the breast cancer detection scheme, (a) breast adaptive average pooling layer is applied to reduce the spatial
cancer data preprocessing, (b) a FastLeakyResNet-CIR, and (c) parameters dimension of the feature map, enhance the robustness of
updating strategy.
the model and make it suitable for the image with varying
size. The LeakyReLU as an activation function enables the
model to have good nonlinearity and generalization abilities.
2) GAMMA TRANSFORM
In addition, the ResNet residual blocks with skip connection
The Gamma transform is employed for image enhancement, help to train the model efficiently and prevent the vanishing
which boosts the overall brightness of the image in the darker gradient when increasing the network depth. The output layer
detail and increases the contrast at lower grey level, making it has two states, benign and malignant, which are represented
easier to distinguish image detail at lower grey value. During in binary as 1 or 0. The structure of FastLeakyResNet-CIR is
the experiment, the gray value of the images in the training shown in Figure 3.
sets is low, but the results can be effectively enhanced by
setting the gamma value to 0.7, so as to improve the training C. PARAMETERS UPDATING
accuracy of the model. The comparison before and after
1) LOSS FUNCTION
Gamma transform is shown in Figure 2.
In binary classification task, the loss function based on
cross-entropy is widely used, which is defined as follows:
3) DATA AUGMENTATION
Data augmentation is a powerful means that helps expose the N
1 X
model to various transformations and variations commonly L(y, ŷ) = − (yi log(ŷi ) + (1 − yi ) log(1 − ŷi )) (1)
N
found in real-world data. By randomly rotating, translating, i=1

VOLUME 12, 2024 70827

 R. Zeng et al.: FastLeakyResNet-CIR: A Novel Deep Learning Framework

FIGURE 3. Proposed FastLeakyResNet-CIR framework designed for breast cancer detection.

where N represents the number of samples, L is loss weights ωmalignant and ωbenign in the loss function to account
function, y and ŷ denotes the true value and predicted value, for class imbalance between malignant and benign samples.
respectively. The motivation is to give more influence to minority
However, the traditional loss function suffers from some malignant samples during training. By assigning a higher
drawbacks, for example, 1) it does not cope well with weight ωmalignant to losses on malignant samples, we increase
imbalanced data, resulting in a bias towards the majority their contribution to update the model parameters. This
samples and far from the minority samples; 2) it does not counters the tendency of overwhelmed malignant examples.
make full use of confidence information, lacking targeted The exact values can be tuned as a hyperparameter, with
penalty for uncertain samples; 3) it does not impose common choices being the inverse class frequencies or a fixed
the constraints on the complexity of the model, leading rebalancing ratio. The λ weight for the confidence penalty
to potential overfitting. The above-mentioned issues may is set to 0.1 based on preliminary experiments. As a result,
degrade the performance of the model when dealing with we obtain a detection model for breast cancer that focuses
imbalanced data and uncertain samples, and the ability of more on important class, possesses robustness, and performs
generalizing to new data. well on unseen data.
In the paper, we introduce the category weights, confidence
penalty and L1 regularization into the above loss function,
thus, Eq. (1) can be rewritten as:
2) OPTIMIZER
L(y, ŷ, c) Adam (Adaptive Moment Estimation), proposed by Kingma
N and Ba [27], is an optimizer for optimizing stochastic gradient
1 X
=− (ωmalignant • yi log(ŷi ) descent algorithm for neural network, which can be seen as a
N combination of Momentum and RMSProp.
i=1
+ ωbenign (1 − yi ) log(1 − ŷi )) + λH (c, t) + α • ∥w∥1 The traditional Adam algorithm ignores the stability of
(2) gradient and only considers the first-order and the second-
order moment estimation. It may cause suboptimal learning
where the weights of malignant samples and benign samples rate when dealing with gradient that exhibits high variance or
are ωmalignant and ωbenign , respectively, c is the predicted instability. Secondly, the Adam lacks personalized learning
confidence, t is the threshold, H (•) is the penalty term and rate adjustment, because it utilizes fixed decay rate or step
the corresponding weight is λ, ∥•∥1 denotes L1 regularization to adjust the learning rate without taking into account the
term and the corresponding weight is α, w is the weight vector difference in parameter importance or sensitivity. Thirdly,
parameter. the Adam relies only on global statistical information and
By employing category weights, we can improve the fails to adequately consider the feature of each individual
problem of class imbalance, and enhance the impact of breast batch.
cancer samples in the optimization process. Considering In the paper, we bring some improvements to the traditional
confidence information, we can strengthen the robustness Adam algorithm as follows.
and stability of the proposed model. Introducing a L1 regu- 1) Computing the gradient gt for the current step.
larization term, we can promote the model’s generalization 2) Calculating the stability term st , which is the standard
ability and prevent the overfitting. We introduce category deviation of the gradient plus a small constant.

70828 VOLUME 12, 2024

R. Zeng et al.: FastLeakyResNet-CIR: A Novel Deep Learning Framework

Algorithm 1 Improved Adam

Input :
α : Step size;
β1 ,β2 ∈ [0, 1): Exponential decay rates for the moment estimate;
(θ ): Stochastic objective function with parameter θ .
Output : θt : Resulting parameter.
1 Initial parameter vectorθ0
2 m0 ← 0 (Initialize first moment vector)
3 v0 ← 0 (Initialize second moment vector)
4 t ← 0 (Initialize time step)
5 while θt not converge do
6t ←t +1
7 gt ← ∇θ ft (θt−1 ) (Get the gradient of the current step)
8 mt ← β1 mt−1 + (1 − β1 )gt (Update first moment estimate)
9 vt ← β2 vt−1 + (1 − β2 )g2t (Update second moment estimate)
10 m′t ← mt /(1 − β1t ) (Compute bias-corrected first moment
estimate)
11 v′t ← vt /(1 − β2t ) (Compute bias-corrected second moment
estimate)p
12 st ← v′t + ε (Compute the stability term)
13 ρt ← gstt (Compute the gradient stability ratio) FIGURE 4. Confusion matrix of FastLeakyResNet-CIR on test set.
14 αt ← α • max(0, 1 − ρt ) (Compute the adjusted learning rate)
15 θt ← θt−1 −αm′t (Update parameters with adjusted learning rate) TABLE 1. Comparison with some classic models.
end while

3) Calculating the gradient stability ratio ρt by dividing the

gradient by the stability term.
4) Computing the adjusted learning rate αt based on the
gradient stability ratio.
5) Updating the parameter θt by means of the adjusted
learning rate. III. RESULT AND EVALUATION
Algorithm 1 describes the improved Adam algorithm. This section provides a detailed description of the exper-
We compute st as the standard deviation of gradient gt plus iment results to verify the performance of the proposed
a small constant ε. Using the standard deviation can quantify FastLeakyResNet-CIR breast cancer detection method.
the dispersion of gradients, allowing us to gauge the level
of noise. The constant ε (set to 1e − 8) ensures numerical A. EVALUATION INDICATORS
stability when the variance is near zero. By incorporating st , In order to quantitatively assess the performance of the
we can adjust the learning rate according to the stability of proposed model, the classification results are evaluated using
the gradients. Parameters with noisy gradients are assigned four metrics including accuracy, recall, precision, and F1
lower learning rates to smooth the updates. This provides score [7], which are defined as:
a personalized learning rate adaptation for each parameter TP + TN
tailored to the stability of its gradients. We introduce an Accuracy = (3)
TP + FP + FN + TN
adjusted learning rate αt based on the gradient stability
TP
ratio ρt . The motivation is to scale the learning rate for Recall = (4)
each parameter according to the stability of its gradient. TP + FN
TP
Parameters with noisy gradients get assigned a lower learning Precision = (5)
rate for that iteration to dampen the update steps. This allows TP + FP
2 × Precision × Recall
for more personalized and adaptive tuning of the learning F1score = (6)
rate, as opposed to using a single global value. Setting the Precision + Recall
learning rate based on ρt provides a data-driven way to where TP is the true positive, FP is the false positive, FN
stabilize training and accelerate convergence by smoothing is the false negative, and TN is the true negative. It is
out fluctuations in the gradient. noted that the precision and recall is a pair of contradictory
In this experiment, we use CUDA in the PyTorch quantities, the former denotes the probability of predicting
framework based on RTX 3090 GPU for training to ensure a positive case as a positive sample and the latter is the
the speed and accuracy of the model. The epoch is set to probability of predicting a true positive case as a positive
60, batch size is 64, and the initial learning rate is set to sample. Thus, we employ the F1score to examine the
0.03. The training process will take earlystop action when no precision and recall of the model in a comprehensive
improvement occurs in the loss of the validation set. way.

VOLUME 12, 2024 70829

 R. Zeng et al.: FastLeakyResNet-CIR: A Novel Deep Learning Framework

such as ResNet18, ResNet50, InceptionV3 and VGG16.

We calculate the accuracy of each model on the test set
and the corresponding convergence rate, and the results are
shown in Figure 5. And, the accuracy, recall, precision,
and F1 score values are listed Table 1. The comparison
results indicate that the FastLeakyResNet-CIR has higher
classification accuracy compared to other neural network
models, therefore, it will have strong applicability in the field
of physiological signal analysis. Meanwhile, it is noteworthy
that the FastLeakyResNet-CIR converges around the 15th
epoch because of improved loss function and improved Adam
algorithm, which is significantly faster than other approaches,
in other words, it can effectively reduce the time needed to
train the model.

IV. CONCLUSION
In this study, we present a novel FastLeakyResNet-CIR
model for breast cancer detection and classification. The
proposed framework is composed of convolutional blocks,
adaptive average pooling layer, linear layer and output layer.
It reaches the highest classification accuracy of 98.94%. Also,
the proposed breast cancer prediction scheme achieves an
impressive score in a performance comparison against several
classic breast cancer detection approaches. In addition, the
FastLeakyResNet-CIR effectively solves the problem of class
imbalance, avoids the overfitting and has good generalization
ability and faster convergence speed by introducing the
improved loss function and improved Adam algorithm, thus,
it is a very promising and application-worthy deep learning
model in the field of medical image analysis. In future
work, data enhancement methods and breast cancer severity
diagnosis will be explored.
FIGURE 5. Accuracy (a) and loss (b) comparison results obtained by five
methods.
ACKNOWLEDGMENT
The authors would like to thank an associate editor Prof.
B. RESULT AND COMPARISON Shashikant Patil, Humaira Nisar, and anonymous reviewers,
In this experiment, the iterative process of FastLeakyResNet- for their constructive comments that improved the article
CIR is relative stable, and it converges at the 15th Epoch. greatly.
The whole training process ends early at the 31st Epoch.
The FastLeakyResNet-CIR achieves better classification REFERENCES
accuracy, recall, precision, and F1-score values 98.94%, [1] T. Mahmood, J. Li, Y. Pei, F. Akhtar, A. Imran, and K. U. Rehman, ‘‘A
98.36%, 99.43% and 98.89%, respectively, which illustrates brief survey on breast cancer diagnostic with deep learning schemes using
multi-image modalities,’’ IEEE Access, vol. 8, pp. 165779–165809, 2020.
the effectiveness of the proposed method for detecting breast [2] Y. Gan, L. Li, L. Zhang, S. Yan, C. Gao, S. Hu, Y. Qiao, S. Tang,
cancer. Figure 4 shows the confusion matrix for the test C. Wang, and Z. Lu, ‘‘Association between shift work and risk of prostate
phase of the proposed breast cancer detection framework. cancer: A systematic review and meta-analysis of observational studies,’’
Carcinogenesis, vol. 39, no. 2, pp. 87–97, Feb. 2018.
It can be clearly observed that the 3280 images are tested, [3] X. Wang, G. Chen, Y. Zhang, W. M. Ghareeb, Q. Yu, H. Zhu, X. Lu,
1590 samples are classified as normal, and 1690 samples Y. Huang, S. Huang, D. Hou, and P. Chi, ‘‘The impact of circumferential
are classified as abnormal. Among the normal samples, tumour location on the clinical outcome of rectal cancer patients managed
with neoadjuvant chemoradiotherapy followed by total mesorectal exci-
1564 samples are correctly predicted as normal (TP), while sion,’’ Eur. J. Surgical Oncol., vol. 46, no. 6, pp. 1118–1123, Jun. 2020.
26 samples are incorrectly predicted as abnormal (FN). [4] X. Xie, X. Wang, Y. Liang, J. Yang, Y. Wu, L. Li, X. Sun, P. Bing,
Furthermore, 9 abnormal samples are wrongly predicted B. He, G. Tian, and X. Shi, ‘‘Evaluating cancer-related biomarkers based
on pathological images: A systematic review,’’ Frontiers Oncol., vol. 11,
as normal (FP), and 1681 abnormal samples are correctly Nov. 2021, Art. no. 763527.
predicted as abnormal (TN). [5] P. Peng, Y. Luan, P. Sun, L. Wang, X. Zeng, Y. Wang, X. Cai, P. Ren, Y. Yu,
In addition, for further clarifying the superiority of Q. Liu, H. Ma, H. Chang, B. Song, X. Fan, and Y. Chen, ‘‘Prognostic
factors in stage IV colorectal cancer patients with resection of liver
the FastLeakyResNet-CIR, a comparison is made between and/or pulmonary metastases: A population-based cohort study,’’ Frontiers
the proposed method with some classical neural networks Oncol., vol. 12, Mar. 2022, Art. no. 850937.

70830 VOLUME 12, 2024

R. Zeng et al.: FastLeakyResNet-CIR: A Novel Deep Learning Framework

[6] N. Wang, J. Chen, W. Chen, Z. Shi, H. Yang, P. Liu, X. Wei, X. Dong, [25] B. Asadi and Q. Memon, ‘‘Efficient breast cancer detection via cascade
C. Wang, L. Mao, and X. Li, ‘‘The effectiveness of case management for deep learning network,’’ Int. J. Intell. Netw., vol. 4, pp. 46–52, 2023.
cancer patients: An umbrella review,’’ BMC Health Services Res., vol. 22, [26] F. A. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte, ‘‘A dataset
no. 1, p. 1247, Oct. 2022. for breast cancer histopathological image classification,’’ IEEE Trans.
[7] Y. Yari, T. V. Nguyen, and H. T. Nguyen, ‘‘Deep learning applied Biomed. Eng., vol. 63, no. 7, pp. 1455–1462, Jul. 2016.
for histological diagnosis of breast cancer,’’ IEEE Access, vol. 8, [27] D. P. Kingma and L. J. Ba, ‘‘Adam: A method for stochastic optimization,’’
pp. 162432–162448, 2020. in Proc. Int. Conf. Learn. Represent. (ICLR), 2015.
[8] A. U. Haq, J. P. Li, A. Saboor, J. Khan, S. Wali, S. Ahmad, A. Ali,
G. A. Khan, and W. Zhou, ‘‘Detection of breast cancer through clinical data
using supervised and unsupervised feature selection techniques,’’ IEEE
Access, vol. 9, pp. 22090–22105, 2021.
[9] U. Naseem, J. Rashid, L. Ali, J. Kim, Q. E. U. Haq, M. J. Awan,
and M. Imran, ‘‘An automatic detection of breast cancer diagnosis and
prognosis based on machine learning using ensemble of classifiers,’’ IEEE RUIMING ZENG received the master’s degree
Access, vol. 10, pp. 78242–78252, 2022. in arts from Changsha University of Science and
[10] R. Sood, A. F. Rositch, D. Shakoor, E. Ambinder, K.-L. Pool, E. Pollack, Technology, Changsha, China, in 2011. He is
D. J. Mollura, L. A. Mullen, and S. C. Harvey, ‘‘Ultrasound for breast currently a Lecturer with Changsha Medical Uni-
cancer detection globally: A systematic review and meta-analysis,’’ J. versity. His research interests include linguistics,
Global Oncol., no. 5, pp. 1–17, Dec. 2019. philosophy, and medical science.
[11] C.-X. Chen, J. Zhu, and Z. Zeng, ‘‘Use of ultrasound to observe mycosis
fungoides: A case report and review of literature,’’ Current Med. Imag.
Formerly Current Med. Imag. Rev., vol. 18, no. 7, pp. 771–775, Jul. 2022.
[12] R. Huo, Y. Liu, H. Xu, J. Li, R. Xin, Z. Xing, S. Deng, T. Wang,
H. Yuan, and X. Zhao, ‘‘Associations between carotid atherosclerotic
plaque characteristics determined by magnetic resonance imaging and
improvement of cognition in patients undergoing carotid endarterectomy,’’
Quant. Imag. Med. Surg., vol. 12, no. 5, pp. 2891–2903, May 2022.
[13] S. Radhakrishna, S. Agarwal, P. M. Parikh, K. Kaur, S. Panwar, S. Sharma,
A. Dey, K. K. Saxena, M. Chandra, and S. Sud, ‘‘Role of magnetic
resonance imaging in breast cancer management,’’ South Asian J. Cancer, BOAN QU is currently pursuing the bachelor’s
vol. 7, no. 2, pp. 69–71, 2018. degree in safety engineering with Beijing Uni-
[14] X. Yi, X. Guan, C. Chen, Y. Zhang, Z. Zhang, M. Li, P. Liu, A. Yu, X. Long, versity of Chemical Technology. His research
L. Liu, B. T. Chen, and C. Zee, ‘‘Adrenal incidentaloma: Machine learning- interests include fault diagnosis and monitoring
based quantitative texture analysis of unenhanced CT can effectively of chemical equipment and solving engineering
differentiate sPHEO from lipid-poor adrenal adenoma,’’ J. Cancer, vol. 9, problems based on machine learning.
no. 19, pp. 3577–3582, 2018.
[15] B. He, J. Lang, B. Wang, X. Liu, Q. Lu, J. He, W. Gao, P. Bing,
G. Tian, and J. Yang, ‘‘TOOme: A novel computational framework to infer
cancer tissue-of-origin by integrating both gene mutation and expression,’’
Frontiers Bioengineering Biotechnol., vol. 8, p. 394, May 2020.
[16] B. He, Y. Zhang, Z. Zhou, B. Wang, Y. Liang, J. Lang, H. Lin, P. Bing,
L. Yu, D. Sun, H. Luo, J. Yang, and G. Tian, ‘‘A neural network framework
for predicting the tissue-of-origin of 15 common cancer types based
on RNA-seq data,’’ Frontiers Bioengineering Biotechnol., vol. 8, p. 737,
Aug. 2020. WEI LIU received the B.E. degree in exploration
[17] B. He, Q. Lu, J. Lang, H. Yu, C. Peng, P. Bing, S. Li, Q. Zhou, Y. Liang, and geophysics from China University of Petroleum
G. Tian, ‘‘A new method for CTC images recognition based on machine (East China), Qingdao, China, in 2006, and the
learning,’’ Frontiers Bioeng. Biotechnol., vol. 8, p. 897, Aug. 2020. M.S. degree in exploration geophysics and the
[18] P. Bing, Y. Liu, W. Liu, J. Zhou, and L. Zhu, ‘‘Electrocardiogram Ph.D. degree in geological resources and geo-
classification using TSST-based spectrogram and ConViT,’’ Frontiers
logical engineering from China University of
Cardiovascular Med., vol. 9, Oct. 2022, Art. no. 983543.
Petroleum (Beijing), Beijing, China, in 2009 and
[19] A. Raza, N. Ullah, J. A. Khan, M. Assam, A. Guzzo, and H. Aljuaid,
2016, respectively. He is currently a Lecturer
‘‘DeepBreastCancerNet: A novel deep learning model for breast cancer
detection using ultrasound images,’’ Appl. Sci., vol. 13, no. 4, p. 2082, with Beijing University of Chemical Technology.
Feb. 2023. His research interests include signal analysis and
[20] Y. J. Tan, K. S. Sim, and F. F. Ting, ‘‘Breast cancer detection using processing, seismic data processing and interpretation, and mechanical fault
convolutional neural networks for mammogram imaging system,’’ in Proc. diagnosis.
Int. Conf. Robot., Autom. Sci. (ICORAS), Nov. 2017, pp. 1–5.
[21] S. Mambou, P. Maresova, O. Krejcar, A. Selamat, and K. Kuca, ‘‘Breast
cancer detection using infrared thermal imaging and a deep learning
model,’’ Sensors, vol. 18, no. 9, p. 2799, Aug. 2018.
[22] F. F. Ting, Y. J. Tan, and K. S. Sim, ‘‘Convolutional neural network
improvement for breast cancer classification,’’ Expert Syst. Appl., vol. 120,
pp. 103–115, Apr. 2019. JIANGHAO LI is currently pursuing the bache-
[23] V. A. de Freitas Barbosa, M. A. de Santana, M. S. Andrade, lor’s degree in medical imaging with Changsha
R. C. F. de Lima, and W. P. dos Santos, ‘‘Deep-wavelet neural networks Medical University. His research interests include
for breast cancer early diagnosis using mammary termographies,’’ in Deep data extraction, data verification, and statistical
Learning for Data Analytics: Foundations, Biomedical Applications and analysis.
Challenges. Amsterdam, The Netherlands: Elsevier, 2020, pp. 99–124.
[24] A. Mohiyuddin, A. Basharat, U. Ghani, V. Peter, S. Abbas, O. B. Naeem,
and M. Rizwan, ‘‘Breast tumor detection and classification in mammogram
images using modified YOLOv5 network,’’ Comput. Math. Methods Med.,
vol. 2022, pp. 1–16, Jan. 2022.

VOLUME 12, 2024 70831

 R. Zeng et al.: FastLeakyResNet-CIR: A Novel Deep Learning Framework

HONGSHEN LI is currently pursuing the bach- SHUANGNI DUAN received the master’s degree
elor’s degree in computer science with Changsha from the University of South China, Hunan, China,
Medical University. His research interests include in 2011. She is currently a Lecturer with the
deep learning and data analysis. School of Public Health, Xiangnan University. Her
research interests include public health and health
policy.

PINGPING BING received the Ph.D. degree in LEMEI ZHU received the master’s degree from the
geophysics from China University of Petroleum University of South China, China, in 2011. She is
(Beijing), Beijing, China, in 2012. She is currently currently with the School of Public Health, Chang-
an Associate Professor with Changsha Medical sha Medical University. Her research interests
University. Her research interests include opti- include molecular pharmaceutics, novel nanofab-
mization algorithms, signal analysis and model rication, and processing technologies.
prediction, and deep learning.

70832 VOLUME 12, 2024