2020 IEEE International Women in Engineering (WIE) Conference on Electrical and Computer

Engineering (WIECON-ECE)

Detection and Classification of Skin Cancer by

Using a Parallel CNN Model
Noortaz Rezaoana Mohammad Shahadat Hossain Karl Andersson
Department of Computer Science Department of Computer Science Department of Computer Science,
and Engineering and Engineering Electrical and Space Engineering
University of Chittagong University of Chittagong Luleå University of Technology
Chittagong -4331,Bangladesh Chittagong -4331,Bangladesh Skellefteå, Sweden
[email protected] hossain [email protected] [email protected]

Abstract—Skin cancer is one of the most lethal of all cancers. to a study showed in 2017 that skin cancer contribute 1.79% to
When it is not diagnosed and handled at the beginning, it is the worldwide disease burden estimated in disability-adjusted
supposed to extend to other areas of the body.It also occurs life years. [1]. Around 7 percent of new cancer cases world-
while the tissue is revealed to light from the sun, mainly due
to the rapid development of skin cells. For early detection, wide are caused by skin cancer. [2], with a 2011 expense
a dependable automated system for skin lesion recognition is of more than 8 billion to the US medicare project. There is
absolutely mandatory in order to minimize effort , time and clinical evidence of such result differences with respect to race
human life.Both Image processing and deep learning are used in in cases of skin cancer: while people with different skin tones
the technique for successful treatment of skin cancer.The paper are around 20 to 30 percent less likely to contract melanoma
suggest an automated techinic for skin cancer classification. The
classification of 9 types of skin cancer has been done in this than those with lighter skin, they have also been discovered
study.Also the performance and ability of deep convolutional to have a lower or higher mortality risk for certain melanoma
neural networks (CNN) is observed..The dataset contains nine types.
clinical types skin cancer ,such as -actinic keratosis,basal cell
carcinoma, benign keratosis, dermatofibroma, melanoma, nevus,
seborrheic keratosis, squamous cell carcinoma, vascular lesions.
The objective is to establish a model that diagnoses skin cancer as
well as classifies it into various classes through the Convolution
Neural Network. The diagnosing methodology uses concept of
image processing and deep learning.Through using different tac-
tics of image augmentation, the number of images has also been
enriched.Finally, the transfer learning approach is used to further
improve the accuracy of the classification tasks.Approximately
0.76 weighted average precision, 0.78 weighted average recall ,
0.76 weighted average f1-score, and 79.45 percent accuracy are
shown by the proposed CNN method.
Index Terms—Skin cancer, CNN, Data Augmentation, Deep
learning,Transfer learning Fig. 1. Skin cancer rates in various country according to world cancer research
foundation

I. I NTRODUCTION
In recent time, Convolutional Neural Networks have widely
Skin cancer has been the most widespread disease globally been used for various classification [3] as well as to classify
.The occurrence of either non-melanoma or melanoma skin skin cancer lesions.In the classification of skin cancers, many
cancers has grown in subsequent decades. As per the World CNN models have dramatically outpaced highly skilled health
Health Organisation (WHO), skin cancer can be detected in care professionals. The performance of these models has
every three cases of cancer, and one out of five Americans, been further increased by many approaches, such as transfer
according to Skin Cancer Foundation Statistics, will face learning using massive datasets.The convolutional networks
skin cancer throughout their lifetime.In regions including the VGG-16 and VGG-19 has 16 and 19 convolutional layers
united states, canada and Australia, the million individuals are , respectively.The pretrained network could identify photos
infected with skin cancer has grown at a relatively steady pace such as keyboard , mouse, pen, and creatures into 1000
in past few centuries. skin diseases appear to have significant object levels.The networks have now accumulated rich feature
adverse effects in health of population worldwide.According representations for a large collection of images, and those

978-0-7381-3144-3/20/$31.00 ©2020 IEEE

networks also have an input data resolution of 224-by-224. datasets.The classification efficiency was associated with Seg-
Net and Residual U-Net (ResU-Net). In the ISIC 2017 dataset
II. R ELATED W ORK against the SegNet and ResU-Net schemes for skin cancer
image classification tasks, the analytical and methodological
In the field of medical image classification, CNNs have results indicated a significant improvement. [20]. In 2018, an-
already been used extensively and so forth [4]. Also it has other notion, called LadderNet, was developed. This modified
showed inspiring findings, such as clinical epidermal 2 cell U-Net model structure was simply a sequence of many U-
[5] ,diabetic retinopathy fundus [6], cervical cell image clas- Nets with several encrypting and decrypting subsystems.This
sification and skin cancer identification [7] [8] [9] [10]. method can be seen as rippling numerous different FCNs
Yuet al [11] established a very deep CNN and a collection of and has been evaluated for retinal blood vessel categorization
learning frameworks with minimal training data. To establish functions.
and obtain a dermatologist-level diagnosis of more than 120 And from the other hand, numerous DL-based dermoscopic
thousand photographs, Esteva et al. [12] used a pre-trained imaging techniques for the ISIC 2016 and 2017 datasets for
CNN technique.CNN models that have been advantageous or the skin cancer categorization agenda have been examined.
outshone by dermatologists have been presented by Haenssle [17].Ensemble approach is done by a DL with validated
et al. [13]. Some other methods, such as the ensemble model deep learning techniques.In the first classification step, lesions
[14], the feature aggregation of multiple models, were devel- are extracted where different classical machine-learning ap-
oped to diagnose skin cancer using deep learning [15] [16]. proaches (such as color histogram, edge histogram, sparse
Several techniques for skin cancer segmentation and clas- coding) and DCNN-based techniques such as ResNet, Ima-
sification problems have been proposed over the last few geNet, and U-Net have been implemented to retrieve input
decades. Different approaches were suggested and submitted image features. Finally, non-linear SVM is used for catego-
for skin cancer’s feature extraction, and categorization in the rization purposes. Such strategies are evaluated in the ISIC
International Skin Imaging Challenge (ISIC) in 2016. The con- 2016 dataset and demonstrated how the approach is superior
clusions of the classification were based on the identification to 8 trained dermatologists in 100 research data subgroups.
of only two cancer types, benign and malignant. [17].In the Mechanisms relying on svm and DCNN were used for the
subsequent year (2017), for segmentation, identification, and identification of skin cancer in a dataset that was obtained from
classification tasks (called ISIC 2017), a significantly larger the web and demonstrated significant identification accuracy.
dataset was released relative to the 2016 version. With different [21]. Interpretation of skin lesions for deep-learning melanoma
DCNNs and Support Vector Machine (LSVM) approaches, diagnosis was presented in 2018. Such DL-based approaches
various methodologies were suggested and exhibit excellent were tested for image labeling in the ISIC 2017 dataset. [22].
recognition performance. [17]. Throughout this paper , we introduce a model that is used
An automated melanoma recognition method was designed for the classification and diagnosis of skin cancer. Figure 2
in the same year, which had been added to ISBI(2016) shows a pictorial depiction of the model. In the next part,
dataset there a Complete Convolutionary Residual Network the model is discussed below. This proposed model varies
(FCRN) was intended to apply for specific lesions catego- considerably from the current VGG-16 and VGG-19 models.
rization [11].The lesion areas were then separated from the Previous work has been conducted to identify a number of 7
input images and a Deep Residual Network (DRN) was used types of categories of skin cancer, but we classify 9 types of
to identify melanoma and non-melanoma lesions.In compar- categories of skin cancer in this paper.
ison, numerous DCNN approaches have been experimented
with, including VGG-16, GoogleNet, FCRN-38, RCRN-50 III. DATA P REPROCESSING
and FCRN-101 [11]. The efficiency of VGG-16 and Inception-
v3 was checked on the ISIC 2016 skin lesion segmentation First, to facilitate the preparation of the proposed model,the
datasets, that provides the best outcome of 61.3 percent and data set and the development plan of the model will be
69.3 percent test accuracy, respectively. [18]. presented in detail. Next, we will analyze the proposed model
In 2015, however, specifically for medical image segmen- architecture modeling approach in depth and the training
tation operations, an updated and modern architecture, called technique addressing the optimum modification of parameters.
”U-Net”, was proposed. U-Net became very common from Finally, we will incorporate simulation methods to highlight
then on and was effectively used in numerous medical imaging important problems with graphic markers in order to make
and the Pathology in Computers modalities. U-Net deals for a suspected illness diagnosis more persuasive.
smaller sample size of training while producing reliable data
A. About Dataset
for segmentation [19]. The Recurrent Residual Convolutional
(RRN) U-Net, which is also known as R2U-Net, was intro- This dataset is from kaggle.com and consists of 25,780
duced in 2018. [20].This improved version of the U-Net model photographs of benign and malignant Every image was cate-
was assessed on three separate datasets for clinical image gorized according to the description taken from ISIC, and all
classification purposes, along with a retinal blood vessel, subgroups were separated by the following class names: actinic
ISIC 2017 skin cancer categorization, and lung classification keratosis, basal cell carcinoma, dermatofibroma, melanoma,
 TABLE I
I MAGES AUGMENTATION SETTINGS

Augmentation setting Range

rotation range 90
(a) actinic keratosis (b) basal cell carcinoma (c) benign keratosis shear range 0.1
zoom range 0.14
horizontal flip T rue

vision.Different neurons partly cross the receptive fields so that

(d) dermatofibroma (e) melanoma (f) nevus they occupy the whole visual field CNN consists of three tiers
of neural layers to assemble the architectures: Convolutional
layer, Pooling layer, and Fully-Connected layer.
Throughout this section, we will clearly explain our sug-
gested model architecture.There are three fundamental ele-
ments in the suggested system, such as feature extraction,
(g) seborrheic keratosis (h) squamous cell carci- (i) vascular lesion detection, and classification. Firstly, in our proposed model, a
noma concurrent layers of convolution, activation, and max-Pooling
Fig. 2. Some image labels available in the dataset. is used.After this, we connect parallel layers at the func-
tion level.Afterwards, the flattened characteristics or features
have Multi-Layer Perceptrons (MLP) into two levels, but
nevus, pigmented benign keratosis, seborrheic keratosis, squa- the number of neurons at each layer that drops has to be
mous cell carcinoma, and vascular lesion.In addition , the calculated through an alteration to prevent overfitting.After
dataset is public and metadata circulates in such a manner this, the classification task has been done by the final layer
that sufficient document input can be generated to allow including softmax layer.There a class activation maps is also
comparisons to each image.One of the main stages of data generated those layers.Created class activation maps function
preprocessing is resizing so that is why all files are resized to as a classification translator coupled to the last convolution
224 x 224 pixels. Some of the images from the benchmark layer. The proposed system’s full system infrastructure, as
dataset are seen in figure-2. The dataset have been divided shown in Figure 3.We split the flow of work in two sections.
in train, validation and test in an acceptable ratio.The traning The first section is the extraction and the last section called the
,validation and testing dataset contains 19537,2167 and 3799 classification & the detection one is for feature extraction and
respectively. another is for image classification and detection respectively.
Feature extraction section is clearly explaind in the next
B. Data Augmentation section.
It is also advantageous to maximize the volume of data
. For training a CNN model properly the data augmentation A. Feature Extraction Phase
technique plays an important role. [23] [24] [25]This technique To achieve an appropriate network architecture,in each
avoids distortion and retains the original consistency of input convolution layer,filter with various size, various MLP and
and output data. Also during training phase, this process is diverse performance evaluation metrics has been evalu-
performed just-in-time, so the model ’s output can be enhanced ated.Throughout the first phase, the proposed CNN model
by solving the overfitting problem. We have several choices contains five dilated convolutionary units, defined as CBr =
for image augmentation to select values from various sizes iconv(n = 64,128,256,512,512), which are alternately max-
such as horizontal flip,rotation range, shear range and zooming pooled. In our suggested model, dilated convolution is inte-
etc.During the training process, every other alternative has the grated. The input picture provides two CBr=iConv blocks in
potential to depict images in a variety of ways and to provide parallel, as seen in Figure 3,Only changing the amount of
essential characteristics, thereby maximizing the utility of dilation, such as di = 1, 2,3,..., N. Convolution layer with
the model. The settings for image augmentation used in our a dilatation rate of 1 is identical to the typical convolution.
experiment can be seen in Table-I. But when the rate of dilation is more then 1 ,it improves
the convolution layer performance while processing data at a
IV. M ETHODOLOGY maximum magnification and achieves fine picture details.The
Biological systems have motivated convolutionary neural section of image in which the filter generates a function
networks. The communication structure between a network’s even without modifying the filter magnitude is known as
neurons matches the arrangement of the visual cortex of the receptive field, which is basically a constant range entry, i.e.
living being. The response of a certain cortical neuron is if there’s any dilation, then every input will lose (di 1)
known as a receptive field in a particular area of the field of pixels.The dilation rate of 1 is a normal convolution.As per
 Fs = f1j⇥j f2j⇥j ...fn=x
j⇥j
(3)

here ,j × j is filter size.

For layer l,if dilation rate is di and mf x mf is denoted as
input feature map then the proposed convolution network will
m ⇥m
generates Yl f f feature maps from the inputs.Which can
be calculated by the equation given below:

m ⇥m
Ylo⇥o = Yl1,df i =k f ⇥ lf + lb (4)

here,lb is the bias of the layer, and lf is filter.

Just after the convolution layer, the generated features are
used by the added activation function in the initial phases to
create a new feature map as output.Towards the context of
the activation unit, a was picked since this nonlinear layer
and the rectification layer can be merged into CNN. ReLU
has a range of benefits, and most notably, it can easily
spread gradients.Consequently, the probability of gradient dis-
appearance can be reduced if the preliminary mass calculates
the fundamental features of CNN.Note that the activation
function executes component-by-element operation upon this
input feature diagram, as the output appears to be the similar
size of the input.
(a) (b)
The developed scheme does have a block of CBr = iconv,
Fig. 3. System architecture of the proposed model. supported by a layer of max-Pooling, which is designed in
incremental form five times in tandem.An effective approach
for downscaling the strained images is maxpooling,it demands
this definition,if two becomes the value of dilation rate and maximum rate for every layer,as most of the features produced
aspects of the input picture data is two dimensional then each are overlooked in each layer by using a 2 x 2 filter size with 2
input data will skip a pixel for such dilation rate. level, which dramatically increases the performance of the next
When the kernel size is specified, the consequence ofdi step of computation The maxpooling layer used in our study
upon rf supports the explanation of correlation between the was 2x2 with stride 2 since, as mentioned through previous
dilation ratedi and receptive field rf .If the kernel size is study, overlapped maxpooling windows won’t enhance dra-
ks , which has been dilated by the dilation rate di then the matically over the windows that has not been overlapped.The
representation of receptive filed is shown in the Equation no. last convolution layer, with kernel size jxj, a total of 512 filters
1 [?] is used. The layer has dilation rate 1, which is followed by the
ReLU activation function unit for creating low-level features.
rf = di (k1) + 1 (1) The parallel branches of the convolution layer that produce
the features are concatenated, this can be seen in Figure-3 and
If the output is O, where i× i input with dilation factor, di , figure-4.
padding pd & the stride isst respectively then the Equation
The purpose of this concatenation operation is that every
no.- 2 refers the size of output O.
other branch generates features that have specific features from
different layers, thereby concatenating the last level collateral
O = i + 2pd rf st + 1 (2)
branch features in the model to analyze the relationship of
Significant features in the observation area are reported at dilated convolution blocks so that the ultimate convolution
various scales by utilizing two receptive fields of different layer can recognise superior classification features.
size.In the current introduced model, each first two CBr=iconv Next, in the flattened layer, the model revealed to transforms
block consist of two convolution layers and activation layers the feature maps.For completing the classification task a
and last three CBr=iconv block consist of four convolution one dimensional feature vector has been generated from the
layer and activation layer and also each of them contains features maps. The last phase is the method of classification,
various filter (3 × 3) with stride and dilation rate with 1 which during the next segment will be illustrated briefly.
and di , respectively. Convolution layer can be define with For the classification task a one dimensional feature vector
filtersFs ✏R1⇥n ,the equation is given below, has been generated from the features maps
 In Keras, our studies was accomplished with the TensorFlow
backend.
A. Tuning of the hyper-parameters
Hyper-parameters are important when they directly influ-
ence the attitudes of the system, since fine-tuned hyperparame-
ters have a significant effect on the model’s output.To train 100
epochs , we have used Adam[40] optimizer,and for the batch
size of 32, we set a learning rate of 0.0001.Furthermore, The
categorical cross entropy loss function calculates the loss of
class probability predicted by the softmax function.And finally
measure the probability of each category’s.
B. Performance Evaluation Metrics
To utilized performance criteria accuracy , precision , recall,
and F1 score are calculated for the proposed model. Since
our dataset is unbalanced, F1 score could be a more accurate
metric.
C. Result
In Figure 5 and Table-II the complete results are displayed,
Fig. 4. Detailed layer representation of the proposed CNN model
where the classification results on various measurement criteria
such as precision,recall,F1-Score and accuracy are listed in
Figure 5, and Table-II displays the weighted average precision,
B. Classification Phase recall , F1-Score results of the proposed model, VGG-16 and
VGG-19 model. From the figure-5 we can see that the accuracy
Via two neural layers,in order to achieve the classification of proposed model, VGG-16 and VGG-19 are 79.45% ,
process, the output produced by flattened level were fed by 69.57%, 71.19% respectively. So we can say that our model
a two-layer MLP (also known as a completely connected gives more accurate result than the traditional VGG-19 and
(FC) layer) during that stage. Besides that,Aftereach FC layer, VGG-16 model.
a dropout layer is connected.Throughout training process,
this dropout layers will discard certain FC layer weights at
random to minimize overfit.The dropout range which can be
0 to 100 percent, measures the amount of arbitrarily selected
weight drops. For our analysis purpose the dropout range of
0.5 is picked. And the size of two fully connected layer is
4096 & 4096 FC respectively.The softmax activation function
has been used after the fully connected layer for classifying
the skin cancer images such as actinic keratosis, basal cell
carcinoma, benign keratosis, dermatofibroma, melanoma, ne-
vus, seborrheic keratosis, squamous cell carcinoma, vascular
lesions. Finally, figure-4 displays the details of the layers of
the proposed model.

V. E XPERIMENTAL E VALUATION

In this segment , we are going to discuss the performance

of implemented model to distinguish images of skin cancer,
primarily classified into nine categories they are actinic ker-
atosis,basal cell carcinoma, benign keratosis, dermatofibroma,
melanoma, nevus, seborrheic keratosis, squamous cell carci-
noma, vascular lesion.It discusses a concise summary of the
dataset used in the process of analysis and augmentation.The Fig. 5. Class-wise classification results of individual model
evaluation sets the training, validation and test data in a
standard ratio.By comparing two VGG models called VGG- The result of detecting skin cancer has been shown in the
16 and VGG-19 with the model described, we did not take figure-6 .We tested 20 images randomly and 9 of them have
into account any pre-trained weights for example ImageNet. been shown in the figure-6 and also the detection rate is quite
 (a) (b) (c)

(d) (e) (f)

Fig. 6. Detection result of the proposed model.

satisfactory for the suggested model than the VGG-16 and

VGG-19 model.
From the figure 5, we can see how the classification of Fig. 7. Confusion matrix of the proposed model
most classes continues to rise in nearly all cases. However,
according to the VGG-16 and VGG-19 models, the best score
belongs to the model we have proposed .The accuracy of all VI. E PILOGUE & F UTURE W ORK
evaluation methods is presented in figure-5, and it can be Throughout the study,a new CNN-based approach has been
shown that the introduced system performs better than the suggested to detect skin cancer using images. It is clearly
other models.Figure-5 outlines the accuracy of all assessment illustrated that the method can successfully capture features
models and it can be shown that the model is superior over of skin cancer by using the parallel convolution the feature
certain models. or characteristic of skin cancer through utilizing parallel
convolution blocks.The model has outstanding classification
TABLE II performance relative to the two well-known CNN architectures
D IFFERENT MODEL’ S WEIGHTED AVERAGE RESULTS VGG-16 and VG19.
In this study,we classified skin cancer in nine types which
Method Precision Recall F1-Score
is the most classification category of skin cancer till now.
Proposed Method 76.17 78.15 76.92 Due to the requisition of vast data for the effective training
VGG-16 65.67 68.89 67.77
and implementation of CNN-based architecture, we have used
VGG-19 68.54 69.45 68.95
data augmentation techniques for the existing dataset. In this
process we achieved the desired outcome.
The exploratory analysis reveals that the proposed approach
The weighted average of evaluated performance has been significantly outperforms-of-the-art models with a substantial
represented in Table-II. For the proposed model the weighted improvement in precision , recall and F1 scores of 76.16%,
average precision, recall, F1-score are 76.17%,78.15% and 78.15% and 76.92% respectively. Via several performace
76.92% respectively. While the weighted average precision ,re- matrics such as the weighted average and the overall accuracy
call and F1-score for the VGG-16 model are 65.67% ,68.89% The model also illustrates its ability.
, 67.77% also for the VGG-19 model are 68.54% ,69.45% Finally, we think that new results will ideally overcome
,68.95% respectively. Better outcomes have been achieved for intellectual difficulties in identifying further cases of skin
the proposed method. In the weighted average analysis, the cancer and using them to test for skin cases in AI-based
findings of the model we suggested are much higher than other systems, especially in clinical practice.In addition, the skin
methods; this can be demonstrated by the fact that maps with cancer classification process can be determined under uncer-
various sizes can be extracted from photographs of skin cancer tainty by using sophisticated methodology like Belief Rule
in the proposed model. Based Expert Systems (BRBES) in an integrated framework.
Figure-7 displays the confusion matrices of the suggested [26] [27] [28] [29] [30]
model. The diagonal value of the confusion matrix for the 9 Further research can be carried out to analyze and accom-
classes are higher .That means it can classify those number of plish a diversified data collection of substantial amounts of
test sample accurately from our test dataset. evidence regarding skin cancer to make our proposed model
more stable and validated. Moreover, there are also room International Symposium on Biomedical Imaging (ISBI 2018). IEEE,
gaining new horizons of knowledge by using several other 2018, pp. 168–172.
[18] J. Burdick, O. Marques, J. Weinthal, and B. Furht, “Rethinking skin
CNN models such as ResNet, DenseNet, InceptionNet, etc. lesion segmentation in a convolutional classifier,” Journal of digital
imaging, vol. 31, no. 4, pp. 435–440, 2018.
R EFERENCES [19] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks
[1] C. Karimkhani, R. P. Dellavalle, L. E. Coffeng, C. Flohr, R. J. Hay, S. M. for biomedical image segmentation,” in International Conference on
Langan, E. O. Nsoesie, A. J. Ferrari, H. E. Erskine, J. I. Silverberg et al., Medical image computing and computer-assisted intervention. Springer,
“Global skin disease morbidity and mortality: an update from the global 2015, pp. 234–241.
burden of disease study 2013,” JAMA dermatology, vol. 153, no. 5, pp. [20] M. Z. Alom, C. Yakopcic, M. Hasan, T. M. Taha, and V. K. Asari,
406–412, 2017. “Recurrent residual u-net for medical image segmentation,” Journal of
[2] T. Gansler, P. A. Ganz, M. Grant, F. L. Greene, P. Johnstone, M. Ma- Medical Imaging, vol. 6, no. 1, p. 014006, 2019.
honey, L. A. Newman, W. K. Oh, C. R. Thomas Jr, M. J. Thun et al., [21] U.-O. Dorj, K.-K. Lee, J.-Y. Choi, and M. Lee, “The skin cancer
“Sixty years of ca: a cancer journal for clinicians,” CA: a cancer journal classification using deep convolutional neural network,” Multimedia
for clinicians, vol. 60, no. 6, pp. 345–350, 2010. Tools and Applications, vol. 77, no. 8, pp. 9909–9924, 2018.
[3] M. Akter, M. S. Hossain, T. Uddin Ahmed, and K. Andersson, [22] S. Sreena and A. Lijiya, “Skin lesion analysis towards melanoma detec-
“Mosquito classication using convolutional neural network with data tion,” in 2019 2nd International Conference on Intelligent Computing,
augmentation,” in 3rd International Conference on Intelligent Computing Instrumentation and Control Technologies (ICICICT), vol. 1. IEEE,
& Optimization 2020, ICO 2020, 2020. 2019, pp. 32–36.
[4] G. Kasinathan, S. Jayakumar, A. H. Gandomi, M. Ramachandran, [23] T. U. Ahmed, S. Hossain, M. S. Hossain, R. Ul Islam, and K. Andersson,
S. J. Fong, and R. Patan, “Automated 3-d lung tumor detection and “Facial expression recognition using convolutional neural network with
classification by an active contour model and cnn classifier,” Expert data augmentation,” in 2019 Joint 8th International Conference on
Systems with Applications, vol. 134, pp. 112–119, 2019. Informatics, Electronics & Vision (ICIEV) and 2019 3rd International
[5] Z. Gao, L. Wang, L. Zhou, and J. Zhang, “Hep-2 cell image classification Conference on Imaging, Vision & Pattern Recognition (icIVPR). IEEE,
with deep convolutional neural networks,” IEEE journal of biomedical 2019, pp. 336–341.
and health informatics, vol. 21, no. 2, pp. 416–428, 2016. [24] M. Z. Islam, M. S. Hossain, R. ul Islam, and K. Andersson, “Static hand
[6] P. Wang, L. Wang, Y. Li, Q. Song, S. Lv, and X. Hu, “Automatic cell gesture recognition using convolutional neural network with data aug-
nuclei segmentation and classification of cervical pap smear images,” mentation,” in 2019 Joint 8th International Conference on Informatics,
Biomedical Signal Processing and Control, vol. 48, pp. 93–103, 2019. Electronics & Vision (ICIEV) and 2019 3rd International Conference
[7] K. M. Hosny, M. A. Kassem, and M. M. Foaud, “Classification of skin on Imaging, Vision & Pattern Recognition (icIVPR). IEEE, 2019, pp.
lesions using transfer learning and augmentation with alex-net,” PloS 324–329.
one, vol. 14, no. 5, p. e0217293, 2019. [25] R. R. Chowdhury, M. S. Hossain, R. ul Islam, K. Andersson, and
[8] X. He, Z. Yu, T. Wang, B. Lei, and Y. Shi, “Dense deconvolution net: S. Hossain, “Bangla handwritten character recognition using convo-
Multi path fusion and dense deconvolution for high resolution skin lesion lutional neural network with data augmentation,” in 2019 Joint 8th
segmentation,” Technology and Health Care, vol. 26, no. S1, pp. 307– International Conference on Informatics, Electronics & Vision (ICIEV)
316, 2018. and 2019 3rd International Conference on Imaging, Vision & Pattern
[9] B. Harangi, “Skin lesion classification with ensembles of deep convolu- Recognition (icIVPR). IEEE, 2019, pp. 318–323.
tional neural networks,” Journal of biomedical informatics, vol. 86, pp. [26] T. Uddin Ahmed, M. N. Jamil, M. S. Hossain, K. Andersson, and M. S.
25–32, 2018. Hossain, “An integrated real-time deep learning and belief rule base
[10] T. J. Brinker, A. Hekler, A. H. Enk, J. Klode, A. Hauschild, C. Berking, intelligent system to assess facial expression under uncertainty,” in 9th
B. Schilling, S. Haferkamp, D. Schadendorf, T. Holland-Letz et al., International Conference on Informatics, Electronics & Vision (ICIEV).
“Deep learning outperformed 136 of 157 dermatologists in a head- IEEE Computer Society, 2020.
to-head dermoscopic melanoma image classification task,” European [27] M. S. Hossain, F. Ahmed, K. Andersson et al., “A belief rule based
Journal of Cancer, vol. 113, pp. 47–54, 2019. expert system to assess tuberculosis under uncertainty,” Journal of
[11] L. Yu, H. Chen, Q. Dou, J. Qin, and P.-A. Heng, “Automated melanoma medical systems, vol. 41, no. 3, p. 43, 2017.
recognition in dermoscopy images via very deep residual networks,” [28] R. Ul Islam, K. Andersson, and M. S. Hossain, “A web based belief
IEEE transactions on medical imaging, vol. 36, no. 4, pp. 994–1004, rule based expert system to predict flood,” in Proceedings of the
2016. 17th International conference on information integration and web-based
[12] A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, applications & services, 2015, pp. 1–8.
and S. Thrun, “Dermatologist-level classification of skin cancer with [29] S. Kabir, R. U. Islam, M. S. Hossain, and K. Andersson, “An integrated
deep neural networks,” nature, vol. 542, no. 7639, pp. 115–118, 2017. approach of belief rule base and deep learning to predict air pollution,”
[13] H. A. Haenssle, C. Fink, R. Schneiderbauer, F. Toberer, T. Buhl, Sensors, vol. 20, no. 7, p. 1956, 2020.
A. Blum, A. Kalloo, A. B. H. Hassen, L. Thomas, A. Enk et al., [30] A. A. Monrat, R. U. Islam, M. S. Hossain, and K. Andersson, “A belief
“Man against machine: diagnostic performance of a deep learning rule based flood risk assessment expert system using real time sensor
convolutional neural network for dermoscopic melanoma recognition in data streaming,” in 2018 IEEE 43rd Conference on Local Computer
comparison to 58 dermatologists,” Annals of Oncology, vol. 29, no. 8, Networks Workshops (LCN Workshops). IEEE, 2018, pp. 38–45.
pp. 1836–1842, 2018.
[14] N. C. Codella, Q.-B. Nguyen, S. Pankanti, D. A. Gutman, B. Helba,
A. C. Halpern, and J. R. Smith, “Deep learning ensembles for melanoma
recognition in dermoscopy images,” IBM Journal of Research and
Development, vol. 61, no. 4/5, pp. 5–1, 2017.
[15] M. Attia, M. Hossny, S. Nahavandi, and A. Yazdabadi, “Skin melanoma
segmentation using recurrent and convolutional neural networks,” in
2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI
2017). IEEE, 2017, pp. 292–296.
[16] N. Nida, A. Irtaza, A. Javed, M. H. Yousaf, and M. T. Mahmood,
“Melanoma lesion detection and segmentation using deep region based
convolutional neural network and fuzzy c-means clustering,” Interna-
tional journal of medical informatics, vol. 124, pp. 37–48, 2019.
[17] N. C. Codella, D. Gutman, M. E. Celebi, B. Helba, M. A. Marchetti,
S. W. Dusza, A. Kalloo, K. Liopyris, N. Mishra, H. Kittler et al.,
“Skin lesion analysis toward melanoma detection: A challenge at the
2017 international symposium on biomedical imaging (isbi), hosted by
the international skin imaging collaboration (isic),” in 2018 IEEE 15th