Deep Learning for Skin Disease Diagnosis with

End-to-End Data Security

1st Shriya Pingulkar 2nd Diti Divekar 3rd Aryaman Tiwary
Dept. of Information Technology Dept. of Information Technology line 2: Dept. of Information Technology
K. J. Somaiya College of Engineering K. J. Somaiya College of Engineering K. J. Somaiya College of Engineering
Mumbai, India Mumbai, India Mumbai, India
shriya.pingulkar@somaiya.edu diti.divekar@somaiya.edu aryaman.tiwary@somaiya.edu

Abstract—Skin diseases, particularly skin cancer, pose This research aims to develop a comprehensive skin lesion
significant health risks, necessitating accurate and timely classification system using CNNs to address these issues. The
diagnosis. Traditional manual analysis methods, however, are objectives are to:
subject to errors and can be time-consuming. To address these
challenges, we propose an automated classification system for  Design and train an accurate CNN model using
skin disease detection that leverages deep learning algorithms to publicly available skin image datasets.
accurately diagnose skin lesions. This system facilitates secure
sharing of diagnostic information among healthcare  Develop a user-friendly diagnostic platform to
professionals, promoting collaborative treatment planning. integrate the model implement security measures for
Recognizing the critical importance of data security and encrypting and sharing diagnostic reports while
confidentiality in the healthcare industry, our system employs maintaining patient privacy and confidentiality.
state-of-the-art security measures to protect sensitive patient
 The results of this study could streamline
data whilst medical data transfer. By maintaining the integrity
of patient information and adhering to regulatory requirements,
dermatological workstreams, improve diagnosis of
our skin disease classification system provides healthcare potentially life-threatening conditions, and
professionals with fast and accurate diagnosis, and the ultimately contribute to better patient care and
confidence that patient privacy is upheld at all times. outcomes.

Keywords—Skin disease classification, skin lesions, image II. LITERATURE REVIEW

classification, deep learning, data privacy and security, CNN, Skin cancer poses a significant public health concern, and
Firebase cloud database, SMTP, medical documentation timely detection and prevention are essential for effective
treatment. It can become hazardous when they progress to an
I. INTRODUCTION invasive stage. Deep learning and computer vision techniques
Accurate and timely diagnosis of skin diseases is crucial have demonstrated potential in enhancing skin cancer
given the potential health implications. However, traditional diagnosis and early detection by understanding complex
visual examination methods rely heavily on the expertise of patterns precisely. This includes the application of
dermatologists and pathologists, which can lead to errors, convolutional neural networks (CNNs), which have achieved
variability in diagnosis, and delays in treatment. The expert-level precision in diagnosing skin cancer[15].
development of an automated skin lesion classification system
has the potential to address these issues by providing a more
Nonpigmented and non-melanotic skin cancers are the most
reliable, efficient and standardized approach to diagnosing
prevalent types and are more challenging to diagnose. A
skin cancer and other diseases.
study compared the accuracy of a CNN-based classifier with
Deep learning algorithms such as convolutional neural physicians at varying experience levels. The CNN-based
networks (CNNs) have shown promising results in automating model was trained on over 7,800 dermoscopic and close-up
medical image analysis tasks. By learning visual features from images of skin lesions excised at a primary skin cancer clinic.
large skin image datasets, CNNs can potentially achieve The combined CNN (cCNN) was evaluated on a set of 2,072
diagnostic accuracy on par with or exceeding human experts. unknown cases and compared to the results from 95 human
An automated CNN-based system could significantly improve raters, including 62 board-certified dermatologists, with
disease detection workflows by augmenting rather than differing levels of experience. The cCNN achieved a higher
replacing the skills of medical professionals. By quickly percentage of correct specific diagnoses compared to human
analyzing images and producing standardized reports,
raters but not compared to experts[1]. Another study utilized
pathologists may be able to devote more attention to complex
cases and other clinical responsibilities. deep neural networks to categorize skin lesion images as
benign or malignant. The researchers trained a CNN on a
However, developing such a system presents technical and dataset of 129,450 clinical images, representing 2,032
practical challenges. Large annotated image datasets suitable distinct diseases, which is two orders of magnitude larger
for training deep learning models are required but scarce in the than earlier datasets. The CNN’s performance was tested
dermatology domain. Ensuring data security, privacy and against 21 board-certified dermatologists using biopsy-
regulatory compliance is also imperative given the sensitive proven clinical images. This system’s accuracy in detecting
nature of medical information. malignant melanomas and carcinomas was comparable to
that of trained dermatologists[2].

979-8-3503-4383-0/23/$31.00 ©2023 IEEE

Another research effort proposed two deep learning methods face"class, with an impressive accuracy benchmark of
to tackle three primary tasks in skin lesion image processing: 88%[10]. In a unique stride, a novel multi-classification
lesion segmentation, dermoscopic feature extraction, and method using a convolutional neural network (CNN) with a
lesion classification. A deep learning framework comprising nested residual structure was unveiled. This was aimed at
two fully convolutional residual networks was proposed to classifying skin diseases from dermoscopy images. By
generate both segmentation and coarse classification results harnessing the power of transfer learning, this model
simultaneously. Additionally, a lesion index calculation unit commendably classified six types of skin diseases and
was employed to refine the results. The proposed framework achieved accuracies of 65.8% and 90% in multi-class and
demonstrated promising accuracies across all tasks[3]. binary classifications, respectively[11].
A further research project proposed a convolutional neural
network (CNN) model with two convolutional layer blocks Notably, a diagnostic system for melanoma was developed
and a dense layer for classification. The model was trained on by researchers using hyperspectral data (HSD) and
a dataset of skin lesion images and achieved high accuracy in GoogLeNet, a known convolutional neural network. The
skin cancer classification[4]. In another study, a hybrid inherent challenge was the compatibility of traditional CNNs
approach was proposed for detecting melanoma skin cancer like GoogLeNet with three input channels. To surmount this,
using a combination of deep learning and traditional machine a "Mini Network" layer was innovated to condense the HSD's
learning techniques. The proposed system relied on a 84 channels to a mere three. With the added advantage of data
convolutional neural network and two classical machine augmentation, this system achieved a 72.3% sensitivity,
learning classifiers trained with a set of features describing 81.2% specificity, and a 77.2% accuracy, with aspirations for
the borders, texture, and colour of a skin lesion. These future enhancements to the Mini Network and an expansion
methods were then combined using majority voting to in the sample size[14]. Another study made a mark by
enhance performance. This system can be used to examine presenting a deep learning approach utilizing a Convolutional
suspicious lesions and can assist doctors and even normal Neural Network (CNN) dedicated to the early and precise
people in determining specific diseases[5]. diagnosis of melanoma types. With an accuracy rate of
88.83%, this CNN model expertly distinguished between
Another paper presented a computer-aided method for lesion maligna, superficial spreading, and nodular melanoma.
detecting melanoma skin cancer using image processing It was observed that this model outstripped other machine
techniques. The proposed system takes a skin lesion image as learning classifiers like DT, RF, and GBT in diagnostic
input and, through novel image processing techniques, accuracy, albeit with the limitation of its method being tested
analyses it for the presence of skin cancer. The system checks on only one dataset, signifying the potential for broader
for various melanoma parameters, and the extracted feature validation[13]. Lastly, the field witnessed the inception of a
parameters are used to classify the image as normal skin or method to classify skin diseases using advanced deep
melanoma cancer lesion. The system’s performance learning networks. The unique combination of MobileNet V2
demonstrated promising accuracy in detecting melanoma with Long Short-Term Memory (LSTM) was deployed for
skin cancer[6]. Another study suggests a new method for the this purpose. Evaluated using the HAM10000 dataset, this
early detection of skin cancer using an optimal Convolutional approach stood out, outpacing other models with an accuracy
Neural Network (CNN) optimized with an improved whale exceeding 85% while requiring significantly fewer
optimization algorithm. The whale optimization algorithm computations. A highlight of this paper was the introduction
can be employed to optimize the network’s hyperparameters, of a mobile application designed for early-stage skin disease
such as the number of filters, filter size, learning rate, and identification. This invaluable tool aims to assist both patients
dropout rate. This study compared the proposed method with and dermatologists in obtaining efficient diagnoses, thereby
different methods on two datasets, finding that it had superior forestalling further complications[12].
performance to the other methods[7].

An advanced research introduces a system for classifying III. METHODOLOGY

skin diseases using 25,331 clinical images, focusing on eight A. Dataset
specific skin conditions. By leveraging the capabilities of the
The HAM10000 dataset[16] is a widely used collection of
Residual Neural Network (ResNet), this study achieved high- skin lesion images that serves as a valuable resource for
accuracy diagnosis reports, representing them as confidence training and evaluating Convolutional Neural Network (CNN)
scores[8]. Furthermore, the rise of Teledermatology was models in the field of skin lesion classification. This dataset
marked by the introduction of a system classifying skin contains a total of 10,015 images, each representing a skin
diseases from dermoscopic images using the MNIST lesion from various individuals. The primary goal of utilizing
HAM10000 dataset. This system employed two pre-trained the HAM10000 dataset is to develop accurate and robust
CNN models, MobileNet v1 and Inception V3. It was algorithms that can aid dermatologists and healthcare
observed that the Inception V3 model boasted a predictive professionals in diagnosing skin conditions. The dataset is
accuracy of 72%, distinctly outperforming the MobileNet v1 categorized into seven different classes, with each class
model, which settled at 58%[9]. Facial skin disease corresponding to a specific type of skin lesion. These classes
identification also took a significant step with the are:
introduction of an automated method that used a pre-trained
 Melanocytic nevi (NV): Commonly referred to as
deep convolutional neural network (CNN). This model moles, melanocytic nevi are benign growths that
displayed proficiency in classifying images into eight distinct
facial skin diseases, a "normal skin" class, and a "no-
 occur due to the accumulation of melanocytes in the B. Data Preprocessing and Augmentation
skin. 1) Addressing Class Imbalance: The visualization
 Melanoma (MEL): Melanoma is a malignant form of revealed a glaring imbalance, especially pronounced in class
skin cancer that arises from melanocytes, the cells number 4. Such imbalances can detrimentally impact the
responsible for producing the skin pigment melanin. model's training, leading to biased outcomes. To counteract
 Benign keratosis-like lesions (BKL): This category this, the 'RandomOverSampler' technique was harnessed.
includes various non-cancerous growths that exhibit This method, rather than merely duplicating
features similar to keratosis, such as seborrheic underrepresented classes, intelligently augments them by
keratosis or solar lentigo. creating synthetic samples based on existing data, ensuring a
balanced and diverse dataset.
 Basal cell carcinoma (BCC): Basal cell carcinoma is
another type of skin cancer that originates from the
2) Data Reshaping for Model Compatibility: Post-
basal cells within the skin's outermost layer.
augmentation, the dataset underwent a transformation to
 Actinic keratoses (AKIEC): Actinic keratoses are align with the CNN model's requirements. The Data variable
precancerous growths that develop on skin that has was reshaped into a 4-dimensional array, specifically (-1, 28,
been frequently exposed to the sun's ultraviolet (UV) 28, 3). This transformation ensures that the data is not only
radiation. compatible with the CNN model but also optimally structured
 Vascular lesions (VASC): This class encompasses to leverage the model's capabilities.
vascular skin lesions, including angiomas and
hemangiomas, which involve abnormal blood vessel
C. Designing the CNN Model
growth.
1) Architectural Overview: The CNN model, crafted
 Dermatofibroma (DF): Dermatofibromas are benign using the TensorFlow Keras library, is a testament to modern
skin tumors that arise in the deeper layers of the skin. machine learning design. Its multi-layered structure includes:
 Input Layer: Tailored to process a 28x28 pixel
single-channel image, this layer serves as the
gateway to the model.
 Convolutional Layers: These layers, equipped with
ReLU activation functions and He normal weight
initialization, are the workhorses of the model,
extracting features from the input data. The 'same'
padding ensures spatial consistency across layers.
 Max Pooling Layers: These layers condense the
data, reducing spatial dimensions by half, thereby
focusing on the most salient features.
 Batch Normalization Layers: Integrated to stabilize
Fig 1: Plotting initial dataset classes. and expedite the learning process, these layers
normalize the activations of neurons in the network.
Each image in the dataset is accompanied by pixel
data. This data is represented in a structured CSV file  Dense Layers: These layers, with their ReLU
where every row corresponds to an individual image, and activation functions and He normal weight
columns denote specific pixels. These pixels are further initialization, serve as the neural network's core,
categorized into their respective red, green, and blue color driving its learning capability.
channels. A preliminary visualization of the dataset was  Dropout Layer: A regularization technique, set at a
executed to discern potential patterns, trends, or rate of 0.2, it \item randomly deactivates certain
anomalies. This involved plotting the distribution of neurons during training to prevent overfitting.
images across the seven classes, which was instrumental  Output Layer: A Softmax layer with seven units, it
in identifying class imbalances. provides the final classification output,
corresponding to the seven distinct classes in the
dataset.

Fig 3: Example of a Convolution Layer

Fig 2: Image classification in dataset after application of

‘RandomOverSampler’
 2) Compilation, Training, and Optimization: The model's  Accuracy and Loss Graphs: These plots, charting the
compilation was a meticulous process, leveraging the model's performance trajectory, offer insights into
categorical cross-entropy loss function paired with the Adam its consistency and reliability across training and
optimizer. The training regimen, spanning 25 epochs with testing phases.
batches of 128 images, was designed for optimal learning.  Advanced Metrics Table: The table I showcases
The 'learning\_rate\_reduction' callback, a dynamic tool, was metrics like F1-score, recall, precision, and support
integrated to adjust the learning rate in real-time, ensuring for each skin cancer class. This granular breakdown
that the model's training remained on an optimal trajectory. offers a deep dive into the model's performance
nuances, highlighting its strengths and areas of
improvement.

Lesions and avg precision Recall F1-score support

score
('akeic', 'Actinic 0.99 1.00 1.00 1667
Keratosis and
intraepithelial
carcinomae')
('bcc', 'basal cell 1.00 1.00 1.00 1689
carcinoma')
('bkl', 'benign 0.97 0.99 0.98 1651
Fig 4: Final architecture of CNN model keratosis-like
lesions')
('df', 1.00 1.00 1.00 1629
'dermatofibroma')
D. Model Evaluation
('nv', 0.99 0.92 0.95 1663
1) Key Performance Metrics: Once training concluded, 'melanocytic
the model underwent rigorous testing. It showcased a loss of nevi')
('vasc', 'pyogenic 1.00 1.00 1.00 1680
0.0558 and an accuracy of 0.9881. These metrics, while granulomas and
standalone impressive, were further dissected to understand hemorrhage')
their implications on the model's predictive capabilities. ('mel', 0.96 0.99 0.97 1751
'melanoma')
micro avg 0.99 0.99 0.99 11734
2) Visualization and In-depth Analysis: A series of macro avg 0.99 0.99 0.99 11734
visualizations were employed to provide a holistic weighted avg 0.99 0.99 0.99 11734
understanding: samples avg 0.99 0.99 0.99 11734

TABLE I
TABLE FOR F1-SCORE, PRECISION, RECALL AND SUPPORT

E. Data Security and Encryption in Medical Reports

Given the sensitive nature of personal images and data
involved in the implementation of the CNN model, especially
in the medical domain, ensuring data security is paramount.
The methodology incorporates robust measures to safeguard
the privacy and integrity of medical reports.

1) Password Generation and Encryption System: A

sophisticated system is proposed that randomly generates an
Fig 5: Plot of training and testing accuracy alphanumeric password for each medical report. This
password then serves a dual purpose: as a unique identifier
and as an encryption key. By using this dynamically
generated key, each medical report is encrypted, ensuring its
contents remain confidential and inaccessible to unauthorized
individuals. The encryption algorithm chosen is of the highest
standard, ensuring that without the correct password,
decryption is virtually impossible.

2) Secure Email Transmission: Once encrypted, the

medical reports are prepared for transmission. They are
attached to an email and dispatched to the intended recipient,
typically the consulting doctor. The encryption ensures that
even if an unauthorized entity were to intercept or access the
Fig 6: Plot of training and testing loss
email, the contents of the report would remain secure and model achieved high F1 scores for melanocytic nevi, even
unreadable. though they can exhibit similarities to melanomas.
The model excelled in identifying the 'akiec' (actinic
3) Password Delivery via API: To bolster security keratoses), 'bcc' (basal cell carcinoma), 'df' (dermatofibroma),
further, the delivery of the decryption password is handled and 'vasc' (pyogenic granulomas) classes, achieving flawless
separately from the email. The system integrates the Twilio precision, recall, and F1-scores of 1.0, thereby correctly
API, a secure communication platform, to transmit the identifying all cases within the test set. The 'bkl' (benign
password directly to the doctor's registered phone number. keratosis-like lesions) class demonstrated strong performance
This method ensures that the password reaches only the with scores of 0.97 precision, 0.99 recall, and 0.98 F1-score,
intended recipient, minimizing the risk of interception or indicating near-optimal performance.
unauthorized access. Although slightly weaker, the model still exhibited
commendable performance for the 'nv' (melanocytic nevi) and
4) Decryption and Access by Authorized Personnel: 'mel' (melanoma) classes. For 'nv', precision was 0.99, recall
Upon receiving the encrypted medical report and the was 0.92, and the F1-score reached 0.95. For 'mel', precision
decryption password, the doctor can then decrypt the report was 0.96, recall was 0.99, and the F1-score was 0.97.
to view its contents. This dual-layered security measure Moreover, the model achieved a high F1-score of 0.97 for
ensures that only those with the correct password (i.e., melanocytic nevi, demonstrating its ability to differentiate
authorized individuals) can access the sensitive medical them from melanomas. The model displayed a precision of
information within the report. 0.96 and recall of 0.99 for melanomas, indicating effective
identification with minimal false positives.
IV. RESULTS
The confusion matrix, as illustrated in Fig 7, offers a
The Convolutional Neural Network (CNN) model designed comprehensive analysis of the model's efficacy. It reveals that
for classifying skin lesions into seven distinct categories the model proficiently categorized the majority of images,
exhibited remarkable accuracy and robustness, as evidenced exhibiting only a minimal number of errors that were
by the precision, recall, and F1-scores presented in Table II. proportionally dispersed among various classes. This suggests
Lesions and avg precision Recall F1-score support a consistent performance across different categories without
score any significant bias or difficulty in particular classes.
('akeic', 'Actinic 0.99 1.00 1.00 1667
Keratosis and
intraepithelial
carcinomae')
('bcc', 'basal cell 1.00 1.00 1.00 1689
carcinoma')
('bkl', 'benign 0.97 0.99 0.98 1651
keratosis-like
lesions')
('df', 1.00 1.00 1.00 1629
'dermatofibroma')
('nv', 0.99 0.92 0.95 1663
'melanocytic
nevi')
('vasc', 'pyogenic 1.00 1.00 1.00 1680
granulomas and
hemorrhage')
('mel', 0.96 0.99 0.97 1751
'melanoma')
micro avg 0.99 0.99 0.99 11734
macro avg 0.99 0.99 0.99 11734 Fig 7: Confusion Matrix
weighted avg 0.99 0.99 0.99 11734
samples avg 0.99 0.99 0.99 11734 The results underscore the model's capability to learn
discriminative features capable of capturing nuanced
TABLE I distinctions among various skin lesion types. The employed
convolution blocks effectively extracted features at multiple
TABLE FOR F1-SCORE, PRECISION, RECALL AND SUPPORT scales, while the incorporation of dropout and batch
normalization mitigated overfitting. The model's architectural
The micro-average F1-score achieved by the model was design was well-suited for this multi-class classification task.
an impressive 0.99, signifying its exceptional accuracy in
In comparison to traditional machine learning methods such
categorizing diverse skin lesions. This superior performance
as Support Vector Machines (SVMs), K-Nearest Neighbors
can be attributed to several factors, including the sophisticated
(KNNs), and Random Forests (RFs), the CNN exhibited
model architecture, incorporation of dropout layers, batch
significantly superior performance. Unlike these methods that
normalization, and prudent choices of activation and weight
necessitate extensive feature engineering, the CNN
initialization functions.
automatically learned discriminative features from input
The model demonstrated adeptness in accurately classifying images, reducing the reliance on handcrafted features and
lesions with visual resemblances, notably actinic keratoses, enhancing classification accuracy.
basal cell carcinomas, and benign keratosis. Remarkably, the
Furthermore, when compared to other deep learning-based [7] Thurnhofer-Hemsi, K., & Domínguez, E. (2021). A convolutional
approaches such as ResNet, VGG, and Inception, the CNN neural network framework for accurate skin cancer detection. Neural
Processing Letters, 53(5), 3073-3093.
showcased comparable or superior performance. This feature
[8] Gouda, N., & Amudha, J. (2020, October). Skin cancer classification
rendered the CNN computationally efficient and conducive using ResNet. In 2020 IEEE 5th International Conference on
for deployment in resource-constrained environments. Computing Communication and Automation (ICCCA) (pp. 536-541).
IEEE.
The exceptional performance of the CNN model can be
[9] Purnama, I. K. E., Hernanda, A. K., Ratna, A. A. P., Nurtanio, I.,
attributed to its deep architecture, utilization of dropout layers Hidayati, A. N., Purnomo, M. H., ... & Rachmadi, R. F. (2019,
for regularization, batch normalization, and astute choices in November). Disease classification based on dermoscopic skin images
activation and weight initialization functions. The precision using convolutional neural network in teledermatology system. In 2019
and recall scores underscore the model's robust predictive international conference on computer engineering, network, and
capacity and clinical applicability. The model demonstrated intelligent multimedia (CENIM) (pp. 1-5). IEEE.
exceptional classification prowess across all categories, [10] El Saleh, R., Bakhshi, S., & Amine, N. A. (2019, October). Deep
convolutional neural network for face skin diseases identification. In
making it a promising tool for facilitating the diagnosis and 2019 Fifth International Conference on Advances in Biomedical
classification of skin lesions. Engineering (ICABME) (pp. 1-4). IEEE.
[11] Zhou, H., Xie, F., Jiang, Z., Liu, J., Wang, S., & Zhu, C. (2017,
V. CONCLUSION October). Multi-classification of skin diseases for dermoscopy images
In conclusion, the skin lesion disease detection system using deep learning. In 2017 IEEE international conference on imaging
systems and techniques (IST) (pp. 1-5). IEEE.
developed using CNN technology, along with its associated
[12] Srinivasu, P. N., SivaSai, J. G., Ijaz, M. F., Bhoi, A. K., Kim, W., &
medical data transfer security system, is an innovative and Kang, J. J. (2021). Classification of skin disease using deep learning
reliable solution to the challenges faced by medical neural networks with MobileNet V2 and LSTM. Sensors, 21(8), 2852.
professionals in diagnosing skin lesions accurately and [13] Allugunti, V. R. (2022). A machine learning model for skin disease
securely. The use of CNN technology enables pathologists to classification using convolution neural network. International Journal
diagnose skin lesions with high probability, improving patient of Computing, Programming and Database Management, 3(1), 141-
outcomes and reducing the time needed for diagnosis. The 147.
system also allows pathologists to generate instant [14] Hirano, G., Nemoto, M., Kimura, Y., Kiyohara, Y., Koga, H.,
standardized medical reports, making the system more user- Yamazaki, N., ... & Nagaoka, T. (2020). Automatic diagnosis of
melanoma using hyperspectral data and GoogLeNet. Skin Research
friendly and allowing for seamless integration into existing and Technology, 26(6), 891-897.
healthcare systems. Furthermore, the system ensures the [15] Wu, Z. H. E., Shuang Zhao, Yonghong Peng, Xiaoyu He, Xinyu Zhao,
security of sensitive medical data through encrypted storage Kai Huang, Xian Wu et al. ”Studies on different CNN algorithms for
and secure password-protected email communication between face skin disease classification based on clinical images.” IEEE Access
pathologists and doctors. This approach to medical data 7 (2019): 66505-66511.
transfer security is crucial in maintaining patient privacy and [16] https://towardsdatascience.com/beginners-guide-to-understanding-
confidentiality and adheres to ethical and regulatory convolutional-neural-networks-ae9ed58bb17d
standards. The system has been designed and developed using [17] https://www.kaggle.com/datasets/kmader/skin-cancer-mnist-
ham10000
a range of technologies and their use has allowed the system
to be developed to the highest standards, with attention paid
to regulatory and ethical requirements. Overall, this project
has the potential to improve the quality of healthcare
significantly, and its impact will be substantial. We believe
that it will make a positive impact on the lives of patients and
medical professionals alike.
REFERENCES

[1] Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H.
M., & Thrun, S. (2017). Dermatologist-level classification of skin
cancer with deep neural networks. nature, 542(7639), 115-118.
[2] Tschandl, P., Rosendahl, C., Akay, B. N., Argenziano, G., Blum, A.,
Braun, R. P., ... & Kittler, H. (2019). Expert-level diagnosis of
nonpigmented skin cancer by combined convolutional neural networks.
JAMA dermatology, 155(1), 58-65.
[3] Dildar, M., Akram, S., Irfan, M., Khan, H. U., Ramzan, M., Mahmood,
A. R., ... & Mahnashi, M. H. (2021). Skin cancer detection: a review
using deep learning techniques. International journal of environmental
research and public health, 18(10), 5479.
[4] Daghrir, J., Tlig, L., Bouchouicha, M., & Sayadi, M. (2020,
September). Melanoma skin cancer detection using deep learning and
classical machine learning techniques: A hybrid approach. In 2020 5th
international conference on advanced technologies for signal and
image processing (ATSIP) (pp. 1-5). IEEE.
[5] Jain, S., & Pise, N. (2015). Computer aided melanoma skin cancer
detection using image processing. Procedia Computer Science, 48,
735-740.
[6] Zhang, N., Cai, Y. X., Wang, Y. Y., Tian, Y. T., Wang, X. L., &
Badami, B. (2020). Skin cancer diagnosis based on optimized
convolutional neural network. Artificial intelligence in medicine, 102,
101756.