Подолання викликів у глибокому навчанні для виявлення раку шкіри
Skin cancer is one of the most prevalent malignancies worldwide. A critical factor in reducing mortality rates is the early detection. It underscores the need for accessible Computer-Aided Diagnostic (CAD) systems. Recent advancements in Deep Learning (DL) have shown great promise in addressing this...
Saved in:
| Date: | 2025 |
|---|---|
| Main Authors: | , |
| Format: | Article |
| Language: | English |
| Published: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2025
|
| Subjects: | |
| Online Access: | https://journal.iasa.kpi.ua/article/view/320423 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | System research and information technologies |
| Download file: |  |
Institution
System research and information technologies| _version_ | 1866302994721538048 |
|---|---|
| author | Nikitin, Vladyslav Danilov, Valery |
| author_facet | Nikitin, Vladyslav Danilov, Valery |
| author_sort | Nikitin, Vladyslav |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2025-07-25T15:56:08Z |
| description | Skin cancer is one of the most prevalent malignancies worldwide. A critical factor in reducing mortality rates is the early detection. It underscores the need for accessible Computer-Aided Diagnostic (CAD) systems. Recent advancements in Deep Learning (DL) have shown great promise in addressing this challenge. Despite this progress in the field of machine learning, researchers encounter numerous obstacles when it comes to skin cancer classification. This article examines the current state of DL-based skin cancer diagnostics. Critical aspects of system development, including data preprocessing, model training, and performance evaluation, are addressed. Moreover, the article highlights opportunities for innovation that could significantly advance the field. By providing a comprehensive overview, this article aims to guide researchers and practitioners in optimizing DL models, addressing existing limitations, and exploring emerging trends to enhance diagnostic accuracy and accessibility. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2025.2.03 |
| first_indexed | 2025-07-27T04:04:05Z |
| format | Article |
| fulltext |
Publisher IASA at the Igor Sikorsky Kyiv Polytechnic Institute, 2025
42 ISSN 1681–6048 System Research & Information Technologies, 2025, № 2
TIДC
ТЕОРЕТИЧНІ ТА ПРИКЛАДНІ ПРОБЛЕМИ
ІНТЕЛЕКТУАЛЬНИХ СИСТЕМ ПІДТРИМАННЯ
ПРИЙНЯТТЯ РІШЕНЬ
UDC 004.855.5
DOI: 10.20535/SRIT.2308-8893.2025.2.03
NAVIGATING CHALLENGES IN DEEP LEARNING FOR SKIN
CANCER DETECTION
V. NIKITIN, V. DANILOV
Abstract. Skin cancer is one of the most prevalent malignancies worldwide. A criti-
cal factor in reducing mortality rates is the early detection. It underscores the need
for accessible Computer-Aided Diagnostic (CAD) systems. Recent advancements in
Deep Learning (DL) have shown great promise in addressing this challenge. Despite
this progress in the field of machine learning, researchers encounter numerous ob-
stacles when it comes to skin cancer classification. This article examines the current
state of DL-based skin cancer diagnostics. Critical aspects of system development,
including data preprocessing, model training, and performance evaluation, are ad-
dressed. Moreover, the article highlights opportunities for innovation that could sig-
nificantly advance the field. By providing a comprehensive overview, this article
aims to guide researchers and practitioners in optimizing DL models, addressing ex-
isting limitations, and exploring emerging trends to enhance diagnostic accuracy and
accessibility.
Keywords: skin cancer, deep learning, classification, transformers, CNN, GAN,
data preprocessing, data augmentation.
INTRODUCTION
Skin cancer affects millions of individuals every year. Malignancy originates in
the skin’s primary layers — the epidermis, dermis, and hypodermis. Due to its
exposure to ultraviolet (UV) radiation, the epidermis is the most common site for
skin cancers. Keratinocytes and melanocytes in these layers can undergo muta-
tions due to prolonged exposure to UV radiation, which can lead to malignancy.
The three major types of skin cancer include Basal Cell Carcinoma (BCC),
Squamous Cell Carcinoma (SCC), and Malignant Melanoma [1].
BCC is the most common skin cancer, typically presenting as a slow-
growing, pearly lesion. While it rarely metastasizes, untreated BCC can cause
significant local tissue damage. SCC, the other type, is more aggressive and capa-
ble of metastasis. It often appears as scaly, crusty lesions, especially on sun-
exposed skin, SCC requires timely intervention to prevent systemic spread. Ma-
lignant melanoma, though rare, is the most aggressive form of skin cancer and is
prone to rapid metastasis [1].
Regardless of the type of skin cancer, early detection is crucial for successful
treatment. Still, there are many barriers preventing timely diagnosis. According to
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 43
the World Health Organization, there is a global shortage of dermatologists, par-
ticularly in low- and middle-income countries [2]. Studies show evidence that
lower-income populations are more likely to have advanced stages of skin cancer
[3]. Hence, obtaining professional help to determine whether a lesion is malignant
can be challenging. Even in developed regions, busy lifestyle and/or financial
constraints can prevent individuals from seeking medical attention.
Computer Aided Diagnostics (CAD) system is a solution for assisted or
automated diagnostics. They can be especially helpful in reducing human errors
and expanding access to advanced medical care. Deep Learning (DL) with its ad-
vancements in recent years is a promising solution for skin cancer CAD. Machine
Learning (ML) models can help to detect malignant lesions early via specialized
or common hardware, potentially covering gap in dermatological services. Such
systems can be particularly useful as prescreening tools, indicating cases where
professional help is necessary.
DL has revolutionized various areas over the recent years, including medical
imaging. During the past decade, researchers have developed numerous DL mod-
els for skin cancer classification with varying levels of success. This diversity of
models and methods creates a sophisticated landscape to navigate. The article
aims to underline key challenges and obstacles that researchers face developing
skin cancer DL model as of 2024, providing examples of solutions proposed in
existing researcher. Finally, it highlights areas for improvement that could take
current DL-based CAD systems to a new level.
To simplify navigation, the article is divided into sections, each focusing on
a critical component in the development of DL-based CAD systems for skin can-
cer detection.
1. Researchers and Communities: An overview of labs, teams, and organiza-
tions that have made significant advancements in DL-based skin cancer tasks in
recent years.
2. Datasets and Data Challenges: A list of widely used datasets with their
descriptions. This section also includes an overview of data based obstacles.
3. Development Pipeline: Addressing data normalization, augmentation, and
techniques to fix data challenges highlighted in the previous section.
4. Model Training and Optimization: This section explores the complexities
of selecting model architectures, leveraging transfer learning, and optimizing hy-
perparameters.
5. Evaluation and Validation: Various metrics suitable for the task are ex-
plored in detail.
6. Opportunities for Innovation: Directions and areas that could significantly
enhance the quality of DL-based CAD systems.
7. Guidance for Emerging Researchers: Practical advice for newcomers to
the field, including best practices in research design, navigating the literature, and
identifying emerging trends.
By exploring these themes, the article provides a concise resource highlight-
ing the current challenges, solutions, and future directions in the application of
DL to skin cancer detection. Ultimately, the aim is to contribute to the develop-
ment of more accurate, accessible diagnostic tools that can improve patient out-
comes and reduce healthcare disparities.
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 44
RESEARCHERS AND COMMUNITIES
Stanford University Team. In 2017, the study “Dermatologist-Level Classifica-
tion of Skin Cancer with Deep Neural Networks” was published in Nature by
a team from Stanford University. They researched a dataset of approximately
130,000 clinical images representing over 2,000 different skin diseases. They took
Google Inception v3 CNN architecture, pre-trained on the ImageNet dataset, and
fine-tuned it for skin lesion classification. Performance was validated against a
group of 21 dermatologists, ensuring clinical relevance. It turned out the model
achieved performance on par with board-certified dermatologists [4].
International Skin Imaging Collaboration. The International Skin Imag-
ing Collaboration (ISIC) is the largest consortium that made a significant contri-
bution to the skin cancer classification problem. The collaboration created the
largest publicly available skin cancer dataset and conducted multiple challenges
to encourage the development of advanced algorithms for skin lesion analysis. In
mid-2024, ISIC released its latest dataset [5]. ISIC promotes transparency and
reproducibility in dermatological research while safeguarding patient privacy. Its
extensive use in numerous studies has made it a cornerstone for developing and
validating DL models for skin cancer diagnostics, significantly accelerating pro-
gress in the field.
Memorial Sloan Kettering Cancer Center. Memorial Sloan Kettering
Cancer Center (MSKCC) is one of the world’s leading institutions dedicated to
cancer treatment, research, and education. In dermatology, MSKCC implemented
AI to support the detection of melanoma. The institution is actively working on
implementing CAD systems to improve the effectiveness of diagnostics. They
consider AI a complementary tool for medical experts [6].
Google Health and DeepMind. Google’s research teams actively worked to
apply DL to medical images. For instance, Yuan Liu and Peggy Bui published a
paper titled “A Deep Learning System for Differential Diagnosis of Skin Dis-
eases” in 2019 [7]. They developed a system that was able to accurately differen-
tiate 26 common skin conditions.
Individual Researchers. There are numerous ML enthusiasts and experts
working on the problem all over the world. For example, Yilmaz et al. achieved
an accuracy of 82% with on the ISIC 2017 dataset with the NASNetMobile model
[8]. Baygin et al. integrated textural and deep features to develop a pyramidal hy-
brid model receiving a classification accuracy of 91.54% with 10-fold cross-
validation [9]. Agarwal and Singh were able to get 86.65% accuracy on ISIC Ar-
chive utilising convolutional neural networks with transfer learning [10].
Many more studies report remarkable results with the models and the data
used. Each year, hundreds of new articles on the topic are being published, some
of them making breakthroughs through usage of new technology or metric. The
article mentions key pioneering research as examples of successful problem-
solving for medical imaging. The main idea of this article is to guide through the
current landscape of skin cancer images with DL leading to even more advances
in the field.
DATASETS AND DATA CHALLENGES
Data is the base of any ML algorithm. The success of any DL project heavily de-
pends on access to robust and comprehensive datasets as well as the correct usage
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 45
of those. By processing the data, the model extracts the most important features
from it so that it can make correct predictions on unseen data of a similar nature.
D. Wen et al. (2022) made an exhaustive list of datasets that were publicly avail-
able at the time [11]. The section below mentions key datasets from that list, add-
ing the data that was released since 2022.
Datasets. ISIC Archive. The ISIC Archive [5] is the largest data repository
with cancerous skin images so far. As of today, the number of open-for-usage
samples is nearing half a million. More than 400,000 of those are images released
as part of the latest ISIC Challenge — ISIC 2024 [12]. The archive includes a di-
verse range of skin lesions: melanomas, nevi, BCCs, SCC, etc. It includes both
annotated and unannotated data sourced from various international centers. Some
of the images contain precise borders of the lesion, allowing not only classifica-
tion but segmentation as well. It might be the most impactful contribution to DL-
based skin cancer CAD systems in terms of data.
HAM10000. The dataset [13], also known as “Human Against Machine with
10,000 training images” (HAM10000), published by Philipp Tschandl et al.
(2018), is the most popular dataset for DL skin cancer classification. It contains a
total of 10,015 dermoscopic images, categorized into seven common types of skin
lesions: actinic keratoses, basal cell carcinoma, benign keratosis-like lesions, der-
matofibroma, melanoma, melanocytic nevi, and vascular lesions. Each image is
labeled with a confirmed diagnosis, verified through histopathology, follow-up
examinations, or expert consensus. It also includes metadata detailing patient
demographics such as age, gender, and lesion location, which adds context for
model training.
Due to its high-quality data, HAM10000 was used in numerous articles,
normally combined with transfer learning to compensate for the lack of size. For
instance, A.T. Priyeshkumar et al. (2024) developed an ensemble DL model for
skin lesion classification using HAM10000, achieving high accuracy in differen-
tiating between lesion types [14]. T.M. Alam et al. (2022) proposed a novel CNN
architecture trained on HAM10000, focusing on improving classification per-
formance through data augmentation techniques [15]. H.A. Owida et al. (2024)
investigated the use of transfer learning with pre-trained models on HAM10000 to
enhance melanoma detection [16]. A. Ameri et al. (2020) explored feature extrac-
tion methods using HAM10000 to improve the interpretability of DL models [17].
S.S. Chaturvedi et al. (2020) conducted a comparative study of DL algorithms on
the HAM10000 dataset to identify the most effective approaches for skin lesion
classification [18].
Non-ISIC Data. While the ISIC Archive and HAM10000 datasets are widely
used in recent skin lesion research, several other datasets have historically con-
tributed to the field, particularly before the availability of ISIC data.
The PH² dataset [19], developed by Pedro Hispano Hospital in Portugal,
contains 200 dermoscopic images focusing on melanocytic nevi and melanomas.
Despite its small size, the dataset is highly valued for its detailed manual segmen-
tation masks and clinical annotations that include colors and dermoscopic struc-
tures. It is often used in studies emphasizing precise lesion segmentation and the
analysis of dermoscopic features.
The MED-NODE dataset [20], created by the University of Twente in the
Netherlands, includes 1,700 clinical (non-dermoscopic) images of melanomas and
benign melanocytic nevi. Diagnoses in this dataset are confirmed through histopa-
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 46
thology. It is particularly useful for research focusing on classification based on
clinical photographs rather than dermoscopic images. This dataset presents chal-
lenges like variations in lighting, skin tone, and image quality, making it a valu-
able resource for developing models that can handle real-world conditions.
The Dermofit Image Library [21], developed by the University of Edin-
burgh, includes approximately 1,300 high-resolution images spanning 10 skin
lesion categories. However, it is accessible only under a commercial license,
which can limit its availability for some researchers. The library is often used in stud-
ies requiring high-resolution images and advanced feature extraction techniques.
The SD-198 dataset [22], compiled from dermatology atlases, contains 6,584
images across 198 different skin disease classes. It is particularly valuable for
multi-class classification research and for developing models that can identify
rare skin conditions.
These datasets, while smaller in scale compared to ISIC, offer unique advan-
tages by addressing diverse imaging conditions and focusing on specific research
challenges, continuing to play a crucial role in dermatology research.
Data Challenges. Class Imbalance. Datasets often have a disproportionately
large number of benign lesions compared to malignant ones. Such skewed data
causes model bias and poor sensitivity in detecting rare but critical cases like
melanomas. This problem is typical for skin cancer data. For instance, in the ISIC
2020 dataset, the benign nevi-to-melanoma ratio is approximately 9 to 1 [5]. If not
handled properly, this imbalance can lead to high overall accuracy but poor per-
formance when detecting minority classes, which are crucial for early diagnosis.
Underrepresentation of Diverse Skin Types. In recent years, it was pointed
out that available skin cancer data is insufficient in terms of darker skin tone rep-
resentation. It can affect a model’s accuracy across different ethnicities and exac-
erbate health disparities [11], [23]. The research community considers this a seri-
ous problem as, even though skin cancer is less prevalent among individuals with
darker skin, it is often detected at later stages, leading to drastic outcomes. For
instance, Black individuals are three times more likely to be diagnosed with late-
stage melanoma [24]. One of the reasons is that melanoma for that part of the
population sometimes appears on less visible areas, such as the palms, soles, or
under the nails, which are less frequently examined [25].
Privacy Concerns and Data Sharing Limitations. Skin cancer imaging data
is part of a person’s personal information. Consequently, it falls under many laws
and acts aimed at protecting users’ data. As this data is of medical nature, it com-
plicates the situation even more. Among the commonly known regulations are:
Health Insurance Portability and Accountability Act [26] (HIPAA) in the
United States.
The General Data Protection Regulation [26] (GDPR) in the European Union.
These regulations impose strict guidelines on the handling, sharing, and
processing of medical data. Understanding these restrictions is crucial for re-
searchers aiming to access and utilize dermatological datasets while ensuring
compliance with legal and ethical standards.
Reluctance to Share Proprietary Datasets. Self-collected datasets offer re-
searchers distinct competitive advantages. This method of data acquisition allows
influencing the data flow from the very start, forming a more precise understand-
ing of the data nature. At the same time, exclusive access to this data lets institu-
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 47
tions uniquely fine-tune algorithms for enhanced performance, potentially leading
to breakthroughs in skin cancer detection and diagnosis. Sulaiman Khan et al.
(2022) [28] found that researchers often use private datasets exclusively or com-
bine them with open data to achieve superior results in skin cancer classification
tasks. Other reasons to withhold data might range from ethical to licensing con-
cerns. Another issue is data privacy. This practice highlights the advantages of
proprietary data but also underscores a significant barrier to progress in the field.
When models are trained on private datasets, it becomes challenging to replicate
studies, validate findings, or compare the effectiveness of different methodologies.
Data Labeling. Many of the publicly available skin cancer image samples
are currently unlabeled, which makes them impossible to use with supervised DL
algorithms. For instance, the ISIC 2020 dataset [5] includes 10,982 images with-
out assigned classes. Meanwhile, accurately annotating medical images requires a
high level of domain expertise, making the process both time-consuming and costly.
Artefacts in Image Acquisition. The most common way in which skin cancer
images are obtained is dermatoscopy. It is a non-invasive imaging technique that
involves examining and capturing skin lesions via a dermatoscope. Nevertheless,
images of a dataset are often collected in different centers and institutions that
utilize instruments with varying characteristics. However, when all the data is col-
lected by a single facility, it may cause variations in resulting images, leading to
models overfitting to device-specific artifacts instead of focusing on generic le-
sion features. Studies have shown that models trained on homogeneous datasets
perform poorly on images from diverse sources [29], [30].
Ethical Considerations. Besides norms and regulations, one should realize
that medical data is deeply personal. Processing it unavoidably raises some ethical
concerns. Syed F.H. Shah in “Ethical implications of artificial intelligence in skin
cancer diagnostics: use-case analyses” (2024) found that existing skin cancer
analysis solutions require a great deal of transparency and collaboration to avoid
potential ethical problems and misuse [31].
To resolve the ethical issues, the research community must show awareness
when handling such sensitive data. In the paper “Ethical considerations for artifi-
cial intelligence in dermatology: a scoping review,” Emily R. Gordon (2024)
identified key principles to follow to avoid ethics-related risks like fairness, inclu-
sivity, transparency, accountability, etc. [32].
Additional Clinical Criteria. While lesion images serve as the primary
source of information, various studies suggest that including more contextual data
can positively impact classification performance. Esteva et al. (2017) noted that
including clinical data could further enhance performance [4]. Haenssle et al. (2018)
demonstrated that combining dermatoscopic images with patient information im-
proved melanoma detection rates [33]. Pacheco and Krohling (2019) found that
integrating clinical metadata with images in DL models led to higher accuracy in
skin lesion classification [34]. These additional criteria may vary: patient age is
important because certain age groups are more inclined to have some types of
skin cancer than others, evolution over time is one of the key characteristics that
experts use to make a decision on lesion state, as well as the size of the lesion.
Embedding mathematical measures derived from the image, such as fractal di-
mension, may also contain insights useful for correct classification [35].
Incorporating these features into a classification model is a new challenge, as
image and contextual data normally have different modalities. Making a model
more complex to handle this may cause unintended overfitting. Therefore, the
challenge is to balance model complexity while integrating features effectively.
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 48
Preprocessing Stage. In the previous section the key data challenges were
identified. Many of them such as class imbalance, ethnic groups underrepresenta-
tion, restricted data, unlabeled data, inconsistency in the data gathering tools and
noise can be addressed on a data preprocessing stage. In this section the key steps
of data preparation are discussed.
Balancing Data. Some datasets do not suffer from class imbalance issues.
For example, Kaggle dataset for malignant melanoma classification [36] contains
approximately 5000 samples of each class which makes it a good option for fine
tuning lightweight melanoma classifiers for exploration and education purposes.
Nevertheless, most of the available data samples are severely imbalanced
with the largest amount of lesions being benign. For instance, HAM1000 dataset
contains around 60% of benign nevus samples. Provided that this dataset is used
for multi class classification, it makes it very imbalanced
Data Augmentation. A typical method for handling imbalanced data involves
performing augmentations on the training data. Shen (2022) proposes an effective
way of augmenting data for DL skin cancer classification with a significantly re-
duced search space of 60 possible transformations, compared to existing methods
like AutoAugment and RandAugment [37]. Himel et al. (2024) applied data aug-
mentation to 3,295 malignant images in the HAM10000 dataset, increasing the
number to 5,000 [38]. These augmentations included rotation, flipping, and zoom-
ing. However, their approach may seem somewhat straightforward, as they simul-
taneously disregarded more than 1,500 benign lesion images to balance the data.
This was also risky, as it essentially balanced the data through data loss. Polat et
al. (2020) augmented images of the same dataset with noise, scaling, and rotation,
increasing the total number of images to more than 33,000 [39]. Walker et al.
(2019) used cropping in addition to the aforementioned methods while working
with the ISIC 2017 Challenge [5], [40]. Ali et al. (2021) applied color-shifting
using principal component analysis to create augmented images [41].
Another way of augmenting images is application of Generative adversarial
network (GAN) architecture. Wu et al. (2020) did a review of GAN application
for augmenting skin cancer images to solve data imbalance problem [42]. In arti-
cle they note that various GAN architectures, including DCGANs [43], style-
based GANs [44], TED-GAN [45], SPGGAN [46], and conditional GANs
(CGANs) [47], have been employed to generate high-quality, diverse synthetic
images. Enhancements such as artifact removal, attention modules, informative
noise vectors, and stability improvements like the Two-Timescale Update Rule
have further optimized GAN performance. Wu state that these methods have consis-
tently improved classification metrics — including accuracy, sensitivity, specificity,
and F1 scores — by providing richer training data and more reliable image genera-
tion, ultimately enhancing the effectiveness of skin cancer classification models.
While this approach offers benefits, it also comes with potential risks and
drawbacks. Excessive noise, extreme scaling, or unusual rotation angles can dis-
tort images to the point where essential features are obscured or altered. If the
augmentation methods produce images that are too similar to each other, the
model might overfit to these synthetic variations rather than learning generalized
features. This reduces the model’s ability to perform well on truly unseen data.
Adding noise can sometimes degrade image quality to a level where the model’s
ability to extract meaningful features is compromised. Drawbacks also include
increased computational cost of creating and processing larger amount of data.
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 49
Fixing Ethnic Representation Issues. To mitigate the issue of underrepresen-
tation, several strategies can be employed to improve both the diversity of data
and the fairness of DL models.
Сolor jittering, which adjusts the brightness, contrast, and saturation of
images, can simulate a broader range of skin tones.
Histogram equalization can improve the visibility of features across dif-
ferent skin tones.
Style transfer can modify images to appear as if they belong to varied skin
types.
Pope et al. (2024) compared training model on imbalanced and sampled
datasets [48]. However, their results show that, albeit the sampled model tends to
be less biased towards dark-toned skin, the overall accuracy decreases.
Beyond augmentation, synthetic data generation can further expand datasets,
particularly for underrepresented skin tones. GANs are useful for generating syn-
thetic images that mimic lesions on darker skin, while domain adaptation tech-
niques can align feature representations to ensure models perform well across dif-
ferent skin tones. Rezk et al. (2022) composed a skin cancer dataset from existing
open access data and applied style transferring to make it more diverse in terms of
skin tone [49].
There also were instances of less traditional approach to solve this problem.
Continuing their study of 2019, Walker et al. (2024) used sonification of skin
cancer data to decrease influence of skin tone on classification [50]. The analysis
demonstrated high and comparable diagnostic accuracy for both fair skin (FS) and
darker skin (DS), with ROC curve AUCs of 0.858 and 0.856, respectively, indi-
cating no significant difference between the two groups. Sensitivity and specific-
ity values were similar (around 80–85%), confirming the model’s consistent per-
formance in detecting true positives and negatives across diverse skin tones. The
results suggest that the classifier maintains equivalent diagnostic reliability for
both FS and DS, even with the limitations of smartphone-based imaging.
To ensure fairness in model performance, algorithmic adjustments can be
utilized. Reweighting samples can help balance training by assigning higher
weights to underrepresented classes, while bias correction layers can be integrated
into models to correct inherent biases.
Data Cleaning. Data cleaning is a crucial step in DL, involving the prepara-
tion and preprocessing of data to improve its quality for model training [51]. This
process includes various data processing techniques that enable more effective
feature extraction, ultimately enhancing model performance.
Normalization. To address inconsistencies in the conditions under which im-
ages of skin lesions are captured, data normalization and denoising are essential.
Normalization techniques, such as Gray World, Shades of Gray, or max-RGB
algorithms, adjust the color balance of images to compensate for lighting differ-
ences [41], [52]. Global or adaptive histogram equalization methods, like Con-
trast Limited Adaptive Histogram Equalization (CLAHE), can be applied to im-
prove contrast and standardize the intensity distribution across images [53], [54].
Gamma correction is another option, adjusting non-linear luminance or color val-
ues to standardize brightness and contrast.
Standard normalization methods, such as scaling pixel values to have zero
mean and unit variance (z-score normalization) or scaling between 0 and 1 (min-
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 50
max normalization), are commonly used in image preprocessing [55]. For exam-
ple, 0–1 normalization was utilized in [41]. These steps ensure that images fed
into DL models are consistent and of high quality.
Noise Reduction. Removing noise from images is critical to enhance quality
and improve model accuracy. Common strategies include applying filters like
median or Gaussian filters to smooth images and reduce noise. Midasala et al.
(2024) applied a top-hat transform to remove thick hairs, while filters effectively
eliminated noise and thin hairs, albeit with limitations on contrast-based histo-
gram equalization [56]. Morphological operations, such as opening and closing,
assist in removing small artifacts.
Body hair presents a specific challenge in analyzing skin images, as it can
obscure important lesion features. Algorithms like the DullRazor detect hair pix-
els using edge detection and inpaint these regions to remove hair from the image
[57]. Thresholding methods identify hair regions so that they can be cleaned out.
Another approach to denoising is the use of autoencoders. These models
learn to create a compressed representation of the input and reconstruct it, effec-
tively removing noise in the process. Maurya et al. (2024) used autoencoders for
denoizing and feature extraction [58].
Segmentation. Segmentation is another crucial preprocessing step, involving
the selection of the region of interest (ROI) from the image—in this case, the skin
lesion. Accurate segmentation focuses analysis on the lesion and removes irrele-
vant background information. This topic is well-researched, and researchers often
utilize large pretrained models.
For instance, Himel et al. (2024) [38] used Meta’s Segment Anything Model
[59] to perform segmentation on skin cancer images from the HAM10000 dataset.
They then used the segmented images to pass only the ROI to feature extraction
and classification models, achieving 96% accuracy using Google’s Vision Trans-
former (ViT) [60]. This approach demonstrates the effectiveness of combining
advanced segmentation models with powerful classification architectures.
TRAINING AND OPTIMIZATION
Model training and optimization are pivotal in developing robust DL models for
skin cancer classification. These processes involve selecting suitable architec-
tures, optimizing learning algorithms, fine-tuning hyperparameters, and employ-
ing strategies to enhance model performance while addressing overfitting.
Transfer Learning. Transfer learning is extensively utilized in medical im-
age analysis due to the scarcity of labeled datasets [11]. By leveraging models
pre-trained on large-scale datasets like ImageNet, researchers can fine-tune these
models for specific tasks such as skin cancer classification [4]. This approach not
only accelerates convergence but also often yields superior performance com-
pared to training models from scratch.
In skin cancer classification, the data available is limited for many reasons.
Therefore, usage of pretrained models is popular approach. Ali et al. (2021) [41],
while developing a custom NN for skin cancer classification, applied transfer
learning to the task with pre trained ResNet [61], AlexNet [62], VGG-16 [63],
DenseNet [64] and MobileNet [65] achieving top accuracy of 86.09. Another ex-
ample is using ViT in [60] with 0.96 accuracy result.
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 51
Optimization Techniques. Learning Rate Scheduling. Adjusting the learn-
ing rate during training can significantly influence model convergence. Tech-
niques such as Step Decay, Exponential Decay, and more advanced methods like
Cyclical Learning Rates and Warm Restarts are employed to optimize training
efficiency [66, 67].
Optimizers. Selecting an appropriate optimization algorithm is crucial for
training deep neural networks. Optimizers like Stochastic Gradient Descent
(SGD) with momentum, Adam, and RMSProp are widely used. Adaptive opti-
mizers like Adam combine the advantages of AdaGrad and RMSProp, providing
efficient training for deep models [68]. In skin cancer classification tasks, Adam
is often preferred for its ability to handle sparse gradients and noisy data (for in-
stance in [69]).
Regularization Methods. Regularization techniques prevent overfitting by
adding constraints to the model. L1 and L2 regularization add penalties to the loss
function based on the magnitude of weights. Dropout randomly deactivates neu-
rons during training, reducing interdependent learning among neurons. Batch
normalization normalizes layer inputs, accelerating training and improving gener-
alization.
It is also worth mentioning that on this stage is possible to address class im-
balance problem if it was not solved completely on data preprocessing stage. Le
et al. (2020) used focal loss to train on imbalanced data of HAM10000 and re-
ceived 0.94 top accuracy with pretrained EfficientNetB1 [70].
Hyperparameter Tuning. Hyperparameters such as learning rate, batch size,
network depth, and activation functions significantly impact model performance.
Systematic methods like Grid Search and Random Search explore combinations
of hyperparameters, while Bayesian Optimization offers a more efficient search
by modeling the performance as a probabilistic function.
Same time hyperparameter tuning can always be costly no matter the tech-
nique chosen. In [69] it was done on only 10% of training dataset which allowed
to select appropriate values more efficiently.
Early Stopping. Early stopping halts training when the validation loss stops
decreasing, preventing the model from overfitting to the training data. This
method is especially useful when training deep networks on limited datasets, as is
common in medical imaging.
Ensemble Learning. Ensembling combines predictions from multiple models
to improve robustness and accuracy. Techniques like averaging, majority voting, or
stacking can enhance performance in skin cancer classification. Ghosh et al (2024)
[71] utilized ensemble learning with DCNN [72], Caps-Net [73] and ViT [60].
Non Supervised Learning. Semi-supervised learning. Semi-supervised
learning combines a small amount of labeled data with a large amount of unla-
beled data during training. This approach is particularly beneficial in medical im-
aging, where labeling is expensive and time-consuming. Liu et al. (2020) em-
ployed a semi-supervised learning approach using a mean teacher model to
leverage unlabeled skin lesion images, improving classification performance [7].
Techniques:
Pseudo-Labeling: Assigning labels to unlabeled data using the model’s
own predictions and then retraining the model with this expanded dataset. This
iterative process can improve performance but may propagate errors if the initial
model is not accurate.
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 52
Consistency Regularization: Encouraging the model to produce similar
outputs when input images are perturbed or augmented, leveraging unlabeled data
to learn robust features.
Mean Teacher Model: Utilizing a student-teacher framework where the
teacher model generates targets for the student model, which learns from both
labeled and unlabeled data. A noisy stutent algorithm was successfully used in
ISIC Kaggle competition in melanoma classification challenge [74].
Graph-Based Methods: Modeling data as graphs where nodes represent
samples and edges represent similarities, propagating labels through the graph to
infer labels for unlabeled data.
Self-Supervised Learning. Self-supervised learning is a form of unsupervised
learning where the model learns representations by solving pretext tasks created
from unlabeled data. This approach is gaining traction in medical imaging.
Chaitanya et al. (2020) showed that self-supervised learning on unlabeled medical
images significantly improves model performance on downstream tasks with lim-
ited labeled data [75]. Techniques:
Contrastive Learning: Learning representations by maximizing agree-
ment between differently augmented views of the same data sample. The key idea
is to develop a non-trivial network that preserves similar semantic structure for
two (somewhat modified) versions of the same image and keeps two different
images distinguishable [76]. Models like SimCLR have been adapted for medical
images to learn robust features from unlabeled data [77].
Pretext Tasks: Designing tasks such as predicting image rotations, solv-
ing jigsaw puzzles, or reconstructing distorted images to force the model to learn
meaningful features. Haggerty and Chandra (2024) showed that models pre-
trained using SSL (Barlow Twins) significantly outperformed those pre-trained
with SL on ImageNet in scenarios with limited labeled data specifically using
skin cancer images. Moreover, by applying additional SSL pre-training on
smaller, task-specific datasets (like skin lesion images), SL-pre-trained models
could achieve performance equivalent to SSL models. This demonstrates that
even minimal further SSL pre-training can be as effective as extensive pre-
training on large datasets [78].
EVALUATION AND VALIDATION
Next important step in the DL pipeline is assessing the model’s effectiveness. It
involves measuring the model’s performance on previously unseen data. Testing
dataset is distinct from the validation dataset used during training. The validation
dataset helps evaluate the model’s progress and detect overfitting, enabling tech-
niques like early stopping to be applied.
To accurately identify model quality, an appropriate evaluation strategy and
metrics must be chosen. For instance, when working with relatively small training
and testing datasets, it’s crucial to focus on validating whether observed im-
provements stem from the new approach rather than statistical noise.
Selecting suitable metrics is equally important. In medical diagnostics, for
example, both type I and type II errors carry significant consequences. However,
it is generally considered “better” to classify a healthy patient as sick than to miss
diagnosing an ill patient. Moreover, medical datasets, such as those related to skin
cancer, are often highly imbalanced. In such cases, relying solely on accuracy as
the primary metric can be misleading.
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 53
Validation Strategies. Selecting fitting strategy involves deep understand-
ing of the data. It normally involves investigating the variability of model results
so that we can distinguish a “luck” from “an actual improvement”. However, in
medical imaging, datasets are often limited in size, making this method suscepti-
ble to high variance in performance estimates.
Cross-Validation. K-fold cross-validation (CV) divides the dataset into k
subsets (folds). The model is trained on k-1 folds and validated on the remaining
fold, a process repeated k times. The results are averaged to produce a perform-
ance estimate [79].
Stratified K-fold ensures that each fold maintains the same class distribu-
tion as the original dataset, which is crucial for imbalanced datasets common in
skin cancer classification. Mahesh et al. (2023) employed stratified K-Fold CV to
handle class imbalance problem in [80].
Nested cross-validation addresses the bias in performance estimation due to
hyperparameter tuning by having an inner loop for model selection and an outer
loop for model assessment.
External validation. This approach involves testing the model on entirely in-
dependent datasets from different institutions or populations. This approach pro-
vides a robust assessment of the model’s generalizability [7]. Brinker et al. (2019)
performed an external validation of a DL model for melanoma detection across
different populations, emphasizing the necessity of external validation for assess-
ing generalizability [81].
Evaluation Metrics. Accuracy measures the proportion of correct predic-
tions but can be misleading with imbalanced datasets. Precision shows how many
identified class samples were identified correctly. Recall (Sensitivity) is how
many samples of a class the model was able to identify. Specificity measures the
proportion of true negatives among all actual negatives
F1-Score is harmonic mean of precision and recall:
RecallPrecision
F
11
2
1
.
Balanced accuracy accounts for class imbalance by averaging the recall ob-
tained on each class:
2
RecallPrecision
AccuracyBalanced
.
These metrics provide a balanced view of performance, especially important
in medical diagnosis where false negatives and false positives have different im-
plications. They are recommended to be used together to get a different perspec-
tives in performance.
Another set of effective metrics is Receiver Operating Characteristic (ROC)
Curve and Area Under the Curve (AUC). Receiver Operating Characteristic is a
graph with a y-axis representing Sensitivity and an x-axis representing 1 – Speci-
ficity. It represents relation of classification threshold and correctly classified
samples. In order to easily compare different models AUC is employed — an area
under ROC Curve.
Han et al. (2020) utilized ensemble learning and evaluated their model using
ROC-AUC, precision-recall curves, and conducted statistical significance testing
to demonstrate improvements over previous methods [82].
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 54
OPPORTUNITIES FOR INNOVATION
In the domain of DL-based skin cancer classification, there are several promising
avenues for innovation that can significantly enhance diagnostic accuracy, patient
outcomes, and system scalability. As the field continues to mature, the integration
of cutting-edge technologies can address existing challenges, optimize model per-
formance, and broaden access to dermatological diagnostics.
Multimodal Data Integration. Integrating clinical data (e.g., patient demo-
graphics, lesion history, and genetic markers) with imaging data has the potential
to improve diagnostic accuracy. While current models primarily rely on dermo-
scopic images, including non-visual patient information can provide additional
context, leading to more accurate predictions. For instance, factors like age, lesion
location, and personal/family history of skin cancer are crucial in assessing mela-
noma risk. Models that combine imaging with clinical data have demonstrated
improved sensitivity and specificity, particularly for complex cases [83]. Moldo-
vanu et al. (2021) used surface fractal dimensions and statistical color cluster fea-
tures to improve model quality. Future research should focus on developing archi-
tectures capable of effectively fusing diverse data types. Nikitin et al. utilised
fractal dimension with ViT focusing on different ways of integration of the metric
[84]. Future research should focus on developing architectures capable of effec-
tively fusing diverse data types.
Explainable AI. Building trust in AI-driven diagnostics is essential for clini-
cal adoption. Techniques like Gradient-weighted Class Activation Mapping
(Grad-CAM) [85] and Layer-wise Relevance Propagation (LRP) [86] offer ex-
plainability by highlighting regions in dermoscopic images that contributed to the
model’s decision. These visual explanations can help clinicians understand the
model’s reasoning, enabling them to verify the output and diagnose more confi-
dently. Explainable AI can also aid in identifying potential biases in the model,
particularly concerning underrepresented skin tones, thereby addressing dispari-
ties in diagnostic outcomes. Future work should explore enhancing the interpret-
ability of DL models while maintaining high classification accuracy, especially in
high-stakes settings like oncology.
Edge Computing. Deploying lightweight DL models on edge devices, such
as smartphones, can facilitate real-time skin cancer detection, especially in re-
source-constrained settings. Advances in model optimization, such as pruning,
quantization, and using architectures like MobileNet and EfficientNet, can reduce
the computational load while maintaining accuracy. This is particularly relevant
for underserved regions where access to dermatologists is limited. Edge-based AI
systems can provide preliminary assessments, encouraging users to seek medical
consultation if a lesion is flagged as suspicious. Research in this area should pri-
oritize developing robust, efficient models that can handle diverse image quality
and environmental conditions typical of mobile device usage.
GUIDANCE FOR EMERGING RESEARCHERS
The field of DL for skin cancer classification is both challenging and rewarding,
offering ample opportunities for innovation. However, newcomers to the field
may face a steep learning curve due to the complexity of data, algorithms, and
clinical considerations. Here are some practical tips for emerging researchers to
navigate this evolving landscape:
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 55
Guidance for Emerging Researchers. The field of DL for skin cancer clas-
sification is both challenging and rewarding, offering ample opportunities for in-
novation. However, newcomers to the field may face a steep learning curve due to
the complexity of data, algorithms, and clinical considerations. Here are some
practical tips for emerging researchers to navigate this evolving landscape:
1. Focus on Data Quality and Preprocessing. One of the biggest hurdles
in dermatological AI research is access to high-quality, annotated datasets. Begin
by familiarizing yourself with widely used datasets like ISIC and HAM10000.
Pay special attention to data preprocessing, including normalization, augmenta-
tion, and segmentation techniques, to optimize model performance. Addressing
challenges like class imbalance and image noise is crucial for achieving reliable
results. However, the real experiments to achieve top metrics results must be con-
ducted on larger data since it is available — as of now ISIC Archive includes
more than 400,000 images available to use.
2. Start with Transfer Learning and Fine-Tuning. Given the limited
availability of labeled medical data (ISIC Archive is not nearly as big as Ima-
geNet), leveraging transfer learning from pre-trained models on large-scale data-
sets like ImageNet can accelerate progress. Fine-tuning these models on skin le-
sion datasets can yield competitive results with relatively less training time.
Explore architectures such as ResNet, EfficientNet, and ViT to identify what
works best for your specific use case. Also, empirical results show that ensemble
models do great in the task.
3. Embrace Explainability from the Start. Building interpretability into
your models is essential, especially in healthcare applications where clinicians
require transparent and understandable outputs. Experiment with tools like Grad-
CAM and SHAP [87] to visualize your model’s decision-making process. Priori-
tizing explainability will not only help you gain the trust of the medical commu-
nity but also ensure that your models can be safely deployed in clinical settings.
4. Keep Ethics and Privacy at the Forefront. When dealing with sensitive
medical data, ethical considerations and compliance with regulations like GDPR
are paramount. Consider approaches like federated learning and differential pri-
vacy to ensure patient confidentiality. Being aware of these considerations early
on will help you design ethically responsible research projects that can be more
easily translated into real-world applications.
5. Stay Updated on Emerging Trends. The field of AI in healthcare is rap-
idly evolving, with new techniques like self-supervised learning, multimodal
models, and quantum ML showing potential. Regularly reviewing the latest re-
search, participating in conferences, and engaging with collaborative research
communities like the ISIC Challenge can keep you at the forefront of the field.
Additionally, exploring adjacent fields like personalized medicine and bioinfor-
matics can open up interdisciplinary opportunities.
By focusing on these areas, emerging researchers can build a strong foundation
and contribute meaningfully to the development of AI-driven skin cancer diagnostics.
CONCLUSION
DL has revolutionized the field of skin cancer diagnostics, offering tools that can
potentially match, or even surpass, dermatologist-level performance. However,
the journey from research to clinical application is fraught with challenges. Our
analysis highlights the importance of high-quality data, rigorous preprocessing,
and robust model evaluation in developing reliable diagnostic systems. Address-
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 56
ing issues like data privacy, class imbalance, and underrepresentation of diverse
skin tones remains critical to ensuring equitable healthcare outcomes.
The integration of clinical metadata with imaging data, along with techniques
such as federated learning and edge computing, presents promising avenues to
enhance model performance and broaden access to diagnostics, particularly in
resource-constrained regions. Additionally, incorporating explainable AI methodologies
can help gain the trust of clinicians, paving the way for real-world adoption.
For emerging researchers, focusing on ethical considerations, leveraging
transfer learning, and embracing advancements in multimodal integration are vital
steps toward impactful contributions in this domain. As AI continues to evolve,
its role in skin cancer diagnostics will likely expand, enabling more accurate, ac-
cessible, and personalized care. Future research should aim at bridging the gap
between technological capabilities and clinical needs, ultimately transforming the
landscape of dermatological diagnostics and improving patient outcomes.
REFERENCES
1. Cancer Research UK. Available: https://www.cancerresearchuk.org/
2. Health workforce: The health workforce crisis. Available: https://www.who.int/news-room/
questions-and-answers/item/q-a-on-the-health-workforce-crisis
3. L.G. Wich et al., “Impact of Socioeconomic Status and Sociodemographic Factors on
Melanoma Presentation Among Ethnic Minorities,” Journal of Community Health, vol.
36, no. 3, pp. 461–468, Nov. 2010. doi: 10.1007/s10900-010-9328-4.
4. A. Esteva et al., “Dermatologist-level classification of skin cancer with deep neural net-
works,” Nature, vol. 542, no. 7639, pp. 115–118, Jan. 2017. doi: 10.1038/nature21056.
5. ISIC International Skin Imaging Collaboration. Available: https://www.isic-archive.com/
6. “Can artificial intelligence detect melanoma?” Memorial Sloan Kettering Cancer Center, Jul.
24, 2024. Available: https://www.mskcc.org/news/can-artificial-intelligence-detect-melanoma
7. Y. Liu et al., “A deep learning system for differential diagnosis of skin diseases,” Nature
Medicine, vol. 26, no. 6, pp. 900–908, May 2020. doi: 10.1038/s41591-020-0842-3.
8. A. Yilmaz, M. Kalebasi, Y. Samoylenko, M.E. Guvenilir, H. Uvet, “Benchmarking of
Lightweight Deep Learning Architectures for Skin Cancer Classification using ISIC 2017
Dataset,” arXiv (Cornell University), Jan. 2021. doi: 10.48550/arxiv.2110.12270.
9. M. Baygin, T. Tuncer, S. Dogan, “New pyramidal hybrid textural and deep features
based automatic skin cancer classification model: Ensemble DarkNet and textural feature
extractor,” arXiv (Cornell University), Jan. 2022. doi: 10.48550/arxiv.2203.15090.
10. K. Agarwal, T. Singh, “Classification of Skin Cancer Images using Convolutional Neural
Networks,” arXiv (Cornell University), Jan. 2022. doi: 10.48550/arxiv.2202.00678.
11. D. Wen et al., “Characteristics of publicly available skin cancer image datasets: a sys-
tematic review,” The Lancet Digital Health, vol. 4, no. 1, pp. e64–e74, Nov. 2021. doi:
10.1016/s2589-7500(21)00252-1.
12. “ISIC 2024 — Skin Cancer Detection with 3D-TBP,” Kaggle. Available:
https://www.kaggle.com/competitions/isic-2024-challenge
13. P. Tschandl, C. Rosendahl, H. Kittler, “The HAM10000 dataset, a large collection of
multi-source dermatoscopic images of common pigmented skin lesions,” Scientific Data,
vol. 5, no. 1, Aug. 2018. doi: 10.1038/sdata.2018.161.
14. A.T. Priyeshkumar, G. Shyamala, T. Vasanth, Selvan V. Ponniyin, “Transforming Skin
Cancer Diagnosis: A Deep Learning Approach with the Ham10000 Dataset,” Cancer In-
vestigation, vol. 42, no. 10, pp. 801–814, Nov. 2024. doi: 10.1080/07357907.2024.2422602.
15. T.M. Alam et al., “An efficient Deep Learning-Based skin cancer classifier for an imbalanced
dataset,” Diagnostics, vol. 12, no. 9, p. 2115, Aug. 2022. doi: 10.3390/diagnostics12092115.
16. H.A. Owida et al., “Improved deep learning architecture for skin cancer classification,”
Indonesian Journal of Electrical Engineering and Computer Science, vol. 36, no. 1, p.
501, Jul. 2024. doi: 10.11591/ijeecs.v36.i1.pp501-508.
17. A. Ameri, “A deep learning approach to skin cancer detection in dermoscopy images,”
Journal of Biomedical Physics and Engineering, vol. 10, no. 6, Nov. 2020. doi:
10.31661/jbpe.v0i0.2004-1107.
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 57
18. S.S. Chaturvedi, J.V. Tembhurne, T. Diwan, “A multi-class skin Cancer classification
using deep convolutional neural networks,” Multimedia Tools and Applications, vol. 79,
no. 39–40, pp. 28477–28498, Aug. 2020. doi: 10.1007/s11042-020-09388-2.
19. ADDI - Automatic computer-based Diagnosis system for Dermoscopy Images. Available:
https://www.fc.up.pt/addi/ph2%20database.html)
20. I. Giotis, N. Molders, S. Land, M. Biehl, M.F. Jonkman, N. Petkov, “MED-NODE: A
computer-assisted melanoma diagnosis system using non-dermoscopic images,” Expert
Systems With Applications, vol. 42, no. 19, pp. 6578–6585, May 2015. doi:
10.1016/j.eswa.2015.04.034.
21. Dermofit Image Library available from The University of Edinburgh. Available:
https://licensing.edinburgh-innovations.ed.ac.uk/product/dermofit-image-library
22. X. Sun, J. Yang, M. Sun, K. Wang, “A benchmark for automatic visual classification of
clinical skin disease images,” in Lecture notes in computer science, 2016, pp. 206–222.
doi: 10.1007/978-3-319-46466-4_13.
23. “Datasets used to train AI to detect skin cancer lack information on darker skin and often
incomplete,” Haiku, Nov. 10, 2021. Available: https://www.cancer.ox.ac.uk/ news/lack-
of-data-available-to-detect-skin-cancer-in-darker-skin
24. S. Hu, R.M. Soza-Vento, D.F. Parker, R.S. Kirsner, “Comparison of stage at diagnosis of
melanoma among Hispanic, Black, and white patients in Miami-Dade County, Florida,”
Archives of Dermatology, vol. 142, no. 6, Jun. 2006. doi: 10.1001/archderm.142.6.704.
25. H.M. Gloster, K. Neal, “Skin cancer in skin of color,” Journal of the American Academy
of Dermatology, vol. 55, no. 5, pp. 741–760, Oct. 2006. doi: 10.1016/j.jaad.2005.08.063.
26. “Health Insurance Portability and Accountability Act (HIPAA)”, U.S. Department of
Health and Human Services. Available: https://www.hhs.gov/hipaa/index.html
27. “General Data Protection Regulation (GDPR) – Legal Text,” General Data Protection
Regulation (GDPR), Apr. 22, 2024. Available: https://gdpr-info.eu/
28. S. Khan, H. Ali, Z. Shah, “Identifying the role of vision transformer for skin cancer —
A scoping review,” Frontiers in Artificial Intelligence, vol. 6, Jul. 2023. doi:
10.3389/frai.2023.1202990.
29. P. Tschandl et al., “Human–computer collaboration for skin cancer recognition,” Nature
Medicine, vol. 26, no. 8, pp. 1229–1234, Jun. 2020. doi: 10.1038/s41591-020-0942-0.
30. T.J. Brinker et al., “Deep neural networks are superior to dermatologists in melanoma
image classification,” European Journal of Cancer, vol. 119, pp. 11–17, Aug. 2019. doi:
10.1016/j.ejca.2019.05.023.
31. S.F.H. Shah, D. Arecco, H. Draper, S. Tiribelli, E. Harriss, R.N. Matin, “Ethical implica-
tions of artificial intelligence in skin cancer diagnostics: use-case analyses,” British Jour-
nal of Dermatology, Nov. 2024. doi: 10.1093/bjd/ljae434.
32. E.R. Gordon et al., “Ethical considerations for artificial intelligence in dermatology: a
scoping review,” British Journal of Dermatology, vol. 190, no. 6, pp. 789–797, Feb.
2024. doi: 10.1093/bjd/ljae040.
33. H.A. Haenssle et al., “Man against machine: diagnostic performance of a deep learning
convolutional neural network for dermoscopic melanoma recognition in comparison to
58 dermatologists,” Annals of Oncology, vol. 29, no. 8, pp. 1836–1842, May 2018. doi:
10.1093/annonc/mdy166.
34. A.G.C. Pacheco, R.A. Krohling, “Recent advances in deep learning applied to skin cancer
detection,” arXiv (Cornell University), Jan. 2019. doi: 10.48550/arxiv.1912.03280.
35. P. Popecki, M. Kozakiewicz, M. Ziętek, K. Jurczyszyn, “Fractal Dimension Analysis of
Melanocytic nevi and melanomas in Normal and Polarized Light—A Preliminary Re-
port,” Life, vol. 12, no. 7, p. 1008, Jul. 2022. doi: 10.3390/life12071008.
36. “Melanoma Skin Cancer Dataset of 10000 Images,” Kaggle, Mar. 29, 2022. Available:
https://www.kaggle.com/datasets/hasnainjaved/melanoma-skin-cancer-dataset-of-10000-
images
37. S. Shen et al., “A Low-Cost High-Performance data augmentation for Deep Learning-
Based skin lesion classification,” BME Frontiers, 2022. doi: https://doi.org/
10.34133/2022/9765307
38. G.M.S. Himel, Md.M. Islam, Kh.A. Al-Aff, S.I. Karim, Md.K.U. Sikder, “Skin Cancer
Segmentation and Classification Using Vision Transformer for Automatic Analysis in
Dermatoscopy-Based Noninvasive Digital System,” International Journal of Biomedical
Imaging, vol. 2024, pp. 1–18, Feb. 2024. doi: 10.1155/2024/3022192.
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 58
39. K. Polat, K.O. Koc, “Detection of Skin Diseases from Dermoscopy Image Using the
combination of Convolutional Neural Network and One-versus-All,” Journal of Artificial
Intelligence and Systems, vol. 2, no. 1, pp. 80–97, Jan. 2020. doi: 10.33969/ais. 2020.21006.
40. B. Walker et al., “Dermoscopy diagnosis of cancerous lesions utilizing dual deep learning
algorithms via visual and audio (sonification) outputs: Laboratory and prospective obser-
vational studies,” EBioMedicine, vol. 40, pp. 176–183, 2019. doi:
https://doi.org/10.1016/j.ebiom.2019.01.028
41. M.S. Ali, M.S. Miah, J. Haque, M.M. Rahman, M.K. Islam, “An enhanced technique of skin
cancer classification using deep convolutional neural network with transfer learning models,”
With Applications, vol. 5, 100036, 2021. doi: https://doi.org/ 10.1016/j.mlwa.2021.100036
42. Y. Wu, B. Chen, A. Zeng, D. Pan, R. Wang, S. Zhao, “Skin Cancer Classification With
Deep Learning: A Systematic Review,” Frontiers in Oncology, vol. 12, Jul. 2022. doi:
10.3389/fonc.2022.893972.
43. D. Bisla, A. Choromanska, R.S. Berman, J.A. Stein, D. Polsky, “Towards Automated
Melanoma Detection With Deep Learning: Data Purification and Augmentation,” IEEE
Computer Society Conference on Computer Vision and Pattern Recognition Workshops
(CVPRW), Jun. 2019. doi: 10.1109/cvprw.2019.00330.
44. Z. Qin, Z. Liu, P. Zhu, Y. Xue, “A GAN-based image synthesis method for skin lesion
classification,” Computer Methods and Programs in Biomedicine, vol. 195, p. 105568,
May 2020. doi: 10.1016/j.cmpb.2020.105568.
45. B. Ahmad, S. Jun, V. Palade, Q. You, L. Mao, M. Zhongjie, “Improving Skin Cancer
Classification Using Heavy-Tailed Student T-Distribution in Generative Adversarial
Networks (TED-GAN),” Diagnostics, vol. 11, no. 11, p. 2147, Nov. 2021. doi:
10.3390/diagnostics11112147.
46. I.S.A. Abdelhalim, M.F. Mohamed, Y.B. Mahdy, “Data augmentation for skin lesion us-
ing self-attention based progressive generative adversarial network,” Expert Systems With Ap-
plications, vol. 165, p. 113922, Sep. 2020. doi: 10.1016/j.eswa.2020.113922.
47. R. Kaur, H. Gholam Hosseini, R. Sinha, “Synthetic Images Generation Using Conditional
Generative Adversarial Network for Skin Cancer Classification,” TENCON 2021 - 2021
IEEE Region 10 Conference (TENCON), vol. 27, pp. 381–386, Dec. 2021. doi:
10.1109/tencon54134.2021.9707291.
48. J. Pope, M. Hassanuzzaman, M. Sherpa, O. Emara, A. Joshi, N. Adhikari, “Skin Cancer
Machine Learning Model Tone Bias,” arXiv (Cornell University), 2024. doi:
https://doi.org/10.48550/arxiv.2410.06385
49. E. Rezk, M. Eltorki, W. El-Dakhakhni, “Improving Skin Color Diversity in Cancer De-
tection: Deep Learning Approach,” JMIR Dermatology, 5(3), e39143, 2022. doi:
https://doi.org/10.2196/39143
50. B.N. Walker et al., “Skin cancer detection in diverse skin tones by machine learning
combining audio and visual convolutional neural networks,” Oncology, 103(5), pp. 113–
120, 2024. doi: https://doi.org/10.1159/000541573
51. J. Han, M. Kamber, J. Pei, Data Mining: Concepts and Techniques; 3rd edition, 2011.
Available: https://www.scholartext.com/book/88809627?_locale=fr
52. T. Majtner, B. Bajić, S. Yildirim, J.Y. Hardeberg, J. Lindblad, N. Sladoje, “Ensemble of
Convolutional Neural Networks for Dermoscopic Images Classification,” arXiv (Cornell
University), 2018. doi: https://doi.org/10.48550/arxiv.1808.05071
53. S.M. Pizer et al., “Adaptive histogram equalization and its variations,” Computer Vision,
Graphics, and Image Processing, 39(3), pp. 355–368, 1987.
54. K. Zuiderveld, “Contrast limited adaptive histogram equalization,” in Graphics gems IV.
Academic Press Professional, Inc., 1994, pp. 474–485.
55. C.M. Bishop, Pattern recognition and machine learning. Springer, 2006.
56. V.D. Midasala, B. Prabhakar, J.K. Chaitanya, K. Sirnivas, D. Eshwar, P.M. Kumar,
“MFEUsLNet: Skin cancer detection and classification using integrated AI with multi-
level feature extraction-based unsupervised learning,” Engineering Science and Technol-
ogy an International Journal, 51(10), 101632, 2024. doi: https://doi.org/
10.1016/j.jestch.2024.101632
57. T.K. Lee et al., “Dullrazor®: A software approach to hair removal from images,” Com-
puters in Biology and Medicine, 27(6), pp. 533–543, 1997.
58. R. Maurya et al., “Skin cancer detection through attention guided dual autoencoder ap-
proach with extreme learning machine,” Scientific Reports, vol. 14, no. 17785, Aug.
2024. doi: 10.1038/s41598-024-68749-1.
Navigating challenges in deep learning for skin cancer detection
Системні дослідження та інформаційні технології, 2025, № 2 59
59. A. Kirillov et al., “Segment Anything,” arXiv (Cornell University), Apr. 2023. doi:
https://doi.org/10.48550/arxiv.2304.02643
60. A. Dosovitskiy et al., “An Image is Worth 16x16 Words: Transformers for Image Recog-
nition at Scale,” arXiv.org, Jun. 03, 2021. Available: https://www.arxiv.org/
abs/2010.11929v2
61. S. Targ, D. Almeida, K. Lyman, “Resnet in Resnet: Generalizing Residual Architec-
tures,” arXiv (Cornell University), Jan. 2016ю doi: 10.48550/arxiv.1603.08029.
62. X. Han, Y. Zhong, L. Cao, L. Zhang, “Pre-Trained AlexNet Architecture with Pyramid
Pooling and Supervision for High Spatial Resolution Remote Sensing Image Scene Clas-
sification,” Remote Sensing, vol. 9, no. 8, p. 848, Aug. 2017. doi: 10.3390/rs9080848.
63. Q. Guan et al., “Deep convolutional neural network VGG-16 model for differential diag-
nosing of papillary thyroid carcinomas in cytological images: a pilot study,” Journal of
Cancer, vol. 10, no. 20, pp. 4876–4882, Jan. 2019. doi: 10.7150/jca.28769.
64. P. Carcagnì et al., “Classification of Skin Lesions by Combining Multilevel Learnings in
a DenseNet Architecture,” in Lecture notes in computer science, pp. 335–344, 2019. doi:
10.1007/978-3-030-30642-7_30.
65. D. Sinha, M. El-Sharkawy, “Thin MobileNet: An Enhanced MobileNet Architecture,”
IEEE Annual Ubiquitous Computing, Electronics & Mobile Communication Conference
(UEMCON), pp. 0280–0285, Oct. 2019. doi: 10.1109/uemcon47517. 2019.8993089.
66. L.N. Smith, “Cyclical learning rates for training neural networks,” in 2017 IEEE Winter
Conference on Applications of Computer Vision, pp. 464–472.
67. I. Loshchilov, F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” arXiv
preprint, 2016. doi: https://doi.org/10.48550/arXiv.1608.03983
68. D.P. Kingma, J. Ba, “Adam: A method for stochastic optimization,” in 3rd International
Conference on Learning Representations (ICLR), 2015.
69. R. Ali, A. Manikandan, R. Lei, J. Xu, “A novel SpaSA based hyper-parameter optimized
FCEDN with adaptive CNN classification for skin cancer detection,” Scientific Reports,
vol. 14, no. 1, Apr. 2024. doi: 10.1038/s41598-024-57393-4.
70. Duyen N.T. Le, Hieu X. Le, Lua T. Ngo, Hoan T. Ngo, “Transfer learning with class-
weighted and focal loss function for automatic skin cancer classification,” arXiv (Cornell
University), Jan. 2020. doi: https://doi.org/10.48550/arxiv.2009.05977
71. S. Ghosh, S. Dhar, Raktim Yoddha, S. Kumar, Abhinav Kumar Thakur, Nanda Dulal
Jana, “Melanoma Skin Cancer Detection Using Ensemble of Machine Learning Models
Considering Deep Feature Embeddings,” Procedia Computer Science, vol. 235, pp.
3007–3015, Jan. 2024. doi: https://doi.org/10.1016/j.procs.2024.04.284
72. J. Atwood, D. Towsley, “Diffusion-Convolutional Neural Networks,” arXiv.org, Jul. 08,
2016. Available: https://arxiv.org/abs/1511.02136
73. S. Sabour, N. Frosst, G.E. Hinton, “Dynamic Routing Between Capsules,” arXiv (Cornell
University), Jan. 2017. doi: https://doi.org/10.48550/arxiv.1710.09829
74. “SIIM-ISIC Melanoma Classification,” Kaggle, 2024. Accessed on: Dec. 28, 2024.
Available: https://www.kaggle.com/competitions/siim-isic-melanoma-classification/ dis-
cussion/181830
75. K. Chaitanya, E. Erdil, N. Karani, E. Konukoglu “Contrastive learning of global and lo-
cal features for medical image segmentation with limited annotations,” in Advances in
Neural Information Processing Systems, 2020.
76. T. Chen, S. Kornblith, M. Norouzi, G. Hinton, “A simple framework for contrastive learning
of visual representations,” in International Conference on Machine Learning, 2020.
77. S. Azizi et al., “Big self-supervised models advance medical image classification,” in
Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.
78. H. Haggerty, R. Chandra, “Self-supervised learning for skin cancer diagnosis with lim-
ited training data,” arXiv (Cornell University), Jan. 2024. doi: https://doi.org/
10.48550/arxiv.2401.00692
79. P. Refaeilzadeh, L. Tang, H. Liu, “Cross-validation,” in Encyclopedia of Database Sys-
tems, pp. 532–538. Springer, 2009.
80. T.R. Mahesh, V.V. Kumar, D.V. Kumar, O. Geman, M. Margala, M. Guduri, “The strati-
fied K-folds cross-validation and class-balancing methods with high-performance en-
semble classifiers for breast cancer classification,” Healthcare Analytics, vol. 4,
p. 100247, Sep. 2023. doi: 10.1016/j.health.2023.100247.
V. Nikitin, V. Danilov
ISSN 1681–6048 System Research & Information Technologies, 2025, № 2 60
81. T.J. Brinker et al., “Deep learning outperformed 136 of 157 dermatologists in a head-to-
head dermoscopic melanoma image classification task,” European Journal of Cancer,
vol. 113, pp. 47–54, 2019.
82. S.S. Han et al., “Deep neural networks show an equivalent and often superior perform-
ance to dermatologists in onychomycosis diagnosis: Automatic construction of onycho-
mycosis datasets by region-based convolutional deep neural network,” PLOS One, vol.
13, no. 1, p. e0191493, Jan. 2018. doi: https://doi.org/10.1371/ journal.pone.0191493
83. S. Moldovanu, F.A. Damian Michis, K.C. Biswas, A. Culea-Florescu, L. Moraru, “Skin
Lesion Classification Based on Surface Fractal Dimensions and Statistical Color Cluster
Features Using an Ensemble of Machine Learning Techniques,” Cancers, vol. 13, no. 21,
p. 5256, Oct. 2021. doi: https://doi.org/10.3390/cancers13215256
84. V. Nikitin, V. Danylov, “Integration of Fractal Dimension in Vision Transformer for
Skin Cancer Classification,” Electronics and Control Systems, vol. 2, no. 76, pp. 15–20,
Jun. 2023. doi: https://doi.org/10.18372/1990-5548.76.17662
85. R.R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, “Grad-CAM:
Visual Explanations from Deep Networks via Gradient-based Localization,” Interna-
tional Journal of Computer Vision, vol. 128, no. 2, pp. 336–359, Feb. 2020. doi:
https://doi.org/10.1007/s11263-019-01228-7
86. A. Binder, G. Montavon, S. Bach, K.-R. Müller, and W. Samek, “Layer-wise Relevance
Propagation for Neural Networks with Local Renormalization Layers,” arXiv (Cornell
University), Apr. 2016. doi: https://doi.org/10.48550/arxiv.1604.00825
87. S. Lundberg, S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,”
arXiv.org, Nov. 24, 2017. Available: https://arxiv.org/abs/1705.07874v2
Received 09.01.2025
INFORMATION ON THE ARTICLE
Vladyslav O. Nikitin, ORCID: 0009-0001-9921-0213, Educational and Research Insti-
tute for Applied System Analysis of the National Technical University of Ukraine “Igor
Sikorsky Kyiv Polytechnic Institute”, Ukraine, e mail: nvo63911@gmail.com
Valery Ya. Danilov, ORCID: 0000-0003-3389-3661, Educational and Research Institute
for Applied System Analysis of the National Technical University of Ukraine “Igor Sikor-
sky Kyiv Polytechnic Institute”, Ukraine, e mail: danilov1950@ukr.net
ПОДОЛАННЯ ВИКЛИКІВ У ГЛИБОКОМУ НАВЧАННІ ДЛЯ ВИЯВЛЕННЯ
РАКУ ШКІРИ / В.О. Нікітін, В.Я. Данилов
Анотація. Рак шкіри є одним із найпоширеніших злоякісних новоутворень у
світі. Раннє виявлення є критичним фактором зниження рівня смертності. Це
підкреслює необхідність доступних систем комп'ютерної діагностики. Нещо-
давні досягнення в глибокому навчанні показали великі перспективи у вирі-
шенні цієї проблеми. Незважаючи на цей прогрес у галузі машинного навчан-
ня, дослідники стикаються із численними перешкодами, коли йдеться про
класифікацію раку шкіри. Розглянуто сучасний стан діагностики раку шкіри на
основі глибокого навчання, критичні аспекти розроблення системи, включно
з попереднім обробленням даних, навчанням моделей та оцінкою продуктив-
ності. Крім того, висвітлюються можливості для інновацій, які можуть значно
просунути цю галузь. Надаючи вичерпний огляд, стаття має на меті допомогти
дослідникам та практикам в оптимізації моделей глибокого навчання, усуненні
існуючих обмежень та дослідженні нових тенденцій для підвищення точності
та доступності діагностики.
Ключові слова: рак шкіри, глибоке навчання, класифікація, трансформатори,
CNN, GAN, попереднє оброблення даних, доповнення даних.
|
| id | journaliasakpiua-article-320423 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-09-17T09:26:02Z |
| publishDate | 2025 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/d6/7b565060864f5ddec76e75a1c2e01dd6.pdf |
| spelling | journaliasakpiua-article-3204232025-07-25T15:56:08Z Navigating challenges in deep learning for skin cancer detection Подолання викликів у глибокому навчанні для виявлення раку шкіри Nikitin, Vladyslav Danilov, Valery рак шкіри глибоке навчання класифікація трансформатори CNN GAN попереднє оброблення даних доповнення даних skin cancer deep learning classification transformers CNN GAN data preprocessing data augmentation Skin cancer is one of the most prevalent malignancies worldwide. A critical factor in reducing mortality rates is the early detection. It underscores the need for accessible Computer-Aided Diagnostic (CAD) systems. Recent advancements in Deep Learning (DL) have shown great promise in addressing this challenge. Despite this progress in the field of machine learning, researchers encounter numerous obstacles when it comes to skin cancer classification. This article examines the current state of DL-based skin cancer diagnostics. Critical aspects of system development, including data preprocessing, model training, and performance evaluation, are addressed. Moreover, the article highlights opportunities for innovation that could significantly advance the field. By providing a comprehensive overview, this article aims to guide researchers and practitioners in optimizing DL models, addressing existing limitations, and exploring emerging trends to enhance diagnostic accuracy and accessibility. Рак шкіри є одним із найпоширеніших злоякісних новоутворень у світі. Раннє виявлення є критичним фактором зниження рівня смертності. Це підкреслює необхідність доступних систем комп'ютерної діагностики. Нещодавні досягнення в глибокому навчанні показали великі перспективи у вирішенні цієї проблеми. Незважаючи на цей прогрес у галузі машинного навчання, дослідники стикаються із численними перешкодами, коли йдеться про класифікацію раку шкіри. Розглянуто сучасний стан діагностики раку шкіри на основі глибокого навчання, критичні аспекти розроблення системи, включно з попереднім обробленням даних, навчанням моделей та оцінкою продуктивності. Крім того, висвітлюються можливості для інновацій, які можуть значно просунути цю галузь. Надаючи вичерпний огляд, стаття має на меті допомогти дослідникам та практикам в оптимізації моделей глибокого навчання, усуненні існуючих обмежень та дослідженні нових тенденцій для підвищення точності та доступності діагностики. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2025-06-28 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/320423 10.20535/SRIT.2308-8893.2025.2.03 System research and information technologies; No. 2 (2025); 42-60 Системные исследования и информационные технологии; № 2 (2025); 42-60 Системні дослідження та інформаційні технології; № 2 (2025); 42-60 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/320423/324673 |
| spellingShingle | рак шкіри глибоке навчання класифікація трансформатори CNN GAN попереднє оброблення даних доповнення даних Nikitin, Vladyslav Danilov, Valery Подолання викликів у глибокому навчанні для виявлення раку шкіри |
| title | Подолання викликів у глибокому навчанні для виявлення раку шкіри |
| title_alt | Navigating challenges in deep learning for skin cancer detection |
| title_full | Подолання викликів у глибокому навчанні для виявлення раку шкіри |
| title_fullStr | Подолання викликів у глибокому навчанні для виявлення раку шкіри |
| title_full_unstemmed | Подолання викликів у глибокому навчанні для виявлення раку шкіри |
| title_short | Подолання викликів у глибокому навчанні для виявлення раку шкіри |
| title_sort | подолання викликів у глибокому навчанні для виявлення раку шкіри |
| topic | рак шкіри глибоке навчання класифікація трансформатори CNN GAN попереднє оброблення даних доповнення даних |
| topic_facet | рак шкіри глибоке навчання класифікація трансформатори CNN GAN попереднє оброблення даних доповнення даних skin cancer deep learning classification transformers CNN GAN data preprocessing data augmentation |
| url | https://journal.iasa.kpi.ua/article/view/320423 |
| work_keys_str_mv | AT nikitinvladyslav navigatingchallengesindeeplearningforskincancerdetection AT danilovvalery navigatingchallengesindeeplearningforskincancerdetection AT nikitinvladyslav podolannâviklikívuglibokomunavčannídlâviâvlennârakuškíri AT danilovvalery podolannâviklikívuglibokomunavčannídlâviâvlennârakuškíri |