DEEP LEARNING-BASED MELANOMA CLASSIFICATION ENHANCED BY FRACTAL DIMENSION ANALYSIS

40 2025 / 4 KPI Science News DOI: https://doi.org/10.20535/kpisn.2025.4.343191 UDC 004.855.5 V.O. Nikitin1*, V.Ya. Danilov1 1 Igor Sikorsky Kyiv Polytechnic Institute, Kyiv, Ukraine *corresponding author: DEEP LEARNING-BASED MELANOMA CLASSIFICATION ENHANCED BY FRACTAL DIMENSION ANALYSIS Background. Melanoma is a malignant skin lesion that is prone to metastasise aggressively, leading to an almost guaranteed lethal outcome if left unchecked. In contrast, early-stage detection allows for the tumour to be removed via a harmless surgical procedure that may not even leave a scar. However, the availability of competent diagnostics are often limited due to a shortage of healthcare specialists and technologies. Deep Learning models such as Visual Transformer (ViT) have demonstrated strong performance, but researchers continuously seek to improve the results by incorporating new features. Since human skin exhibits fractal-like characteristics, it is theorised that metrics quantifying this complexity can act as valuable supplementary features for DL models, leading to increased classification accuracy. Objective. We investigated the impact of the integration of fractal dimension (FD) on a Vision Transformer deep learning model used for melanoma classification. A comparison was conducted between the model that was exposed to random noise and the models that were provided with computed FD values. Methods. Vision Transformer was used as a feature-extracting backbone pre-trained on the ImageNet dataset. Fine-tuning was done on this backbone in combination with a classification head targeted to distinguish melanoma vs. nevus classes. Along with extracted features, the classification head received FD value. An identical model received random noise instead of FD. Statistical testing and FD impact analysis were conducted to validate the significance of the new feature. Results. Integrating FD into ViT showed noticeable improvement in test metrics. SHAP analysis confirmed the meaningfulness of the new feature. McNemarʼs test validated that the difference in model predictions was statistically significant. Conclusions. The results suggest that FD can serve as a valuable supplementary feature for DL models, and the integration of biomarkers such as FD provides a basis for more robust melanoma classification. Keywords: deep learning; vision transformer; melanoma; fractal dimension; XAI; skin cancer. Introduction Malignant skin lesions, such as squamous cell carcinoma and melanoma, can metastasise at advanced stages, significantly reducing the chances of successful treatment. Among those, melanoma is considered the most aggressive skin cancer type. Late-stage melanoma has a low likelihood of a positive outcome. In contrast, early-stage lesions can often be surgically removed with just minimal or no scarring [1]. However, the availability of competent diagnostics is often limited due to a shortage of healthcare specialists and technologies. Consequently, there is ongoing research focused on developing robust computer-aided diagnostic (CAD) systems leveraging deep learning (DL) techniques, which is being undertaken by various teams, including ISIC [2]. Data feature engineering is a significant part of any machine learning (ML) pipeline. Since human skin exhibits fractal-like characteristics, it is hypothesized that Fractal Dimension (FD) may serve as a valuable feature for enhancing DL-based skin lesion classification models [3]. FD is a metric that quantifies the complexity of fractal-like structures. To investigate this hypothesis, we employed the Vision Transformer (ViT) [4] model as a feature extractor, as it has demonstrated strong performance in skin cancer classification tasks [5]. Пропозиція для цитування цієї статті: В.О. Нікітін, В.Я. Данилов, “Модель глибокого навчання для класифікації меланоми, покращена за допомогою фрактальної розмірності”, Наукові вісті КПІ, № 4, с. 40–45, 2025. doi: https://doi.org/10.20535/kpisn.2025.4.343191 Offer a citation for this article: V.O. Nikitin, V.Ya. Danilov, “Deep learning-based melanoma classif ication en-hanced by fractal dimension analysis”, KPI Science News, no. 4, pp. 40–45, 2025. doi: https://doi.org/10.20535/ kpisn.2025.4.343191 © The Autor(s). The article is distributed under the terms of the license CC BY 4.0 СИСТЕМНИЙ АНАЛІЗ ТА НАУКА ПРО ДАНІ Contributions We hypothesise that integrating FD as a feature to DNN Skin Cancer Classifier can improve the results. The main contributions of this study are as follows: We developed a deep neural network (DNN) classifier that combines ViT-extracted features with FD as an additional input. We conducted statistical analysis, including McNemar’s test, to confirm the significance of the observed performance improvements after incorporating FD. Evaluation of the FD impact was performed using SHAP (SHapley Additive exPlanations). Related Work Fractal Dimension in Skin Lesion Analysis. Fractal Dimension has been explored as a quantitative metric to capture the complexity and irregularity of skin lesion boundaries. Studies have demonstrated that FD can potentially serve as a discriminative feature in differentiating between benign and malignant lesions [3]. For instance, research utilising the Higuchiʼs method for computing surface FD, combined with colour features, achieved classification accuracy of approximately 80 % [6]. Despite these promising results, the integration of FD into DL architectures for skin lesion classification remains fairly underexplored. We previously made efforts on the integration of FD to DL models such as Vision Transformer. The study showed that FD can positively impact a modelʼs output [7]. Another of our studies explored approximating FD for skin lesions. In this study, the box counting dimension and its modification were compared against the Hausdorff dimension of real fractals [8]. Vision Transformers in Skin Cancer Classification. ViTs leverage self-attention mechanisms to capture global contextual information, which is particularly beneficial in analysing complex skin lesion patterns. Recent studies have demonstrated the efficacy of ViTs in skin cancer classification tasks, achieving high accuracy rates [5]. For example, a study employing a ViT model on the HAM10000 dataset reported an accuracy of 96.15 % [9]. Explainable AI in Medical Image Classification. The integration of explainable AI (XAI) techniques in medical image classification has become increasingly important to enhance model transparency and trustworthiness. SHapley Additive exPlanations (SHAP) is one such technique that provides insights into feature contributions towards model predictions. In the context of skin lesion classification, SHAP has been utilised to interpret 41 model decisions, thereby aiding in the validation and acceptance of AI systems in clinical settings [10]. Methods Model Selection and Study Design. The Vision Transformer architecture was selected as the base model, buil (...truncated)