Hybrid multiclass classification of dermatoscopic images based on clinically interpretable features and convolutional neural networks

Abstract

This study presents a hybrid approach to multiclass classification of dermoscopic images by combining clinically interpretable ABCD features with deep features extracted by convolutional neural networks. The experiments were conducted using the publicly available HAM10000 dataset and a local clinical dataset from the Samara Regional Clinical Oncology Dispensary. Since the local images were underexposed and had a green tint, color normalization based on histogram matching was applied to reduce differences between the data sources.
A software module was developed to automatically calculate asymmetry, border sharpness, color heterogeneity, dermoscopic structures, and the total dermoscopy score. Classical machine learning models trained only on these features achieved a best macro-averaged F1 score of approximately 0.26, confirming that ABCD and TDS alone are insufficient for reliable multiclass discrimination. Therefore, hybrid classifiers were constructed by combining the clinical feature vector with embeddings, logits, or class probabilities extracted from DenseNet, MobileNet, ResNet, and EfficientNet models. The fused representations were classified using support vector machines or multilayer perceptrons.
In most experiments, hybridization improved classification performance by 1-2.5 percentage points. The best result was obtained with the ABCD + EfficientNet Logits + MLP configuration, which achieved an accuracy of 0.851 and a macro-F1 score of 0.734. A desktop application was also developed to demonstrate image loading, ABCD and TDS calculation, lesion classification, probability output, and visual analysis. The proposed approach may serve as a basis for a clinically interpretable decision-support system for dermoscopic image analysis.

Speaker

Parfenova Ekaterina
Samara National Research University
Russia

Discussion

Ask question