Skin Lesion Diagnosis via ResNet50 Deep Features and Handcrafted Texture/Color Fusion

Abstract

Diagnosis of skin cancer needs to be performed with accurate and early detection, but most deep learning models do not account for clinically significant visual features like texture and color variation. In the present work, we introduce a hybrid framework that combines deep features obtained with the help of a pretrained ResNet50 model with handcrafted texture (Local Binary Patterns) and color descriptors, including both data-driven and clinically meaningful information. A target-driven resampling strategy is used to balance learning in all lesion categories in order to deal with the problem of class imbalance in the HAM10000 dataset. Experimental evaluation using stratified 5-fold cross-validation demonstrates robust performance, achieving a mean accuracy of 90.62% (±0.81), precision of 90.66% (±0.79), recall of 90.62% $(\pm 0.81)$, and F1-score of $90.61 \% (\pm 0.80)$. Also, the hold-out test set evaluation results in an accuracy of 89.87% which proves the ability of the model to generalize. The results indicate that, by incorporating clinically inspired handcrafted features with deep learning, classification robustness can be enhanced, without losing computational efficiency.