TRANSFER LEARNING FOR SEVEN-CLASS SKIN LESION CLASSIFICATION ON A HAM10000-LIKE SYNTHETIC DATASET: A COMPARATIVE STUDY

Abstract

Automated dermoscopic image analysis has become an active research direction for assisting early triage of skin lesions. The aim of this work is to evaluate the performance of two ImageNet pre-trained convolutional backbones, DenseNet121 and EfficientNetB0, with respect to seven classes of skin lesion classification on a procedurally generated corpus of HAM10000-like dermoscopic images. An RGB-image dataset of 2,340 images of various skin lesions was compiled: actinic keratoses/intraepithelial carcinoma (akiec), basal cell carcinoma (bcc), benign keratosis-like lesions (bkl), dermatofibroma (df), melanoma (mel), melanocytic nevi (nv) and vascular lesions (vasc). Then the images were divided into 60% training, 20% validation and 20% test sets. Both backbones were frozen and trained with a shallow dense classification head, Adam optimization, sparse categorical cross-entropy, and geometric data augmentation. DenseNet121 obtained 100.0% test accuracy and a macro F1-score of 1.00 on the synthetic test set, whereas EfficientNetB0 obtained 57.26% accuracy, a macro F1-score of 0.10, and collapsed to the majority nv class. The findings show that transfer-learning behavior can vary substantially across architectures under class imbalance and simplified synthetic visual patterns. However, the perfect DenseNet121 score should be interpreted as evidence of task simplicity and distribution bias rather than clinical readiness. The manuscript therefore presents a transparent simulation-based baseline and identifies the validation steps required before journal submission or clinical interpretation