Transfer Learning for Dermatological Image Classification: A Comparative Study of Adaptation Strategies under Leakage-Aware Evaluation

The main contributions of this paper are:
• A unified framework for comparing transfer-learning
strategies and ImageNet-pretrained CNNs under identical
experimental conditions.
• A dataset auditing method that identifies duplicate and
near-duplicate images using feature similarity and geometric verification.
• An analysis of how many images in a widely used hair
and scalp disease dataset are genuinely unique.
• A comparison of naive image-level splitting against
lesion-grouped splitting on HAM10000.
• An examination of how model-specific input preprocessing affects reported results, showing why consistent
settings matter when comparing transfer-learning models.