DEEP LEARNING-POWERED SKIN DISEASE CLASSIFICATION: OPTIMIZING TRANSFER LEARNING FOR IMPROVED ACCURACY

Skin diseases represent a prevalent global health challenge, with prevalence rates in India ranging from 7.9% to 60%. While deep learning approaches show promise for automated diagnosis, existing methods exhibit limitations in robustness and generalizability due to: (1) inadequate feature selection methodologies, (2) suboptimal data augmentation strategies, (3) unexplored hybrid frameworks and (4) insufficient validation of attention mechanisms in clinical settings. These study addresses these gaps through three key innovations: comprehensive optimizer-architecture benchmarking, a novel dual-dataset validation framework and an integrated preprocessing pipeline combining seven enhancement techniques. We systematically evaluate four deep learning architectures (custom CNN, VGG16, DenseNet-121, Inception-ResNet-v2) with three optimization algorithms (Adam, SGD, RMSprop) on partitioned HAM10000 datasets containing 10,015 dermatoscopic images across seven disease categories. Our approach reveals new knowledge: (1) VGG16 with Adam achieves state-of-the-art 93.14% accuracy- the highest reported for single-model HAM10000 classification; (2) RMSprop unexpectedly outperforms Adam for DenseNet121 (83.86% vs 81.43%); and (3) dataset-specific optimizer behaviors critically impact clinical applicability. These findings establish that systematic evaluation of the model optimizer data set significantly improves diagnostic robustness. Research provides a foundation for affordable and accessible diagnostic tools with clinically actionable insights for deployment optimization, which could benefit populations with limited healthcare access.