Adaptive Fuzzy-Reinforced Deep Learning Framework Advancing Precision Dermatology Through Intelligent Skin Disease Classification

This paper presents a hybrid fuzzy convolutional neural network with reinforcement learning (HFCNN-RL) for automated skin disease classification, achieving state-of-the-art performance with 88.01% accuracy and 0.9738 AUC-ROC. The proposed framework integrates fuzzy logic to handle uncertainty in dermatological images while reinforcement learning optimises classification policies through adaptive reward mechanisms. Comprehensive evaluation across four skin disease categories (contact_polarized, contact_non_polarized, non_contact_polarized, and others) demonstrates statistically significant results ( p < 10 ^-70 ) across all statistical tests. Systematic analysis of filter configurations reveals that 64 filters achieve optimal performance while 32 filters provide the best balance between accuracy (85.22%) and computational efficiency. The reinforcement learning component shows remarkable resilience, recovering from a 14.29% accuracy dip during training to achieve peak performance. Despite limitations including computational demands and occasional optimisation instability, the HFCNN-RL framework demonstrates significant potential for clinical deployment in dermatology practice.