This study introduces an innovative approach to skin disease classification by integrating Hierarchical Dropout into a Convolutional Neural Network (CNN) architecture. This innovation strategy exhibits several advantages, addressing the complexities of skin disease classification effectively. Hierarchical Dropout, operating hierarchically, enables adaptive adjustment of dropout rates across hidden layers, accommodating the diverse spectrum of human skin disease conditions in the dataset. Focusing on network layer outputs rather than modifying input data, it applies a dropout mechanism, dedicated for preventing overfitting by reducing feature-specific dependencies. Variable dropout rates, linked to various human skin disease conditions, facilitate model adaptability to different conditions and types. The tailored regulation of dropout extends to both convolutional and fully connected layers, ensuring comprehensive feature learning while guarding against overfitting. Moreover, the study incorporates Weighted Ensembling, combining predictions from various models with weights assigned based on validation set performance. This technique enhances classification accuracy by capitalizing on the strengths of multiple models. The adoption of Probabilistic Output Layers, employing a Bayesian neural network approach, produces probabilistic predictions presented as probability distributions over classes. This captures the intrinsic uncertainty in skin disease classification, essential for clinical diagnostics. The proposed tailored regularization results in a robust, adaptive, and reliable approach essential for clinicians and dermatologists relying on accurate skin disease diagnoses.

Dermatological Diagnosis; Skin Disease Classification; Hierarchical Dropout; Convolutional Neural Networks