Skin Disease Classification System

Abstract

Our approach at skin disease classification system aims to classify seven types of skin diseases using the HAM10000 dataset. In this study, we integrated deep learning techniques, specifically leveraging VGG19 pre-trained on ImageNet for feature extraction and GLCM for texture analysis in medical imaging. VGG19 provided intricate pattern recognition capabilities through its convolutional layers, while GLCM computed texture features based on pixel intensity relationships, enhancing the characterization of skin disease. These extracted features were fused into a unified feature vector, combining semantic richness from VGG19 with fine-grained texture details from GLCM. The model was fine-tuned using transfer learning, initially freezing the pre-trained weights to preserve learned features, then unfreezing specific layers to adapt to the HAM10000 dataset of skin diseases. Additional custom layers including ReLU activations, batch normalization for stable gradients, and dropout regularization for mitigating overfitting were applied. Training utilized the Adam optimizer with a categorical cross-entropy loss function to optimize model parameters, achieving classification accuracy as the primary evaluation metric. After training for 100 epochs with batch size 64, the model achieved training accuracy of 81.18%, validation accuracy of 79.69%, training loss of 0.5880 and validation loss of 0.6170.