An ensemble deep learning framework for skin disease classification with explainable and generative AI support

Accurate and timely identification of skin disorders from dermoscopic images is essential to support effective clinical decisions. In this paper, a deep learning-based approach is proposed to classify multi-class skin disorders using pre-trained convolutional neural networks. Four popular models, including EfficientNetB0, MobileNetV2, DenseNet121, and ResNet50, are trained and compared to evaluate their performance differences. The total number of classes considered for this study was five and one as no class, and the images were taken from the HAM10000 and ISIC2019 datasets. In addition, to improve the robustness of the model performance, a maximum confidence ensemble technique is used to combine the performance results of each individual model. Experimental results indicate that individual models can reach high training accuracy with values over 96%, but their test accuracy values are distributed between 85 and 88%. However, the proposed test accuracy reached 90.74%, showing a significant improvement to reduce model-specific deviation. Apart from the classification, the model also encompasses a generative AI module for the automated generation of structured and disease-specific information regarding the symptoms, causes, and treatment for the predicted class. The model will therefore be discriminative and generative to provide predictions and information to the user. The model has shown the efficiency of the use of ensemble deep learning and generative AI in designing an intelligent health assistance system for the analysis of skin diseases.