Enhancing Interpretability of Skin Lesion Classification Using Grad-CAM and Weighted Grad-CAM

This paper presents a comprehensive study on the application of deep learning techniques, particularly transfer learning with the ResNet50 model, to the classification of skin lesions and provides insights into the model's decision. The research emphasizes both model accuracy and interpretability, employing Gradient-weighted Class Activation Mapping (GradCAM) and its weighted variant to enhance the transparency of the model's decision-making process. We provide a detailed exploration of the dataset, model architecture, training process, and results. Each result is analyzed with a focus on what the visualizations reveal about the model's behavior. The paper aims to bridge the gap between complex deep-learning methodologies and the need for practical, interpretable applications in the medical field, specifically in dermatology. The model achieved a final training accuracy of 97.83 % and a validation accuracy of 81.83 % on the HAM10000 dataset, with Grad-CAM and novel weighted Grad-CAM visualizations offering significant insights into the model's focus areas during classification tasks.