Two-stage CNN with weakly supervised segmentation for skin lesion classification

The accurate classification of skin lesions is critical for the early detection and diagnosis of potentially serious skin conditions. This research presents a novel approach that employs a two-stage CNN methodology for developing an advanced skin lesion classification system. In the first stage, we enhance the CNN model training process with weakly supervised segmentation, using Grad-CAM to extract essential features from skin lesion images. The second stage utilizes the CNN to categorize skin lesions based on these extracted features. Our two-stage approach demonstrates significant improvements in the accuracy of feature identification within skin images, leading to better interpretability and increased confidence in decision-making. The system’s effectiveness was evaluated using comprehensive classification metrics on a large dataset of skin lesions. Our results show that this method significantly enhances sensitivity, specificity, and accuracy in skin lesion classification. While the proposed approach advances current techniques, there remains potential for further improvements in the precision of skin lesion classification systems. We anticipate that this method will contribute to more effective and accurate dermatological diagnoses in clinical settings.