Developing a trustworthy and explainable framework for classifying skin lesions through transfer learning and...

Detection of high precision skin lesions, especially melanoma, are still a major challenge in medical imagination due to their close visual equality and lack of reliably labeled datasets. In this study, we introduce a deep learning sketch aimed at balancing clinical accuracy with clinical interpretation. The workflow starts with a series of preprosaresing steps: removing hair from dermoscopic images, correction with the cow and separating the wound area using a U-NET segmentation model. On top of that, a skilled-B4 network was properly set and increased with a competition block meditation module (CBAM) to focus the model on clinically important properties. In order to further strengthen performance, this spine was integrated into a dress with its -201 and Renex -50, where predictions are added through a soft poll. The model output was interpreted by the use of character comb and lime, which provides visual clarification of areas affecting the final decision. The training was held on him10000 datasets and valid against ISIC-2019 and pH, which demonstrated the contour's ability to generalize in different wound categories. The model reached 98.95 % accuracy, 98.7 % balanced accuracy and 99.6 % sensitivity to melanoma, improvement in recent benchmarks. By combining efficiency, interpretation and design of privacy and inconvenience, the framework gives a realistic step towards safe and reliable integration of AI units into dermatology practice.