ABSTRACT
This chapter presents an extensive experimental evaluation of the proposed artificial intelligence (AI)-based framework for automated skin cancer detection using dermoscopic images. The analysis integrates segmentation, classification, and explainability within a unified architecture. A Modified SegNet model achieved precise lesion localization with an IoU of 0.98 and a Dice coefficient of 0.95, enhancing subsequent classification accuracy by 5.6%. Multiple deep learning architectures—including CNN, VGG16, VGG19, and Xception—were benchmarked against the proposed Attention-Based CNN (ABCNN), which achieved the highest accuracy (98.9%), AUC (0.98), and fastest convergence (35 epochs). Statistical validation confirmed significant performance improvement (p < 0.01). Explainable AI methods such as Grad-CAM, LIME, and Layer-wise Relevance Propagation (LRP) validated the model’s interpretability and clinical reliability. Overall, the results demonstrate dermatologist-level accuracy, scalability, and transparency, establishing the SegNet–ABCNN pipeline as a reproducible, efficient, and explainable framework for real-world skin cancer detection.
