Explainable AI for skin cancer detection using convolutional neural networks(CNNs) with SE and grad-CAM

Summary: Abstract Body: Early detection of melanoma,a life-threatening skin cancer,is crucial as it spreads rapidly if untreated.Dermoscopic patterns help dermatologists distinguish benign from malignant lesions.CNN has shown promise in automating and enhancing diagnostic accuracy for skin cancer classification.This study reviews AI-driven approaches for skincancer classification,focusing on models with attention mechanisms to improve diagnostic performance.Our goal is to bridge gaps in computational efficiency and model explainability using the Squeeze and Excitation(SE) mechanism and the Grad-CAM EAI technique.The SE mechanism enhances the model's focus on critical image features,while Grad-CAM provides heatmap visualizations,allowing clinicians to understand the model's predictions.A systematic review of 650 articles from 2012–2024 from PubMed,PMC and Scopus was conducted.Studies on image-based skin cancer classification were analyzed,emphasizing segmentation and classification of dermoscopic patterns.The dataset included dermoscopic images from ISIC,HAM10000,and PH2 repositories,ensuring diversity in global skin types and conditions.The proposed CNN-SE model demonstrated superior performance metrics compared to traditional CNN approaches.Across a dataset of dermoscopic images,the model achieved 92%accuracy,94% sensitivity,and 90%specificity.Grad-CAM visualizations highlighted lesion regions critical to diagnosis, aligning with dermatological expertise and reinforcing the model's reliability.Compared to existing models such as ResNet,the CNN-SE model achieved 10%improvement in diagnostic accuracy and reduced computational time by 15%.Integrating this model into teledermatology and resource-limited settings can improve early diagnosis and reduce global melanoma detection disparities.By utilizing diverse datasets,it has the potential to enhance outcomes and lower melanoma-related mortality.Future work will focus on validating the model with larger,multi-ethnic datasets and real-time clinical integration. Balakrishnan Kamaraj¹, Hrithik Dakssesh Putta Nagarajan¹, Gurunathan Srinivasan¹, Eshana Kaur¹, Jagjot Singh¹, Aswath Sreeman Saravanan¹, Miruthula Murugan¹, Naveen Vishwanath¹ 1. Madurai Medical College, Madurai, TN, India. Bioinformatics, Computational Biology, and Imaging