An interpretable skin cancer detection and classification framework using meta-heuristic driven symmetric convolution-based adaptive MobileNet with LSTM layer.

Abstract

Skin cancer requires early and accurate detection for effective treatment. Automated image analysis can help identify skin lesions and support clinical diagnosis. This study aims to develop an efficient framework for accurate skin cancer detection using lesion segmentation, deep learning, and optimization. Skin lesion images are first segmented using MobileUNet++ with Spatial Attention (MUNet++-SA) to identify lesion regions accurately. The segmented images are then classified using the proposed SAMNet-LSTM model. The model parameters are optimized using Rationalized Masterpiece Optimization (RMO) to improve classification performance. The proposed method is compared with existing approaches. The proposed MUNet++-SA achieves improved segmentation performance. Its mean IoU is 16.47%, 15.88%, 14.11%, and 8.23% higher than U-Net, U-Net3+, ResUNet, and DenseUNet, respectively. The RMO-optimized SAMNet-LSTM also provides improved and efficient skin cancer classification. The proposed framework combines accurate lesion segmentation with optimized deep learning for reliable skin cancer detection. It can support faster and more effective clinical decision-making.