Skin cancer, the leading type of cancer, poses a serious risk to public health, most notably melanoma, which is fatal if not treated. Early diagnosis is essential, but traditional diagnosis has low precision due to the poor image quality and the challenges of visual discrimination. Using the publicly available HAM10000 dataset, which is suitable for segmentation and multi-class classification, we propose a robust deep learning-based hybrid system for classification, segmentation, and severity analysis of skin cancer. The model begins with advanced preprocessing in the ELPWF module to pre-process the image and eliminate noise. Enhanced images are then processed with the TS-HCaps feature extraction algorithm to capture complex temporal and hidden features while reducing the dimensionality problem. The best features are selected using the TCWOA module to reduce computational complexity before segmentation by the PA-HRST model, achieving an HD analysis of 4%, and an ASSD of 0.008078, which is higher than existing schemes. The extracted features are then forwarded to the GA-MSKAD hybrid classification model, employing global attention to extract channel and spatial features and accurately classify skin cancer types including AKIEC, BCC, BKL, DF, MEL, NV, and VASC with a 99.18% accuracy, precision of 99.09%, recall of 99.13%, specificity of 99.03% and an F1-score of 99.11%. Finally, the severity is forecast by applying the RLLM regression model with residual and lasso analysis to achieve RMSE of 0.282, MAE of 0.08, and MSE of 0.08. This complete approach from image enhancement and feature extraction to hybrid classification and severity analysis is far superior to conventional diagnostic techniques. To enhance interpretability, stability, and clinical practical utility, future studies will prioritize the inclusion of Explainable Al (EAI), multiple datasets, and clinical data.
Patil et al. (Mon,) studied this question.