Deep Learning-Based Skin Cancer Classification Using Hybrid S-ResNet and Explainable AI

Abstract

Skin cancer is one of the most common and life-threatening diseases in the world. Early detection is essential in order to enhance treatment outcome and patient survival. In this work, a Hybrid S-ResNet framework with Grad-CAM is proposed for automated skin cancer classification using dermoscopic images. In the proposed approach, a modified architecture based on ResNet18 is used as a feature extractor to learn discriminative characteristics of lesions using residual learning. The input images are preprocessed using resizing, normalization and data augmentation techniques to improve the robustness of the model. Weighted Random Sampling, label smoothing and dropout regularization are employed during training to address class imbalance and improve generalization. The model is trained by AdamW optimizer with cosine annealing learning rate scheduling for better convergence and less overfitting. In addition, the proposed framework utilizes Gradient-weighted Class Activation Mapping (Grad-CAM) to produce visual explanations via highlighting the important parts of the image that lead to the prediction. The proposed model was compared with Resnet18, Resnet50, DenseNet121 and EfficientNetB0 under similar experiment settings. Experimental results show that the proposed Hybrid S-ResNet achieved better results with an accuracy of 84.90%, recall of 93.41%, F1-score of 90.69%, and AUC of 90.83%, proving the efficiency of the Hybrid S-ResNet to classify skin cancer. The combination of deep learning and explainable artificial intelligence offers a reliable framework to assist clinical decision making in dermatological diagnosis.