Deep Learning Models-Based Skin Cancer Classification and Early Detection

Skin cancer diagnosis requires accurate and interpretable classification systems to support clinical decision-making. This study systematically compares conventional convolutional neural networks (CNNs) with advanced transfer learning models (ResNet50, VGG16, EfficientNetB4) for multi-class skin lesion classification using the HAM10000 dataset. To address dataset imbalance and enhance interpretability, this research integrated Squeeze-and-Excitation attention modules, Focal Loss optimization, and an improved Gradient-weighted Class Activation Mapping (Grad-CAM) technique enabling multi-scale visualization. Experimental results demonstrate that transfer learning models outperform custom CNNs, with ResNet50 achieving 77.11% accuracy and EfficientNetB4 attaining the highest AUC-ROC (0.943) while using 4.3× fewer parameters than ResNet50. Despite these advancements, rare lesion categories such as dermatofibroma showed persistently low recall (≤25%), highlighting the need for targeted approaches to minority class learning. The enhanced Grad-CAM method improved localization accuracy by 18.7% compared to conventional implementations, offering clinicians clearer insights into model decision patterns. These findings highlight the potential of efficient transfer learning architectures combined with interpretability enhancements to advance AI-assisted dermatological diagnostics, particularly in resource-constrained environment.