Soft Attention-Enhanced ResNet101 for Robust Skin Lesion Classification in Dermoscopic Images

This paper proposes an advanced deep learning method for automated skin lesion classification based on a ResNet101 modified with a soft attention mechanism. The model was evaluated against the HAM10000 dataset—an unbalanced and challenging set of dermoscopic images. Stratified 5-fold cross-validation was used for evaluation. For the class imbalance problem, RandomOverSampler was used together with data augmentation techniques, resulting in a high diversity of training samples. This configuration allowed the model to obtain a 97.76% mean validation accuracy with low variance across folds. The soft attention mechanism in the ResNet101framework directed the model to focus on key regions of the images to improve the recognition of subtle differences between similar lesion types. Key performance metrics such as accuracy, loss, precision, and recall were analysed, which are in line with the training and validation phases. EarlyStopping/ReduceLROnPlateau allowed stable convergence and reduced overfitting risks. These results validate the model and suggest its use as a robust diagnostic tool to detect skin cancer at an early stage to support clinical decision-making in dermatology.