A Three Layer Attention Enhanced Deep Learning Framework for Automated Skin Lesion Classification Using ISIC 2016 and ISIC 2018 Dermoscopic Image Dataset

Abstract Background Automated classification of dermoscopic skin lesions is an important direction in computer aided dermatological diagnosis, since deep learning can learn discriminative visual patterns directly from images, though performance is limited by small datasets, class imbalance, inter class similarity, and imaging artifacts. Methods This study proposes TLANet, a three layer CNN enhanced with a channel spatial attention mechanism, evaluated on two independent ISIC tasks: binary benign/malignant classification on ISIC 2016 (900 train / 379 test images) and seven class diagnosis on ISIC 2018/HAM10000 (10,015 images). The network combines three convolutional feature extraction blocks, attention, global average pooling, and a task specific classification head, alongside standardized preprocessing, augmentation, baseline comparisons (no attention CNN, ResNet, EfficientNet, MobileNet), ablation, and a full metric suite. Results The reported accuracy figures approximately 88.3% on ISIC 2016 and a macro F1 of 0.681 on ISIC 2018.