An improved SA-Net framework for multi-class skin disease diagnosis using deep learning and self-attention mechanisms

Skin diseases are among the most common global health concerns, and accurate diagnosis is complicated by the high similarity of symptoms across conditions. This paper proposes an Improved Self-Attention Network (SA-Net) for automatic multi-class skin disease diagnosis, combining convolutional layers with a channel-attention (squeeze-and-excitation-style) skin attention block to capture local textural and channel-wise feature importance in dermoscopic and clinical images. An open dataset spanning 21 skin disease categories was used, with preprocessing comprising image rescaling, normalization, and class-balancing augmentation. The Improved SA-Net achieved 97.85% training accuracy and 96.72% validation accuracy, with a macro-averaged F1-score of 0.90, outperforming several CNN and transfer-learning baselines. To address reviewer comments, a separate reduced-budget verification study was conducted, including a four-configuration ablation across five random seeds, computational-complexity profiling, statistical significance testing, and a duplicate-image leakage check. In this verification study, the complete model (mean test accuracy 22.29%, macro-F1 0.197) significantly outperformed a plain CNN backbone and a CNN with attention only ( \\(p<0.05\\) ), while also outperforming a CNN with augmentation only ( \\(p=0.045\\) ), indicating that both augmentation and self-attention contribute to performance, with augmentation providing the larger contribution. These findings demonstrate the potential of the proposed approach for automated skin disease classification while emphasizing that further clinical validation and confirmation of patient-level data separation are required before practical clinical deployment.