ZACP: enhancing skin lesion classification ability using zero attention and complete perception.

With the increasing number of dermatology patients owing to ecological deterioration and unhealthy lifestyles, hospitals are experiencing significant diagnostic pressure, and patients are suffering from delayed treatment and poor medical experience. Therefore, there is an urgent need to develop accurate skin lesion classification tools. To address this issue, this study proposed the zero attention and complete perception (ZACP), a novel skin lesion classification system based on EfficientNet (EN). ZACP integrates two key modules: ZA and CP. ZA is an improved multi-head attention mechanism that injects zero vectors into key/value vectors, introduces a benchmark to suppress invalid attention to non-semantic regions (e.g., normal skin texture), and enhances the focus on lesions. Derived from the squeeze and excitation (SE) module, the CP embeds the ZA to achieve multiscale attention-retaining SE channel-weighting ability while enabling the ZA to capture local subtle features (e.g., melanoma boundaries) via patch segmentation and positional encoding. Additionally, the original activation function of EN was replaced with Mish to stabilize the training. Evaluated on the authoritative HAM10000 and ISIC2017 datasets, ZACP outperformed the comparative models in terms of accuracy (ACC) and other metrics, achieving 95% and 91% ACC on the two datasets, respectively. The source code is available at: http://github.com/ZACP-DERMATOLOGY .