A Geometric algebra-enhanced network for skin lesion detection with diagnostic prior

Automatic intelligent skin lesion recognition is crucial for elevating detection accuracy, enhancing diagnostic efficiency, and mitigating the risk of melanoma mortality. Despite advances, current methods often fall short in accuracy and acceptance among clinicians and patients. To improve clinical credibility, we propose an integrated system that incorporates medical diagnostic priors, structured into four distinct modules: skin area focus module, skin feature extraction module, anomaly feature attention module, and skin diagnostic interpretation module. These modules correspond to the clinical diagnostic steps in real-world settings. To obtain more structural information and richer features to assist in skin lesion classification, we developed neural networks and an attention module grounded in geometric algebra. Our skin lesion detection network is designed to be parameter-efficient and computationally light, with the parameters of the geometric algebra-based network and the attention mechanism reduced to 25% of their original size. Evaluations on the ISIC2020 public skin image dataset reveal that our approach outperforms existing methods. The method achieves an accuracy of 98.5% and an area under the curve of 98.8%, improving by 2.5% and 1.2% over the baseline.