A novel cupping mark recognition method base on multi-scale feature fusion network

Cupping therapy is widely utilized in China and other Asian countries, and has gained adoption among Olympic athletes for therapeutic purposes. The color, shape, and texture characteristics of cupping marks serve as critical indicators of therapeutic efficacy and health status. However, accurate classification of cupping marks requires overcoming challenges associated with color gradient characteristics and skin background interference. In this study, we propose a novel multi-scale feature fusion framework based on the EfficientNet-B0 backbone network. This framework enhances classification performance through systematic feature extraction, pattern refinement and cross-scale feature representation fusion. Experimental findings demonstrate that our model exhibits superior overall performance compared to all evaluated models, achieving an accuracy of 0.9484, an AUC score of 0.9966, an F1 score of 0.9485, and a sensitivity of 0.9484, alongside the lowest Brier score (0.0164) and log loss (0.1854). Our research establishes an objective analytical approach for evaluating cupping therapy, and offers methodological insights for the design of feedback-driven intelligent cupping systems.