RD‐SACSeg: Region‐Detail Guided Superficial Auricular Collateral Segmentation Network and Its Applications in Traditional Chinese Medical Auricular Diagnosis

ABSTRACT

To address the subjectivity and poor reproducibility of traditional auricular diagnosis in Traditional Chinese Medicine (TCM), this study introduces the Region‐Detail guided Superficial Auricular Collateral Segmentation network (RD‐SACSeg), a novel deep learning‐based network for the automated and precise segmentation of auricular collaterals. The proposed architecture innovatively incorporates a dual‐constraint mechanism—integrating regional priors and edge details—alongside a multi‐scale feature fusion strategy to bridge the gap between low‐level encoder features and high‐level semantic information in the UNet architecture. This approach specifically overcomes the challenges of low contrast and the structural complexity of fine peripheral branches. Validated on a clinical dataset of 97 samples, RD‐SACSeg significantly outperformed state‐of‐the‐art CNN and Transformer‐based models in Intersection over Union (IoU) and Dice coefficient metrics. Furthermore, clinical correlation analysis demonstrates that integrating superficial auricular collateral segmentation with acupoint localisation reveals distinct quantitative patterns between collateral distribution and disorders such as insomnia‐anxiety, gastrointestinal disorder, and lower back pain. These results suggest that RD‐SACSeg provides a reliable, objective tool for TCM diagnostics, offering substantial potential for the modernisation of disease screening and pattern‐based therapeutic assessment.

Trial Registration : ITMCTR2025001298