CNN-Based Tongue Image Segmentation in Traditional Chinese Medicine

Tongue image segmentation is a key component of the intelligent diagnosis in Traditional Chinese Medicine (TCM), and its accuracy directly affects the subsequent classification and diagnostic results. However, current research faces challenges due to data scarcity and limited model adaptability. The lack of publicly available tongue image datasets restricts the model’s generalization ability, while traditional algorithms are highly sensitive to lighting and posture changes. Moreover, general deep learning models have not been optimized for the characteristics of blurry tongue edges and low contrast background, which often leads to loss of details or over-segmentation. To address these issues, this study has constructed a high-quality dataset containing 1405 tongue images and proposes a lightweight tongue image segmentation network (TSNet). TSNet employs depthwise separable convolutions to reduce computational cost, introduces a Parallel Atrous Spatial Pyramid Pooling (PASPP) module to extract multi-scale features and handle blurred boundaries, and incorporates a Boundary Adjustment (BA) module to enhance edge segmentation accuracy. Experimental results show that TSNet achieves a mean Intersection over Union (mIoU) of 97.20% while using fewer parameters than mainstream models. While preserving tongue details, it effectively reduces the number of parameters, providing an efficient solution for tongue image segmentation in TCM.