ENHANCING TONGUE REGION SEGMENTATION THROUGH SELF-ATTENTION AND TRANSFORMER BASED

Abstract

As an essential component of traditional Chinese medicine diagnosis, tongue diagnosis has faced limitations in clinical practice due to its subjectivity and reliance on the experience of physicians. Recent advancements in deep learning techniques have opened new possibilities for the automated analysis and diagnosis of tongue images. In this paper, we collected 500 tongue images from various patients. These images were initially preprocessed and annotated, resulting in the dataset used for this experiment. This project is based on the previously proposed segmentation method using Harnessing Self-Attention and Transformer, which is divided into three key stages: feature extraction, feature fusion, and segmentation prediction. By organically combining these three key stages, our tongue region segmentation model is better equipped to handle complex tongue images and provides accurate segmentation results. The segmentation DICE coefficient reaches 0.953, which is of significant importance for the automation and objectivity of tongue diagnosis in traditional Chinese medicine.