Semi-supervised Lip-Tongue segmentation with Boundary Region Contrast Sampling

In Traditional Chinese Medicine, the accurate segmentation mask of the tongue and lip is the key of inspection. Although deep learning has made remarkable progress in medical image segmentation, a lot of manual annotations are still required for training. Semi-supervised learning (SSL) is used to reduce annotation work, but its performance often suffers when applied to tongue and lip segmentation, which is because tongue and lip images have noisy background information and unique boundary regions. To alleviate the problem, we propose a semi-supervised framework named Lip-Tongue segmentation with Boundary Region Contrast Sampling (Lip-Tongue-BReCoSample). We first preprocess the data, roughly locating the target and filtering out noisy background information. Then we generate the key boundary regions and sample to carry out contrast learning, which alleviates the problem that SSL cannot make fine modeling of the boundary regions of the target with limited information. After a lot of experiments, our method has achieved good results in SSL, and makes it reach or even exceed the performance of many traditional supervised methods, which can improve MIOU performance from 88.09 to 90.43 (+2.34) in SSL specifically. Our method is also better than the latest large-dataset pre-trained model (e.g., SegGPT). To the best of our knowledge, it is the first application of SSL in tongue and lip semantic segmentation.