Automatic Segmentation of Tongue Image Based on Res-Unet Approach

Abstract

Tongue diagnosis occupies an important position in the field of traditional Chinese medicine and has been developed for thousands of years. Doctors diagnose disease based on tongue images of patients stored in hospital databases. Hence, segmenting the tongue area of the tongue image facilitates the diagnosis and saves space for storing the tongue image. In order to solve such a challenging problem, we put forward a method combing Unet and Res-net for tongue image segmentation and implements the end-to-end form. In our Res-Unet architecture, including four encoder blocks and four decoder blocks, and the residual network (Res-net) block used as the backbone for each block. The upsampling layer restores the features extracted by the sampling layer. We use our own datasets named TongueSet1 (TS1) and Tongueset2 (TS2) that collected from the hospital. The collection methods of these two datasets are different; TS1 is collected by professionals while TS2 is taken by nurses. This method obtained the latest results on both data sets. We used accuracy (acc) and mean intersection (mIoU) as the evaluation indicators of our model. Among them, the acc and mIoU of the model on TongueSet1 reached 0.984 and 0.925, on TongueSet2 reached 0.985 and 0.925.