ABSTRACT
The main method to diagnose COVID-19 is a nucleic acid test from a throat swab. Routine manual collection methods expose medical personnel to high-risk environment, which has a high risk of cross-infection. A throat swab sampling robot was developed to take the place of medical staff. The automatic segmentation of M-region in the pharyngeal swab image, which plays a core guiding role when the robot takes a throat swab sample. Aiming at the problem of discontinuous or fuzzy boundary in M-region of oral cavity, the segmentation accuracy is affected. An improved U-Net model is proposed and a new multi-scale feature fusion module with channel attention mechanism is presented. The ability of adaptive learning is enhanced and the segmentation precision of M-region with discontinuous or fuzzy edges is increased. Oral images of 45 volunteers were collected for training and testing. Experimental results showed that the model could accurately segment M-region in pharyngeal swab images, and compared with other segmentation networks, it has better indexes of segmentation precision.