Image segmentation using deep learning for tongue diagnosis in traditional Chinese medicine

Abstract

Deep learning has the advantages of high efficiency, high speed, high accuracy, and strong objectivity, and is widely used in the fields of pathology and laboratory diagnosis. The diagnostic techniques of traditional Chinese medicine are world-famous, and the four basic methods for diagnosing diseases, namely inspection, auscultation- olfaction, inquiry, and palpation, are collectively referred to as ”four diagnostics”. Tongue diagnosis is an important part of inspection, and it is also an effective diagnosis and treatment method for doctors to understand the changes of the patient’s body through the tongue image. In order to realize automatic tongue diagnosis, one of the important tasks is to implement the automatic segmentation of tongue images. However, using feature engineering to segment tongue images requires a lot of work, and only hand-crafted features cannot represent the features of the tongue well. Therefore, this paper designs a tongue segmentation network (TSN). TSN consists of three parts: feature encoding extraction module, context-aware module and feature decoding module. This model can fully extract tongue feature vector and perform information fusion through context-aware module, so that Effectively segment the tongue from the image. Compared with various deep learning image segmentation methods, the TSN proposed in this paper achieves the best performance results with 97.20% mean intersection over union (mIoU) and 98.83% pixel accuracy (PA).