Tongue Image Segmentation Based on Transformer Feature Channel Fusion
To address challenges in tongue image segmentation,such as discontinuous tongue edges and interference from complex backgrounds,this paper proposes a traditional Chinese medicine tongue image segmentation method based on Transformer feature channel fusion.First,multi-level feature maps containing both positional and feature information are obtained through a multi-stage convolution module.Next,an inverted pyramid network is introduced to match the dimensions of the multi-level feature maps.Finally,the skip connections of the traditional U-Net network are replaced with the CTrans module of UCTransNet to capture contextual information in the image better and achieve accurate segmentation of medical images.Dice coefficient and mean intersection-over-union(MIoU)are selected as evaluation criteria.Training,evaluation,and validation on a self-collected dataset of tongue images yield a Dice value of 96.81％and an MIoU value of 93.89％,indicating strong segmentation performance.The proposed method has a good segmentation effect on the tongue image dataset,and can accurately extract the features of the tongue body.This method can be used for standardized research in tongue diagnosis,improving the accuracy and reliability of tongue diagnosis.Additionally,it demonstrates strong generalization capabilities on other medical image datasets.
deep learningimage segmentationTransformertongue diagnosis
万方数据
