文章摘要

文章信息/Info

Title:: Tongue segmentation method based on dual encoding feature extraction path

作者:: 封晓燕1; 2; 田琪1; 2; 徐云峰1; 2; 丛金玉1; 2; 刘坤孟1; 2; 王苹苹1; 2△; 魏本征1; 2△; （1.山东中医药大学青岛中医药科学院，青岛 266112；2.山东中医药大学医学人工智能研究中心，青岛 266112）

Author(s):: FENG Xiaoyan1; 2; TIAN Qi1; 2; XU Yunfeng1; 2; CONG Jinyu1; 2; LIU Kunmeng1; 2; WANG Pingping1; 2; WEI Benzheng1; 2; （1.Qingdao Academy of Chinese Medical Sciences， Shandong University of Traditional Chinese Medicine, Qingdao 266112, China; 2. Center for Medical Artificial Intelligence, Shandong University of Traditional Chinese Medicine, Qingdao 266112）

Keywords:: Tongue segmentation; Objectification of tongue diagnosis; Deep learning; Transformer; Multi-scale feature extraction

摘要:: 针对舌图像舌体边缘分割模糊、小区域分割错误等问题，本研究设计了一种双编码特征提取路径的方法,以获取丰富的信息特征，辅助舌体精确分割。首先，设计双编码特征提取路径，其中，空间信息路径保留空间信息并生成高分辨率特征图，上下文信息路径提高网络提取多尺度特征能力；其次，采用一种特征融合模块，融合空间信息路径和上下文信息路径的输出特征；最后，采用轻量级解码器模块减少模型参数量，提高模型计算效率。结果显示，该方法精确率、召回率、F1分数和平均交并比（mean intersection over union, MIoU）分别达98.82%、98.53%、98.60%和97.67%，模型总参数量和每秒浮点运算次数（floating point operations per second, FLOPs）为7.54 M和67.09 G。结果表明，该方法可有效提高舌体的分割精度，显著改善舌体小区域分割错误和边缘模糊性，为中医舌象智能辅助分析提供必要支撑。

Abstract:: Aiming at the problems such as blurred tongue edge segmentation and small domain segmentation errors in tongue images, a segmentation method with dual encoding feature extraction paths was designed to obtain rich information features and assist tongue segmentation accurately. Firstly, a dual encoding feature extraction pathway was designed，in which the spatial information path preserved spatial information and generated high-resolution feature maps，and contextual information pathway enhanced the network′s ability to extract multi-scale features. Then, a feature fusion module was adopted to merge the output features from spatial information paths and contextual information paths. Finally, a lightweight decoder module was adopted to reduce the number of network model parameters and improve the computational efficiency of the model. The results showed that the precision, recall, F1 score, and mean intersection over union(MIoU) of the algorithm reached 98.82%, 98.53%, 98.60%, and 97.67%, respectively. The total parameter counts and floating-point operations per second (FLOPs) of the model were 7.54 M and 67.09 G.The results demonstrate that this algorithm effectively enhances the accuracy of tongue body segmentation, significantly improves the segmentation errors and edge fuzziness in small areas of the tongue body. This method can provide essential support for intelligent auxiliary analysis of traditional Chinese medicine tongue images.