基于多流神经网络的痤疮证候分类  

Acne syndrome classification based on multiflow neural network

作　　者：张可晔 张新峰[1] 边浩南 ZHANG Keye;ZHANG Xinfeng;BIAN Haonan(Information department,Beijing University of Tecnology,Beijing 100124)

机构地区：[1]北京工业大学,北京100124

出　　处：《北京生物医学工程》2023年第5期456-463,共8页Beijing Biomedical Engineering

基　　金：国家重点研发项目(2017YFC1703300)资助。

摘　　要：目的以为中医提供贴合实际需求的痤疮辅助分类方案为目的,针对传统神经网络在痤疮证候分类任务中计算复杂、准确率低,且无法实现将面图像与舌图像相结合进行同时分析的问题,提出了基于多流神经网络的痤疮证候分类模型。方法首先,针对面部图像,引入预处理模块Faster-RCNN识别出痤疮患者五官区域并对其进行遮挡,完整保留面部皮肤。其次,针对舌图像,引入预处理模块基于K-means聚类算法的舌图像苔质分离方法。接着,将面部皮肤、舌质和舌苔图像分别作为多流神经网络分类模块的三个并行对称网络流的输入,提取更充分有效的图像特征。最后,将三个分类结果相加融合得到证候类型输出。结果实验完成面图像五官自动识别、舌图像苔质分离的图像预处理任务,搭建多流神经网络完成痤疮证候分类,整体模型针对痤疮证候的最高分类准确率为97.13%。结论基于多流神经网络的痤疮证候分类模型实现了将面图像与舌图像相结合进行分析的任务,同时,模型测试效率及准确率较经典网络有所提高。Objective The purpose of this paper is to provide TCM(traditional Chinese medicine)with an auxiliary acne classification scheme that fits the actual needs.Aiming at the problems of complex calculation and low accuracy of traditional neural network in the acne syndrome classification task,and the difficult simultaneous analysis of face image and tongue image,this paper proposed an acne syndrome classification model based on multi-flow neural network.Methods Firstly,for facial images,the preprocessing module Faster-RCNN was introduced to identify facial regions of acne patients and block them,so as to completely preserve facial skin.Secondly,for the tongue image,a preprocessing module based on k-means clustering algorithm was introduced to separate the tongue and the tongue coating.Then,the facial skin,tongue and tongue coating images were used as the inputs of the three parallel symmetric network flows of the multi-flow neural network classification module respectively to extract more sufficient and effective image features.Finally,the three classification results were added and fused to obtain the output of syndrome type.Results The pre-processing tasks of automatic recognition of facial features and moss separation of tongue images were completed,and the multi-flow neural network was built to complete the classification of acne syndrome.The highest classification accuracy of the overall model for acne syndrome was 97.13%.Conclusions The model of acne syndrome classification based on multi-flow neural network not only realized the task of combining face image and tongue image for analysis,but also improved the efficiency and accuracy of model testing compared with classical network.

关 键 词：中医辅助诊断 中医望诊 痤疮证候分类 多流神经网络 图像特征 

分 类 号：R318.04[医药卫生—生物医学工程]