基于区域定位和特征融合策略的Williams-Beuren综合征人脸识别模型构建与验证  

Construction and evaluation of automatic facial recognition for Williams-Beuren syndrome based on local phenotype navigator and feature fusion

作　　者：莫梓华 王树水[2] 刘惠 玉今珒 黄飚 MO Zi-hua;WANG Shu-shui;LIU Hui;YU Jin-jin;HUANG Biao(School of Medicine,South China University of Technology,Guangzhou 510110,China;Department of Pe-diatric Cardiology,Guangdong Cardiovascular Institute,Guangdong Provincial People′s Hospital,Guangdong Academy of Medical Sciences,Guangzhou 510100,China)

机构地区：[1]华南理工大学医学院,广州510110 [2]广东省心血管病研究所心儿科广东省人民医院(广东省医学科学院),广州510100

出　　处：《岭南心血管病杂志》2021年第3期302-307,共6页South China Journal of Cardiovascular Diseases

基　　金：国家自然科学基金面上项目(项目编号:82070321)。

摘　　要：目的构建基于区域定位和特征融合策略的Williams-Beuren综合征(Williams-Beuren syndrome,WBS)人脸识别模型并进行验证,从而通过面容来诊断WBS。方法本研究纳入2018年1月至2020年12月广东省人民医院收治的面容异常的WBS儿童104例。以同期伴有特殊面容的其他遗传综合征患儿91例及正常儿童145名作为对照组。每位受试者选取脸部正位照片一张。以残差网络34(residual network 34,ResNet-34)为面部特征提取器从人脸照片提取人脸深度特征,利用区域定位和特征融合策略的深度学习方法建立WBS人脸识别模型。同时,单独使用ResNet-34网络构建人脸识别模型作为基准模型进行比较。邀请4位临床医生通过面部照片进行WBS识别。采用Bootstrapping方法分别将基于区域定位和特征融合策略的模型对WBS儿童的识别结果与ResNet-34模型及临床医生的识别结果进行比较。结果本研究构建的基于区域定位和特征融合策略的模型对WBS诊断的准确率为0.911(0.888~0.933),敏感性为0.846(0.785~0.923),特异性为0.939(0.909~0.967),曲线下面积(AUC)为0.933(895~0.967),均高于ResNet-34模型及4位临床医师。结论基于区域定位和特征融合策略的WBS人脸识别模型在WBS的诊断中具有重要作用,有助于对WBS进行临床诊断。Objectives To establish and verify an automatic facial recognition model for Williams-Beuren syndrome(WBS)based on local phenotype navigator and feature fusion,aiming at diagnosing WBS by its distinctive facial appearance.Methods The study enrolled 104 WBS patients with facial dysmorphism admitted to Guangdong Provincial People′s Hospital between January 2018 and December 2020,a total of 91 patients with other genetic syndromes with facial dysmorphism and 145 normal children from the same period.One frontal facial image was taken from each individual,then it was input into the residual network 34(ResNet-34)facial feature extractor to extract depth feature of facial phenotype to establish the automatic facial recognition model based on local phenotype navigator and feature fusion.Meanwhile,a facial recognition model based solely on ResNet-34 was built as a benchmark model.Additionally,4 physicians were invited to distinguish WBS patients based solely on the facial image.The recognition results of our proposed model were compared with those of solely ResNet-34 and of physicians by Bootstrapping,respectively.Results The accuracy,sensitivity,specificity and area under the cuirve(AUC)of our proposed model based on local phenotype navigator and feature fusion were 0.911(0.888-0.933),0.846(0.785-0.923),0.939(0.909-0.967)and 0.933(0.895-0.967),respectively.This model outperformed both solely ResNet-34 model and human experts.Conclusions The facial recognition model based on local phenotype navigator and feature fusion could play a prominent role in WBS diagnosis,which is helpful in clinical practice.

关 键 词：Williams-Beuren综合征 人工智能 人脸识别 遗传综合征 肺动脉瓣上狭窄 

分 类 号：R541.1[医药卫生—心血管疾病]