摘要
目的探讨卷积神经网络(VGG19)模型在辅助肝细粒棘球蚴病(HCE)超声分型诊断的应用价值。方法回顾性收集2012年1月至2022年12月在新疆医科大学第一附属医院就诊的肝细粒棘球蚴病(HCE,含CE1~CE5分型)患者及非棘球蚴病性肝局灶性病变(NHFLL)患者的超声图像,使用VGG19模型对超声图像进行6个类型肝局灶性占位病变(CE1~CE5、NHFLL)的诊断判定,比较各类型的判定百分比。当VGG19误判HCE与NHFLL时,比较患者的一般人口学信息和相关临床资料。将超声图片按照各分型占比基本一致的原则随机均分为两组,2名低年资超声医师和2名高年资超声医师各随机选择1组进行人工分类汇总,对比低、高年资超声医师与模型诊断正确率。采用混淆矩阵、精确率、召回率、特异度和F1分数评估VGG19的诊断性能。用构成比对计数资料进行描述性分析,使用卡方检验、Fisher确切概率法、配对卡方检验进行差异比较分析。结果871例HCE中包括203例CE1、227例CE2、110例CE3、159例CE4、172例CE5;600例NHFLL中包括300例肝囊肿、150例肝钙化灶、150例肝实性占位病变(100例肝血管瘤、25例肝癌、25例肝脓肿)。VGG19模型整体精准率为82.0%,召回率为87.9%,F1分数为84.3%。VGG19模型整体正确率为86.2%(1268/1471),各类型正确率由高到低依次为CE5(95.3%,164/172)、CE4(91.2%,145/159)、CE3(89.1%,98/110)、CE1(84.7%,172/203)、CE2(84.6%,192/227)、NHFLL(82.8%,497/600)。共有203例误诊,误诊率为13.8%(203/1471),其中100例为HCE分型间的误诊,包括31例CE1、35例CE2、12例CE3、15例CE4、8例CE5;103例为NHFLL被误诊为HCE,包括68例肝囊肿、17例肝钙化灶、17例肝血管瘤、1例肝脓肿,无肝癌病例被误诊。VGG19误诊的HCE患者和NHFLL患者的年龄、住址类别、有无犬类接触史、文化程度、省份差异有统计学意义(χ^(2)=55.116、24.197、35.834、14.069、11.918,均P<0.05),性别差异无统计学意义(χ^(2)=0.047,P>0.05)。VGG19模型诊断整体正确率(86.2%,1268/1471)高于低年资医师(81.2%,1195/1471)(P<0.05),低于高年资医师(92.3%,1358/1471)(P<0.05)。结论VGG19模型能够较好地识别HCE的5种分型及NHFLL,其诊断正确率低于高年资超声医师,但高于低年资超声医师,有望推广到基层医院,联合临床信息后可辅助修正超声诊断。
Objective To evaluate the application value of convolutional neural network(VGG19)model in the ultrasound diagnosis of hepatic cystic echinococcosis(HCE).Methods The ultrasound images of patients with he‑patic cystic echinococcosis(HCE,including CE1-CE5 types)and patients with non‑echinococcosis focal liver lesion(NHFLL)in the First Affiliated Hospital of Xinjiang Medical University from January 2012 to December 2022 were retrospectively collected.The VGG19 model was used to determine the image diagnosis of 6 types of hepatic focal space occupying lesions(CE1-CE5,NHFLL),and the percentage of each type determined was compared.When VGG19 misclassified HCE and NHFLL,the general demographic information and relevant clinical data of patients were compared.The ultrasound images were randomly divided into two groups based on the principle of essentially consistent proportion of different types,of which manual classification summary were performed by 2 junior ultrasound physicians and 2 senior ultrasound physicians on one group each selected randomly.The diagnostic accuracy of the model made by junior and senior ultrasound physicians were compared.The diagnostic performance of VGG19 was evaluated using confusion matrix,precision,recall,specificity,and F1 score.Descriptive analysis was conducted using contingency tables to describe the count data,and chi‑square test,Fisher’s exact probability method and paired card square test were used for comparative analysis of differences.Results Among the 871 HCE cases,there were 203 cases of CE1,227 of CE2,110 of CE3,159 of CE4,and 172 of CE5.The 600 NHFLL cases include,300 cases of hepatic cysts,150 of hepatic calcified lesions,and 150 solid hepatic occupying lesions(100 cases of hepatic heman‑gioma,25 of hepatoma,and 25 of liver abscess).The overall accuracy of the VGG19 model was 82.0%,the recall rate was 87.9%,and the F1 score was 84.3%.The overall accuracy rate of VGG19 model was 86.2%(1268/1471),and the accuracy rates of each type from high to low were CE5(95.3%,164/172),CE4(91.2%,145/159),CE3(89.1%,98/110),CE1(84.7%,135/159),CE2(84.6%,192/227)and NHFLL(82.8%,497/600),respectively.A total of 203 cases were mis‑diagnosed,and the misdiagnosis rate was 13.8%(203/1471).Among them,100 cases were misdiagnosed between HCE types,including 31 cases of CE1,35 cases of CE2,12 cases of CE3,15 cases of CE4,and 8 cases of CE5.103 cases of NHFLL were misdiagnosed as HCE,including 68 cases of hepatic cyst,17 cases of liver calcification,17 cases of liver hemangioma,and 1 case of liver abscess.No case of liver cancer was misdiagnosed.There were statisti‑cally significant differences in age,residential area,history of contact with dogs,education level and province between HCE patients and NHFLL patients misdiagnosed by VGG19(χ^(2)=55.116,24.197,35.834,14.069,11.918,all P<0.05),but there was no significant difference in gender(χ^(2)=0.047,P>0.05).The overall diagnostic accuracy of VGG19 model(86.2%,1268/1471)was higher than that by junior doctors(81.2%,1195/1471)(P<0.05),lower than that by senior physicians(92.3%,1358/1471)(P<0.05).Conclusion The VGG19 model can identify the 5 types of HCE and non echinococcal NHFLL,its diagnostic accuracy is lower than that by senior sonographers,but higher than that by junior sonographers.The model is expected to be promoted to primary hospitals to assist correction of ultrasound diagnosis in combination with clinical information.
作者
任艳
宋涛
尚丰
吴淼
王正业
王晓荣
REN Yan;SONG Tao;SHANG Feng;WU Miao;WANG Zhengye;WANG Xiaorong*(Abdominal Ultrasonography Department,The First Affiliated Hospital of Xinjiang Medical University,State Key Laboratory of Pathogenesis Prevention and Treatment of High Incidence Diseases in Central Asia,Urumqi 830000,Xinjiang,China;College of Medical Engineering Technology of Xinjiang Medical University,Urumqi 830000,Xinjiang,China;Center for Disease Control and Prevention of Xinjiang Production and Construction Corps,Urumqi 830000,Xinjiang,China)
出处
《中国寄生虫学与寄生虫病杂志》
CSCD
北大核心
2024年第4期454-460,474,共8页
Chinese Journal of Parasitology and Parasitic Diseases
基金
省部共建中亚高发病成因与防治国家重点实验室开放课题(SKL‑HIDCA‑2020‑YG2)。
关键词
肝细粒棘球蚴病
超声
卷积神经网络
肝脏局灶性病变
Hepatic cystic echinococcosis
Ultrasound
Convolutional neural network
Focal liver lesions