深度学习识别痛风患者第一跖趾关节异常超声图像的方法初探  

Preliminary study on deep learning method in identifying abnormal ultrasound images of the first metatarsophalangeal joint in gout patients

作　　者：杜梅霞 刘家栋 肖丽珊 王凌涛 闫萌萌 李昱臣 赵诚[1] 丁建睿[2] 宁春平[1] Du Meixia;Liu Jiadong;Xiao Lishan;Wang Lingtao;Yan Mengmeng;Li Yuchen;Zhao Cheng;Ding Jianrui;Ning Chunping(Department of Abdominal Ultrasound,Affiliated Hospital of Qingdao University,Qingdao 266071,China;School of Computer Science and Technology,Harbin Institute of Technology,Weihai 264209,China)

机构地区：[1]青岛大学附属医院腹部超声科,青岛266071 [2]哈尔滨工业大学(威海)计算机科学与技术学院,威海264209

出　　处：《中华超声影像学杂志》2024年第4期335-340,共6页Chinese Journal of Ultrasonography

基　　金：科技部国家重点研发计划项目(2022YFC2503302)。

摘　　要：目的探索深度学习方法识别痛风患者第一跖趾关节(MTP1)异常超声图像的可行性。方法前瞻性收集2023年2-10月于青岛大学附属医院就诊行MTP1超声检查的患者351例,共收集超声图像4032张,包括2040张阳性图像和1992张阴性图像,根据不同采集切面分为足背图像集、足内侧图像集、足底图像集。所有图像集均按照6∶4分为训练集和测试集。采用ResNet-50网络,分别在不同训练集上进行模型训练,建立单模型(MA)和多模型(MB)。分别在各测试集上进行效能检测,绘制两种方法识别MTP1异常超声图像的ROC曲线,计算曲线下面积(AUC)、准确性、敏感性、特异性,AUC的比较采用De-long检验。结果成功建立两种深度学习方法,且两种方法在总体测试集上表现优异,识别异常MTP1超声图像的AUC分别为0.92、0.92,准确性分别为85.95%、86.57%,敏感性分别为83.60%、82.37%,特异性分别为88.36%、90.86%,且二者比较差异无统计学意义(Z=-0.50,P=0.62)。与MB相比,MA在足背和足底图像测试集中AUC略低(0.95比0.96;0.83比0.86),在足内侧测试集中AUC略高(0.90比0.89),但差异无统计学意义(均P>0.05)。结论深度学习方法可以有效区分MTP1正常和异常超声图像,且对不同切面超声图像具有较高的包容度,当训练样本量足够大时,无需对不同解剖切面单独建模。Objective To explore the feasibility of deep learning method in identifying abnormal ultrasound images of the first metatarsophalangeal joint(MTP1)in gout patients.Methods A total of 351 patients who underwent MTP1 ultrasound examination in the Affiliated Hospital of Qingdao University from February to October 2023 were prospectively enrolled.A total of 4032 ultrasound images were collected,including 2040 positive images and 1992 negative images.All the images were divided into dorsal image set,medial image set and plantar image set according to the anatomic characters.All image sets were divided into training set and test set by 6∶4.ResNet-50 network was used in different training sets to establish deep learning models,including a single model and a multiple model.Efficiencies of different models were tested on each test set,ROC curves of the two methods for identifying abnormal MTP1 ultrasonic images were plotted.The area under the curve(AUC),accuracy,sensitivity and specificity were calculated,and the De-long test was used to compare the AUC.Results Two deep learning methods were successfully established,and both methods performed well on the whole test set.The AUC of the two methods in identifying abnormal MTP1 ultrasound images was 0.92 and 0.92,the accuracy was 85.95%and 86.57%,the sensitivity was 83.60%and 82.37%,and the specificity was 88.36%and 90.86%,respectively.There was no significant difference between the two methods(Z=-0.50,P=0.62).Compared with the multiple model,the AUC of single model was slightly lower in the dorsal and plantar test sets(0.95 vs 0.96;0.83 vs 0.86),and slightly higher in the medial test set(AUC:0.90 vs 0.89),but there were no significant differences(P>0.05).Conclusions Deep learning method can effectively distinguish abnormal MTP1 ultrasonic images from normal ones,and the tolerance for different anatomical views quite high.Multiple model is not necessary when the training sample size is large enough.

关 键 词：超声检查 人工智能 深度学习 第一跖趾关节 痛风性关节炎 

分 类 号：R589.7[医药卫生—内分泌] TP18[医药卫生—内科学] R445.1[医药卫生—临床医学]