基于T2WI的ResNet深度学习模型术前预测膀胱癌病理分级的研究  

Preoperative prediction of pathological grading in bladder cancer by deep residual network model based on T2WI

作　　者：黄翔 曹康养 邹玉坚[1] 邓磊 张蔚菁 杨水清[1] 张坤林 朱玉容 李建鹏[1] HUANG Xiang;CAO Kangyang;ZOU Yujian;DENG Lei;ZHANG Weijing;YANG Shuiqing;ZHANG Kunlin;ZHU Yurong;LI Jianpeng(Department of Radiology,the Tenth Affiliated Hospital of Southern Medical University(Dongguan People's hospital),Dongguan 523000,China;Medical AI Lab,School of Biomedical Engineering,Health Science Centre,Shenzhen University,Shenzhen 518000,China;Department of Radiology,Sun Yat-sen University Cancer Centre,Guangzhou 510000,China;Department of Pathology,the Tenth Affiliated Hospital of Southern Medical University(Dongguan People's hospital),Dongguan 523000,China)

机构地区：[1]南方医科大学第十附属医院(东莞市人民医院)放射科,东莞523000 [2]深圳大学医学部生物医学工程学院人工智能实验室,深圳518000 [3]中山大学肿瘤防治中心影像科,广州510000 [4]南方医科大学第十附属医院(东莞市人民医院)病理科,东莞523000

出　　处：《磁共振成像》2024年第1期125-131,共7页Chinese Journal of Magnetic Resonance Imaging

基　　金：东莞市社会发展科技(重点)项目(编号:20211800905212)。

摘　　要：目的 本研究旨在构建并验证基于T2加权成像(T2-weighted imaging, T2WI)的50层深度残差网络(50-layer deep residualnetwork,ResNet-50)深度学习(deeplearning,DL)模型术前预测膀胱癌(bladder cancer, BCa)病理分级的效能。材料与方法 回顾性分析来自南方医科大学第十附属医院(中心1)和中山大学肿瘤防治中心(中心2)共169名BCa患者的211个肿瘤病灶数据。以病理组织学诊断作为金标准,以肿瘤病灶为单位进行分析,其中高级别尿路上皮癌(high grade urothelial carcinoma, HGUC)为111个,低级别尿路上皮癌(low grade urothelial carcinoma, LGUC)为100个。采用DL模型的ResNet-50方法,基于中心1内部训练集构建模型,所得出的模型在中心1的内部测试集中测试后筛选出最优模型,随后在中心2的外部测试集上进行独立验证。采用受试者工作特征(receiver operating characteristic, ROC)曲线下面积(area under the curve, AUC)、准确率、敏感度和特异度对模型性能进行评估,并进行特征可视化展示。结果 DL模型在内部测试集的AUC为0.856(95%CI:0.723~0.941),准确率为80.9%(95%CI:69.6%~92.1%),敏感度为77.8%(95%CI:65.9%~89.7%),特异度为82.8%(95%CI:72.0%~93.6%);在外部测试集的AUC为0.814 (95%CI:0.686~0.906),准确率为78.2%(95%CI:67.3%~89.1%),敏感度为77.3%(95%CI:66.2%~88.3%),特异度为81.8%(95%CI:71.6%~92.0%)。特征可视化结果显示DL模型较高激活区域与BCa病灶基本重叠,可正确识别BCa靶区域,同时对HGUC与LGUC的特征有一定区分度。结论 本研究首次采用DL方法在术前建立基于T2WI的BCa病理分级预测模型,并在双中心进行验证。该模型无创、客观,泛化性及可重复性强,具有较高的预测准确性,有助于临床术前更精准地诊断。Objective:To develop and validate the efficacy of a deep learning(DL)model by 50-layer deep residual network(ResNet-50)based on T2WI for preoperative prediction of preoperative pathological grading of bladder cancer(BCa).Materials and Methods:A total of 211 tumors in 169 BCa patients[109 for training and 47 for internal test,from the Tenth Affiliated Hospital of Southern Medical University(centre 1);55 for external test,from Sun Yat-sen University Cancer centre(centre 2)]were enrolled,including 111 tumors of high grade uroepithelial carcinoma(HGUC)and 100 tumors of low grade uroepithelial carcinoma(LGUC).Grade determination was confirmed by pathological examination.ResNet-50 was used to construct the models based on the internal training set from centre 1.The optimal model was selected from the resulting models after being tested on the internal test set from centre 1 and validated independently on the external test set from centre 2.The performance of the model was evaluated using the area under the receiver operating characteristic curve(AUC),accuracy,sensitivity,and specificity,with feature visualization images presented.Results:In the internal test set,we achieved an AUC of 0.856[95%confidence interval(CI):0.723-0.941],accuracy of 80.9%(95%CI:69.6%-92.1%),sensitivity of 77.8%(95%CI:65.9%-89.7%),and specificity of 82.8%(95%CI:72.0%-93.6%).In the external test set,we achieved an AUC of 0.814(95%CI:0.686-0.906),accuracy of 78.2%(95%CI:67.3%-89.1%),sensitivity of 77.3%(95%CI:66.2%-88.3%),and specificity of 81.8%(95%CI:71.6%-92.0%).Feature visualization showed that the activated regions overlapped with the BCa lesions largely,indicating the DL model identified the target area of BCa correctly.And the t-distributed stochastic neighbor embedding(T-SNE)helped to distinguish HGUC from LGUC in a certain extent.Conclusions:This study is the first to establish a preoperative BCa pathological grading prediction model based on T2WI using DL methods and be validated across two centres.With high prediction accuracy,the model

关 键 词：膀胱癌 深度学习 磁共振成像 肿瘤分级 

分 类 号：R445.2[医药卫生—影像医学与核医学] R737.14[医药卫生—诊断学]