基于注意力区域不同组分特征的磁共振成像前列腺癌包膜侵犯诊断研究  

Radiomics features of sub-attention region on the diagnosis of the extracapsular extension of the prostate cancer on magnetic resonance imaging

作　　者：张一鸿 侯莹 包婕[3] 王成龙[1] 宋阳 张玉东[2] 杨光[1] ZHANG Yihong;HOU Ying;BAO Jie;WANG Chenglong;SONG Yang;ZHANG Yudong;YANG Guang(Shanghai Key Laboratory of Magnetic Resonance,East China Normal University,Shanghai 200062,China;Department of Radiology,the First Affiliated Hospital of Nanjing Medical University,Nanjing 210029,China;Department of Radiology,the First Affiliated Hospital of Soochow University,Suzhou 215006,China)

机构地区：[1]华东师范大学上海市磁共振重点实验室,上海200062 [2]南京医科大学第一附属医院,南京210029 [3]苏州大学第一附属医院,苏州215006

出　　处：《磁共振成像》2021年第12期39-43,66,共6页Chinese Journal of Magnetic Resonance Imaging

基　　金：国家自然科学基金重点项目(编号:61731009);上海市浦江人才计划资助(编号:2020PJD016)。

摘　　要：目的从磁共振图像中分别提取可能发生包膜侵犯区域的不同组分的组学特征,用来帮助诊断前列腺癌的包膜侵犯。材料与方法一共选取了718例前列腺癌患者的T2和ADC数据,分为训练集574例和测试集144例。手动勾画腺体和癌灶ROI,并根据两者位置关系计算注意力ROI。将注意力ROI分为背景、腺体和癌灶三个部分分别提取组学特征组合建模。通过ROC、AUC、混淆矩阵和决策曲线对模型进行分析。结果使用腺体和癌灶ROI建立的模型在训练集上AUC为0.740和0.742,测试集上的AUC为0.746和0.755。注意力ROI模型AUC在训练集上为0.732,在测试集上提高到0.766。注意力子区域特征组合建模的结果最好,可以将训练集和测试集AUC提升到0.794和0.792。结论相较于通过腺体或癌灶建立的影像组学模型,使用可能侵犯区域子区域建模能更好地预测前列腺癌的包膜侵犯,可以为临床包膜侵犯的诊断提供帮助。Objective:Extract radiomics features from the sub-regions of the generated attention region on magnetic resonance images to help diagnose the extracapsular extension (ECE) of the prostate cancer.Materials and Methods:Seven hundred and eighteen cases with prostate cancer diagnosis including T2 weighted images and apparent diffusion coefficient maps were selected in this study,and divided into 574 training cases and 144 test cases.An attention ROI was generated according to the ROIs of the prostate gland and the prostate cancer lesion.Further,sub-regions of the attention ROI were split into the background,prostate gland and lesion to be used for feature extraction.Radiomics models were built based on features from prostate gland (Model_(Pro)),prostate cancer (Model_(PCa)),attention ROI (Model_(Att)),and sub-regions of attention ROI (Model_(Region)),respectively.The area under the receiver operating characteristic (ROC)curve (AUC),confusion matrix and the decision analysis curve were used for statistical analysis.Results:The AUCs of the Model_(Pro)were0.740 and 0.746,and that of Model_(PCa)were 0.742 and 0.755 on the training and the test cohorts,respectively.The AUC of Model_(Att)was higher and was achieved of 0.732 and 0.766 on the training and test cohorts.Compared to the above models,Model_(Region)performed best to achieved an AUC of 0.794 and 0.792 on the training and test cohorts.Conclusion:The radiomics model based on the attention ROI and the sub-regions performed more accurately than the usual prostate gland and cancer lesion,and could provide aids in the ECE diagnosis in the clinics.

关 键 词：影像组学 磁共振成像 前列腺癌 包膜侵犯 注意力感兴趣区 

分 类 号：R445.2[医药卫生—影像医学与核医学] R737.25[医药卫生—诊断学]