孙兆男,王可欣,黄文鹏,吴鹏升,张晓东,王霄英.3D ResNet深度学习模型自动甄别前列腺多参数MR扫描序列:多中心研究[J].中国医学影像技术,2024,40(5):769~773
3D ResNet深度学习模型自动甄别前列腺多参数MR扫描序列:多中心研究
3D ResNet deep learning model for automatically identifying sequences of prostate multi-parametric MRI:A multicenter study
投稿时间:2023-08-08  修订日期:2023-12-05
DOI:10.13929/j.issn.1003-3289.2024.05.028
中文关键词:  前列腺肿瘤  磁共振成像  人工智能
英文关键词:prostatic neoplasms  magnetic resonance imaging  artificial intelligence
基金项目:
作者单位E-mail
孙兆男 北京大学第一医院医学影像科, 北京 100034  
王可欣 首都医科大学基础医学院, 北京 100069  
黄文鹏 北京大学第一医院医学核医学科, 北京 100034  
吴鹏升 北京赛迈特锐医疗科技有限公司, 北京 100011  
张晓东 北京大学第一医院医学影像科, 北京 100034  
王霄英 北京大学第一医院医学影像科, 北京 100034 wangxiaoying@bjmu.edu.cn 
摘要点击次数: 164
全文下载次数: 125
中文摘要:
      目的 构建基于前列腺多参数MRI(mpMRI)自动甄别其主要扫描序列的3D ResNet深度学习模型,并评估其价值。方法 收集于3个医疗中心接受超声引导下前列腺穿刺的1 086例患者穿刺前1 153次前列腺mpMRI资料,并按不同扫描序列加以拆分,分别将T2WI、弥散加权成像(DWI)及表观弥散系数(ADC)图归入相应数据集,共获得5 151组图像,并将归类为非脂肪抑制T2WI(T2WI_nan,n=1 000)、脂肪抑制T2WI(T2WI_fs,n=1 188)、高b值DWI(DWI_High,b值≥500 s/mm2,n=1 045)、低b值DWI(DWI_Low,b值<500 s/mm2,n=1 012)及ADC图(ADC map,n=906)。按8 ∶ 1 ∶ 1比例将全部图像分为训练集(n=4 122)、验证集(n=513)和测试集(n=516)。行预处理及扩增后,采用3D ResNet于训练集及验证集训练及优化自动甄别图像类别模型,以测试集评估模型分类效能。结果 所获模型分类测试集不同序列图像的准确率、敏感度、特异度、阳性预测值、阴性预测值、F1值及Kappa值分别为0.995~1.000、0.990~1.000、0.998~1.000、0.990~1.000、0.998~1.000、0.995~1000、0.994~1.000。结论 3D ResNet深度学习模型能有效自动甄别前列腺mpMRI所涉主要扫描序列。
英文摘要:
      Objective To construct a 3D ResNet deep learning model based on multi-parametric prostate MRI (mpMRI), and to observe its value for automatically identifying the main MR sequences. Methods Totally 1 153 sets pre-biopsy prostate mpMRI data of 1 086 patients who underwent ultrasound-guided prostate biopsy in 3 hospitals were collected and divided into different image datasets, i.e. T2WI, diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps with a total of 5 151 images. Then the images were categorized into non-fat-suppressed T2WI (T2WI_nan, n=1 000), fat-suppressed T2WI (T2WI_fs, n=1 188), high b-value DWI (DWI_High, b-value≥500 s/mm2, n=1 045), low b-value DWI (DWI_Low, b-value<500 s/mm2, n=1 012) or ADC map (n=906), also divided into training set (n=4 122), verification set (n=513) and test set (n=516) at the ratio of 8 : 1 : 1. After preprocessing and augmentation, a 3D ResNet model for automatically identifying image categories was trained and optimized in the training and verification sets, and its classification efficiency was evaluated in the test set. Results The identifying accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score and Kappa value of the obtained model for automatically identifying categories of images in the test set was 0.995—1.000, 0.990—1.000, 0.998—1.000, 0.990—1.000, 0.998—1.000, 0.995—1.000 and 0.994—1.000, respectively. Conclusion The obtained 3D ResNet deep learning model could effectively and automatically identify the main sequences of prostate mpMRI.
查看全文  查看/发表评论  下载PDF阅读器