基于超快速动态对比增强MRI的深度学习模型诊断乳腺恶性病变的价值  

The value of deep learning models based on ultrafast dynamic contrast-enhanced MRI for diagnosing malignant breast lesions

作　　者：王文琪 马文娟 郭一君 王静博 路红 Wang Wenqi;Ma Wenjuan;Guo Yijun;Wang Jingbo;Lu Hong(Department of Breast Imaging,Tianjin Medical University Cancer Institute and Hospital,National Clinical Research Centre for Cancer,Tianjin′s Clinical Research Centre for Cancer,Key Laboratory of Breast Cancer Prevention and Therapy,Tianjin Medical University,Ministry of Education,Key Laboratory of Cancer Prevention and Therapy,Tianjin 300060,China)

机构地区：[1]天津医科大学肿瘤医院乳腺影像诊断科、国家恶性肿瘤临床医学研究中心、天津市恶性肿瘤临床医学研究中心、乳腺癌防治教育部重点实验室、天津市肿瘤防治重点实验室,天津300060

出　　处：《中华放射学杂志》2025年第3期307-312,共6页Chinese Journal of Radiology

基　　金：国家自然科学基金(82172025,82072004,82302302);天津市医学重点学科(专科)建设项目(TJYXZDXK-009A)。

摘　　要：目的探讨基于超快速动态对比增强MRI(UF-DCE MRI)的深度学习模型预测乳腺恶性病变的价值。方法该研究为横断面研究。回顾性分析2023年3月至2024年1月在天津医科大学肿瘤医院接受诊治的347例乳腺病变患者的临床和影像资料。347例患者共347个病变,其中良性75例,恶性272例。采用随机数字法按7∶3的比例分为训练集243例和验证集104例。所有患者均接受乳腺UF-DCE MRI及常规动态增强MRI(DCE-MRI)检查。基于ImageNet上预训练的ResNet18深度学习模型建立27通道模型(输入UF-DCE MRI的27期增强图像)、3通道模型(输入DCE-MRI的3期增强图像)、1通道模型(DCE-MRI的第1期增强图像)。采用受试者操作特征曲线和曲线下面积(AUC)分析各模型预测乳腺病变良恶性的效能,采用DeLong检验比较AUC的差异。结果训练集和验证集中,27通道模型诊断乳腺恶性病变的AUC最高,分别为0.848(95%CI 0.818~0.877)、0.784(95%CI 0.752~0.817)。DeLong检验显示验证集中3种模型诊断乳腺恶性病变的AUC值两两比较差异无统计学意义(P>0.05)。UF-DCE MRI扫描27期,共81 s,时间分辨率3 s/期;DCE-MRI扫描3期,共270 s,时间分辨率90 s/期。结论UF-DCE MRI深度学习模型诊断乳腺恶性病变与DCE-MRI深度学习模型效能相当,但UF-DCE MRI具有时间分辨率高、扫描时间短等优势,在乳腺癌的精准诊疗中具有良好的应用价值。ObjectiveTo explore the value of deep learning models based on ultrafast dynamic contrast-enhanced MRI(UF-DCE MRI)in predicting malignant breast lesions.MethodsThe study was a cross-sectional study.Clinical and imaging data of 347 patients with breast lesions who received treatment at Tianjin Medical University Cancer Institute and Hospital from March 2023 to January 2024 were analyzed retrospectively.A total of 347 lesions were observed in the 347 patients,including 75 benign and 272 malignant lesions.The random number method was used to divide into the training set with 243 cases and the validation set with 104 cases in a ratio of 7∶3.All patients underwent breast UF-DCE MRI and conventional dynamic-enhanced MRI(DCE-MRI).A 27-channel model(27-phase enhancement images of input UF-DCE MRI),a 3-channel model(3-phase enhancement images of input DCE-MRI),and a 1-channel model(1st-phase enhancement images of DCE-MRI)were built based on the pre-trained ResNet18 deep learning model on ImageNet.The efficacy of each model in predicting breast malignant lesions was analyzed using receiver operating characteristic curves and area under the curve(AUC).The differences of AUC were compared using DeLong test.ResultsIn the training and validation sets,the 27-channel model had the highest AUC for diagnosing malignant breast lesions,which were 0.848(95%CI 0.818-0.877)and 0.784(95%CI 0.752-0.817),respectively.DeLong test showed no statistically significant difference in the AUC values of the three models in the validation set for the diagnosis of malignant lesions of the breast in a two-by-two comparison(P>0.05).UF-DCE MRI scans were 27 phases totaling 81 s with a temporal resolution of 3 s/phase;DCE-MRI scans were 3 phases totaling 270 s with a temporal resolution of 90 s/phase.ConclusionsThe model combining UF-DCE MRI with deep learning demonstrates comparable efficacy to DCE-MRI deep learning model in diagnosing breast malignant lesions.However the UF-DCE MRI has the advantages of high temporal resolution and short scanning t

关 键 词：乳腺肿瘤 超快速动态对比增强磁共振成像 深度学习 诊断 

分 类 号：R73[医药卫生—肿瘤]