深度学习模型在布加综合征患者增强3D-MRA图像分割中的初步应用  

Preliminary study on deep learning-based enhanced 3D-MRA segmentation in patients with Budd-Chiari syndrome

作　　者：朱磊 邱祥龙 沙俊诚 王朝 苟亚博 张雷鸣 张庆桥 ZHU Lei;QIU Xianglong;SHA Juncheng;WANG Chao;GOU Yabo;ZHANG Leiming;ZHANG Qingqiao(Department of Interventional Radiology,the Affiliated Hospital of Xuzhou Medical University,Budd-Chiari Syndrome Diagnosis and Treatment Center of Jiangsu Province,Xuzhou 221006,China;Intervention Center of the First Affiliated Hospital of Bengbu Medical University,Bengbu 233004,China)

机构地区：[1]徐州医科大学附属医院介入放射科,江苏省布加综合征诊疗中心,徐州221006 [2]蚌埠医科大学第一附属医院介入中心,蚌埠233004

出　　处：《磁共振成像》2025年第4期54-59,80,共7页Chinese Journal of Magnetic Resonance Imaging

基　　金：徐州市重点研发计划(社会发展)项目(编号:KC22239)。

摘　　要：目的评估深度学习(deep learning,DL)模型在布加综合征(Budd-Chiari syndrome,BCS)患者增强三维磁共振血管成像(three dimensional-magnetic resonance angiography,3D-MRA)图像中对肝脏和血管的分割性能,并评估不同影像医生对其分割结果的一致性。材料与方法回顾性分析220例BCS患者的MRA图像,两名分别具有8和12年经验的影像医生对图像进行手动分割,DL模型通过对手动分割结果的特征提取后自动分割图像。使用Dice相似性系数(Dice similarity coefficient,DSC)、敏感度、特异度和准确度评估DL模型的分割性能,通过受试者工作特征(receiver operating characteristics,ROC)曲线下面积(area under the curve,AUC)比较不同影像医生分割结果与DL模型之间的一致性,通过组内相关系数(intra-class correlation coefficient,ICC)和Wilcoxon配对检验评估不同影像医生对DL模型分割结果的一致性。结果DL模型分割肝脏、下腔静脉和肝静脉的DSC分别为0.93、0.84、0.65;敏感度分别为92%、81%、73%;特异度分别为93%、93%、76%;准确度分别为95%、94%、86%,AUC分别为0.95、0.87、0.71。两名影像医生和DL模型在BCS患者的肝脏和血管识别中AUC值差异无统计学意义(P>0.05),两名影像医生对DL模型分割结果的主观评价结果差异无统计学意义(P>0.05),总分ICC值为0.94(95%CI:0.92~0.95)。结论DL模型对BCS患者的增强3D-MRA图像有较好的分割性能,且对于DL模型所分割的图像,不同影像医生之间的判定结果表现出很好的一致性。Objective:To evaluate the segmentation performance of a deep learning(DL) model in the analysis of enhanced three dimensional-magnetic resonance angiography(3D-MRA) images of patients with Budd-Chiari syndrome(BCS),and assess the inter-observer agreement among radiologists in the evaluation of the DL model's segmentation outcomes.Materials and Methods:A retrospective analysis was conducted on MRA images from 220 BCS patients.Manual segmentation was performed by two radiologists with 8 and 12 years of experience,respectively.The DL model was trained on the features extracted from these manual segmentations to enable automatic segmentation.The performance of the DL model was assessed using the Dice similarity coefficient(DSC),sensitivity,specificity,and accuracy.Consistency comparison between the segmentation results of different radiologists and the DL model was used by the area under the curve(AUC) of receiver operating characteristics(ROC).Inter-observer agreement regarding the DL model's segmentation results was evaluated using the intra-class correlation coefficient(ICC) and Wilcoxon paired test.Results:The DL model achieved DSC values of 0.93,0.84,and 0.65 for the liver,inferior vena cava,and hepatic veins,respectively;sensitivity values were 92%,81%,and 73%;specificity values were 93%,93%,and 76%;accuracy values were 95%,94%,and 86%,and AUC values were 0.95,0.87,0.71,respectively.There was no statistically significant difference(P > 0.05) in AUC values between two radiologists and DL model in liver and vascular recognition of BCS patients.The subjective assessments of the DL model's segmentation results by the two radiologists showed no statistically significant differences(P > 0.05).The overall ICC was 0.94(95% CI:0.92 to 0.95).Conclusions:The DL model exhibited robust segmentation performance in enhanced 3D-MRA images of BCS patients.Furthermore,there was excellent inter-observer agreement among radiologists of the images segmented by the DL model.

关 键 词：布加综合征 深度学习 磁共振血管成像 磁共振成像 分割 

分 类 号：R445.2[医药卫生—影像医学与核医学] R575[医药卫生—诊断学]