摘要
目的探讨基于肝特异性对比剂钆塞酸二钠增强MRI图像, 采用深度学习的自动分割技术在肝细胞癌(HCC)病灶分割中的效能, 并探讨通过影像组学分析预测患者预后的价值。方法本研究为横断面研究, 回顾性收集2015年6月至2023年5月在哈尔滨医科大学附属肿瘤医院352例单发HCC患者资料, 采用加权随机抽样法按3∶2比例分为训练集(n=213)和验证集(n=139)。由2名放射科医师共同对病灶进行手动标注。对钆塞酸二钠增强肝胆期MRI图像进行标准化处理, 使用6种深度学习模型nnU-Net、nnFormer、UnetR、Swin-UnetR、UnetR++和MedNeXt, 对训练集进行自动分割训练, 并在验证集中评估分割效果, 分割效能通过相似系数(Dice系数)、95%豪斯多夫距离(HD_(95))进行评估, 获得最优模型。提取手动与自动分割区域的影像组学特征, 计算影像组学评分(Radscore), 对患者进行分层(高风险组和低风险组), 采用Kaplan-Meier曲线和log-rank检验分析不同分层患者无复发生存期(RFS)和总生存期(OS)的差异。结果在自动分割模型的评估中, MedNeXt模型在验证集中表现最佳, Dice系数为76.0%, HD_(95)为7.2, 分割成功率为90.6%(126/139)。nnFormer模型次之, Dice系数为75.3%, HD_(95)为10.1, 分割成功率为89.9%(125/139)。其他模型的Dice系数为66.3%~74.1%。联合MedNeXt和nnFormer模型建立Mednext-nnF融合模型, 其在验证集中Dice系数为78.2%, HD_(95)为5.9, 分割成功率为92.1%(128/139)。构建自动分割影像组学预后模型后, 通过Radscore对患者进行预后分层, 无论使用手动分割还是自动分割模型, Kaplan-Meier曲线分析均显示, 不同分层患者RFS和OS的差异有统计学意义(P均<0.001)。结论 Mednext-nnF融合模型可实现肝特异性增强MRI中HCC病灶的高效自动分割, 基于自动分割构建的影像组学模型具备良好的预测预后风险分层效果。
Objective To explore the efficacy of deep learning-based automatic segmentation technology in the segmentation of hepatocellular carcinoma(HCC)lesions using gadoxetate disodium-enhanced MRI(EOB-MRI),and to investigate the prognostic value of radiomics analysis in predicting patient outcomes.Methods This was a cross-sectional,retrospective study that collected data from 352 patients with solitary HCC who underwent imaging at the Harbin Medical University Cancer Hospital between June 2015 and May 2023.The patients were randomly divided into a training set(n=213)and a validation set(n=139)in a 3:2 ratio using weighted random sampling.Two radiologists manually annotated the lesions.Hepatobiliary-phase EOB-MRI images were standardized,and six deep learning models,nnU-Net,nnFormer,UnetR,Swin-UnetR,UnetR++and MedNeXt,were trained for automatic segmentation on the training set.The segmentation performance was evaluated on the validation set,and the segmentation efficacy was assessed using the Dice coefficient and 95%Hausdorff distance(HD_(95)),identifying of the optimal model.Radiomics features were extracted from both manual and automatic segmentation regions,and the radiomics score(Radscore)was calculated to stratify patients into high-risk and low-risk groups.Kaplan-Meier curves and log-rank tests were used to analyze the differences in relapse-free survival(RFS)and overall survival(OS)between the different stratified groups.Results Among the automatic segmentation models,the MedNeXt model performed best in the validation set,with a Dice coefficient of 76.0%,HDgs of 7.2,and a segmentation success rate of 90.6%(126/139).The nnFormer model was the second-best,with a Dice coefficient of 75.3%,HDos of 10.1,and a segmentation success rate of 89.9%(125/139).Other models showed Dice coefficients ranging from 66.3%to 74.1%.A MedNext-nnF model was established by combining the MedNeXt and nnFormer models,achieving a Dice coefficient of 78.2%,HDgs of 5.9,and a segmentation success rate of 92.1%(128/139)in the validation group.After constructing the automatic segmentation radiomics prognostic model,patients were stratified by Radscore.Both manual and automatic segmentation models showed statistically significant differences in RFS and OS between different risk groups(P<0.001).Conclusions The Mednext-nnF fusion model enables efficient and automated segmentation of HCC lesions in EOB-MRI.The radiomics model constructed based on the automated segmentation demonstrates strong performance in predicting and stratifying prognostic risk.
作者
俞灿
张奇
王玥琪
范恬湉
厉惠滢
丛山
周洋
Yu Can;Zhang Qi;Wang Yueqi;Fan Tiantian;Li Huiying;Cong Shan;Zhou Yang(Imaging Center,Harbin Medical University Cancer Hospital,Harbin 150001,China;Qingdao Innovationn and Development Base,Harbin Engineering University,Qingdao 266000,China;Department of Medical Oncology,Harbin Medical University Cancer Hospital,Harbin 150001,China)
出处
《中华放射学杂志》
北大核心
2025年第6期681-687,共7页
Chinese Journal of Radiology
基金
哈尔滨医科大学附属肿瘤医院攀登计划(PDYS2024-10)
黑龙江省自然科学基金(LH2022H067)。
关键词
癌
肝细胞
磁共振成像
钆塞酸二钠
人工智能
深度学习
影像组学
预测模型
Carcinoma,hepatocellular
Magnetic Resonance Imaging
Gadoxetic acid disodium
Artificial intelligence
Deep learning
Radiomics
Prognostic model
