深度学习抑制静脉信号改善下肢淋巴水肿MR淋巴成像
本研究基于四家医学中心1022例下肢淋巴水肿患者的MR淋巴成像数据,开发并验证深度学习模型LympClear。该模型可抑制静脉信号干扰,提高淋巴管显示与阅片效率。结果显示图像质量、静脉抑制评分及淋巴反流可见性均显著提升,漏检率下降。
要点
- LympClear可显著抑制MR淋巴成像中的静脉信号,减少淋巴管误判
- 静脉抑制图像提高淋巴反流区检出,减少病例漏检
- 动态图像较静态图像更清晰显示对比剂流动,缩短阅片时间
核心结论
- LympClear可有效抑制MR淋巴成像中的静脉信号,提升下肢淋巴水肿患者的影像诊断质量与阅片效率
研究设计与样本
- 回顾性与前瞻性收集2007年7月至2024年3月四家医学中心1022例下肢淋巴水肿患者的6162次动态MR淋巴成像扫描
- 三名放射科医师标注静脉信号强度,共识标签作为训练与评估的参考标准
- LympClear基于nnU-Net框架训练,采用亮度匹配技术抑制静脉信号强度
关键影像表现/征象
- LympClear对静脉分割的Dice评分为0.94
- 静脉抑制图像较原始图像显著改善图像质量、静脉信号抑制评分及淋巴反流可见性
- 75例评估病例中,静脉抑制图像帮助阅片者多识别36个(14.9%)反流区域
- 病例层面漏检率从75例中11例(14.7%)降至5例(6.7%)
- 动态图像较静态图像提高对比剂流动可见性,淋巴形态分级与病因相关
诊断标准与关键数值
- 图像质量、静脉信号抑制评分、淋巴反流可见性均采用10分制,静脉抑制图像分别为7.5±0.9、8.8±0.7、8.3±1.0,原始图像分别为6.7±1.1、5.8±1.2、6.0±1.3,均P<.001
- 静脉抑制图像平均阅片时间为3.6分钟±1.7,原始图像为16.7分钟±3.3,P<.001
- 动态图像对比剂流动可见性评分为8.68±0.54,静态图像为6.85±0.69,P<.001
临床应用与阅片要点
- 对下肢淋巴水肿患者行MR淋巴成像时,可考虑使用LympClear进行静脉信号后处理以提升淋巴管显示
- 阅读静脉抑制图像可减少因静脉信号遮盖导致的淋巴反流区域漏检
- 动态序列有助于追踪对比剂流动,建议结合静态图像进行综合判读
局限性
- 研究基于四家中心数据,外部验证仅来自三个测试集,模型跨中心泛化能力需进一步验证
- 参考标准由放射科医师共识标注构建,可能存在主观差异
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
Background Venous signal can obscure lymphatic structures, leading to misinterpretation and diagnostic errors in MR lymphangiography for lower extremity lymphedema (LEL). Purpose To develop and evaluate LympClear for venous signal suppression in MR lymphangiography. Materials and Methods Retrospective and prospective data from MR lymphangiography from four medical centers were collected (July 2007-March 2024) for individuals with LEL. Three radiologists annotated venous signal intensity on images, and consensus labels served as the reference standard for training and evaluation. LympClear, based on the nnU-Net framework, was trained with a brightness-matching technique to suppress venous signal intensity. Dynamic images were generated to depict contrast agent flow, and a grading system for lymphatic morphologic features was introduced. Model performance and diagnostic impact were evaluated using quantitative and reader-based analyses, including paired t tests, χ2 tests, κ statistics, and linear mixed-effects models. Results A total of 1022 individuals with LEL (6162 dynamic scans) (mean age, 49.3 years ± 14.0 [SD]; 560 female individuals) were included in the training, validation, prospective internal test, and three external test sets. LympClear had a Dice score of 0.94 for venous segmentation. Compared with original images, venous-suppressed images showed improved image quality, venous signal suppression score, and lymphatic reflux visibility (all 10-point scales; mean, 7.5 ± 0.9 vs 6.7 ± 1.1, 8.8 ± 0.7 vs 5.8 ± 1.2, and 8.3 ± 1.0 vs 6.0 ± 1.3, respectively; all P < .001) and improved reading efficiency (mean, 3.6 minutes ± 1.7 vs 16.7 minutes ± 3.3; P < .001). Of 241 reference-standard reflux regions identified in 75 evaluated cases, readers identified 36 (14.9%) more reflux regions on venous-suppressed images. At the case level, the missed detection rate decreased from 11 of 75 (14.7%) to five of 75 (6.7%). Dynamic images enhanced contrast agent flow visibility compared with static images (mean, 8.68 ± 0.54 vs 6.85 ± 0.69; P < .001), and lymphatic morphologic grade was associated with etiology. Conclusion LympClear suppressed venous signal at MR lymphangiography and improved image diagnostic quality in individuals with LEL. © RSNA, 2026 Supplemental material is available for this article. See also the editorial by Nezami and Habibollahi in this issue.
引用
- Ma, Gong, Zhang, et al. MR Lymphangiography for Diagnosis of Lower Extremity Lymphedema: Suppressing Venous Signal Interference Using Deep Learning. Radiology. 2026. PMID: 42334351. doi:10.1148/radiol.251225
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。