PO.BCS02.04 · 生物信息与计算
使用AI模型对肝硬化患者对比增强MRI上的肝细胞癌病灶进行自动分割
Automated segmentation of hepatocellular carcinoma lesions on contrast-enhanced MRI using an AI model in patients with cirrhosis
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
目的：在多期相对比增强T1 MRI图像上开发一种自动肝细胞癌（HCC）检测模型。方法：从两家机构（中心1：n=106；中心2：n=87）获取了多期相（对比前、动脉期、静脉期和延迟期）对比增强T1 MRI，采集时间为2007年至2023年。一名放射科专家在586次扫描中手动勾画了1794个局灶性肝病灶，并使用肝脏影像报告和数据系统（LI-RADS）为每个病灶评分。六个3D全分辨率nnU-Net模型在多期相（对比前、动脉期、静脉期和延迟期）或仅动脉期对比增强T1图像上训练。应用病灶大小（最大病灶直径≥1 cm）和病灶评分（LI-RADS≥3）标准以评估其对这些nnU-Net模型训练的影响。在扫描和病灶层面评估模型性能。性能指标包括Dice相似系数（DSC）、敏感性、特异性、准确性和阳性预测值（PPV）。结果：不使用病灶标准的动脉期被确定为在所开发的模型中具有最佳总体性能。在扫描层面，动脉期模型实现了93%（40/43）敏感性、46.6%（7/15）特异性和81%（47/58）准确性。在动脉期模型的病灶层面，性能略有下降，敏感性为65.9%（126/191）、PPV为68.1%（126/185）、准确性为51.8%（133/257）。在失败分析中，观察到45.2%（57/126）的真阳性病灶LI-RADS评分≥4，而9.2%（6/65）的假阴性病灶LI-RADS评分≥4。结论：在多机构数据上开发了一个动脉期对比增强MRI nnU-Net模型，用于检测肝硬化患者的HCC，产生了有前景的结果。本研究表明AI模型可改善高风险患者的检测。
基于MRI期相、病灶评分和大小阈值的纳入对模型的分类 模型 使用的MRI期相 病灶纳入 大小阈值 A 对比前、动脉期、静脉期和延迟期 全部 无 B 动脉期 全部 无 C 对比前、动脉期、静脉期和延迟期 LI-RADS≥3 无 D 动脉期 LI-RADS≥3 无 E 对比前、动脉期、静脉期和延迟期 LI-RADS≥3 ≥1.0 cm F 动脉期 LI-RADS≥3 ≥1.0 cm
查看英文原文 English abstract
Purpose: Develop an automated hepatocellular carcinoma (HCC) detection model on multi-phasic contrast-enhanced T1 MRI images. Methods: Multi-phase (pre-contrast, arterial, venous, and delayed phases) contrast-enhanced T1 MRIs were obtained from two institutions (Center 1: n=106; Center 2: n=87) and acquired between 2007 and 2023. An expert radiologist manually contoured 1794 focal liver lesions across 586 scans and assigned each lesion a score using the Liver Imaging Reporting and Data System (LI-RADS). Six 3D full-resolution nnU-Net models were trained on multi-phase (pre-contrast, arterial, venous, and delayed) or only arterial-phase contrast-enhanced T1 images. Lesion size (maximum lesion diameter ≥ 1 cm) and lesion score (LI-RADS ≥ 3) criteria were applied to assess the impact on the training of these nnU-Net models. Model performance was evaluated at a scan and lesion level. Performance metrics included dice similarity coefficient (DSC), sensitivity, specificity, accuracy, and positive predictive value (PPV). Results: The arterial phase without lesion criteria was determined to have the best overall performance among the models developed. At the scan level, the arterial phase model achieved 93% (40/43) sensitivity, 46.6% (7/15) specificity, and 81% (47/58) accuracy. At a lesion level for the arterial phase model, the performance dropped slightly, with 65.9% (126/191) sensitivity, 68.1% (126/185) PPV, and 51.8% (133/257) accuracy. In a failure analysis, it was observed that 45.2% (57/126) of true-positive lesions had a LIRADS score ≥ 4, while 9.2% (6/65) of false-negative lesions had a LIRADS score ≥ 4. Conclusion: An arterial phase contrast-enhanced MRI nnU-Net model was developed on multi-institutional data to detect HCC in patients with cirrhosis, producing promising results. This study indicates that AI models can improve detection in high-risk patients.
Categorization of Models based on the Inclusion of MRI Phases, Lesion Score, and Size Threshold Model MRI Phase(s) Used Lesion Inclusion Size Threshold A Pre-Contrast, Arterial, Venous, and Delayed All None B Arterial All None C Pre-Contrast, Arterial, Venous, and Delayed LIRADS ≥ 3 None D Arterial LIRADS ≥ 3 None E Pre-Contrast, Arterial, Venous, and Delayed LIRADS ≥ 3 ≥ 1.0 cm F Arterial LIRADS ≥ 3 ≥ 1.0 cm
利益披露 Disclosure
E. J. Stevenson, None..
N. Lay, None..
S. A. Harmon, None..
H. Zhang, None..
F. Haque, None..
P. Choyke, None..
T. Heller, None..
R. Filice, None..
B. Turkbey, None..
C. Hsu, None.