摘要目的:评估三维no new U-Net（3D nnU-Net）深度学习（DL）网络基于腹部CT图像自动分割结直肠癌（CRC）的可行性。方法:本研究为横断面研究，回顾性收集2018年1月至2023年5月广东省中医院（中心1， n=777）、南方医科大学南方医院（中心2， n=732）及中山大学孙逸仙纪念医院（中心3， n=671）共2 180例经病理证实的原发性CRC患者，其基线腹部CT检查于以上3个中心、4种品牌、7种不同机型的CT设备完成，包括动脉期及静脉期增强扫描。由2名放射科医师对动脉期和静脉期CT图像上全瘤病灶进行手动标注。合并中心1和中心3的数据，按照4∶1的比例使用加权随机抽样法分为训练集（ n=1 159）和验证集（ n=289），中心2的数据作为独立的外部测试集（ n=732），对3D nnU-Net分割模型进行训练和验证。以手动标注的标签数据作为评价标准，分别基于不同期相和不同肿瘤部位计算分割覆盖率（SCR）、相似系数Dice（DSC）、召回率（REC）、精确度（PRE）、F1分数及95%豪斯多夫距离（HD 95），对模型的分割性能进行评价。采用独立样本 t检验比较肿瘤自动分割时间与手动标注时间的差异。 结果:在独立的外部测试集中，3D nnU-Net模型于动脉期自动分割肿瘤的效能优于静脉期，动脉期SCR、DSC、REC、PRE、F1分数及HD 95分别为0.865、0.714、0.716、0.736、0.714及27.228；静脉期分别为0.834、0.679、0.710、0.675、0.679及29.358。模型对右半结肠癌的分割效果最好，基于动脉期图像的SCR、DSC、REC、PRE、F1分数及HD 95分别为0.901、0.775、0.780、0.787、0.775及21.793，其次为左半结肠癌和直肠癌，对横结肠癌的分割效果最差（SCR、DSC、REC、PRE、F1分数及HD 95分别为0.731、0.631、0.641、0.630、0.631及38.721）。每例肿瘤单期相的自动分割时间为（1.0±0.3）min，手动标注时间为（17.5±6.0）min，差异有统计学意义（ t=128.24， P<0.001）。 结论:经多中心多机型的数据集训练和验证后，3D nnU-Net DL网络基于腹部CT图像可实现对CRC的自动分割，且模型具有较好的鲁棒性及泛化能力。

abstractsObjective:To investigate the feasibility of a three-dimensional no new U-Net (3D nnU-Net) deep learning (DL) network for the automatic segmentation of colorectal cancer (CRC) based on abdominal CT images.Methods:This was a cross-sectional study. From January 2018 to May 2023, a total of 2180 primary CRC patients, confirmed by pathology at the Guangdong Provincial Hospital of Traditional Chinese Medicine (center 1, n=777), Nanfang Hospital, Southern Medical University (center 2, n=732), and Sun Yat-sen Memorial Hospital (center 3, n=671), were enrolled in this retrospective study. The baseline abdominal CT examination of each patient was conducted using CT equipment from 7 different models across 4 vendors, at the 3 centers, encompassing both the arterial phase (AP) and venous phase (VP). Two radiologists manually delineated the volume of interest to circumscribe the entire tumors in dual-enhanced phase CT images. The CT data of CRC patients from center 1 and center 3 were merged and divided into a training set ( n=1 159) and a validation set ( n=289) using a weighted random method with a ratio of 4∶1. The patients from center 2 were used as an independent external test set ( n=732). The 3D nnU-Net segmentation model was trained and tested. Using manually annotated label data as the benchmark, segmentation performance of the model was evaluated based on different phases and tumor locations. The segmentation coverage rate (SCR), Dice similarity coefficient (DSC), recall (REC), precision (PRE), F1-score, and 95% Hausdorff distance (HD 95) were calculated. The mean manual segmentation time and the mean automatic time were compared using independent samples t-test. Results:In the independent external test set, the performance of the 3D nnU-Net model based on the AP CT images was superior to that based on the VP CT images. On the AP images, the SCR, DSC, REC, PRE, F1-score, and HD 95 were 0.865, 0.714, 0.716, 0.736, 0.714, and 27.228, respectively; on the VP images, they were 0.834, 0.679, 0.710, 0.675, 0.679, and 29.358, respectively. The model achieved the best performance on right-sided colon cancer, with SCR, DSC, REC, PRE, F1-score, and HD95 on the AP CT images at 0.901, 0.775, 0.780, 0.787, 0.775, and 21.793, respectively. Next were left-sided colon cancer and rectal cancer, while the segmentation performance for transverse colon cancer was the worst (SCR, DSC, REC, PRE, F1-score, and HD 95 were 0.731, 0.631, 0.641, 0.630, 0.631 and 38.721, respectively). The automatic segmentation time on a single phase was (1.0±0.3) min, while the manual segmentation time was (17.5±6.0) min ( t=128.24, P<0.001). Conclusions:After training and validating on a dataset from multiple centers with various CT scanner vendors, the 3D nnU-Net DL model demonstrates the capability to automatically segment CRC based on abdominal CT images, while also showcasing commendable robustness and generalization ability.