The application of deep learning in medical imaging faces challenges such as time-consuming and labor-intensive data annotation, which hinders their clinical translation efficiency.

To investigate the feasibility and efficacy of a deep learning network (TransUNet-Cascade) that integrates 3D U-Net and Transformer for the automatic segmentation of colorectal cancer (CRC) lesions in abdominal CT images.

A retrospective analysis was conducted on contrast-enhanced abdominal CT images from 2 180 CRC patients at Guangdong Provincial Traditional Chinese Medicine Hospital (Center 1), Nanfang Hospital, Southern Medical University (Center 2), and Sun Yat-sen Memorial Hospital, Sun Yat-sen University (Center 3) between January 2018 and May 2023. The dataset was divided into a training set (n=1 159), a validation set (n=289), and an external test set (n=732) using weighted random sampling. A cascaded network framework (TransUNet-Cascade) based on 3D U-Net and Swin Transformer was proposed, optimizing segmentation accuracy through a multi-stage learning strategy. Model performance was evaluated using manual annotations as the gold standard, with metrics including Dice similarity coefficient (DSC), F1 score, 95% Hausdorff distance (HD₉₅), intersection over union (IoU), precision (PRE), and recall rate (REC). This study selected the 3D no new U-Net (3D nnU-Net) as the baseline model for comparison. Under the same dataset and evaluation criteria, a systematic training and performance comparison was conducted with the proposed TransUNet-Cascade network to comprehensively verify its effectiveness in the CRC segmentation task.

In the external test set, the performance of the two deep learning models based on the AP CT images was superior to that based on the VP CT images. TransUNet-Cascade achieved average DSC, F1 score, HD₉₅, IoU, PRE, and REC values of 0.740, 0.839, 34.084, 0.656, 0.737, and 0.767, respectively, which were overall superior to 3D nnU-Net (average DSC, F1 score, HD₉₅, IoU, PRE, and REC values of 0.724, 0.838, 35.954, 0.642, 0.730, and 0.744, respectively). The model demonstrated the best segmentation performance for right-sided colon cancer (DSC=0.784), while segmentation of rectal cancer was relatively less effective (DSC=0.622).

By combining the strengths of convolutional neural networks and Transformers, TransUNet-Cascade improves the accuracy of automatic CRC lesion segmentation, demonstrating potential for clinical application.

深度学习在医学图像中的应用面临数据标注耗时、费力等挑战，制约了其临床转化效率。

探讨一种融合三维U-Net（3D U-Net）与Transformer的深度学习网络（TransUNet-Cascade）在腹部CT图像中自动分割结直肠癌（CRC）病灶的可行性与效能。

回顾性纳入2018年1月—2023年5月来自广东省中医院（中心1）、南方医科大学南方医院（中心2）及中山大学孙逸仙纪念医院（中心3）的2 180例CRC患者腹部CT增强图像，采用分层加权随机抽样法划分为训练集（n=1 159）、验证集（n=289）和外部测试集（n=732）。本研究提出一种新型深度学习网络模型——TransUNet-Cascade，通过多阶段学习策略优化分割精度。以手动标注为基准，采用Dice相似系数（DSC）、F1值、95%豪斯多夫距离（HD₉₅）、交互比（IoU）、精确度（PRE）以及召回率（REC）评估模型性能。本研究选取三维no new U-Net（3D nnU-Net）作为对比基线模型，在统一的数据集和评估标准下，与所提出的TransUNet-Cascade网络进行系统训练与性能对比，以全面验证其在CRC分割任务中的有效性。

在独立的外部测试集中，两种深度学习网络基于动脉期图像的分割效能总体优于静脉期图像；TransUNet-Cascade动脉期平均DSC、F1值、HD₉₅、IoU、PRE及REC分别为0.740、0.839、34.084、0.656、0.737及0.767，总体优于3D nnU-Net（平均DSC、F1值、HD₉₅、IoU、PRE及REC分别为0.724、0.838、35.954、0.642、0.730及0.744）。模型对右半结肠癌的分割效果最佳（DSC=0.784），直肠癌分割效果相对较差（DSC=0.622）。

TransUNet-Cascade通过结合卷积神经网络与Transformer的优势，提升了CRC病灶的自动分割精度，具有一定的临床应用潜力。