DeepCRC: Colorectum and Colorectal Cancer Segmentation in CT Scans via Deep Colorectal Coordinate Transform

Lisha Yao, Yingda Xia, Haochen Zhang, Jiawen Yao, Dakai Jin, Bingjiang Qiu, Yuan Zhang, Suyun Li, Yanting Liang, Xian-Sheng Hua, Le Lu, Xin Chen, Zaiyi Liu, Ling Zhang

Abstract: We propose DeepCRC, a topology-aware deep learning-based approach for automated colorectum and colorectal cancer (CRC) segmentation in routine abdominal CT scans. Compared with MRI and CT Colonography, regular CT has a broader application but is more challenging. Standard segmentation algorithms often induce discontinued colon prediction, leading to inaccurate or completely failed CRC segmentation. To tackle this issue, we establish a new 1D colorectal coordinate system that encodes the position information along the colorectal elongated topology. In addition to the regular segmentation task, we propose an auxiliary regression task that directly predicts the colorectal coordinate for each voxel. This task integrates the global topological information into the network embedding and thus improves the continuity of the colorectum and the accuracy of the tumor segmentation. To enhance the model’s architectural ability of modeling global context, we add self-attention layers to the model backbone, and found it complementary to the proposed algorithm. We validate our approach on a cross-validation of 107 cases and outperform nnUNet by an absolute margin of 1.3% in colorectum segmentation and 8.3% in CRC segmentation. Notably, we achieve comparable tumor segmentation performance with the human inter-observer (DSC: 0.646 vs. 0.639), indicating that our method has similar reproducibility as a human observer.