Abstract: Objective: To evaluate the feasibility of a CT-based semi-automatic segmentation method for target volume segmentation of colorectal cancer. Methods: Patients (n = 147) with postoperative pathological diagnosis were retrospectively analyzed. The radiologists used MITK to manually segment the lesions as the reference standard. In addition, DeepCRC, a deep learning method based on topology awareness, was used to segment the lesions semi-automatically. Radiomics features, including gross tumor volume (GTV), were extracted from the region of interest (ROI), and the intraclass correlation coefficient (ICC) was used to evaluate the consistency between GTV and radiomics. The Dice similarity coefficient (DSC) and volume similarity (VS) between the semi-automatic and manual segmentation results were analyzed to evaluate the repeatability of DeepCRC in tumor region segmentation. Results: There was good agreement between DeepCRC and manual segmentation ICC = 0.922 (95% CI: 0.896–0.942). Among the 1834 extracted radiomics features, 88.4% (n = 1621) showed a high consistency (ICC > 0.75). Compared with the DSC (0.831 ± 0.153) and VS (0.852 ± 0.079) of manual segmentation among human observers, DeepCRC had good repeatability with DSC (0.939 ± 0.134) and VS (0.950 ± 0.110). Conclusions: DeepCRC shows a high consistency and repeatability in colorectal cancer segmentation, and its influence on radiomics feature extraction is controllable, which provides technical support for standardized radiomics analysis and clinical application.