A Deep Learning-Enabled Computational Pathology Pipeline for Comparative Assessment of Intratumoral and Peritumoral Budding in Colorectal Cancer

Tumor budding is a well-established adverse prognostic factor in colorectal cancer, but current guideline-based assessment mainly emphasizes peritumoral budding (PTB) at the invasive margin, whereas the significance of intratumoral budding (ITB) remains less clear. Manual evaluation is also labor-intensive, subjective, and difficult to standardize. In this retrospective study, 245 colorectal cancer cases were analyzed to investigate the clinicopathological relevance of both ITB and PTB and to develop an AI-assisted computational pathology workflow for objective budding assessment. The workflow integrated multimodal whole-slide image registration, deep learning-based tissue segmentation with a Wide ResNet-101 backbone, and morphology-guided tumor budding detection on CAM5.2-stained whole-slide images. A total of 2,045 budding events were identified, including 1,310 PTB and 735 ITB. High-grade ITB was significantly associated with adverse pathological features, including poor differentiation, lymph node metastasis, and vascular invasion. The automated workflow showed strong concordance with expert assessment in both budding quantification and risk stratification. These findings suggest that ITB is closely associated with PTB and may serve as a complementary indicator of tumor aggressiveness. More importantly, the proposed workflow provides a robust and standardized framework for automated tumor budding assessment in computational pathology.