A semi-supervised deep learning framework (TBMNet) for automatic tumor budding quantification: real-world validation and clinical utility in stage II colorectal cancer

Abstract Background Tumor budding (TB) is a clinically important prognostic marker in colorectal cancer (CRC), but its assessment remains subjective and its biological basis is incompletely understood. We developed an artificial intelligence (AI)-based TB evaluation framework and explored the molecular characteristics of distinct TB phenotypes through spatial transcriptomic analysis. Methods A retrospective dual-cohort study including 256 CRC patients was performed. A semi-supervised deep-learning framework (TBMNet) was developed to automatically quantify tumor budding from whole-slide images and generate an AI-derived tumor budding (AI-TB) score. The prognostic and risk-stratification value of the AI-TB score in stage II CRC was evaluated using survival analyses, and an interaction analysis was performed to test whether AI-TB modified chemotherapy-associated outcomes. To characterize the molecular features of different TB phenotypes, external public transcriptomic and single-cell RNA sequencing (scRNA-seq) datasets were independently analyzed. Key findings were subsequently validated in clinical tissue samples using immunohistochemistry (IHC), and selected candidate genes were further assessed by exploratory RT-qPCR. Results TBMNet achieved superior TB detection under limited annotation conditions, improving mAP@50 from 0.535 to 0.682. Automated grading demonstrated strong agreement with pathologist assessment, with accuracies of 89.7% and 87.5% in the internal and external cohorts, respectively. In the exploratory stage II CRC analysis, high AI-TB status was associated with poorer recurrence-free survival, whereas the AI-TB x chemotherapy interaction did not reach statistical significance (HR = 0.702; 95% CI: 0.420–1.173; P = 0.188). These findings support the AI-TB score as an exploratory risk-stratification marker, but do not establish it as a predictive biomarker for chemotherapy benefit. Molecular analyses revealed marked heterogeneity among TB phenotypes: TB1 exhibited fibroblast/endothelial enrichment with activation of Wingless/Integrated (Wnt) and hypoxia pathways, TB2 displayed myeloid-cell accumulation with suppressed immune activity and enhanced neural signaling, whereas TB3 showed epithelial predominance, inflammatory activation, and metabolic suppression. IHC analyses confirmed the distinct molecular characteristics of these TB subgroups. Exploratory RT-qPCR further supported MDK upregulation in high-grade TB, whereas REG1A and SRPX2 did not show statistically significant overall differences at the mRNA level. Conclusions An AI-based TB assessment system was developed and validated for automated and reproducible TB evaluation in CRC. The AI-TB score showed exploratory prognostic value in stage II disease, while molecular analyses, supported by IHC and exploratory RT-qPCR validation, revealed distinct microenvironmental features across TB phenotypes, including MDK upregulation in high-grade TB. However, the non-significant AI-TB × chemotherapy interaction does not support AI-TB as a predictive biomarker for chemotherapy benefit. These findings highlight TB as a biologically heterogeneous process linking histopathological phenotypes with tumor microenvironmental states.