A new study published in the Journal of Translational Medicine tackles that problem with artificial intelligence. Researchers led by Yi Wang and Hongzhi Chang, working across Xi’an No.9 Hospital, Xi’an Jiaotong University, and collaborating institutions, developed a semi-supervised deep learning framework called TBMNet that automatically detects and quantifies tumor budding from whole-slide histopathology images. The team then tested the resulting AI-derived tumor budding score, or AI-TB, in a retrospective dual-cohort study of 256 colorectal cancer patients, asking whether a machine’s count of tumor buds could meaningfully stratify patients with stage II disease, a group in which decisions about adjuvant chemotherapy remain genuinely contentious.

The technical challenge the authors confronted is familiar to anyone working in computational pathology. Annotated medical images are expensive: every labeled tumor bud on a training slide represents minutes of expert effort, and the fine-grained, small-object nature of budding makes annotation particularly demanding. Semi-supervised learning offers a way around this bottleneck by letting a model extract structure from large volumes of unlabeled data while relying on only a modest set of expert annotations. Under these limited annotation conditions, TBMNet improved the mean average precision at an overlap threshold of 0.5 (mAP@50) from 0.535 to 0.682, a substantial gain in detection performance that translates directly into more reliable counts of these tiny cell clusters.

Detection accuracy alone, however, is not what makes a tool clinically useful. What matters is whether the machine’s grading matches what pathologists actually conclude. Here the results were encouraging: automated tumor budding grading agreed with pathologist assessment in 89.7% of cases in the internal validation cohort and 87.5% in the external cohort. That level of concordance, achieved across two independent patient groups, suggests the framework is not merely memorizing the visual quirks of one hospital’s scanning equipment or staining protocols but is capturing something robust about the biology visible on the slide.

The clinical payoff was explored in stage II colorectal cancer, the disease stage where prognosis is generally good but roughly one in five patients relapses, and where clinicians have long lacked a reliable way to identify those who need chemotherapy beyond surgery. In this exploratory analysis, patients whose tumors showed high AI-TB scores had significantly poorer recurrence-free survival, positioning the AI-TB score as a candidate risk-stratification marker for the stage II population. The finding, if confirmed in prospective studies, could eventually help spare low-risk patients unnecessary treatment while directing closer surveillance or adjuvant therapy to those at higher risk of recurrence.

Crucially, the researchers were careful about what their data did not show. A key question in stage II colorectal cancer is not just who will relapse, but who will actually benefit from chemotherapy. Interaction analysis testing whether the AI-TB score modified chemotherapy-associated outcomes did not reach statistical significance, with a hazard ratio of 0.702, a 95% confidence interval of 0.420 to 1.173, and a P value of 0.188. In plain terms, the study supports AI-TB as a prognostic marker, a measure of how the disease is likely to behave, but does not establish it as a predictive biomarker that identifies which patients gain from a specific treatment. The distinction between prognostic and predictive is one of the most important in oncology, and the authors’ restraint on this point strengthens the credibility of the work.

Beyond counting buds, the study probed what tumor budding actually is at the molecular level, and the answer turned out to be: not one thing. By independently analyzing external public transcriptomic datasets and single-cell RNA sequencing data, the team found that tumors cluster into distinct budding phenotypes with strikingly different microenvironments. Tumors in the TB1 group showed enrichment of fibroblasts and endothelial cells along with activation of the Wingless/Integrated (Wnt) and hypoxia signaling pathways, a profile consistent with tumors actively remodeling their surroundings and coping with low oxygen. TB2 tumors accumulated myeloid cells, showed suppressed immune activity, and displayed enhanced neural signaling, a combination that evokes an immunologically cold, nerve-engaged microenvironment. TB3 tumors were dominated by epithelial cells with inflammatory activation and metabolic suppression.

These molecular distinctions were not left as computational abstractions. The team validated the characteristic features of each budding subgroup in clinical tissue samples using immunohistochemistry, confirming that the different phenotypes carry distinct protein-level signatures visible under the microscope. In a further exploratory step, reverse transcription quantitative polymerase chain reaction (RT-qPCR) supported upregulation of the gene MDK, which encodes midkine, a growth factor implicated in tumor progression, in high-grade tumor budding. Two other candidate genes, REG1A and SRPX2, did not show statistically significant overall differences at the mRNA level, illustrating the exploratory nature of this molecular arm of the study and the reality that not every candidate survives scrutiny.

Taken together, the molecular findings reframe tumor budding not as a single pathological event but as a biologically heterogeneous process in which different microenvironmental states converge on a similar histological appearance. A high budding score might reflect Wnt-driven invasion supported by stromal cells in one patient and immune suppression with myeloid infiltration in another. This heterogeneity has practical implications: it hints that budding phenotypes could eventually serve as a bridge between what a pathologist sees and the molecular profiling that increasingly guides oncology, and it suggests that therapies targeting the tumor microenvironment might need to be matched to the specific phenotype underlying a given patient’s high budding score.

The study is not without limits that the authors themselves acknowledge. The clinical cohorts were retrospective, the stage II chemotherapy interaction analysis was underpowered to detect a modest effect, and the molecular subtype work relied on external public datasets with confirmatory immunohistochemistry in a clinical sample set rather than full prospective validation. Still, the core achievement stands: a deep learning system that counts tumor buds reproducibly, agrees closely with experts, and carries exploratory prognostic weight in exactly the patient group where better risk stratification is most needed. As digital pathology slides accumulate in hospital archives worldwide, frameworks like TBMNet point toward a future where quantitative, AI-derived measures of tumor behavior become a routine second opinion, one that never tires, never blinks, and counts every bud on the slide.

Subject of Research: Semi-supervised deep learning for automated tumor budding quantification and molecular profiling in stage II colorectal cancer

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

Article References: A semi-supervised deep learning framework (TBMNet) for automatic tumor budding quantification: real-world validation and clinical utility in stage II colorectal cancer. (n.d.). https://doi.org/10.1186/s12967-026-08944-y

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08944-y

Keywords: colorectal cancer, tumor budding, deep learning, semi-supervised learning, digital pathology, whole-slide imaging, prognostic biomarker, stage II colorectal cancer, tumor microenvironment, spatial transcriptomics, single-cell RNA sequencing, immunohistochemistry