Deep Learning TME Analysis Improves Prognostic Discrimination in Stage III Colon Cancer

Quantifying tumor microenvironment (TME) features via a deep learning algorithm (QuantCRC) significantly enhanced prognostic discrimination when combined with postoperative circulating tumor DNA (ctDNA) status in patients with stage III colon cancer.¹ Findings from a retrospective analysis of the phase 3 NCCTG N0147 trial (NCT00079274) presented at the 2026 American Society of Oncology (ASCO) Annual Meeting suggest that integrating digital pathology-derived TME metrics with minimal residual disease (MRD) status provides a more comprehensive assessment of recurrence risk than clinical factors alone.

Key Prognostic and Histopathologic Data

The deep learning analysis revealed distinct TME profiles associated with MRD status. Patients who were ctDNA-positive exhibited a TME characterized by features including a higher proportion of stromal content, high-grade histology, and immature stroma within the tumor bed. Furthermore, ctDNA positivity was linked to increased tumor budding, poorly differentiated clusters, and the mature stroma within the stromal area.

In contrast, patients who were ctDNA-negative demonstrated TMEs associated with robust immune infiltration. These cases showed significantly higher tumor cellularity, greater tumor-infiltrating lymphocyte (TIL) density, and increased inflammatory stromal features.

Multivariable analysis demonstrated that QuantCRC feature-outcome associations were both ctDNA-independent and ctDNA-dependent. Specific features consistently associated with adverse clinical outcomes across both cohorts included mature stromal features, tumor budding/poorly differentiated clusters, and higher mucin content.

Notably, inflammatory stromal features were associated with improved outcomes regardless of ctDNA status. However, higher TIL density was specifically associated with more favorable outcomes in the ctDNA-negative subgroup, suggesting that the prognostic relevance of certain immune features may depend on the absence of detectable MRD.

Impact on Prognostic Discrimination

The integration of QuantCRC features provided incremental prognostic value beyond standard clinical variables and MRD status. In nested Cox models, adding postoperative ctDNA status to clinical features increased the Harrell’s C-index by 0.062 for disease-free survival (DFS) and 0.044 for overall survival (OS).

The addition of QuantCRC features to a model already containing clinical and MRD data resulted in a further statistical gain, adding 0.018 to the C-index for DFS and 0.022 for OS. Likelihood ratio tests for these additions were highly significant, with P-values ranging from approximately 10^–17 to 10^–13. These data indicate that artificial intelligence (AI)-driven histopathologic quantification captures prognostic information that is not fully accounted for by either traditional clinical staging or molecular MRD detection.

“ctDNA positivity was associated with a more aggressive and stromal-rich TME, whereas ctDNA negativity is associated with more immune-infiltrated tumor,” said presenting study author Frank A. Sinicrope, MD, gastroenterologist at the Mayo Clinic, Rochester, and coauthors.¹ “QuantCRC feature-outcome associations demonstrated both ctDNA-independent and ctDNA-dependent prognostic patterns across DFS, OS, and treatment-related factor analyses, suggesting that the prognostic relevance of QuantCRC features may depend on postoperative ctDNA status.”

Trial Design and Patient Population

The study utilized data from the NCCTG N0147 (Alliance) trial, which evaluated folinic acidc, fluorouracil, and oxaliplatin (FOLFOX)-based adjuvant chemotherapy in patients with resected stage III colon adenocarcinoma. Researchers applied QuantCRC, an AI-based deep learning histopathologic algorithm, to digitized hematoxylin and eosin (H&E)-stained whole-slide images; 15 quantitative histopathologic features were extracted.

Out of the original trial cohort, 1817 patients met the requirements for data quality control. Investigators made associations between the histopathologic features, ctDNA status, and clinical outcomes such as OS and DFS.

The primary objective of the analysis was to quantify TME features and evaluate their association with postoperative ctDNA status. The ctDNA status was determined previously revealed using the Guardant Reveal assay.²

“The combination of ctDNA and QuantCRC features improved prognostic discrimination that further refined risk stratification in stage III colon cancer,” concluded the study authors.

Reference

1. Sinicrope FA, Shi Q, Segovia DI, et al. A deep learning approach to quantify tumor microenvironment features associated with postoperative ctDNA status and outcomes in a phase III FOLFOX-based adjuvant colon cancer trial (N0147; Alliance). J Clin Oncol. 2026;44(suppl 16):3525. doi:10.1200/JCO.2026.44.16_suppl.3525