AI Trained on Colorectal Cancer H&E Slides Also Stratifies Risk in Gastric and Esophageal Cancers

An artificial intelligence (AI) framework can stratify the risk of gastrointestinal cancers and precancerous lesions based solely on cell morphology in H&E-stained pathology slides. Morphologic features learned from colorectal cancer also distinguished outcomes when transferred directly to gastric and esophageal cancers.

According to the international journal Advanced Science on the 28th, a joint research team including Wang Bin of Zhongshan Hospital, Fudan University; Wang Lei of Nanjing Drum Tower Hospital, Nanjing University; and Jian-Hua Mao of Lawrence Berkeley National Laboratory in the United States reported the findings.

The team validated the framework in 2,602 patients across nine cohorts involving colorectal, gastric and esophageal cancers, as well as colorectal adenomas and early esophageal lesions. The final set of 13 cell morphologic biomarkers (CMBs) stratified outcomes across digestive organs and could also be used to assess progression risk in precancerous lesions and early tumors.

The approach follows a complete sequence: “unsupervised discovery—cross-tissue transfer—precancer risk management—multiomics interpretation.” First, the team automatically discovered a morphologic vocabulary through unsupervised learning on colorectal cancer pathology slides. It then transferred the fixed CMB dictionary directly to rectal, gastric and esophageal cancers without retraining, adjusting only the risk thresholds for each institution.

The team first used the TCGA colorectal cancer cohort as a discovery set and trained 128 cell morphologic biomarkers from whole-slide H&E images using an unsupervised CMB-ML framework. Each CMB comprises 15 interpretable morphologic attributes, including cell nucleus size, chromatin density, nuclear contour smoothness and perinuclear background features.

Of the 128 CMBs, 75 were significantly associated with patient outcomes. Multivariable modeling then identified the 13 most representative markers, which were used to create a CMB risk score (CMBRS) and high-, intermediate- and low-risk groups (CMBRG). In the colorectal cancer discovery cohort, CMBRS retained independent prognostic value after adjustment for TNM stage and other clinical factors (hazard ratio [HR]=2.08, p=0.001).

Notably, the model was not retrained for new cancer types. The CMB dictionary learned from colorectal cancer was transferred directly to rectal, gastric and esophageal cancers, with only the thresholds for baseline risk adjusted for each institution.

The 13-CMB features showed clear prognostic discrimination across all three cancer types. HRs were 4.414 for rectal cancer, 2.385 for gastric cancer and 2.693 for esophageal cancer; all had p<0.0001. The team also conducted external validation in cohorts reflecting real-world clinical settings, including 307 patients with colorectal cancer and 208 with gastric cancer, further confirming cross-tissue transferability and reproducibility.

The study’s clinical value lies in applying high-risk morphologic patterns identified in advanced tumors earlier, to the management of precancerous and early-stage lesions. Among patients with colorectal adenomas, CMBRG distinguished recurrence risk. Median recurrence-free survival was 42.5 months in the low-risk group, compared with 23.7 months in the high-risk group. CMBRS remained an independent risk factor after adjustment for clinical factors including age, sex, lesion location and number of adenomas.

This suggests that even among patients with the same or similar conventional pathology diagnoses, morphologic risk scores could help determine who may need longer intervals between examinations and who may need more frequent colonoscopic surveillance.

The difference in risk stratification was even more pronounced for early esophageal lesions. More than 95% of patients with low-risk esophageal inflammation did not progress, and about 70% eventually recovered normal tissue. By contrast, all 28 patients (100%) with high-risk inflammation showed lesion progression. Among patients with low-grade intraepithelial neoplasia (LGIN), none in the low-risk group progressed, while all 11 patients (100%) in the high-risk group did.

These findings point to the potential to replace a management model in which “everyone is followed at the same intervals and receives the same interventions” with more individualized surveillance and early-intervention strategies.

To determine what the AI was actually detecting, the team biologically interpreted the CMBs using bulk RNA sequencing, single-cell RNA sequencing and multiplex immunohistochemistry. Eleven of the 13 features were linked to clear biological functions, including cell proliferation, DNA repair, immune response, cell junctions, the cytoskeleton and the extracellular matrix.

In 17 prospectively collected colorectal cancer specimens, single-cell sequencing analyzed 106,515 cells and classified them into 11 major cell groups. Some high-risk CMBs were associated with myofibroblast accumulation and extensive extracellular matrix remodeling, alongside a typical “immune-excluded, stromal barrier” phenotype that makes it difficult for T cells to enter tumor nests. Other CMBs corresponded to more active immune-stromal cell interactions.

Thus, CMBs can be viewed not merely as statistical prognostic indicators but as “morphologic imprints” that capture the biological state of the tumor microenvironment. Starting with routine H&E pathology slides, this study advances digital pathology from diagnosing “what it is now” toward predicting “how it may progress.”

Because H&E slides are low-cost and widely used in clinical practice, these interpretable morphologic biomarkers could serve as a scalable tool to complement costly molecular testing. However, the team said further prospective validation is needed because the study was based on retrospective, multicenter cohorts.

Wang Bin of Zhongshan Hospital, Fudan University; Zhang Chengpei of Nanjing Drum Tower Hospital, Nanjing University; April W. Mao of Lawrence Berkeley National Laboratory and the University of California, Los Angeles (UCLA); Sun Qi of Nanjing Drum Tower Hospital, Nanjing University; and Luijun of Zhongshan Hospital, Fudan University, are co-first authors. Wang Bin, Wang Lei and Jian-Hua Mao are co-corresponding authors.