by Christos Evangelou, MSc, PhD – Medical Writer and Editor 

In a recent multinational study, researchers developed a foundation model that uses standard H&E-stained histopathology slides to predict prognosis in patients with gastrointestinal (GI) cancer. The study showed that the AI model not only accurately predicts survival outcomes but also identifies patients most likely to benefit from adjuvant chemotherapy. According to the authors, the model could transform treatment decision-making for patients with resected gastric and colorectal cancers. 

The report was published in the Journal of Clinical Oncology. 

DINOPath: A Foundation Model Pretrained on Millions of Histopathology Image Patches

Current staging systems for GI cancers provide general prognostic information but lack precision at the individual patient level. This limitation creates significant challenges for clinicians when deciding whether to recommend adjuvant therapy after surgical resection. The authors highlight an unmet need for more reliable prognostic tools for GI cancers, particularly when distinguishing patients who need additional therapy from those who are likely cured by surgery alone. 

The researchers addressed this challenge by developing DINOPath, a foundation model pretrained on over 130 million histopathology image patches from 104,876 whole-slide images across 25 organs. This approach represents a significant advancement over previous AI models that require large, manually labeled datasets for each application. Foundation models leverage advances in self-supervised learning to first learn patterns from unlabeled data during pretraining, making them more data-efficient than previous approaches. 

“Our results suggest that pretraining the foundation model on multi-center, large-scale datasets across diverse patient populations was key to its superior performance,” explained Ruijiang Li, PhD, Associate Professor at Stanford University School of Medicine and the corresponding author of the study. “This pretraining strategy allowed the model to more effectively learn representations of tissue morphology and generalize to new data.” 

Model Validation

The researchers validated the model across seven cohorts comprising 4,213 patients from three continents. They included 1,619 patients with gastric and esophageal cancers and 2,594 patients with colorectal cancer, all of whom had undergone surgical resection with or without adjuvant therapy. 

Despite the geographic diversity of the validation cohorts, the model performed consistently well across different populations. There were no significant differences in model performance across geographic regions.  

“In general, the prognostication accuracy was similar among patient cohorts, with the exception of gastric cancer, where small variations of less than 4% in model performance occurred between Chinese and US validation cohorts,” said Dr. Li. “These small variations could have arisen from racial differences in disease biology or treatment approaches.” 

The foundation model was first pretrained using self-supervised learning, allowing it to identify intrinsic tissue morphology patterns without human annotation. The researchers then fine-tuned separate prognostic models for gastric and colorectal cancers. For gastric cancer, the model predicted disease-free survival with a concordance index of 0.726-0.797 across validation cohorts. For colorectal cancer, it predicted disease-specific survival with a concordance index of 0.714-0.757. These performance metrics exceed those of existing prognostic tools and other foundation models tested in the study. 

Clinically Meaningful Risk Stratification

The model stratified patients into high and low-risk groups with substantial differences in survival outcomes. For gastric cancer, 5-year disease-free survival (DFS) rates were 49%-52% for high-risk patients versus 76%-92% for low-risk patients. For colorectal cancer, the 5-year disease-specific survival rates were 43%-72% versus 81%-98% for the high- and low-risk groups, respectively. 

In addition, the model was able to refine prognosis within traditional staging categories. For stage II gastric cancer, the model identified patients with 5-year DFS rates of 66% versus 93% for the high- and low-risk groups, respectively. For stage III, the rates were 31% versus 71%. Similar stratification was achieved for colorectal cancer. The model maintained its prognostic value across clinically relevant subgroups defined by age, sex, tumor location, and microsatellite status. 

In multivariable analysis, the AI risk score remained an independent prognostic factor after adjusting for established clinicopathological variables, including TNM stage, grade, age, and sex. When combined with clinical staging information, the model improved prognostic accuracy across all validation cohorts, demonstrating its complementary value to existing risk factors. 

Predicting Benefit From Adjuvant Chemotherapy

After propensity score matching to minimize confounding effects, the researchers found that adjuvant chemotherapy was associated with improved survival only in the high-risk group. 

In high-risk patients with gastric cancer, adjuvant chemotherapy reduced the risk of recurrence by 56% (hazard ratio [HR] = 0.44, P < 0.0001). For high-risk patients with colorectal cancer, adjuvant chemotherapy reduced cancer-specific mortality by 55% (HR = 0.45, P = 0.003). In contrast, low-risk patients showed no benefit from adjuvant therapy for gastric cancer (HR = 1.01) and potentially worse outcomes for colorectal cancer (HR = 1.21).  

Dr. Li envisions that this finding could significantly impact clinical practice: 

“Our AI model may be used to assist oncologists in making personalized, risk-adapted treatment decisions. For instance, high-risk patients would be recommended for adjuvant chemotherapy to improve survival, while low-risk patients could avoid chemotherapy to minimize toxicity and improve their quality of life. Prospective clinical trials will be needed to definitively confirm these findings.” 

Biological Insights

The researchers provided interpretability analysis to understand the histological features that the model identified as prognostically significant.  

“Through visualization of our AI model, we found that important areas for high-risk patients contain diffuse and solid growth, necrosis, tumor budding, and signet ring cells, which are well-established histological features associated with a worse prognosis,” Dr. Li explained. “The model also highlighted adipose infiltration, and its prognostic significance is not yet established.” 

Quantitative analysis revealed that high-risk areas contained significantly fewer lymphocytes and more tumor cells in colorectal cancer, whereas high-risk gastric cancers showed more necrotic tissue and fewer lymphocytes. These findings align with established adverse prognostic features in GI cancers and suggest the model is focusing on biologically relevant tissue characteristics. 

Future Directions

The authors emphasized the need for prospective validation before clinical implementation. 

“One potential application of the model is in stage III colorectal cancer, specifically for patients who are predicted to be at high risk of recurrence, and a prospective trial could be designed to test the benefit of more intensive treatment such as prolonged adjuvant chemotherapy for 6 months versus the standard regimen for 3 months,” Dr. Li said.  

He further explained that a positive trial would demonstrate that administering prolonged 6-month adjuvant chemotherapy to patients with AI-predicted high risk of recurrence improves survival outcomes compared to those treated with the standard 3-month regimen. 

Dr. Li also noted that because the foundation model was pretrained on pan-cancer data, its applicability could extend beyond GI cancers. 

“This would require fine-tuning the foundation model on a dataset consisting of digital pathology images and corresponding clinical outcomes for the new cancer,” he added. 

References

1. Wang X, Jiang Y, Yang S, et al. Foundation Model for Predicting Prognosis and Adjuvant Therapy Benefit From Digital Pathology in GI Cancers. J Clin Oncol. Published online April 1, 2025. doi:10.1200/JCO-24-01501