Title: A Gastric-Specific Pathology Foundation Model Validated in Real-World Clinical Settings

Gastric cancer continues to be a leading cause of cancer-related deaths, but its complex histological and molecular diversity presents significant challenges for accurate diagnosis and risk assessment. While general-purpose pathology foundation models (PFMs) have shown promise, they often struggle with the nuanced endpoints critical to gastric cancer management. Furthermore, such models rarely undergo rigorous prospective validation or clinical reader studies. To address these gaps, we introduce GRACE (Gastric-specific foundation model for Real-world Assessment and Clinical dEcision support).

GRACE was trained on extensive multicenter gastric pathology datasets comprising 48,364 whole-slide images, predominantly stained with hematoxylin and eosin (HE), sourced from 37,493 patients. In evaluations across 28 clinically relevant tasks, GRACE consistently surpassed representative pancancer PFMs, securing a macro-AUC of 0.9188. The model demonstrated particularly robust performance in diagnosing precancerous lesions (macro-AUC 0.9322), assessing tumor histopathology (macro-AUC 0.9119), conducting molecular profiling (macro-AUC 0.8682), and predicting patient prognosis.

The translational potential of GRACE was validated through a comprehensive evidence chain. Under strict safety protocols requiring 100% negative predictive value (NPV) for ruling out conditions and 100% positive predictive value (PPV) for ruling them in, GRACE streamlined the review process for up to 69.6% of malignancy diagnosis cases and triaged 46.8% of requests for MMR-IHC follow-up.

This feasibility was further confirmed by a randomized crossover reader study involving pathologists working in collaboration with AI. The integration of GRACE boosted diagnostic accuracy from 82.0% to 89.9%, nearly doubling the adjusted odds of a correct diagnosis (OR 1.987) while simultaneously enhancing both sensitivity and specificity. Additionally, AI assistance cut diagnostic time by 14.9%, increased diagnostic confidence by 9.0%, and significantly improved inter-rater agreement. When the workflow was calibrated to ensure performance non-inferior to that of senior pathologists, it enabled the triage of 60.7% of atrophy cases and 82.7% of intestinal metaplasia cases.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC