Mucin phenotype-based deep learning framework for intestinal metaplasia-carcinogenesis progression prediction

Area of Science:

• Gastroenterology and Oncology
• Computational Pathology
• Biomedical Imaging

Background:

• Gastric carcinogenesis involves dynamic changes in mucin expression.
• Distinguishing gastric intestinal metaplasia (GIM) is crucial for gastric cancer (GC) risk stratification.
• Current methods for mucin analysis can be labor-intensive and require specialized stains.

Purpose of the Study:

• To decode spatiotemporal mucin dynamics during gastric carcinogenesis.
• To develop and validate an AI model for predicting mucin markers from H&E images.
• To integrate AI-derived mucin information with clinical data for improved GIM risk assessment.

Main Methods:

• Analysis of gastric tissue samples to track mucin marker expression (MUC5AC, MUC6, MUC2, CD10) in GIM and GC.
• Development of a dual-function UNI-pretrained Vision Transformer (ViT) model (MPMR) for predicting mucin markers from H&E whole-slide images.
• Integration of MPMR outputs with clinical variables to create the MPMR-IMCP risk model.

Main Results:

• Gastric-type mucin markers (MUC5AC/MUC6) decreased, while intestinal-type markers (MUC2/CD10) increased in GIM, before declining in GC.
• The MPMR model achieved high accuracy (AUC: 0.921-0.997) in predicting mucin markers from H&E images.
• The MPMR-IMCP risk model significantly improved GIM risk stratification compared to clinical-only models (ΔAUC = 0.050).

Conclusions:

• The study elucidates mucin dynamics in gastric carcinogenesis, highlighting a transition from gastric to intestinal phenotypes.
• A novel AI framework (MPMR) enables accurate, stain-free prediction of mucin markers from H&E images.
• This approach provides an efficient tool for phenotype analysis and GIM risk stratification, particularly in high-incidence regions.