Gastric mucosa intestinal metaplasia segmentation and grading via transformer-CNN fusion architecture: an interpretable digital pathology diagnostic framework.

Abstract

Intestinal metaplasia (IM) is a critical stage in the precancerous lesions of gastric cancer, and its accurate grading is essential for clinical risk stratification and early intervention. However, traditional diagnosis relies on the subjective assessment of pathologists, which suffers from inefficiency and low consistency. To address these limitations, this study proposes a deep learning framework based on whole slide images (WSI), utilizing multi-scale image patch cropping and a Transformer-CNN hybrid network (UDTransNet) to achieve pixel-level segmentation of intestinal metaplasia regions in gastric mucosa. The framework quantifies the area proportion of metaplastic glands according to the Sydney system criteria, enabling automated grading of IM severity. The results show that on the internal test set, the model achieved a segmentation performance with a Dice coefficient of 0.9698 and a grading accuracy of 0.8879 (Kappa value 0.85), demonstrating high consistency with pathological experts' diagnoses. The model significantly outperformed the diagnostic consistency between junior and intermediate pathologists (Kappa 0.67-0.82). Additionally, the model achieved specificity in identifying completely normal tissues, with a recall rate of 92.1% and an F1 score of 95.9%, effectively assisting clinical "negative exclusion" workflows. This study is the first to establish a digital mapping between pathological morphological features and the Sydney system criteria, addressing the challenges of strong subjectivity and low reproducibility in traditional diagnosis. It provides an efficient and interpretable solution for intelligent screening of gastric precancerous lesions.