Abstract
Inflammation, acute activity, intestinal metaplasia, and atrophy are key indicators in gastric biopsy evaluations, with their grading being crucial for assessing gastric cancer progression. However, diagnostic subjectivity among pathologists and the complex interrelationships between these indicators present significant challenges. Additionally, the high resolution of whole slide images (WSIs) complicates large-scale annotation efforts. To address these issues, we propose a multi-task learning model utilizing self-supervised pre-trained weights from extensive pathological datasets. The model integrates four indicators-severity of inflammation, atrophy, acute activity, and intestinal metaplasia-training on WSIs to predict these indicators while accounting for their interrelationships. Our results show that multi-task learning outperforms single-task models, achieving higher accuracy across all indicators. This model can thus serve as an auxiliary tool for evaluating non-tumor gastric diseases and supporting diagnosis.