SamMIL: Semi-Supervised Adaptive Multiple Instance Learning for Precise HP Grading in Gastric Pathology

AI-assisted diagnosis using gastric pathology images is essential for achieving accurate, efficient, and quantitative grading of Helicobacter pylori (HP) infection, which is closely linked to the development and progression of chronic gastritis and gastric cancer. However, current approaches largely depend on pixel-level annotations, which are labor-intensive and struggle to cope with the high resolution, annotation sparsity, and heterogeneous lesion distribution in whole slide images (WSIs). These limitations hinder the scalability and effectiveness of traditional diagnostic models. To address these challenges, we propose SamMIL, a novel semi-supervised adaptive multiple instance learning framework designed for precise HP grading. SamMIL leverages a pseudo-label generation strategy based on attention mechanisms and introduces a curriculum learning-based optimization scheme to adaptively update instance-level encoder parameters. This enables more effective global feature fusion and alleviates the bottlenecks of feature fragmentation. Furthermore, we incorporate a foundation model-based feature extractor to improve representation learning across both labeled and unlabeled WSIs. We evaluate SamMIL on a dataset of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 8 0}$</tex> gastric WSI samples annotated by expert pathologists. The results demonstrate that our method achieves an accuracy of 83.78 % and a macro-AUC of 0.936, outperforming existing state-of-the-art models by 13.51 % in accuracy and 4.4 % in AUC. Notably, SamMIL requires only 10 % of labeled data to reach the performance level of fully supervised approaches, significantly reducing annotation cost. Visualization results further confirm the model's interpretability and ability to focus on clinically relevant regions. In summary, SamMIL offers a powerful and data-efficient solution for automated HP grading in gastric pathology and holds strong potential for generalization to other pathological image classification tasks.