SGAFormer: Score-Guided Attention Transformer for Gastric Section Classification

Gastric section classification serves as an important preprocessing step for downstream endoscopic image analysis. However, endoscopic images of different gastric sections are visually similar, making it challenging to identify appropriate gastric sections and leading to a fine-grained classification problem. To address this challenge, we propose a score-guided attention transformer (SGAFormer), a novel architecture comprising a score-guided attention encoder backbone, a fine-grained-aware token selection module, and a pair-wise feature enhancement module. Specifically, the score-guided attention encoder backbone captures both local textures and global context while explicitly guiding the network to focus on informative regions. To distinguish subtle mucosal differences across visually similar sections and extract the most discriminative fine-grained features, the fine-grained-aware token selection module dynamically filters and aggregates multi-scale patch tokens. Furthermore, the pair-wise feature enhancement module enforces cross-image feature interaction, helping the model extract robust, diverse representations from different gastric sections. The experimental results demonstrate the superiority of the proposed method over competing methods on both public and private gastric section classification datasets.