Edge-guided Multi-scale Attention Fusion Network for gastrointestinal tumor image classification

Automated classification of gastrointestinal tumor images is a pivotal technology in computer-aided diagnostic systems. However, medical images are often affected by high-frequency noise due to limitations of acquisition devices and intrinsic tissue characteristics. In addition, lesion regions frequently exhibit blurred edges and low contrast, posing challenges for accurate extraction of discriminative features. To address these challenges, we propose a novel Edge-guided Multi-scale Attention Fusion Network (EdgeMAF-Net) for gastrointestinal tumor image classification. Specifically, we introduce a cross-stage partial fusion module that dynamically allocates features to both CNN and Transformer branches, enabling simultaneous modeling of local details and global context. This is complemented by a high-frequency attenuation and noise suppression mechanism, as well as a multi-scale edge attention calibration module, which integrates a three-stage enhancement strategy to capture features at different scales and delineate blurred boundaries. The module leverages a feature enhancement attention block to weight multi-source features, combined with a multi-scale edge enhancement block employing multi-scale pooling and edge extraction, and an adaptive gated fusion block to dynamically adjust feature fusion. EdgeMAF-Net outperforms existing methods in terms of accuracy, sensitivity, and boundary localization on the Chaoyang, Kvasir, and GasHisSDB datasets. Our code is available at https://github.com/Bambi-lab/EdgeMAF-Net.