MultiAttenGastro: Multi-Dimensional Attention Augmentation for Gastrointestinal Endoscopy Classification

The researchers proposed MultiAttenGastro, a plug-and-play attention framework consisting of parallel 1-D channels, 2-D space, and 3-D context heads. This framework was systematically evaluated across five common GI datasets on eight CNN and Transformer backbone networks (80 runs in total). Experiments showed that MultiAttenGastro improved the performance of the five backbone networks on the Kvasir-Capsule dataset with large domain differences by 6 units (the best macro F1 score reached 98.33%), while it performed negatively on the Kvasir-v2 benchmark for small-domain differences. Five ablation experiments indicated that although the improvement direction was consistent in the strongest case, ConvNeXt-Tiny, statistical tests did not reach the significance level (paired t-test p=0.47; Wilcoxon p=0.63). The Central Keeper Alignment (CKA) analysis linked this pattern to representation redundancy.