Patch Attention: A Lightweight, Plug-and-Play Input-Level Attention Fusing Static Spatial Priors and Dynamic Saliency
Shukai Ding, Nirattaya Khamsemanan, Cholwich Nattee, Piya Limcharoen · IEEE Access · 2026
Accurate classification of small targets in high-resolution medical images remains challenging because salient signals are diluted by extensive background and existing methods have structural limitations. Manual Region-of-Interest (ROI) selection is not scalable. Standard patch-based strategies are computationally expensive and may disrupt global context. To address these issues, we introduce the Patch Attention (PA) module, a lightweight, plug-and-play component applied at the input level to effectively leverage spatial priors. PA adopts a dual-branch design: a static branch that encodes dataset-level anatomical priors and a dynamic branch that adapts to instance-specific content, and fuses their outputs to produce an input-level attention map.We evaluate PA on four medical imaging datasets and TinyImageNet across five convolutional backbones. The results show consistent and substantial gains in classification performance with negligible computational overhead. Compared with the baseline, the F1 score improved by up to 8.55%. Moreover, PA complements rather than competes with feature-level attention mechanisms (such as SENet and CBAM), and combining the two yields additional improvements. Our sensitivity analysis indicates that PA maintains stable performance across different datasets and backbones, with only limited need for hyperparameter tuning in practice. In summary, these findings position PA as an effective and interpretable paradigm for proactive, input-level attention, providing a practical and efficient means to enhance medical image analysis.