APGFusion : Adaptive PoolFormer and CNN medical image fusion network based on convolutional gated linear units

Medical image fusion combines important information from images taken by different sensors into one image. This process produces clearer fused images and eliminates unnecessary details, thereby enhancing the efficiency of clinical analysis and improving diagnostic accuracy. However, existing multimodal medical image fusion models often lose global and local features when handling complex fusion strategies and fail to ensure visual naturalness in the fused results. Moreover, most current models lack sufficient exploration of multi-channel data, reducing the clinical reliability of fusion outcomes. To solve these issues, we designed a novel deep model named APGFusion. The model uses an adaptive dual-branch network that effectively fuses global and local features with high quality. First, we apply the adaptive feature fusion (AFF) module in the field of medical image fusion for the first time. This module effectively extracts local features and global representations, resulting in a richer semantic representation of the fused images. Second, we propose a TransGLU module that combines the convolutional gated linear units module from TransNeXt for shared feature extraction and reconstruction, implementing an attention mechanism for channel mixing to selectively choose important features and dynamically adjust the feature fusion strategy. TransGLU allows for diverse medical image characteristics and outcomes, addressing the limitations of traditional medical image fusion methods in information integration and feature extraction.The proposed model adaptively adjusts feature weights across modalities, ensuring superior visual and structural fusion performance. Our proposed model demonstrated significant improvements on the Harvard dataset. By leveraging convolutional gated linear units, PoolFormer modules, and adaptive feature fusion, local and global feature extraction are enhanced, resulting in a 4.12% increase in SF, 1.38% improvement in SSIM, 3.662% boost in SD, and a 0.547% rise in Qab/f. Our code is available athttps://github.com/HXY13/APGFusion. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.