MAAFusion: A Multimodal Medical Image Fusion Network Via Arbitrary Kernel Convolution And Attention Mechanism

This study introduces an advanced multimodal medical image fusion technique that overcomes the limitations of traditional CNN-based approaches and attention mechanisms, which often result in computational redundancy and overlook critical channel-specific features. Our innovative dual-branch attention fusion network utilizes Arbitrary Kernel Convolution (AKConv) for efficient feature aggregation and minimizes unnecessary computations. Additionally, the Channel Prior Convolutional Attention (CPCA) module improves the discrimination of salient channel features, enhancing overall feature extraction. The Lightweight Convolutional Module (LCM) further enhances the feature extraction process while reducing the computational load by integrating dynamic gating mechanisms with deep separable convolution. Experimental results confirm the superior performance of our framework in both quantitative and qualitative assessments compared to existing fusion methods.