GRAFNeT: Guided Recursive Attention Fusion Network for Multimodal Medical Imaging

This study introduces GRAFNeT (Guided Recursive Attention Fusion Network for Multimodal Medical Imaging), a novel multi-stage image fusion framework designed to achieve high-quality multimodal integration with minimal computational overhead. The framework employs CNN feature extraction and attention modules, followed by an iterative refinement of fusion parameters at inference time. The methodology begins with adaptive weight normalization to dynamically balance modality contributions and redundancy reduction. Hybrid feature extraction integrates Sobel-based edge detection with ResNet-derived deep features to enhance anatomical structure representation. Present work with guided recursive attention, through metric guided optimization iteratively refine the fused features. Thus, emphasizing diagnostically significant regions. The consistency across modalities is ensured using multi-scale fusion with scale-specific refinement. Moreover, diagnostically reliable reconstruction is achieved using generalized metric integration involving SSIM, gradient preservation, and patch-wise alignment. The qualitative and quantitative evaluation were performed on three benchmark datasets, which shows GRAFNeT's superior ability to preserve spatial fidelity and capture intricate multimodal details, proving it as a robust and clinically relevant fusion approach.