MULTI-MODAL MEDICAL IMAGE FUSION LEVERAGING TRANSFORMER-BASED CROSS-ATTENTION NETWORKS IN CLINICAL APPLICATIONS

Multi-modal medical imaging provides complementary anatomical and functional information essential for accurate diagnosis and treatment planning. However, conventional fusion techniques often fail to retain fine-grained structural and functional details, leading to suboptimal diagnostic quality. Integrating diverse modalities such as MRI, CT, and PET requires an approach capable of capturing complex inter-modal relationships while preserving both spatial structures and functional intensities. Existing convolution-based methods are limited in modeling long-range dependencies, resulting in loss of critical clinical information. This study proposed a transformer-based cross-attention network for multi-modal medical image fusion. Initially, input images underwent preprocessing including normalization, resizing, and noise reduction. Feature representations were extracted via parallel encoder streams for each modality. A cross-attention mechanism then enabled the network to learn modality-specific relationships, highlighting complementary regions while suppressing redundant information. Finally, a fusion module combined the attended features into a single output, followed by reconstruction to generate a high-fidelity fused image. Performance was evaluated using standard metrics including structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and edge preservation (EP) across multiple datasets. The proposed method consistently outperformed conventional and deep learning- based fusion approaches. Quantitative evaluation showed improvements in structural similarity index (SSIM: 0.94), peak signal- to-noise ratio (PSNR: 39.1 dB), edge preservation (EP: 0.89), mutual information (MI: 1.38), and standard deviation (SD: 47.1) on the ADNI dataset. Qualitative analysis demonstrated enhanced visualization of anatomical and functional features, supporting its potential clinical applicability.