<p>Medical image fusion plays a vital role in enhancing diagnostic accuracy by integrating complementary information from multimodal imaging modalities such as MRI, PET, and SPECT. However, many existing methods are limited by information loss and inadequate structural preservation. To overcome these limitations, we propose a novel fusion framework that integrates spatial-frequency saliency analysis with graph-based optimization to improve fusion quality. The process begins with a preprocessing stage designed to enhance image quality through denoising, contrast adjustment, and edge sharpening. Following this, a multi-level decomposition strategy based on the Iterative Least Squares Smoothing Filter (ILSSF) is employed to extract base and detail components while preserving essential anatomical structures. For the fusion of base components, we introduce the Energy-Balanced Visual Saliency Map (EBVSM), which adaptively combines spatial gradients and frequency features through an energy-weighted scheme. To fuse detail components, we present a Patch-Based Graph Optimization (PBGO) method that models both local and global patch relationships using Laplacian regularization, with optimization performed via the Adam algorithm. This approach ensures a balance between visual saliency and structural coherence, resulting in superior detail preservation and enhanced contrast. Experimental results on two benchmark datasets (PET-MRI and SPECT-MRI) demonstrate that the proposed method achieves competitive performance and achieves superior results compared to nine existing techniques across eight quantitative metrics. Furthermore, the fused images exhibit strong clinical potential by effectively maintaining anatomical clarity and functional contrast.</p>