YUV-based SVD-VGG hybrid fusion for multimodal MRI-PET image integration

Abstract

Multimodal medical image fusion enhances diagnostic interpretation by integrating anatomical and functional information into a single image. This work proposes an efficient hybrid framework, termed SVD–VGG Hybrid Fusion, unifying Singular Value Decomposition (SVD) for luminance decomposition and a lightweight VGG-based feature extractor for high-frequency enhancement. Synthetic Gaussian noise ( σ 2 = 0.25 ) is added to MRI and Poisson noise to PET images to simulate representative acquisition degradations, while the SVD and VGG-based feature paths strengthen structural detail and functional contrast. Experiments were conducted on a single public brain dataset with image pairs resized to 256 × 256 for fusion and 224 × 224 for feature extraction. Quantitative evaluation using PSNR, SSIM, CC, and perceptual LPIPS indicates that the proposed method achieves consistent structural fidelity, perceptual quality, and color preservation while maintaining sub-second runtime per case. Although evaluated only on brain MRI–PET data and under synthetic noise conditions, the results suggest that the SVD–VGG hybrid design provides a noise-aware and color-preserving fusion strategy suitable for practical multimodal image fusion workflows.