Transformation of Image Fusion Technologies with Generative AI: Architectures, Metrics, and Clinical Impact

Fusing the complementary information in multiple imaging modalities is a fundamental problem in remote sensing, surveillance and medical imaging. Classical fusion approaches based on multi-resolution transforms, sparse representations and optimization provide interpretability but are not well suited to heterogeneous, high-dimensional data and application-specific requirements. Deep learning systematically provides data-driven feature extraction and end-to-end fusion, but may lead to deterministic results without perceptual realism and uncertainty prediction. Recently emerged Generative AI techniques, e.g., GANs, VAEs, attention-based Transformers now make it possible of probabilistic modeling and perceptual discrimination also long-range multimodal correlation. This review of image fusion from 2010–2025 classifies significant architecture families, reports on objective, perceptual, and clinical evaluation metrics, and explores the translational aspects associated with CT–MRI and PET–MRI fusion. Our review discusses strengths and limitations of paradigms, considering representative clinical case studies, as well as key research gaps related to interpretability, robust evaluation, clinical validation, and effective deployment, providing further research in advanced multimodal fusion.