A methodological assessment of information complementarity between MRI and GAN-based synthetic PET for multimodal medical image analysis

Multimodal medical imaging aims to enhance analysis by combining complementary anatomical and functional information. However, access to functional modalities such as positron emission tomography (PET) is limited in many clinical settings, prompting efforts to synthesize PET-like images from magnetic resonance imaging (MRI) using generative models. Despite growing interest, the degree to which MRI-derived synthetic PET provides independent and complementary information remains unclear. In this study, we systematically evaluate the information complementarity between MRI and GAN-based synthetic PET for glioma classification and MGMT promoter methylation prediction using publicly available datasets. MRI-only and MRI plus synthetic PET pipelines were compared across convolutional and transformer-based architectures using normalized performance metrics, statistical testing, and feature-level analysis. Across all models and tasks, synthetic PET produced performance comparable to MRI alone, with no statistically significant improvement. Feature attribution and error analyses indicated substantial overlap between MRI and synthetic PET representations. This limited contribution likely reflects the fact that synthetic PET is derived from MRI and cannot capture true tracer-driven metabolic or molecular information. These findings highlight the need to validate information complementarity before integrating synthesized modalities into clinical workflows.