Revisiting 3D Multi-Modal Medical Image Generation: Model Configurations for Brain MR Image Synthesis

Multi-modal magnetic resonance imaging (MRI) has become a widely used tool for segmenting subregions of brain glioblastomas. However, acquiring a complete set of multi-modal MRI images is constrained by time, cost, and patient movement during prolonged MRI scans. In this paper, we address this issue by exploring various model configurations to develop an optimal algorithm for synthesizing missing modalities. We validated our model using the Brain MR Image Synthesis Benchmark (BraSyn) dataset. Our results show that the proposed method generates high-quality synthetic images that closely match the ground truth of the missing modalities while maintaining the performance of downstream tumor segmentation in this scenario.