Researchers have developed a method to synthesize phase maps for MRI scans using conditional score-based diffusion models, addressing the scarcity of raw k-space data for training deep learning reconstruction models. This technique combines synthesized phase maps with existing magnitude-only MR images to create large datasets for training accelerated MRI reconstruction models. The resulting models trained with this synthetic data demonstrated superior performance compared to those trained with naive phase synthesis, GAN-generated phase maps, or ground truth data, showing improved quantitative metrics and reduced hallucinated features. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Enables training of more robust and generalizable accelerated MRI reconstruction models by overcoming limitations in raw k-space data availability.
RANK_REASON Academic paper detailing a novel method for synthesizing MRI data for deep learning model training. [lever_c_demoted from research: ic=1 ai=1.0]