Generative model jointly synthesizes complex-valued brain MRI to enhance pathology detection

A generative AI model jointly synthesizing brain MRI magnitude and phase images produced synthetic data that trained a classifier to detect abnormal tissue with AUROC 0.880, beating real-data training (0.842) on fastMRI, and persisted on an external biopsy-confirmed set.

• The autoencoder preserved phase coherence above 0.997. Real-vs-synthetic classifiers yielded AUROCs of 0.53–0.64 for phase (realistic) and 0.53–0.82 for magnitude (some distribution shift).
• On the training split, image-space coverage tests showed slight to moderate undercoverage depending on acquisition sequence and tissue condition.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.