Generative modeling of complex-valued brain MRI data.
Authors
Affiliations (5)
Affiliations (5)
- Department of Physics, TU Dortmund University, Otto-Hahn-Straße 4a, Dortmund, 44227, Germany.
- University Hospital Essen Institute for Artificial Intelligence in Medicine, Girardetstraße 2, Essen, NRW, 45131 , Germany.
- Division of Radiology, German Cancer Research Center, Im Neuenheimer Feld 280, Heidelberg, BW, 69120, Germany.
- Department of Physics, TU Dortmund University, Otto-Hahn-Straße 4a, Dortmund, Nordrhein-Westfalen, 44227, Germany.
- Department of Physics, TU Dortmund University, Otto-Hahn-Straße 4a, Dortmund, NRW, 44227, Germany.
Abstract
Standard Magnetic Resonance Imaging (MRI) reconstruction pipelines discard phase information captured during acquisition, despite its sensitivity to tissue properties that are not represented by magnitude alone. The aim of this study is to develop a generative framework that jointly models magnitude and phase information in complex-valued brain MRI data and enables the conditional generation of MRI samples representing normal and abnormal brain tissue. The proposed generative framework combines a conditional variational autoencoder, which compresses complex-valued MRI scans into latent representations while preserving phase coherence, with a flow-matching-based generative model. Synthetic sample quality is assessed using a real-versus-synthetic classifier, image-space coverage tests, and downstream classifiers trained on synthetic data for abnormal tissue detection. The autoencoder preserves phase coherence above 0.997. Real-vs-synthetic classifiers yielded low AUROCs between 0.53 and 0.64 for generated phase images supporting realistic phase modeling. Magnitude was moderately distinguishable, with AUROCs between 0.53 and 0.82, suggesting appreciable overlap between real and synthetic sample distributions while also indicating a distribution shift for several sequences. Image-space coverage was broad. On the held-out test split, no consistent directional difference from the matched real-to-real references was detected in any evaluated setting, while the training split showed slight to moderate undercoverage depending on acquisition sequence and tissue condition. In downstream normal-versus-abnormal classification, classifiers trained entirely on synthetic data achieve an AUROC of 0.880, surpassing the real-data baseline of 0.842 on the fastMRI dataset. This advantage persists on an independent external test set with biopsy-confirmed labels. The proposed framework demonstrates the feasibility of jointly modeling magnitude and phase information for normal and abnormal complex-valued brain MRI data. Beyond synthetic data generation, the framework enables future investigation of how magnitude and phase jointly encode pathology-specific features once sufficiently large raw MRI datasets with fine-grained, pathology-confirmed labels become available.