Bayesian AI uncovers unreliable parameters in gray matter MRI models
Deep learning framework reveals hidden uncertainty in NEXI and SANDIX brain imaging models
Diffusion MRI (dMRI) biophysical models promise to reveal gray matter microstructure, but parameter reliability remains understudied, especially when water exchange is included. In a new study, researchers led by Maëliss Jallais used μGUIDE—a Bayesian inference framework powered by deep learning—to quantify uncertainty and detect degeneracies in two recent models, NEXI and SANDIX. Using both simulated and in vivo data with standard acquisition protocols, the team systematically evaluated accuracy, precision, and modeling degeneracies.
The results highlight a stark split: parameters like extra-cellular diffusivity and neurite signal fraction were robustly estimated across conditions. In contrast, exchange time and soma radius were often associated with high uncertainty and estimation bias, particularly under realistic noise and reduced acquisition protocols. Comparing with traditional non-linear least squares fitting, the Bayesian approach allowed the team to flag and filter out unreliable estimates. The study concludes that reporting uncertainty and accounting for degeneracies is critical for reproducible, biologically interpretable dMRI analysis. The findings advocate for integrating probabilistic fitting into standard neuroimaging pipelines to avoid misleading conclusions in clinical and research applications.
- μGUIDE Bayesian framework applied to NEXI and SANDIX gray matter dMRI models reveals parameter-specific uncertainty.
- Extra-cellular diffusivity and neurite signal fraction are robust; exchange time and soma radius show high bias and uncertainty.
- Uncertainty-aware fitting outperforms traditional nonlinear least squares by enabling filtering of unreliable estimates.
Why It Matters
Improves reproducibility in brain microstructure imaging by flagging unreliable model parameters in clinical MRI.