NPBDREG: Uncertainty assessment in diffeomorphic brain MRI registration using a non-parametric Bayesian deep-learning...

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Biology

Background:

• Uncertainty quantification in deep neural network (DNN) image registration is vital for clinical applications.
• Current DNN registration methods may yield inaccurate uncertainty estimates, impacting clinical decisions.
• Existing approaches often rely on parametric assumptions for the registration latent space.

Purpose of the Study:

• To introduce NPBDREG, a non-parametric Bayesian framework for robust uncertainty estimation in DNN-based deformable image registration.
• To evaluate NPBDREG's performance against a baseline probabilistic VoxelMorph (PrVXM) model.
• To demonstrate NPBDREG's ability to provide uncertainty estimates correlated with out-of-distribution data.

Main Methods:

• Developed NPBDREG, a fully non-parametric Bayesian framework for uncertainty estimation.
• Combined Adam optimizer with stochastic gradient Langevin dynamics (SGLD) for posterior distribution characterization via sampling.
• Validated on 390 brain MRI pairs from MGH10, CMUC12, ISBR18, and LPBA40 datasets.

Main Results:

• NPBDREG showed a strong correlation between predicted uncertainty and out-of-distribution data (r > 0.95).
• Achieved a 7.3% improvement in registration accuracy (Dice score) and an 18% improvement in registration smoothness.
• Demonstrated superior generalization for noisy data (Dice score of 0.73 vs. 0.69).

Conclusions:

• NPBDREG offers a significant advancement in uncertainty estimation for deformable image registration.
• The non-parametric Bayesian approach enhances reliability for clinical deployment and research.
• NPBDREG provides more accurate and generalizable results compared to existing methods.