Uncertainty mapping and probabilistic tractography using Simulation-based Inference in diffusion MRI: A comparison with classical Bayes

Manzano-Patron, J.P.; Deistler, Michael; Schröder, Cornelius; Kypraios, Theodore; Gonçalves, Pedro J.; Macke, Jakob H.; Sotiropoulos, Stamatios N.

Abstract

Simulation-Based Inference (SBI) has recently emerged as a powerful framework for Bayesian inference: Neural networks are trained on simulations from a forward model, and learn to rapidly estimate posterior distributions. We here present an SBI framework for parametric spherical deconvolution of diffusion MRI data of the brain. We demonstrate its utility for estimating white matter fibre orientations, mapping uncertainty of voxel-based estimates and performing probabilistic tractography by spatially propagating fibre orientation uncertainty. We conduct an extensive comparison against established Bayesian methods based on Markov-Chain Monte-Carlo (MCMC) and find that: a) in-silico training can lead to calibrated SBI networks with accurate parameter estimates and uncertainty mapping for both single- and multi-shell diffusion MRI, b) SBI allows amortised inference of the posterior distribution of model parameters given unseen observations, which is orders of magnitude faster than MCMC, c) SBI-based tractography yields reconstructions that have a high level of agreement with their MCMC-based counterparts, equal to or higher than scan-rescan reproducibility of estimates. We further demonstrate how SBI design considerations (such as dealing with noise, defining priors and handling model selection) can affect performance, allowing us to identify optimal practices. Taken together, our results show that SBI provides a powerful alternative to classical Bayesian inference approaches for fast and accurate model estimation and uncertainty mapping in MRI.

Citation

Manzano-Patron, J., Deistler, M., Schröder, C., Kypraios, T., Gonçalves, P. J., Macke, J. H., & Sotiropoulos, S. N. (in press). Uncertainty mapping and probabilistic tractography using Simulation-based Inference in diffusion MRI: A comparison with classical Bayes. Medical Image Analysis,

• Journal Article Type: Article
• Acceptance Date: Apr 1, 2025
• Deposit Date: Apr 2, 2025
• Journal: Medical Image Analysis
• Print ISSN: 1361-8415
• Electronic ISSN: 1361-8423
• Publisher: Elsevier
• Peer Reviewed: Peer Reviewed
• Keywords: Bayesian inference, dMRI, Artificial Neural Networks, Fibre Orientations, Markov-Chain Monte-Carlo, Ball & Sticks, Parametric Deconvolution
• Public URL: https://nottingham-repository.worktribe.com/output/47278371