Multi-shell dMRI Estimation from Single-Shell Data via Deep Learning

Document Type

Conference Proceeding

Publication Date

1-1-2023

Abstract

Diffusion magnetic resonance imaging (dMRI) data acquired with multiple diffusion gradient directions and multiple b-values (“multi-shell” data) enables compartmental modeling of brain tissues as well as enhanced estimation of white matter fiber orientations via the orientation distribution function (ODF). However, multi-shell dMRI acquisitions are time consuming, expensive and difficult in certain clinical populations. We present a method to estimate high b-value volumes from low b-value volumes via deep learning. A 3-dimensional U-NET architecture is trained from multi-shell dMRI training data to synthesize a high b-value volume from a diffusion gradient direction, given a low b-value volume from that same gradient direction as input. We show that our method accurately synthesizes high b-value (2000 and 3000 s/mm2) volumes from low b-value (1000 s/mm2) input volumes when applied to simulated and real, public-domain human dMRI data. We also show that synthesized multi-shell dMRI data gives rise to accurate compartmental model parameters and ODFs. Finally we demonstrate good out-of-training-sample generalization to previously-unseen diffusion gradient directions and different MRI scanners. Deep learning based estimation of high b-value dMRI volumes has the potential to combine with pulse sequence accelerations to enhance time efficiency of multi-shell dMRI protocols.

This document is currently not available here.

Share

COinS