Highly accelerated, model-free diffusion tensor MRI reconstruction using neural networks | JoVE Visualize

Area of Science:

Medical Imaging
Neuroimaging
Artificial Intelligence in Medicine

Background:

Diffusion Tensor Imaging (DTI) is crucial for neuroimaging, but long scan times limit its clinical utility.
Accelerated DTI acquisition techniques are needed to improve efficiency without compromising image quality.
Reconstructing high-quality DTI data from highly undersampled scans presents a significant technical challenge.

Purpose of the Study:

To develop and validate a novel neural network, termed DiffNet, for accurate DTI reconstruction from highly accelerated scans.
To assess the performance of DiffNet in reconstructing key DTI metrics, fractional anisotropy (FA) and mean diffusivity (MD).
To evaluate DiffNet's efficacy in tumor delineation for glioblastoma multiforme (GBM) patients.

Main Methods:

A retrospective study utilizing DTI data from healthy volunteers (N=10) and GBM patients (N=12) acquired between 2013-2018.
Training a neural network (DiffNet) to reconstruct FA and MD maps from sparse DTI data (3-20 directions).
Comparing DiffNet reconstructions against conventional methods in terms of accuracy, precision, and tumor delineation performance using Area Under the Receiver Operating Characteristic Curve (AUC).

Main Results:

DiffNet demonstrated superior accuracy and precision in FA reconstructions compared to conventional methods across all acceleration factors.
DiffNet enabled reliable reconstruction using only three diffusion-encoding directions, showing significantly lower bias than conventional methods with six directions.
DiffNet achieved comparable or superior tumor delineation performance for GBM, especially FA-based delineation, with reduced scan times.

Conclusions:

DiffNet significantly enhances the accuracy, precision, and clinical utility of DTI reconstruction from accelerated scans.
The neural network facilitates robust DTI analysis with a minimal number of diffusion-encoding directions, addressing current acquisition limitations.
DiffNet shows promise for improving diagnostic capabilities in neuro-oncology, particularly for GBM, by enabling faster and more accurate imaging.