Probabilistic machine learning models for resolution enhancement of diffusion magnetic resonance imaging.

Diffusion magnetic resonance imaging (dMRI) measures the direction and rate at which water molecules diffuse in biological tissues, e.g. brain tissue. Scalar metrics derived from dMRI are used to assess changes in the brain tissue due to different diseases. Also, as diffusion occurs preferentially along the direction of white matter fiber bundles in the brain, dMRI can be used to map structural connectivity, providing unique information with many potential applications, including predicting brain tumour prognosis and for neurosurgical planning.
Due to the limited scan time typically available in the clinical setting, a result of both time/cost and patient comfort considerations, clinical dMRI protocols rarely allow for dMRI acquisitions with sufficient spatial resolution to provide a detailed description of the complex microanatomy of the brain tissue.

Obtaining such a level of detail requires expensive scanners, with enhanced gradient hardware and long acquisition times, both of which are beyond the resources typically available in routine clinical practice.

In this project, our objective is to develop novel machine learning models for increasing the spatial resolution of diffusion images. We are particularly interested on developing new approaches for deep probabilistic models on non-Euclidean manifolds that can be applied to improve resolution on these type of images.

A successful outcome will achieve the dual purpose of pushing the state-of-the-art in resolution, or reducing scan time for a given resolution.