Super-resolution of Multi Dimensional Diffusion MRI data

This is the challenge design document for the "Super-resolution of Multi Dimensional Diffusion MRI data", accepted for MICCAI 2020.

Magnetic Resonance Imaging (MRI) is a fundamental asset for clinical assessment and diagnosis in modern healthcare. A key component to its success is the possibility of non-invasively performing quantitative measurements of physical properties of the living tissue (e.g. the brain) such as the diffusion coefficient and the relaxation times T1 and T2. However, this typically requires lengthy acquisitions to collect high resolution images at different contrasts: e.g. different echo-times (TE) to estimate T2, different inversion-times (TI) to estimate T1, or different b-values to estimate diffusivity. At the same time, in order to guarantee a sufficient signal-to-noise ratio (SNR) the image resolution is often sacrificed. A straightforward way of achieving high resolution and high SNR consists of performing repeated acquisitions and averaging them. However, this would lead to a too long acquisition time which, for MUDI data, is already prohibitive for clinical applications.

In order to mitigate the need of trading off between SNR and resolution, as well as keep the acquisition time unaltered, one possibility is to acquire MRI images at low resolution, e.g. thick slices (2.5x2.5x5 mm3). In this setup, each image, e.g. an axial slice, will share information with that immediately above/below. Such information will then be disentangled with a super-resolution algorithm to obtain the desired high resolution dataset. Similar strategies are currently being used in fetal and neonatal imaging, as they show the additional advantages of being more robust to motion.

The Super-MUDI challenge addresses the critical problem of guaranteeing high SNR and resolution, while keeping the scan time from increasing, by super-resolving MUDI images.

MUDI data: the dataset consists of 1344 volumes comprising 106 unique diffusion gradient directions uniformly spread over four b-shells (500, 1000, 2000, 3000 s/mm2), three TEs, and 28 TIs. These were acquired from five healthy human volunteers (3 f, 2 m, age = 19-46 years), after informed consent was obtained (REC 12/LO/1247), on a clinical 3T Philips Achieva scanner (Best, Netherlands) with a 32-channel adult head coil. Single-shot PGSE EPI with the modifications proposed recently to include relaxometry was employed. Other parameters are TR=7.5s, Resolution=2.5mm isotropic, FOV=220x230x140mm3, SENSE=1.9, halfscan=0.7, Multiband factor 2, TA=52min (including preparation time).

The challenge is comprised of two tasks (see individual task description for further details). Only submissions that attempt all the tasks will be considered for the evaluation. The winner of the challenge will be the submission with the overall best score (lower value) across the two tasks, computed as the sum of the scores in each task (see individual task description for further details about the metrics used for evaluation). Each task explores a different MRI acquisition strategy, leading to two different types of images to super-resolve. Task 1 aims at super-resolving data having high in-plane resolution but thick (axial) slices, whereas Task 2 data having isotropically lower resolution.

Therefore, the outcome of the challenge will be two-fold: 1) evaluating how reliable and stable is a superresolution method; 2) which combination of sub-sampling strategy and super-resolution method is the best alternative. The challenge’s outcome will provide guidelines on how to obtain MR images having high SNR and resolution, with no additional acquisition time prolongation, by adopting a subsampling strategy in combination with superresolution methods.