Challenges

Our BrainHack will make available public MR data that will be used to tackle real MR image processing and reconstruction challenges using machine learning. The computational infra-structure for the development of the challenges is supported by the AWS Cloud Credits for Research program. Each team participating of the workshop will have two senior mentors (at least one speaker at the conference) and they will be given 2-GPU cloud computers to develop their solutions. The teams will be formed during the workshop trying to generate multi-disciplinary groups.

We are proposing three different challenges:

1. MR Compressed Sensing Reconstruction

Reducing magnetic resonance (MR) imaging acquisition times can potentially reduce costs and make MR examinations more widely accessible. Compressed sensing (CS) methods aim to reduce acquisition time by reconstructing high quality images that were originally sampled at rates less than specified by the Nyquist-Shannon sampling theorem. Iterative algorithms employing regularization parameters are the standard approaches to reconstructing these ill-posed inverse problems. These approaches, however, have the drawback of being slow, therefore, preventing real time reconstruction. Recent developments have highlighted the potential of deep-learning methods to solve the CS reconstruction latency problem. Deep-learning methods have the advantage of being able to reconstruct the images in a single-pass using a suitably trained network, (i.e., they are fast). Recent studies have shown that the reconstruction quality can surpass the quality of the conventional iterative approaches. The goal of this challenge is to develop a network to reconstruct undersampled MR acquisition data at different undersampling rates.

2. Hippocampus Subfields Segmentation

Volumetric analysis of the hippocampus and its subfields is believed to be a biomarker for depression. Recently, Thyreau et al. published the paper entitled: "Segmentation of the hippocampus by transferring algorithmic knowledge for large cohort processing", which is a public network called hippodeep that provides accurate hippocampal segmentation under 50 seconds on a modern CPU computer. The proposal of this challenge is to develop a fast and robust network to get the subfields segmentation using hippodeep's output as a starting point. We will provide labeled data generated with FreeSurfer for training the networks. FreeSurfer is a robust tool to segment the hippocampus and its subfields, but it takes several hours just to process one subject.

3. Age and Sex Prediction of Normal Subjects

Being able to differentiate sex differences and understanding normal aging is important because the features used to perform these distinctions can be potentially used to detect brain abnormalities. In this challenge, the goal is to design two classifiers/regressors:

a) A classifier able to determine the sex of a person given a brain MR scan;

b) A classifier/regressor capable of predicting the age of a normal subject from a brain MR scan;

For that we will use the Calgary-Campinas brain MR dataset, which is composed of normal subjects with a sex and age distribution illustrated in the histogram below. Most of the subjects are in the 40-60 years of age range.