Description¶
Update:
Thanks for your interest in the RealNoiseMRI challenge. The challenge is now closed and the data has since been anonymized and is publically available!
The data is distributed on Openneuro: https://openneuro.org/datasets/ds004332 and while there's a description there, there is also a preprint on PsyArxiV with lots of details: https://psyarxiv.com/vzh4g
The scripts to reproduce the results are on our MoCo GitHub:
https://github.com/melanieganz/MoCoProject
There are still minor details missing, but check it out and reach out to me personally under (mleanie.ganz@nru.dk) if you have any questions!
Information about the challenges
The challenge was run in two rounds: The first round was associated with MICCAI 2021 and the second round was associated with the MedNeurIPS 2021 workshop.
In recent years, there is a growing focus on the application of fast magnetic resonance imaging (MRI) based on prior knowledge. In the 1980s and 2000s the community used either purely mathematical models such as the partial Fourier transform or solutions derived through advanced engineering such as parallel imaging to speed up MRI acquisition. Since the mid-2000’s, compressed sensing and artificial intelligence have been employed to speed up MRI acquisition. These newer methods rely on under sampling the data acquired in Fourier (aka k-) space and then interpolating or augmenting k-space data based on training data content. One of the underlying problems for the development of fast imaging techniques, that just as in e.g. [1], it is common to use a fully sampled image as ground truth and then under sample it in k-space in order to simulate under sampled data. The problem with this approach is that in cases were the under sampled data is corrupted, through e.g. motion, this under sampling is unrealistic. This could easily happen in a clinical setting. Hence, the robustness of the new fast MRI reconstruction methods is often not evaluated in a realistic setting.
We propose a challenge that aims at checking the robustness of fast MRI reconstruction methods by providing image and k-space datasets that consist of ground truth as well as motion degraded scans. We aim to closely mimic the ongoing challenge organized by Facebook at NeurIPS: https://fastmri.org/, but provide motion degraded data to realistically test robustness. While the fastMRI challenge provides participants with a large dataset, the way the under-sampled k-space data is constructed is artificial and not realistic, since patient motion corrupts the k-space differently than removing certain lines of k-space.
Our setup would be the following: For 25 healthy volunteers, we are collecting brain imaging data with standard cerebral protocols including axial, 2D-encoded T1- and T2-weighted sequences (Short-TI-Inversion Recovery and Turbo Spin Echo). In our study, we are assessing the utility of motion correction sequences and hence, we acquire data where the participants lie still and where they move in a pre-described fashion. We can therefore provide challenge participants with the ground truth (still) as well as degraded (due to motion) T1- and T2-weighted images and the corresponding k-space data. The participants are expected to train their models on the challenge data provided by the fastMRI challenge.
First round: We will release a set of 5 k-space data and corresponding ground truth images for optimization and validation of the pre-trained models using fastMRI data. For the remaining 20 volunteers, we will withhold the ground truth images and only distribute the degraded k-space data as a test set.
Second round: We will release a set of 10 k-space data and corresponding ground truth k-space data and ground truth images for optimization and validation of the pre-trained models using fastMRI data. For the remaining 20 volunteers, we will withhold the ground truth images and only distribute the degraded k-space data as a test set. For both the validation and test data set motion data will be provided.
The goal of our challenge for participants would then be to send in estimates of the ground truth images. We can evaluate the contributions based on the available ground truth data using the structural similarity measure (SSIM) as quantitative performance metric together with a visual evaluation through radiologists.
References
[1] Duan J, Schlemper J, Qin C, Ouyang C, Bai W, Biffi C, Bello G, Statton B, O’Regan DP, Rueckert D. VS-Net: Variable splitting network for accelerated parallel MRI reconstruction. In International Conference on Medical Image Computing and Computer-Assisted Intervention 2019 Oct 13 (pp. 713-722). Springer, Cham.