Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction

Abstract

Accelerating MRI scans is one of the principal outstanding problems in the MRI research community. Towards this goal, we hosted the second fastMRI competition targeted towards reconstructing MR images with subsampled k-space data. We provided participants with data from 7,299 clinical brain scans (de-identified via a HIPAA-compliant procedure by NYU Langone Health), holding back the fully-sampled data from 894 of these scans for challenge evaluation purposes. In contrast to the 2019 challenge, we focused our radiologist evaluations on pathological assessment in brain images. We also debuted a new Transfer track that required participants to submit models evaluated on MRI scanners from outside the training set. We received 19 submissions from eight different groups. Results showed one team scoring best in both SSIM scores and qualitative radiologist evaluations. We also performed analysis on alternative metrics to mitigate the effects of background noise and collected feedback from the participants to inform future challenges. Lastly, we identify common failure modes across the submissions, highlighting areas of need for future research in the MRI reconstruction community.

Fig. 1.

Examples of 4X submissions evaluated by radiologists with slice-level SSIM scores. All methods reasonably reconstructed T2 and FLAIR images. The ATB and Neurospin methods struggled with a susceptibility region, exaggerating the focus of susceptibility and introducing a few false vessels between the susceptibility and the lateral ventricular wall. In other cases, radiologists observed mild smoothing of white matter regions on T1POST images.

Fig. 2.

Examples of 8X submissions evaluated by radiologists with slice-level SSIM scores. At this level of acceleration fine details are smoothed and obscured for all contrasts. On T1POST images, AIRS Medical was relatively more successful than ATB and Neurospin in showing fine details of the mass, particularly in its periphery. Noticeable on the FLAIR images are horizontal “banding” effects that arise from how neural networks interact with anisotropic sampling patterns.

Fig. 3.

Examples of 4X Transfer submissions evaluated by radiologists with slice-level SSIM scores. The T1POST and T2 examples are from GE scanners, whereas the FLAIR example is from a Philips scanner. All methods introduced blurring to the images. Several methods had trouble adapting to the GE data while performing relatively well on the Philips data, as seen in the form of aliasing artifacts in one of the T1POST images.

Fig. 4.

Summary of SSIM values across contestants. (top) Model perfomance for teams submitting to the main 4X and 8X Siemens competition tracks. (bottom) Model performance for teams submitting to the Transfer track (combination of GE and Philips data). The “AVG” model score for the Transfer track was a simple average across all volumes in the Transfer track.

Fig. 5.

Scatter plot of mean radiologist rank across cases. The horizontal axis has a separate tick for each case evaluated by the radiologist cohort. The scatter plot markers indicate whether that method was from the team with the highest, middle, or lowest SSIM scores. We generally observed radiologists awarding the best ranks to models with the best SSIM score.

Fig. 6.

Examples of reconstruction hallucinations among challenge submissions with SSIM scores over residual plots (residuals magnified by 5). (top) A 4X submission from Neurospin generated a false vessel, possibly related to susceptibilities introduced by surgical staples. (middle) An 8X submission from ATB introduced a linear bright signal mimicking a cleft of cerebrospinal fluid, as well as blurring of the boundaries of the extra-axial mass. (bottom) A submission from ResoNNance introduced a false sulcus or prominent vessel.