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.

Robust Compressed Sensing MRI with Deep Generative Priors

“…We first highlight that an advantage of the proposed approach is the invariance to the sampling scheme during training. In contrast, this is a design choice that must be made for supervised end-to-end methods, which here were trained on equispaced, vertical sampling masks, following the fastMRI 2020 challenge guidelines [91,62]. As our results show, this affords us a significant degree of robustness across a wide distribution of sampling masks during inference.…”

SKM-TEA: A Dataset for Accelerated MRI Reconstruction with Dense Image Labels for Quantitative Clinical Evaluation

“…fastMRI is a large repository of k-space data acquired for ≈1600 2D knee and ≈7000 2D brain MRI scans of varying contrasts, with each imaging volume averaging 30-40 slices per acquisition [25]. The extent of images made available through fastMRI and the global challenges has catalyzed ML research for MRI reconstruction [24,34]. However, as described in the 2019 knee MRI reconstruction challenge, even for some of the challenge winning methods, conventional IQAs did not accurately convey true clinical imaging quality and also clearly obscured pathology [24].…”

Unsupervised Deep Learning Methods for Biological Image Reconstruction and Enhancement

“…( 5b) is implemented by conventional methods that explicitly use H(•), such as gradient descent with the only learnable parameter being the gradient step size. These physics-guided methods have recently become the state-of-the-art in a number of image reconstruction problems, including large-scale medical imaging reconstruction challenges [24],…”

“…Unlike in medical imaging applications, where there is significant interest in the release and use of publicly available raw imaging data [24], there is less momentum for such databases in biological imaging applications. In fact, a number of studies discussed in this article utilized imaging datasets that were released for other purposes, such as segmentation or tracking challenges [34].…”