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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Accelerating magnetic resonance imaging (MRI) scan times is a critical challenge in medical imaging.
• Subsampled k-space data acquisition offers a potential solution but requires advanced reconstruction techniques.

Purpose of the Study:

• To advance MRI image reconstruction from undersampled k-space data through a competition format.
• To evaluate reconstruction algorithms based on image quality and pathological assessment.

Main Methods:

• The second fastMRI challenge provided 7,299 de-identified brain MRI scans for training and 894 for evaluation.
• A new 'Transfer' track assessed model performance on external MRI scanner data.
• Submissions were evaluated using Structural Similarity Index (SSIM) and radiologist assessments.

Main Results:

• One team achieved top performance in both SSIM scores and qualitative radiologist evaluations.
• Analysis explored alternative metrics for noise mitigation and gathered participant feedback.
• Common failure modes in reconstruction algorithms were identified.

Conclusions:

• The fastMRI competition successfully spurred innovation in accelerated MRI reconstruction.
• Identified failure modes highlight key areas for future research in developing robust MRI reconstruction models.