Universal Model for Cardiac MRI Reconstruction Challenge

Cardiac magnetic resonance imaging (CMR) has emerged as a crucial imaging technique for diagnosing cardiac diseases, thanks to its excellent soft tissue contrast and non-invasive nature. However, a notable limitation of MRI is its slow imaging speed, which causes patient discomfort and introduces motion artifacts into the images.

To accelerate image acquisition, CMR image reconstruction (recovering high-quality clinical interpretable images from highly under-sampled k-space data) has gained significant attention in recent years. Particularly, AI-based image reconstruction algorithms have shown great potential in improving imaging performance by utilizing highly under-sampled data. However, the field of CMR reconstruction lacks publicly available, standardized, and high-quality datasets for the development and assessment for AI-based CMR reconstruction.

CMR imaging has the nature of multi-contrast, e.g., cardiac cine, mapping, tagging, phase-contrast, and dark-blood imaging. It also includes imaging of different anatomical views such as long-axis (2-chamber, 3-chamber, and 4-chamber), short-axis, outflow tract, and aortic (cross-sectional and sagittal views). Additionally, accelerated imaging trajectories, including uniformly undersampling and variable-density sampling, are employed. Unfortunately, conventional CNN-based reconstruction models often require training and deployment for each specific imaging scenario (imaging sequence, view, and device vendor), limiting their clinical application in the real world.

The objective of establishing the 'CMRxUniversalRecon' challenge is to provide a benchmark that enables the broader research community to contribute to the important work of accelerated CMR imaging with universal approaches that allow more diverse applications and better performance in real-world deployment in various environments. To achieve this goal, in the first run of the 'CMRxRecon' challenge (MICCAI 2023) we provided training and test data from a total of 200 subjects and the technical infrastructure as well as a baseline model for CMR reconstruction on muti-contrast imaging. The results of 'CMRxRecon' 2023 demonstrated the feasibility of highly sub-sampled k-space reconstruction on dedicated pre-trained models. 

In this second run of the CMR reconstruction challenge we aim to make an important step towards clinical implementation by extending the challenge scope in two directions: 

1. trustworthy reconstruction on multi-contrast CMR imaging using a universal pre-trained reconstruction model;
2. robust reconstruction with diverse and even unseen k-space trajectory and various acceleration factors using a universal model.