Towards a Universal Learning-Based Framework for Accelerated CMR

Presenting Author(s)

• 

  Fanwen Wang, BSc

  PhD student
  Fudan University
  Shanghai, Shanghai, China (People's Republic)

Primary Author(s)

• 

  Chengyan Wang, PhD

  Associate Professor
  Fudan University
  Shanghai, Shanghai, China (People's Republic)

Co-Author(s)

• ZW

  Zi Wang, PhD

  Postdoc Research Associate
  Imperial College London, United Kingdom
• YL

  Yan Li, MSc

  Shanghai, China
  Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine
  Shanghai, Shanghai, China (People's Republic)
• JL

  Jun Lyu, PhD

  Boston, United States
  Harvard Medical School
  Boston, Massachusetts, United States
• 

  Chen Qin, PhD

  Lecturer (Assistant Professor)
  Imperial College London, United Kingdom
• YZ

  Yajing Zhang, PhD

  Suzhou
  GE Healthcare
  Suzhou, Jiangsu, China (People's Republic)
• 

  Hao Li, PhD

  Assistant Professor
  Fudan University
  Shanghai, Shanghai, China (People's Republic)
• XH

  Xihong Hu, MD, PhD

  Chief Physician
  Children’s Hospital, Fudan University, China (People's Republic)
• LW

  Lianming Wu, MD

  Professor
  Ren Ji Hospital, China (People's Republic)
• XQ

  Xiaobo Qu, PhD

  Professor
  Xiamen University-Neusoft Medical Magnetic Resonance Imaging Joint Research and Development Center, China (People's Republic)
• GY

  Guang Yang, PhD

  Senior Lecturer in Bioengineering
  Imperial College London, United Kingdom

Background: Deep learning (DL) reconstruction has demonstrated state-of-the-art performance for specific CMR sequences, but robust generalization across modalities and sampling schemes is less explored. In clinical practice, a complete CMR examination comprises multiple sequences. Relying on separately tuned models for each sequence complicates deployment and disrupts real-time workflow integration. Progress is further limited by the lack of large-scale, multi-contrast raw k-space benchmarks to develop and validate universal reconstruction models.

Methods: The CMRxRecon2024 challenge released a large-scale multi-coil k-space dataset with 330 subjects across six modalities (cine, tagging, T1/T2 mapping, aorta, 2D flow, black-blood)¹.We provided a public CMR benchmark (https://www.synapse.org/Synapse:syn54951257/wiki/627936), tutorial (https://github.com/CmrxRecon/CMRxRecon2024), and platform (https://cmrxrecon.github.io/2024/Home.html). Two tasks assessed model generalization (Fig 1 a):

1. Modality universality: training on four “seen” modalities, testing on both seen and two “unseen” modalities.


2. Sampling universality: single-model reconstruction across Gaussian, uniform, and radial undersampling patterns with acceleration factors of 4–24.

All methods were evaluated on hidden test data using adjusted SSIM/PSNR/NMSE (weighted by successful reconstruction rate) and blinded radiologist quality scores².

Results: From 200 registered teams worldwide, 15 completed the final phase (Fig 1b). Their rankings are shown in Fig 1c,d. Effective strategies include intensity standardization (min-max/z-score), balanced sampling-pattern training, and acceleration-aware augmentation (Fig 2a,c). SSIM is used by all teams as loss function (Fig 2b,d). High-performing network backbones include:

1. Prompt-Unet³ (Fig 2e) embedded learnable priors with adaptive representations.


2. vSHARP⁴ (Fig 2f) integrated iterative optimization as a data consistency.


3. E2E-VarNet⁵ (Fig 2g) dynamically updated the spatial encoding information.

Notably, universal models surpassed GRAPPA⁶-based reference for certain modalities (aorta in Fig 3a,d). A ‘Rad Score’ was calculated by subtracting radiologist’s rating of the reference from the reconstruction, followed by z-score normalization. They showed that SSIM could drop when reconstructions were rated better than reference, highlighting its limits in capturing perceptual quality (Fig 3b,c,e,f).

Conclusion: CMRxRecon2024 established a public, multi-modality, multi-view CMR benchmark to accelerate the translation of AI-based reconstruction into clinical workflows. The challenge identified practical design principles—physics-informed encoding, adaptive prompt learning, and robust coil sensitivity handling—that improve generalization across sequences and sampling schemes. These insights support the development of single-model, clinically versatile CMR reconstruction systems that can streamline acquisition, reduce scan time, and maintain diagnostic quality across the breadth of clinical CMR protocols.

Challenge tasks, contributions, timeline and rankings. (a) This schematic illustrates the two challenge tasks and the data/modalities involved. Task 1 focuses on modality universality, training on seen modalities, and testing on both seen (cine, tagging, T1 mapping, and T2 mapping) and unseen (black-blood and 2D flow) cases. Task 2 emphasizes sampling universality, and training on diverse 3D sampling schemes with a single model for inference. Evaluation includes objective metrics and radiologist ratings. Contributions include platform benchmarking, community evaluation criteria, and a toolbox with code resources. (b) The timeline and summary of participation. (c) Ranking result for Task 1 (p-value: * $ <$ 0.05, ** $ <$ 0.01). "NA" in row of GT indicates they work as the reference, while "NA" in the column of the Rad. Score means the teams beyond the top five are not ranked. (d) Ranking result for Task 2 (p-value: * $ <$ 0.05, ** $ <$ 0.01). "NA" in row of GT indicates they work as the reference, while "NA" in the column of the Rad. Score means the teams beyond the top five are not ranked.
Overview of top-performing team configurations, loss functions and backbones. Abbreviations: flip (F), rotation (R), shift (S), data consistency (DC), gradient descent (GD), conjugate gradient (CG), learning rate (LR), not applicable (NA). (a) Summary of data preprocessing, model architecture, and training configurations in Task 1. (b) Loss function components used in Task 1. (c) Summary of data preprocessing, model architecture, and training configurations in Task 2. (d) Loss function components used in Task 2. (e) Prompt-UNet: CNN-based architecture with prompt-driven feature modulation for MRI reconstruction. (f) vSHARP: Unrolled ADMM reconstruction framework integrating learned denoisers and differentiable conjugate gradient descent (DCGD). (g) E2E-VarNet: End-to-end variational network combining iterative learned refinement (R), sensitivity estimation (SME), and soft data consistency (DC) for multi-coil MRI reconstruction.
Objective metric and radiologist evaluation of top submissions. (a) One reconstructed case demonstrates superior performance compared to the ground truth (GT) image. The images showcase the outputs from the top five teams in Task 1. (b) The SSIM, PSNR, and NMSE against the normalized radiologist rating for Task 1. (c) The radiologist score of different modalities for Task 1. (d) One reconstructed case demonstrates superior performance compared to the ground truth (GT) image. The images showcase the outputs from the top five teams in Task 2. (e) The SSIM, PSNR, and NMSE against the normalized radiologist rating for Task 2. (f) The radiologist score of different modalities for Task 2.