SRE-CNN: A Spatiotemporal Rotation-Equivariant CNN for Cardiac Cine MR Imaging

Yuliang Zhu, Jing Cheng, Zhuo-Xu Cui, Jianfeng Ren, Chengbo Wang, Dong Liang

Introduction

Cardiac cine MRI serves as the clinical gold standard for assessing heart function, providing high-resolution dynamic images of the beating heart. However, the acquisition process is inherently slow, often requiring long breath-holds that are difficult for patients and prone to motion artifacts. To accelerate imaging, clinicians employ undersampling techniques, which necessitate advanced reconstruction algorithms to recover high-quality images from sparse data.

While deep learning, particularly Convolutional Neural Networks (CNNs), has revolutionized MRI reconstruction, standard architectures possess a significant limitation: they are translation-equivariant but not rotation-equivariant. In cardiac imaging, anatomical structures like the myocardium appear in various orientations across different spatial locations and temporal frames. Standard CNNs must learn these variations as separate patterns, requiring more data and often failing to maintain sharp details when faced with rotated features.

The SRE-CNN Framework

To address these challenges, researchers have introduced the Spatiotemporal Rotation-Equivariant CNN (SRE-CNN). This framework is designed to inherently recognize and preserve anatomical rotations across both space and time.

Figure 1: Illustration of spatial and temporal symmetry in dynamic cardiac MRI, where anatomical features appear in various orientations across frames.

The SRE-CNN architecture incorporates three primary innovations to ensure robust reconstruction at high acceleration rates:

1. Temporal-Equivariant Convolutional Module

In traditional (2+1)D convolutional setups, spatial and temporal layers are processed sequentially. However, standard temporal convolutions often disrupt the rotation-aware properties established by spatial layers. SRE-CNN solves this by using a novel module that applies cyclic shifts to filter channels. This mechanism ensures that the temporal layers remain synchronized with the rotation state of the spatial features, maintaining "global equivariance" throughout the 2D+t volume.

2. High-Precision Filter Parametrization

Standard pixel-based filters suffer from interpolation artifacts when rotated, leading to a loss of fine anatomical detail. SRE-CNN utilizes 1D and 2D Fourier series expansion for filter parametrization. This mathematical approach allows filters to be rotated with zero representation error, which is critical for medical imaging where precise boundary detection is essential for diagnosis.

3. Unrolled Iterative Reconstruction

The SRE-CNN is integrated into an "unrolled" network architecture based on the Proximal Gradient Descent (PGD) algorithm. This design combines the physical constraints of MRI (data consistency) with the learned priors of equivariant convolutions across multiple iterations.

Figure 2: The SRE-CNN architecture, showing the unrolled iterative reconstruction process (a) and the detailed structure of the equivariant convolutional layers (b).

Experimental Evaluation

The model was evaluated using a dataset of 800 2D+t multi-coil cardiac MR samples from 29 healthy volunteers, acquired on a 3T Siemens scanner. The researchers tested the model's performance under extreme retrospective undersampling at 12x, 16x, and 20x acceleration factors using the VISTA strategy.

Key Results

SRE-CNN demonstrated significant improvements over state-of-the-art methods, including MoDL and R2plus1D:

  • Quantitative Performance: SRE-CNN consistently achieved the highest Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Notably, at 20x acceleration, it outperformed the next best method by approximately 2dB in PSNR.
  • Qualitative Quality: Visual analysis confirmed that SRE-CNN produces sharper myocardial boundaries and more accurately captures dynamic motion, whereas competing models often exhibited blurring or residual artifacts at high acceleration.
  • Clinical Viability: Despite its mathematical complexity, the reconstruction time remains under one second, making it suitable for real-time clinical workflows.

Conclusion and Future Directions

SRE-CNN represents a significant advancement in cardiac MRI reconstruction by "baking" physical symmetries into the network architecture. By leveraging rotation equivariance, the model achieves superior image quality with less training data. Future work will explore extending this framework to include other transformations, such as scaling and reflection, and validating the model on patients with diverse cardiac pathologies.

Summary updated

BibTeX
@inproceedings{zhu2024sre,
  author = {Yuliang Zhu and Jing Cheng and Zhuo-Xu Cui and Jianfeng Ren and Chengbo Wang and Dong Liang},
  title = {{SRE-CNN: A Spatiotemporal Rotation-Equivariant CNN for Cardiac Cine MR Imaging}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2024},
  doi = {10.1007/978-3-031-72104-5\_65},
}
Figure