MRI Image Reconstruction via Learning Optimization Using Neural ODEs

Eric Z. Chen, Terrence Chen, Shanhui Sun

Abstract

We propose to formulate MRI image reconstruction as an optimization problem and model the optimization trajectory as a dynamic process using ordinary differential equations (ODEs). We model the dynamics in ODE with a neural network and solve the desired ODE with the off-the-shelf (fixed) solver to obtain reconstructed images. We extend this model and incorporate the knowledge of off-the-shelf ODE solvers into the network design (learned solvers). We investigate several models based on three ODE solvers and compare models with fixed solvers and learned solvers. Our models achieve better reconstruction results and are more parameter efficient than other popular methods such as UNet and cascaded CNN. We introduce a new way of tackling the MRI reconstruction problem by modeling the continuous optimization dynamics using neural ODEs.

Introduction

Magnetic Resonance Imaging (MRI) is a critical diagnostic tool, but its acquisition process is inherently slow. To accelerate MRI, technicians often undersample data in the frequency domain (k-space), which unfortunately introduces aliasing artifacts and blurring in the reconstructed images. Recovering a high-fidelity image from this undersampled data is a challenging inverse problem.

Traditional deep learning approaches, such as UNets or cascaded CNNs, typically treat this reconstruction as a discrete mapping or a series of iterative steps. This paper introduces ReconODE, a framework that reimagines MRI reconstruction as a continuous optimization trajectory. By modeling the process as a flow governed by an Ordinary Differential Equation (ODE), the authors move from discrete algorithmic steps to a continuous dynamic system.

Conceptual overview of MRI reconstruction as a continuous optimization trajectory (A) and the ReconODE framework (B).

Methods

The ReconODE framework treats the undersampled image as an initial state $x(t_0)$ and the fully reconstructed image as the final state $x(t_N)$ of a dynamic process. The change in the image over "time" is defined by a learned dynamics function $f(x, t)$, modeled by a convolutional neural network.

The authors explore two primary methodologies for solving the ODE:

  • Fixed Solvers (ReconODE-FT/FA): These models use standard mathematical solvers (e.g., Euler, Runge-Kutta 2nd and 4th order) to integrate the learned dynamics. Training is performed either via standard backpropagation (FT) or the memory-efficient adjoint sensitivity method (FA).
  • Learned Solvers (ReconODE-LT): This novel approach replaces rigid solver formulas with neural network blocks that mimic solver structures. These "learned solvers" can adaptively determine optimal coefficients and step sizes. Furthermore, an attention mechanism replaces traditional weighted sums in the solver equations to better capture multi-scale image features.

To manage the computational demands of these complex architectures, the authors utilize gradient checkpointing, significantly reducing GPU memory usage during training with minimal impact on compute time.

Experimental Results

The models were evaluated on the fastMRI Knee Dataset (Single-Coil), using 34,742 slices for training and 7,135 for validation. The task involved reconstructing images from 4x and 8x undersampled data.

Parameter Efficiency and Accuracy

A standout result of this work is the extreme parameter efficiency of the ReconODE models. The best-performing variant, ReconODE-LT-RK4, utilizes only 0.15 million parameters. In comparison, a standard UNet baseline requires 3.35 million parameters—approximately 22 times more—yet ReconODE achieves higher reconstruction accuracy.

At 4x acceleration, ReconODE-LT-RK4 achieved a PSNR of 32.39 and SSIM of 0.7333. At 8x acceleration, it reached a PSNR of 30.27 and SSIM of 0.6617. These metrics consistently outperformed UNet, KIKI-Net, and the cascaded CNN (D5C5) across both acceleration factors.

Qualitative Performance

Visual analysis confirms that the ReconODE-LT models produce sharper images with fewer artifacts compared to both fixed-solver variants and standard baselines. The learned attention modules within the solvers were particularly effective at recovering high-frequency details that are often lost in traditional reconstruction methods.

Visual comparison of reconstruction results and error maps for 4x and 8x acceleration factors. ReconODE-LT variants show superior detail recovery and lower error.

Memory Optimization

Using gradient checkpointing on the ReconODE-LT-RK4 model reduced GPU memory usage by 73% (from 13.6GB to 3.7GB). This optimization is crucial for training deep ODE-based models, as it only added a negligible 0.13 seconds to the training time per step.

Discussion and Conclusion

The study reveals several key insights into the application of Neural ODEs for inverse problems. Interestingly, while the adjoint method is often touted for its memory efficiency, the authors found that standard backpropagation through the solver (FT) yielded more accurate reconstructions. Additionally, the transition from fixed mathematical solvers to learned, attention-augmented solvers provided a significant boost in image quality.

While the current models are small and were evaluated on validation sets due to leaderboard constraints, they establish a powerful new paradigm. The ReconODE framework demonstrates that modeling optimization as a continuous flow is not only theoretically elegant but also practically superior in terms of parameter efficiency and reconstruction fidelity. Future work will focus on scaling these dynamics functions to larger architectures to further push the boundaries of MRI acceleration.

Summary updated

Abstract

We propose to formulate MRI image reconstruction as an optimization problem and model the optimization trajectory as a dynamic process using ordinary differential equations (ODEs). We model the dynamics in ODE with a neural network and solve the desired ODE with the off-the-shelf (fixed) solver to obtain reconstructed images. We extend this model and incorporate the knowledge of off-the-shelf ODE solvers into the network design (learned solvers). We investigate several models based on three ODE solvers and compare models with fixed solvers and learned solvers. Our models achieve better reconstruction results and are more parameter efficient than other popular methods such as UNet and cascaded CNN. We introduce a new way of tackling the MRI reconstruction problem by modeling the continuous optimization dynamics using neural ODEs.
BibTeX
@inproceedings{chen2020mri,
  author = {Eric Z. Chen and Terrence Chen and Shanhui Sun},
  title = {{MRI Image Reconstruction via Learning Optimization Using Neural ODEs}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  url = {https://openreview.net/forum?id=Y73u4B6utd},
}
Figure