Accelerating MRI reconstruction via active acquisition

June 18, 2019

What the research is:

A novel approach to undersampled magnetic resonance imaging (MRI) reconstruction that restores a high-fidelity image from partially observed measurements. We reduce reconstruction error and uncertainty by dynamically selecting which measurements are best to observe. The goal of this data-driven active acquisition approach is to simultaneously minimize image reconstruction errors and acquisition time (i.e., the number of measurements to acquire). Evaluated on the large-scale fastMRI data set, provided by our research collaborators at NYU Langone Health, our approach successfully outperforms active acquisition baselines.

How it works:

While most deep learning methods for MRI reconstruction are designed to work with a fixed set of measurements, we propose that the sampling trajectory should be adapted on the fly, depending on the difficulty of the reconstruction. We propose a system that, at inference time, actively acquires k-space measurements (kMA) and iteratively refines the prediction with the goal of reducing the error and, thus, the final uncertainty.

To do this, we introduce a novel neural network–based system composed of a reconstruction network and an evaluator network, which are trained jointly to minimize the reconstruction error while maximizing the acquisition speed. The reconstruction network takes a basic zero-filled image reconstruction as input and outputs an improved image reconstruction, along with its uncertainty estimates. We explore a variety of architectural designs for the reconstruction network and present a residual-based model that exploits the underlying characteristics.

The goal of the evaluator network is to rate all the unobserved kMA of a reconstructed image. This rating is used to guide the measurement selection criterion. At inference time, the evaluator scores are used to select the next unobserved measurement to acquire. The input image is then updated accordingly, and the process iterates until all measurements are acquired or a stopping criterion is met, e.g., a low global uncertainty score.

Given an initial trajectory, an MRI scanner (1) acquires measurements. The zero-filled image reconstruction (2) is fed into our neural network-based system (3). The system outputs an MRI reconstruction, an uncertainty map, and the next suggested measurement to scan (4). These steps are repeated until the stopping criterion is met.

 AI Habitat

A video showing active acquisition in action. From left to right: sampling trajectories acquired so far, current image reconstruction, current reconstruction error, current uncertainty estimation, and percentage of acquired kMA.

Why it matters:

The promises of MRI, compared with computed tomography and X-rays, are the superior soft tissue contrast and the lack of ionizing radiation. But the acquisition process is slow. Currently, MRI examinations can take as long as an hour, resulting in uncomfortable examination experiences and long scheduling backlogs. Accelerating the speed of MRI has the potential to substantially improve both accessibility and patient experience. Active acquisition in particular enables optimizing MRI scans to acquire the appropriate subset of measurements for each particular scan or patient, which is faster because fewer kMA are recorded.

Read the full paper:

Reducing uncertainty in undersampled MRI reconstruction with active acquisition

See this work presented at CVPR 2019

We'd like to thank the researchers at NYU Langone Health for collaborating with us on this research.


Adriana Romero

Research Scientist

Michal Drozdzal

Research Scientist

Nafissa Yakubova

Research Scientist