Using reinforcement learning to personalize AI-accelerated MRI scans

October 05, 2020

What the research is:

A method leveraging reinforcement learning to improve AI-accelerated magnetic resonance imaging (MRI) scans. Experiments using the fastMRI data set created by NYU Langone show that our models significantly reduce reconstruction errors by dynamically adjusting the sequence of k-space measurements, a process known as active MRI acquisition.

These k-space measurements are the building blocks of an MRI scan, the raw data from which images are reconstructed. The fastMRI project recently demonstrated that AI can reconstruct diagnostically useful scans from undersampled k-space data, and that these scans are interchangeable with fully sampled scans. That work used a predetermined sequence of these k-space measurements, called an acquisition trajectory.

But in parallel, we have been working on trajectory optimization, using AI to be smarter about which k-space measurements to take in the first place. We do this by formulating the problem as a sequential decision process and using reinforcement learning to solve it.

Our early experiments with the fastMRI data set show that our models outperform the previous active MRI acquisition state of the art over a broad range of acceleration factors. To further foster research on active MRI acquisition, we are releasing a reinforcement learning environment that enables the simulation of Cartesian acquisition trajectories from MRI images, together with simple acquisition heuristics and our pretrained models of acquisition policies.

How it works:

We first demonstrated the value of a data-driven active MRI acquisition approach in 2019. Since then, we have improved on our initial research by using a long-term reconstruction forecast to iteratively decide which k-space measurements to acquire. Our new approach formulates the active MRI acquisition problem as a partially observable Markov decision process (POMDP) and proposes the use of deep reinforcement learning policies to solve it.

To this end, we trained a reconstruction network that produced high-fidelity images from previously acquired k-space measurements and used such images as observations in our POMDP. Given this set of observed measurements and the corresponding high-fidelity reconstruction, we then solved the POMDP by training a value network that estimated the utility of each possible measurement, quantifying its effect on the image reconstruction quality. The measurement with the highest utility is selected and added as the next acquisition in the overall sequence.

Something Went Wrong
We're having trouble playing this video.

The video shows two active MRI acquisition trajectories. The top uses the baseline, while the bottom uses our reinforcement learning approach. Left to right: acquired k-space columns, ground truth image, reconstruction from partial k-space measurements, and error map. Note how the reinforcement learning approach achieves higher-quality reconstructions earlier than the baseline. (Source: fastMRI data set)

Our approach optimizes the reconstruction over the whole range of acceleration factors. The system determines the most expedient trajectory at the moment of inference, as the most accurate reconstruction can require different trajectories and accelerations depending on the patient or disease.

Why it matters:

A cornerstone of modern medicine and a key diagnostic tool, MRIs are nevertheless hampered by the time it currently takes to gather the necessary data for a scan. Our goal with fastMRI is to leverage AI to accelerate this process, reducing patient discomfort, allowing practitioners to serve more patients per day, and potentially even expanding the uses for MRIs.

This is a multifaceted problem, though. The fastMRI project has previously released research demonstrating how AI can be used to improve image reconstruction quality for a fixed acceleration factor and set of measurements. While this is important, our team’s parallel work on active MRI acquisition is equally vital because it has the potential to pave the way for patient-specific acceleration factors. By leveraging AI earlier in the process, we ensure each MRI scan meets the needs of the patient or scenario in question. This personalized approach is widely considered to be the future of medicine.

With fastMRI, we also committed to sharing our research and models with the community, thereby accelerating progress on this challenging problem. As part of this commitment to open science, we are releasing the aforementioned reinforcement learning environment that can be easily configured to add new MRI reconstruction models with minimal coding overhead. We hope this will foster future research by providing researchers the core simulation logic needed to train active MRI acquisition models.

Read the full paper:

Written By

Luis Pineda

Research Engineer

Michal Drozdzal

Research Scientist

Adriana Romero

Research Scientist