Artificial Intelligence Accelerates MRI Parameter Mapping
Key findings
- The clinical utility of magnetic resonance mapping has been limited because of the long times required for image acquisition and reconstruction
- Massachusetts General Hospital researchers previously created Model-Augmented Neural neTwork with Incoherence Sampling (MANTIS), a deep learning model that accelerates T2 parameter acquisition
- The team recently added a generative adversarial network (GAN) to MANTIS that accelerates image reconstruction
- In a feasibility study, MANTIS-GAN was more accurate than conventional reconstruction approaches and provided greater image texture and sharpness than either conventional methods or MANTIS
Subscribe to the latest updates from Radiology Advances in Motion
Magnetic resonance (MR) parameter mapping (e.g., T1 and T2 mapping) is a valuable tool for tissue characterization in a variety of diseases. However, its clinical utility has been limited because of the long times required for image acquisition and reconstruction.
At Massachusetts General Hospital and elsewhere, there is interest in accelerating parameter mapping by using deep learning, a form of artificial intelligence. In the setting of MRI, deep learning is the process of having convolutional neural networks (CNNs) train themselves to analyze visual images.
In Magnetic Resonance Imaging, Georges El Fakhri, PhD, director of the Gordon Center for Medical Imaging in the Department of Radiology at Massachusetts General Hospital, and Fang Liu, PhD, a researcher in the center, and colleagues describe their newest effort, a method that rapidly creates realistic MRI parameter maps by incorporating adversarial training into a previously created deep learning framework.
The Previous Model: MANTIS
As reported in Magnetic Resonance in Medicine, Dr. Liu and colleagues previously developed Model-Augmented Neural neTwork with Incoherence Sampling (MANTIS), a deep learning method for very rapidly mapping T2 parameters. An end-to-end CNN rapidly transforms undersampled images (k-space data) directly into MR parameter maps, guided by a physical model that characterizes the MR signal.
The challenge for MANTIS and similar models is the potential for image blurring and loss of detail while the model is being trained.
MANTIS-GAN
The new model incorporates a generative adversarial network (GAN). Adversarial training is an unsupervised machine learning technique that pits two neural networks against each another to speed up the learning process. The hope in combining MANTIS with a GAN was that adversarial training would improve image reconstruction.
Feasibility Study
The feasibility of MANTIS-GAN for T2 mapping was evaluated using simulated brain imaging datasets and actual knee imaging datasets. For both datasets, the new framework was more accurate than conventional reconstruction approaches. It also provided greater image texture and sharpness than either conventional methods or MANTIS, at a high acceleration rate.
Crisp detail about tissue structure should be very useful for evaluating small lesions and subtle tissue abnormalities caused by early pathologies.
view original journal article Subscription may be required
Learn about the Gordon Center for Biological Imaging
Refer a patient to the Department of Radiology