Generalizable synthetic MRI with physics-informed convolutional networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Magnetic Resonance Imaging (MRI) offers high-quality neuroimaging but requires multiple contrast acquisitions.
• Synthetic MRI aims to reduce scan times by generating various contrasts from a single acquisition.

Purpose of the Study:

• To develop a physics-informed deep learning method for synthesizing multiple brain MRI contrasts.
• To achieve this from a single, accelerated 5-minute MRI acquisition.
• To evaluate the method's generalizability to arbitrary MRI contrasts.

Main Methods:

• A generative adversarial network (GAN) model was trained on 55 subjects' data.
• The model generated quantitative parameter maps (q*-maps) from a 5-min transient-state sequence.
• Synthesized contrasts (PD, T1, T2, T2-FLAIR) were compared to an end-to-end deep learning method and ground truth.

Main Results:

• The physics-informed method achieved comparable quality to standard MRI contrasts.
• Structural similarity and peak signal-to-noise ratios were high, preserving lesion details.
• The method demonstrated generalizability, synthesizing unseen contrasts with similar signal contrast and CNR.

Conclusions:

• Physics-informed deep learning enables high-quality synthetic MRI contrast generation.
• The method successfully generalizes to contrasts beyond the training dataset.
• This technology holds significant potential for accelerating neuroimaging protocols.