Deep Generative Adversarial Neural Networks for Compressive Sensing MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Undersampled magnetic resonance image (MRI) reconstruction is an ill-posed inverse problem with speed-accuracy tradeoffs.
• Current compressed sensing (CS) methods do not consider diagnostic image quality.
• Existing techniques require significant computational resources and time.

Purpose of the Study:

• To develop a novel CS framework using generative adversarial networks (GANs) for improved MRI reconstruction.
• To address the limitations of existing CS methods by incorporating diagnostic quality.
• To accelerate the MRI reconstruction process.

Main Methods:

• A novel CS framework combining least-squares (LS) GANs and pixel-wise l1/l2 cost was developed.
• A deep residual network with skip connections served as the generator, trained to remove aliasing artifacts.
• A convolutional neural network (CNN) discriminator was used for perceptual scoring, trained on high-quality MR images.

Main Results:

• The proposed GANCS method demonstrated superior image quality and fine texture details compared to conventional and deep-learning CS schemes.
• Expert radiologists confirmed the improved diagnostic quality of GANCS reconstructions.
• Reconstruction times were reduced to under a few milliseconds, achieving a two-order-of-magnitude speedup.

Conclusions:

• The GANCS framework effectively models the manifold of high-quality MR images, enabling artifact removal and noise suppression.
• This approach significantly enhances both the speed and diagnostic quality of undersampled MRI reconstruction.
• The method offers a computationally efficient solution for accelerated MRI acquisition and analysis.