Deep Generative Adversarial Networks for Compressed Sensing Automates MRI (1706.00051v1)

Abstract: Magnetic resonance image (MRI) reconstruction is a severely ill-posed linear inverse task demanding time and resource intensive computations that can substantially trade off {\it accuracy} for {\it speed} in real-time imaging. In addition, state-of-the-art compressed sensing (CS) analytics are not cognizant of the image {\it diagnostic quality}. To cope with these challenges we put forth a novel CS framework that permeates benefits from generative adversarial networks (GAN) to train a (low-dimensional) manifold of diagnostic-quality MR images from historical patients. Leveraging a mixture of least-squares (LS) GANs and pixel-wise $\ell_1$ cost, a deep residual network with skip connections is trained as the generator that learns to remove the {\it aliasing} artifacts by projecting onto the manifold. LSGAN learns the texture details, while $\ell_1$ controls the high-frequency noise. A multilayer convolutional neural network is then jointly trained based on diagnostic quality images to discriminate the projection quality. The test phase performs feed-forward propagation over the generator network that demands a very low computational overhead. Extensive evaluations are performed on a large contrast-enhanced MR dataset of pediatric patients. In particular, images rated based on expert radiologists corroborate that GANCS retrieves high contrast images with detailed texture relative to conventional CS, and pixel-wise schemes. In addition, it offers reconstruction under a few milliseconds, two orders of magnitude faster than state-of-the-art CS-MRI schemes.

Deep Generative Adversarial Networks for Compressed Sensing in MRI

The paper "Deep Generative Adversarial Networks for Compressed Sensing (GANCS) Automates MRI" presents a novel framework leveraging Generative Adversarial Networks (GANs) for enhancing the efficiency of MRI reconstruction. The methodology addresses the intrinsic challenges associated with MRI, notably its classification as a severely ill-posed linear inverse problem and the traditional trade-off between speed and diagnostic quality in image reconstruction.

Core Contributions and Methodology

The paper introduces a deep learning-based compressed sensing (CS) framework that integrates GANs to automate MRI reconstruction processes. The approach employs a residual network with skip connections to build a generator that effectively learns and leverages a manifold of diagnostic-quality MR images from historical data, thereby mitigating aliasing artifacts evident in routine CS reconstructions.

1. Network Architecture: The architecture utilizes a mixture of least squares GANs (LSGANs) for high textual fidelity while an $\ell_1$ cost function regulates high-frequency noise. The GAN framework, comprising a generator and discriminator network, facilitates the generation of full-quality images from aliased versions obtained through inverse Fourier transforms of undersampled data.
2. Training Framework: A blend of LSGAN and pixel-wise $\ell_1$ loss functions is adopted, providing a balanced mechanism to fine-tune image texture details, retain diagnostic features, and ensure noise control. This dual mechanism is absent in traditional CS-MRI approaches.
3. Efficiency and Evaluation: One of the significant outcomes reported is a computational speed improvement, with the process completing within milliseconds—offering an acceleration two orders of magnitude faster than traditional methods. Evaluation conducted on a robust pediatric dataset demonstrated superior image quality benchmarked by radiologists' assessments against conventional CS methods.

Implications and Future Directions

The research brings forth considerable implications for clinical practice, largely due to its capacity to expedite real-time MRI visualization critical for practices such as MR-guided neurosurgery. The dramatic reduction in computation time from seconds to mere milliseconds could greatly enhance workflows and throughput in clinical settings.

From a theoretical standpoint, the intersection drawn in the paper between deep learning and traditional image reconstruction paradigms indicates potential advancements in manifold learning and data-consistency integration within generative models.

Future Trajectories: The advancement in GANCS opens avenues for further exploration in several directions:

• Adaptive Training: Exploring the adaptability of the GANCS model to various MRI scanner settings and different physical architectures could universalize its application.
• Enhanced Image Features: Further improving the generator architecture and extending to three-dimensional comprehensive spatial data could enhance the robustness of MRI reconstructions.
• Integration with Anomalies and Diverse Imaging Protocols: Incorporating pathologies and accounting for varied patient pathologies in the modeling process will ensure diagnostic robustness across diverse patient demographics and clinical scenarios.

This paper's stride towards joining GAN's strengths with MRI underscored its importance in navigating the complexities of biomedical image reconstruction, emphasizing an advanced narrative in healthcare imaging technologies through deep learning techniques.