[Paper Review] All-in-One Deep Learning Framework for MR Image Reconstruction

Geunu Jeong, Hyeonsoo Kim+3 more|arXiv (Cornell University)|May 6, 2024

Medical Imaging Techniques and Applications7 citations

TL;DR

A single context-enhanced U-Net model reconstructs MR images across multiple k-space dimensions and vendors, enabling tunable denoising, multi-directional super-resolution, and broad compatibility with diverse protocols and acceleration strategies.

ABSTRACT

We introduce a novel, all-in-one deep learning framework for MR image reconstruction, enabling a single model to enhance image quality across multiple aspects of k-space sampling and to be effective across a wide range of clinical and technical scenarios. This DICOM-based algorithm serves as the core of SwiftMR (AIRS Medical, Seoul, Korea), which is FDA-cleared, CE-certified, and commercially available. We first detail the comprehensive development process of the model, including data collection, training pair preparation, model architecture design, and DICOM inference. We then assess the model's capability to enhance image quality in a multi-dimensional manner, specifically across various aspects of k-space sampling. Subsequently, we evaluate several features of the multi-dimensional enhancement: the accuracy of tunable denoising, the effectiveness of super-resolution in each encoding direction, and the reduction of artifacts that become more prominent at lower spatial resolutions. Additionally, we assess its compatibility with various scan parameter sets and its generalizability across scanner vendors not seen during training. Finally, we present specific cases demonstrating the model's utility in reducing scan time across anatomical regions in conjunction with protocol optimization. The proposed model is compatible with a broad spectrum of scenarios, including various vendors, pulse sequences, scan parameters, and anatomical regions. Its DICOM-based operation particularly enhances its applicability for real-world applications. Given its demonstrated effectiveness and versatility, we expect its use to expand in the field of clinical MRI.

Motivation & Objective

• Motivate and address the need for high-quality MR reconstruction across diverse k-space sampling dimensions and clinical scenarios.
• Develop an all-in-one, DICOM-based DL framework that can adapt to multiple degradation patterns and reconstruction steps.
• Propose a Context Enhanced U-Net (CE U-Net) that integrates contextual data to enable tunable denoising and multi-dimensional improvements.
• Demonstrate broad compatibility with vendors, pulse sequences, scan parameters, and unseen datasets, including potential scan-time reductions.

Proposed method

• Degrade raw k-space data along multiple dimensions (noise, k_max, partial Fourier, elliptical sampling) and simulate reconstruction steps to create diverse input–target pairs.
• Use a standardization module and CE U-Net architecture with a Dynamic Modulation Pathway to inject contextual data into intermediate layers.
• Train separate 2D and 3D networks (3D uses seven-slice stacks to enable slice-direction super-resolution).
• Provide contextual inputs including k-space sampling parameters and an explicitly derived noise-reduction factor to enable tunable denoising.
• Apply L1 loss with ADAM optimizer and a learning rate schedule to optimize network parameters.
• Perform DICOM-based inference by extracting relevant tags for standardization and routing inputs to the appropriate network.

Experimental results

Research questions

• RQ1 Can a single DL model improve MR image quality across multiple k-space sampling dimensions and reconstruction steps?
• RQ2 How does integrating contextual data influence tunable denoising and multi-dimensional enhancement in MR reconstruction?
• RQ3 Does the model generalize across unseen vendors and different scan parameter configurations?
• RQ4 Can the framework contribute to actual scan-time reductions when paired with protocol optimization.

Key findings

• The model enables multi-dimensional enhancement by adding successive dimensions (noise reduction, k_max in frequency, phase partial Fourier, phase k_max, slice k_max).
• Tunable denoising factors yield corresponding reductions in measured noise without changing mean signal levels.
• Edge sharpness in all encoding directions improves by about 2.5× with DLR, indicating slice-direction super-resolution and cross-direction gains.
• DLR reduces low-spatial-resolution artifacts such as truncation and intravoxel dephasing, especially with slice-direction super-resolution.
• The framework remains effective across eight brain scans with varied k-space sampling and reconstruction parameters, and across three unseen vendors, indicating broad compatibility.
• Images from standard vs accelerated protocols across multiple regions show noise reduction and improved overall image quality after DLR, suggesting potential for scan-time reduction.