[Paper Review] Performance characterization of a novel deep learning-based MR image reconstruction pipeline

R. Marc Lebel|arXiv (Cornell University)|Aug 14, 2020

Medical Imaging Techniques and ApplicationsMedicine15 references63 citations

TL;DR

A novel deep learning–based MR reconstruction pipeline is characterized for improving sharpness and denoising, with evaluation on digital references, phantoms, and in-vivo data to support reduced scan time.

ABSTRACT

A novel deep learning-based magnetic resonance imaging reconstruction pipeline was designed to address fundamental image quality limitations of conventional reconstruction to provide high-resolution, low-noise MR images. This pipeline's unique aims were to convert truncation artifact into improved image sharpness while jointly denoising images to improve image quality. This new approach, now commercially available at AIR Recon DL (GE Healthcare, Waukesha, WI), includes a deep convolutional neural network (CNN) to aid in the reconstruction of raw data, ultimately producing clean, sharp images. Here we describe key features of this pipeline and its CNN, characterize its performance in digital reference objects, phantoms, and in-vivo, and present sample images and protocol optimization strategies that leverage image quality improvement for reduced scan time. This new deep learning-based reconstruction pipeline represents a powerful new tool to increase the diagnostic and operational performance of an MRI scanner.

Motivation & Objective

• Address fundamental image quality limitations of conventional MR reconstruction.
• Provide high-resolution, low-noise MR images.
• Convert truncation artifacts into improved image sharpness while jointly denoising.
• Demonstrate performance across digital reference objects, phantoms, and in-vivo data.
• Offer protocol optimization strategies leveraging image quality improvements for faster scans.

Proposed method

• Describe the integration of a deep convolutional neural network to aid in the reconstruction of raw MR data.
• Outline the pipeline features that convert truncation artifact into sharper imagery while performing denoising.
• Evaluate performance using digital reference objects, phantoms, and in-vivo data.
• Provide sample images and protocol optimization guidance based on image quality improvements.

Experimental results

Research questions

• RQ1 How does the deep learning reconstruction pipeline affect image sharpness and noise relative to conventional methods?
• RQ2 How does the pipeline perform across digital reference objects, phantoms, and in-vivo MR data?
• RQ3 Can the pipeline enable reduced scan time through improved image quality?
• RQ4 What are practical protocol optimization strategies that leverage the enhanced image quality?

Key findings

• The pipeline improves image sharpness and reduces noise compared with conventional reconstruction in tested objects.
• Performance is demonstrated across digital reference objects, phantoms, and in-vivo data.
• Sample images and protocol strategies illustrate how image quality gains can support shorter scan times.