Boosting the signal-to-noise of low-field MRI with deep learning image reconstruction

Abstract

Recent years have seen a resurgence of interest in inexpensive low magnetic field (< 0.3 T) MRI systems mainly due to advances in magnet, coil and gradient set designs. Most of these advances have focused on improving hardware and signal acquisition strategies, and far less on the use of advanced image reconstruction methods to improve attainable image quality at low field. We describe here the use of our end-to-end deep neural network approach (AUTOMAP) to improve the image quality of highly noise-corrupted low-field MRI data. We compare the performance of this approach to two additional state-of-the-art denoising pipelines. We find that AUTOMAP improves image reconstruction of data acquired on two very different low-field MRI systems: human brain data acquired at 6.5 mT, and plant root data acquired at 47 mT, demonstrating SNR gains above Fourier reconstruction by factors of 1.5- to 4.5-fold, and 3-fold, respectively. In these applications, AUTOMAP outperformed two different contemporary image-based denoising algorithms, and suppressed noise-like spike artifacts in the reconstructed images. The impact of domain-specific training corpora on the reconstruction performance is discussed. The AUTOMAP approach to image reconstruction will enable significant image quality improvements at low-field, especially in highly noise-corrupted environments.

On Real-time Image Reconstruction with Neural Networks for MRI-guided Radiotherapy

“…While these dense layers enable data-driven learning of the manifold between k-space and the image domain, they also have significant memory requirements that make the translation to 3D reconstruction and tracking more challenging. [51,52] To perform reconstructions above the relatively low resolutions used for motion tracking on an MRI-Linac will require lighter-weight reconstruction networks. [18,53] One light-weight implementation of AUTOMAP is decomposed-AUTOMAP (dAU-TOMAP) which replaces dense layers with orthogonal 'domain transform' layers.…”

“…However, another valuable strength of data-driven MRI reconstruction models such as AUTOMAP is that they implicitly learn to suppress common MRI artifacts, such as spike noise caused by RF leakage, as these inputs typically fall outside the training domain. [52] The incorporation of new adversarial approaches into the training corpus will make neural network reconstructions more robust by identifying nonphysical input perturbations that can negatively impact reconstruction performance. [57,58] Our experiments with motion-encoded k-space demonstrated that as a highly over-parameterized model, AU-TOMAP has significant capacity to learn additional features.…”

Approaches for Image Reconstruction in Low-Field Magnetic Resonance Imaging

“…There is little study in the literature on applying image reconstruction techniques to improve the quality of images from low-eld MRI scanners. These techniques include iterative methods like [14], dictionary learning methods like the approaches proposed in [46], [47], [48], and deep learning methods like AUTOMAP [74]. Experimental results from the literature revealed improved image quality of the reconstructed images.…”

“…However, some studies used synthetic data in their experiments [75], [77], and therefore more experiments are using measured/real-world data. Studies using deep learning approaches in low eld MRI were very limited, during this study, we managed to identify and retrieve only two articles [74], [77]. However, deep learning approaches have been used for image reconstruction in high eld MRI [78].…”

A Streaming Volumetric Image Generation Framework for Development and Evaluation of Out-of-Core Methods

“…For both, the problem of creating not necessarily large, but convincing 3D datasets has already been studied extensively, which also highlights the need for artificial test data in general. VascuSynth [9] is a popular tool (see [15], [16], [20], [36], [39]) for synthesis of arterial vessel trees, corresponding noisy volume data and segmentation based on oxygen demand maps. The software implementation [12], however, does not support out-of-core volume generation since it operates completely in memory.…”