²³Na MRI in ischemic stroke: Acquisition time reduction using postprocessing with convolutional neural networks

Abstract

Quantitative 23Na magnetic resonance imaging (MRI) provides tissue sodium concentration (TSC), which is connected to cell viability and vitality. Long acquisition times are one of the most challenging aspects for its clinical establishment. K‐space undersampling is an approach for acquisition time reduction, but generates noise and artifacts. The use of convolutional neural networks (CNNs) is increasing in medical imaging and they are a useful tool for MRI postprocessing. The aim of this study is 23Na MRI acquisition time reduction by k‐space undersampling. CNNs were applied to reduce the resulting noise and artifacts. A retrospective analysis from a prospective study was conducted including image datasets from 46 patients (aged 72 ± 13 years; 25 women, 21 men) with ischemic stroke; the 23Na MRI acquisition time was 10 min. The reconstructions were performed with full dataset (FI) and with a simulated dataset an image that was acquired in 2.5 min (RI). Eight different CNNs with either U‐Net–based or ResNet‐based architectures were implemented with RI as input and FI as label, using batch normalization and the number of filters as varying parameters. Training was performed with 9500 samples and testing included 400 samples. CNN outputs were evaluated based on signal‐to‐noise ratio (SNR) and structural similarity (SSIM). After quantification, TSC error was calculated. The image quality was subjectively rated by three neuroradiologists. Statistical significance was evaluated by Student’s t‐test. The average SNR was 21.72 ± 2.75 (FI) and 10.16 ± 0.96 (RI). U‐Nets increased the SNR of RI to 43.99 and therefore performed better than ResNet. SSIM of RI to FI was improved by three CNNs to 0.91 ± 0.03. CNNs reduced TSC error by up to 15%. The subjective rating of CNN‐generated images showed significantly better results than the subjective image rating of RI. The acquisition time of 23Na MRI can be reduced by 75% due to postprocessing with a CNN on highly undersampled data.

Compressed Sensing in Sodium Magnetic Resonance Imaging: Techniques, Applications, and Future Prospects

“…Fixed sparsifying transform operators are frequently employed in sodium MRI, such as wavelets, 30 finite differences, 15,33,37 or the orthogonal DCT 30 . Moreover, it is possible to perform sparse representation of sodium images based on a trained dictionary, 33,34,36,37 or others 33,39 …”

“…Lachner et al pioneered the combination of parallel imaging with CS sodium MRI in a study on female breast imaging using a multichannel phased‐array sodium/hydrogen double‐tuned coil and found that the incorporation of parallel imaging improved CS reconstruction with higher image quality 38 . More recently, Adlung et al provided proof for the first time that convolutional neural networks in the field of deep learning were able to reconstruct 4‐fold undersampled sodium MRI images with loss functions acting as regularizations while maintaining SNR and TSC quantification accuracy for ischemic stroke patients 39 . The majority of the research relating to CS‐based sodium MRI have been conducted at an ultra‐high field strength of 7 T using various forms of 3D radial sampling schemes with undersampling factors (USFs) ranging from 2 to 10.…”

“…The majority of the research relating to CS‐based sodium MRI have been conducted at an ultra‐high field strength of 7 T using various forms of 3D radial sampling schemes with undersampling factors (USFs) ranging from 2 to 10. The effects of sparse reconstruction on the quantitation of TSC and relaxometry have been investigated in multiple studies 30,34,39,68–70 . Further to UTE sequences using non‐Cartesian k ‐space trajectories, the chemical shift imaging (CSI) sequence 71 without readout gradients was shown to be a feasible modality for CS sodium MRI by Weingärtner et al, by using iterative reconstruction with the aid of an object support constraint from hydrogen MRI and a phase‐homogeneity constraint 65 .…”

“…Novel CS frameworks with high‐efficient data consistency constraints were proposed by leveraging generative adversarial networks for rapid and accurate image reconstruction from highly undersampled MRI measurements 97,98 . Taking encouragement from the successes of deep learning‐based CS in hydrogen MRI, 94,96–98 recent advances have also addressed sodium MRI 39 . Figure 7 illustrates an example development procedure for a deep learning framework for sodium image reconstruction 39 .…”

“…Ever since Madelin et al demonstrated the applicability of CS in sodium MRI in a study of human knee cartilage, 29 CS has been increasingly applied to sodium imaging of the brain, 30–32 skeletal muscle, 33,34 breasts, 15,35 and human torso 36 . Recent efforts have been made to further advance CS sodium MRI by incorporating methods such as dictionary‐based learning, 36,37 prior hydrogen anatomical constraint, 32 parallel imaging, 38 or deep learning 39 . As an emerging technique, CS has great potential in further facilitating the clinical applicability of sodium MRI by, for example, applying advanced incoherent undersampling methods, 27,40,41 or by accelerating intracellular sodium mapping, 42,43 quantitative relaxometry, 9,10,44 and dynamic sodium MRI 45,46 …”