Pushing the limits of low-cost ultra-low-field MRI by dual-acquisition deep learning 3D superresolution

Purpose: Recent development of ultra-low-field (ULF) MRI presents opportunities for low-power, shielding-free, and portable clinical applications at a fraction of the cost. However,


INTRODUCTION
The last four decades witnessed the increased availability and applications of MRI in modern health care. MRI is now a routine procedure in the diagnosis and management of various diseases and injuries. Over 100 million MRI investigations are performed each year worldwide. MRI is intrinsically superior to other imaging modalities, as it is nonionizing, noninvasive, inherently 3D, quantitative, and multi-parametric. 1 These very properties position MRI as an ideal imaging platform for artificial intelligence-driven medical diagnoses in the future. Despite the clear advantages and clinical impact of MRI in health care, this technological advance remains out of reach for many worldwide, even in developed countries. MRI is not generally available outside radiology departments and centralized large imaging facilities in developed countries. It is also rarely available in developing and underdeveloped countries. There are only about 60 000 installations of MRI scanners worldwide (∼7 per million inhabitants based on the Organization for Economic Cooperation and Development, OECD 2018 statistics 2 ) compared with about 200 000 for CT and about 1 500 000 ultrasound scanners. This scenario occurs partially through the high costs of procurement, installation, maintenance, and operation associated with existing superconducting magnet-based, high-field (1.5T and 3T) MRI scanners. They are mostly unavailable in trauma centers, pediatric/neonatal clinics, acute and primary care, surgical suites, and community clinics, where various unmet clinical needs clearly exist for MRI. Globally, MRI procedures remain mostly out of reach for more than two thirds of the world's population categorized by the World Bank as citizens of developing/underdeveloped countries, 2 presenting an exemplary case of ever-expanding global health disparities.
Recently, several academic and commercial research groups around the world have pursued the development of low-power and portable MRI hardware technologies at ultralow-field (ULF) strength (< 0.1 T) for low-cost clinical imaging applications, particularly in the brain (0.05 T, 3,4 0.05 T, 5 0.064 T, 6-8 0.08 T, 9 and 0.055 T 10 ). Essential neuroimaging protocols have been successfully implemented and demonstrated on these ULF scanners, yielding clinically valuable information for stroke and tumor diagnosis within reasonable scan time, including use in intensive care units and coronavirus disease 2019 wards. [6][7][8]10 Recent developments also eliminate the need for traditional RF shielding room through active sensing and cancelation of electromagnetic interference (EMI) signals using analytical and deep learning approaches, [10][11][12] positioning ULF MRI for truly plug-and-scan point-of-care deployment. These advances demonstrate the clear potential to realize patient-centric low-cost shielding-free MRI scanners and democratize MRI for low-income and middle-income countries.
However, the imaging performance of these emerging portable ULF MRI scanners remains limited despite the low cost and ease of operation. 6,[8][9][10] These limitations manifest as low spatial resolution, high level of image noise and artifacts, and long scan time (especially when images of multiple orientations are needed). They arise primarily from the fundamental physics that, with main field B 0 (e.g., 0.055 T) dramatically lower than that for the prevailing superconducting MRI, ULF signal is three orders of magnitude weaker than that at 3 T (the present standard for brain MRI), as signal is approximately proportional to B 0 7/4 . 13 The relatively poor ULF MRI quality leads to suboptimal diagnostic efficacy compared with high-field MRI, [6][7][8]10 which severely impedes the widespread adoption of portable ULF-MRI scanners in clinical practice. Note that this MR signal versus B 0 relationship has been partly exploited in mainstream MRI development by B 0 increase and parallel imaging, together with a myriad of acquisition sequences and reconstruction algorithms, continuously driving the high-field MRI performance and engineering complexity, as well as the costs in the past four decades. To advance low-cost ULF MRI technology, a different approach should be taken, to ensure its eventual clinical adoption.
Deep learning powers a paradigm shift in computational science and engineering. It has shown promise to improve the performance of existing high-field MRI scanners through image reconstruction from undersampled k-space data for fast MRI, image denoising, artifact suppression, and tissue parameter quantitation. [14][15][16][17][18] We envision that deep learning, together with increasing availability of large-scale anatomy-specific and protocol-specific high-quality high-field MRI data sets (e.g., from HCP Consortium and UK Biobank), [19][20][21][22][23] will catalyze and propel the low-cost ULF MRI technological advances for accessible health care. 10 For example, our deep knowledge of 3D human anatomical structures and their MRI characteristics (e.g., contrast, texture, artifacts) has largely been neglected in the image formation/restoration processes. This untapped knowledge and rapidly advancing deep learning algorithms are expected to revolutionize the next advances in MRI.
Recently, several deep learning attempts have been made to improve ULF MRI image quality, but with limited success so far. [24][25][26][27] Image quality transfer (IQT) 28 is also proposed in a preliminary study to create artificially enhanced contrast and increase through-plane resolution of multislice 2D T 1 -weighted MRI images at low field (0.36 T) with anisotropic U-Net. Some of these methods use synthetically degraded high-field images as training data for deep learning restoration of noisy and low-resolution MRI images. 24,25,28 Nonetheless, they all focus on a single 2D or 3D image data set to solve the restoration problem. Yet, ULF MRI greatly suffers from a high level of noise and artifacts. This renders the extraction of lost information from a single ULF MRI data set difficult, leading to suboptimal results.
In this study, we report a deep learning strategy to push the limits of MRI image quality at ULF. Specifically, we developed a dual-acquisition 3D superresolution model to restore fine structural details while suppressing noise and artifacts. We trained this model with 3D image data sets synthesized from large-scale, high-resolution 3T brain data from the Human Connectome Project for two common brain MRI protocols at ULF: T 1 -weighted (T1W) and T 2 -weighted (T2W) contrasts. Second, we successfully applied this model to the low-resolution 3D image data sets acquired using our recently constructed low-cost 0.055T MRI head scanner 10 with 3-mm isotropic resolution. We achieved high ULF image quality with synthetic 1.5-mm isotropic resolution and noise/artifact suppression, revealing fine 3D anatomical details with intrasubject reproducibility and intercontrast consistency in healthy volunteers, and confirmed by 3T MRI. Finally, we also applied this strategy to ULF MRI of elderly subjects and patients with brain lesions.

Dual-acquisition 3D superresolution model description
We design an end-to-end fully 3D architecture in Figure 1 to produce a 3D superresolution or high-resolution output with an up-sampling factor 2× along all three spatial dimensions from two consecutively acquired 3D MR magnitude image data sets (Acq1 and Acq2) of the same contrast. Note that multiple 3D acquisitions are often needed during ULF MRI to improve SNR. The deep learning model consists of three modules: the feature extractor, attentional fusion module, and reconstruction modules.
The purpose of the 3D pyramid feature extractor with residual groups (RGs), inspired by residual channel attention networks, 29 was to first extract two sets of multiscale, high-level features from two low-SNR, low-resolution input image data sets ( Figure 1B). Residual channel attention blocks (RCABs) were modified here to mRCABs by incorporating the multiscale channel attention module. 30 For the fusion module, we specifically designed a spatial self-attention submodule to first fuse two sets of features, and then modulate the fused features through similarity ( Figure 1C). The resulting features were further modulated in the pyramid spatial attention submodule based on interspatial high-level relationships within features 31,32 Finally, modulated fused features were fed into a cascade of mRCABs, up-sampled to the high-resolution space through a 3D variant of subpixel convolution layer 33 using a 2× scale, and transformed into a single superresolution image residue data set through 3D convolutions in reconstruction module ( Figure 1D). The final superresolution 3D image output data set was formed from the superresolution image residue data set and the average of the trilinearly up-sampled Acq1 and Acq2 data sets ( Figure 1A). We optimized this model and key hyperparameters. Full details of the deep learning model architecture are provided in Supporting Information Methods in Appendix S1.

2.2
Model training and validation

Training data preparation
Training data were derived from publicly available high-resolution 3T brain MRI data sets from the WU-Minn Human Connectome Project (HCP S1200). 22 Specifically, we used 1248 3D T1W data sets and 1182 3D T2W data sets, respectively, from healthy subjects (22-35 years old) to generate training data sets. For T1W or T2W model validation and testing, we used additional 268 and 200 data sets from healthy subjects. For model training, all of these original high-resolution 3D image data sets were down-sampled to 1.5-mm isotropic resolution for generating target output image data sets through local mean, and further down-sampled to 3-mm isotropic resolution for generating input image data sets through k-space truncation. K-space truncation was used here for image down-sampling to avoid inverse crime 34 and to ensure that k-space sampling related artifacts (i.e., Gibbs ringing artifacts in all three directions) were present in the low-resolution (i.e., 3 mm) input image data sets. Furthermore, two independent sets of Rician noises 35 were added to each down-sampled, low-resolution (3 mm) single-channel data set to form the final two input image data sets to the model (Acq1 and Acq2). Noise levels for the T1W and T2W data sets were tuned to approximately match those observed in the original ULF brain images.

2.2.2
Model training and testing Two models were trained for T1W and T2W contrasts, respectively. During the training of each model, we randomly extracted four pairs of 32 × 32 × 32 3D patches and four corresponding 64 × 64 × 64 3D patches within the two low-resolution input data sets and one target

F I G U R E 1
An end-to-end fully 3D superresolution (SR) deep learning model for dual-acquisition ultralow-field (ULF) MRI. (A) Proposed deep learning model first extracts deep cross-scale features separately from two independently acquired low-SNR low-resolution (LR) image data sets (Acq1 and Acq2). The features are then fused, modulated, and reconstructed to form a single 3D image data set in SR space with high SNR. The model functions primarily by learning 3D SR image residue (i.e., the difference between SR image and trilinearly interpolated average of Acq1 and Acq2). (B) Feature extractor module has a three-level asymmetric pyramid structure for extracting multiscale and high-level features from two LR data sets. Inspired by residual channel attention networks, residual groups (RGs) are implemented with residual channel attention (RCABs) in the contracting or down-sampling path. The skip connections ease training difficulty and allow passage of low-frequency information. All RCABs are modified to mRCABs by incorporating multiscale channel attention modules to improve the representational power of the model. (C) Extracted features are fused and modulated based on features' similarity in spatial self-attention submodule, and further modulated in pyramid spatial attention submodule. (D) Reconstruction module is composed of RGs containing mRCABs, and a 3D subpixel convolution layer for processing the modulated fused features and up-sampling them to the SR space. The up-sampled features later form into one 3D-SR image residue through convolutions. high-resolution data set, respectively. This was to train the superresolution model to focus on image local characteristics rather than regional or gross structures. Further augmentation was realized by random rotations (at 0 • , 90 • , 180, and 270 • within axial plane to achieve augmentation factor of 4) and flipping (in right-left direction with augmentation factor of 2) on these 3D patches. We optimized the model based on L1 loss with the AdamW optimizer 36,37 with initial learning rate of 0.00003, 1 = 0.9, and 2 = 0.999 for 250 epochs. The learning rate was reduced every 50 epochs by a factor of 0.8. We used a batch size of 8. Regularization to improve model generalization 38 was introduced through a weight decay of 0.01. The total training time was 37.4 h and 37.4 h for the T1W and T2W models, respectively, on three Nvidia RTX A6000 GPUs running Python 3.9 and PyTorch 1.11 with CUDA 11.3. The application time for each pair of low-resolution 3D data sets on a single GPU was 2.2 s.
We used the 268 validation data sets to optimize the model and key hyperparameters by monitoring the L1 loss and peak SNR. Batch size, number of RGs, and number of mRCABs in each RG were tuned based on the validation performance as well as memory consumption and computational time. Each model was evaluated with 200 test data sets. The output-image data sets were compared with the ground-truth results (i.e., noise-free high-resolution data sets from which noisy low-resolution input test data sets were synthesized). Their similarities were evaluated using 3D structural similarity index map values. 39 Two common upsampling/denoising methods were performed for comparison: (1) tricubic interpolation and (2) tricubic interpolation with the standard 3D denoising by BM4D (block matching with 4D filtering). Furthermore, we investigated the importance of two components of the proposed model (in Figure 1) by evaluating the performance of the models with no pyramid-scale feature extraction (i.e., one level only) or dual-acquisition feature fusion module. We also performed experiments to assess the dependence of model performance on training-set size and training patch size.

0.055T MRI and imaging protocol optimization
All ULF MRI data acquisitions were performed on a low-cost, shielding-free, and acoustically quiet 0.055T MRI head scanner recently developed in our laboratory. 10 The scanner is low-power, operates on a standard AC wall power outlet, and consumes less than 1200 W during scanning.
All human subjects were scanned using 3D fast spin echo (FSE)-based T1W and T2W protocols with two averages (i.e., number of excitations [NEX] = 2) and isotropic resolution. Thus, two 3D low-resolution image data sets (Acq1 and Acq2) with uncorrelated noises were obtained for each contrast as required by the proposed deep learning model. Note that our recently demonstrated T1W and T2W contrasts were relatively weak at 0.055 T. 10 In this study, these 0.055T protocols were redesigned and experimentally optimized further in phantoms and healthy subjects, exhibiting brain gray-matter and white-matter T1W and T2W tissue contrasts very similar to those at 3 T. In particular, T1W contrast was produced by the 3D FSE sequence with inversion preparation, unlike the 3D gradient-echo sequence used earlier. 10 Furthermore, two optimized protocols produced isotropic acquisition resolution of 3 mm with acquisition matrix of 80 × 80 × 70, and encoding FOVs of 240 × 240 × 210 mm. Specifically, T1W data were acquired with the 3D inversion FSE sequence using echo train length (ETL) = 10 and TR/TE/TI = 590/14/190 ms. T2W data were acquired with a 3D-FSE sequence with TR/TE = 1500/207 ms and ETL = 21. A 2D elliptical phase-encoding sampling pattern was used for both T1W and T2W protocols. The data-acquisition time was 8.6 min and 11.2 min for T1W and T2W, respectively.
For each ULF MRI protocol, raw k-space data were first processed for EMI removal through a k-space deep learning scheme 10,12 and then reconstructed to two sets of 3D image data sets with 3-mm isotropic resolution, Acq1 and Acq2, corresponding to NEX = 2. They were then fed into the model trained for specific contrast, producing one single superresolution 3D image data set with synthetic 1.5-mm isotropic resolution. Low-resolution Acq1 and Acq2 were averaged to produce a single image data set with 3-mm resolution. Images from experimental ULF data sets without k-space EMI removal were also reconstructed.

3T clinical MRI for comparison
All subjects were also scanned on a clinical GE 3T MRI scanner (Signa Premier) using standard clinical neuroimaging protocols. T1W data were acquired with a MP-RAGE sequence with TR/TR1/TE/TI = 1800/5.4/2.0/900 ms, flip angle (FA) = 8 • , bandwidth = 390 Hz/pixel, acquisition matrix = 160 × 160 × 160, NEX = 1, and scan time = 3.6 min. T2W data were acquired with a 3D SPACE sequence with TR/TE = 3000/68.4 ms, ETL = 130, bandwidth = 780 Hz/pixel, acquisition matrix = 160 × 160 × 160, NEX = 1, and scan time = 2.6 mins. Two protocols produced 1.5-mm isotropic resolution for direct comparison with 0.055T image data sets. Whenever necessary, 3T image data sets were rendered slightly through a simple rigid 3D co-registration (FSL version 6.0.4) with 3D translations and rotations to match the 0.055T data set locations/orientations, to facilitate the visual comparison between 0.055T and 3T MRI data sets. Note that no nonrigid coregistration was performed here to account for ULF MRI gradient field nonlinearity, as this would confound the comparison between ULF and 3T results in terms of local image structures.

Study participants
Thirty healthy volunteers (24-84 years old) were recruited for comparison between 0.055T and 3T MRI. Ten healthy volunteers were recruited for the 0.055T MRI reproducibility test. In addition, 8 patients were recruited from neurology and neurosurgery clinics at Hong Kong Queen Mary Hospital with screening for eligibility based on admission diagnosis, clinical examination, and the need for a clinical MRI follow-up examination. The study was conducted with two institutional review board research protocol approvals.

Application of deep learning model to 0.055T MRI
To evaluate the deep learning model for T1W and T2W imaging, 0.055T data sets were fed to the respective models trained by data sets synthesized from high-field data. The resulting images with synthetic 1.5-mm isotropic resolution were carefully examined by (1) directly comparing images details between 0.055T and 3T images at different orientations; (2) consistency test (i.e., consistency of structural/anatomical details revealed by 0.055T image data sets of different contrasts acquired from the same subject during the same imaging sessions), as determined by the well-understood image-intensity behaviors of various brain tissues in T1W and T2W images; and (3) reproducibility test (i.e., reproducible structural/anatomical details among 0.055T image data sets of specific contrast acquired from the same subject but different imaging sessions).

Model testing
The model testing results demonstrated significant image-quality improvement with enhanced resolution and extremely low level of noise and artifacts for both contrasts (Figure 2 and Videos S1 and S2). Numerous fine structural details were successfully recovered with high 3D fidelity using the proposed deep learning model. Figure S1 in Appendix S1 shows the quantitative comparison to the high-resolution ground-truth reference and results from simplified models without pyramid-scale feature extraction or dual-acquisition fusion module, highlighting the importance of these two modules in the proposed SR model. It also presents the quantitative comparison to two conventional upsampling and 3D denoising approaches (i.e., tricubic interpolation and tricubic interpolation with BM4D denoising), clearly illustrating the advantage of the proposed deep learning model approach ( Figure S1 in Appendix S1 and corresponding axial images in Figure S2A in Appendix S1). Careful visual inspection of the superresolution results in Figures S2B and S3 in Appendix S1 also revealed some local blurring. Note that, unsurprisingly, some small but incorrect local structures (i.e., hallucinations) were also observed, although they did not occur often. Figures S4 and S5 in Appendix S1 show the dependence of proposed model on training data set size and the patch size, respectively. Model performance diminished with decreased training set size. Small training patch sizes (e.g., below 32 × 32 × 32) could not sufficiently capture the brain 3D features and led to image-quality degradation, whereas the large one (i.e., 40 × 40 × 40) did not improve model performance despite the significantly increased GPU memory use and computational time.

Evaluation of deep learning ULF MRI in volunteers
We applied the two 3D superresolution models trained by synthetic data sets to experimental low-resolution ULF MRI data sets acquired using a low-cost low-power permanent magnet 0.055T MRI head scanner. T1W and T2W models significantly increased 0.055T brain image quality through resolution enhancement, and effective noise and artifact removal in all volunteers studied. Figure 3 shows the typical superresolution results from 1 volunteer, together with corresponding 3T images with 1.5-mm isotropic resolution. Videos S3 and S4 present the complete 3D T1W and T2W results, respectively, from the same volunteer. Note that 3T data sets were coregistered to 0.055T data sets here using simple rigid 3D translations and rotations to facilitate visual comparison. Deep learning transformed the original 3-mm resolution images with dramatically improved clarity through synthetic 1.5-mm resolution and suppression of noise and artifact texture. They exhibited clear demarcations between gray matter, white matter, and CSF that could barely be perceived visually, if not imagined, from the original 0.055T images, even through careful examination of consecutive slice at

F I G U R E 2
Deep learning model testing with dual-acquisition LR noisy 3D image data sets synthesized from publicly available Human Connectome Project (HCP) 3T human brain data. Two models were trained for T 1 -weighted (T1W) and T 2 -weighted (T2W) contrasts, respectively. (A) LR input data sets, LR Acq1 and LR Acq2, are two independent 3D noisy data sets with 3-mm isotropic resolution. Their average, LR, is also shown. The SR output data set has 1.5-mm isotropic resolution. Axial images are shown for each contrast from the same subject, together with a high-resolution, noise-free, ground-truth reference data set. (B,C) Coronal (B) and sagittal (C) images from the same 3D data sets. Deep learning achieved significant image-quality improvement by restoring details, increasing resolution, and suppressing noise and artifacts.

F I G U R E 3
Application of deep learning models to experimental dual-acquisition low-resolution 3D brain data sets acquired from 1 healthy volunteer (33 years old male) using a low-cost shielding-free 0.055T MRI head scanner. Two models were trained using synthetic data for T1W and T2W contrasts, respectively. (A) Low-resolution input data sets, LR Acq1 and LR Acq2, are two independent 3D data sets with 3-mm isotropic resolution. Their average, LR, and counterparts before k-space electromagnetic interference (EMI) removal are also shown. The deep learning superresolution output data set, SR, has synthetic 1.5-mm isotropic resolution. Four axial images are shown for each contrast, together with corresponding reference images acquired from the same volunteer using a standard 3T clinical MRI scanner with 1.5-mm isotropic resolution. (B) Axial images with zoom views indicating the restoration of numerous structural details and dramatic suppression of noise and artifacts. The total scan time for acquiring 0.055T T1W and T2W data sets for two contrasts was 19.8 min. Note that, for comparison, 3T data sets were coregistered to the 0.055T data sets using simple rigid 3D translations and rotations.

F I G U R E 4
Multi-orientation views of the 0.055T and 3T results shown in Figure 3. (A,B) Coronal and sagittal T1W and T2W images (A) and triplanar images (B) from original LR input and deep learning SR output data sets, together high-resolution (HR) 3T MRI data sets. Brain cortical surface renderings are also displayed. These multiplanar comparisons indicate the excellent restoration of numerous 3D structures, as directly confirmed by 3T images with 1.5-mm isotropic resolution. Note that the severe CSF Gibbs ringing artifacts in LR T2W images (pointed by green arrows) were effectively removed and became absent in SR results. various orientations. Such deep learning transformation of low-resolution noisy images was effective for both T1W and T2W protocols. As a result, fine anatomical details, such as cortical gyri/sulci, putamen, caudate nucleus and internal capsule, were readily identified in superresolution T1W and T2W images (Figure 3, Videos S3 and S4). More importantly, the boundaries of these restored structures and their often bilaterally asymmetric distributions were consistently in excellent agreement with high-resolution 3T images. It is worth noting that the hypointensity of globus pallidus (in axial T2W images in Video S4) due to iron deposition was present only at 3 T but not at 0.055 T due to the significantly reduced magnetic susceptibility effect at 0.055 T, as expected. Figure S6 in Appendix S1 compares the results in Figure 3 with those using common 3D interpolation and denoising methods. Figure 4 provides the multiplanar views from the same data sets, together with brain cortical surface renderings. Note that the severe CSF Gibbs ringing artifacts visible in original low-resolution coronal and sagittal T2W images were effectively removed ( Figure 4A). Examining these images and videos supported that deep learning models enabled high-quality restoration of numerous fine 3D structural details throughout the entire 3D brain volume in 0.055T superresolution T1W and T2W results, as confirmed by 3T results.
For normal brains, T1W and T2W images typically provide highly complementary contrasts for identifying main brain tissue types (i.e., gray matter, white matter, and CSF). For example, gray-matter intensity is low in T1W but high in T2W, while white-matter intensity is high in T1W but low in T2W. In this study, T1W and T2W models were trained separately with independently synthesized 3D T1W and T2W data sets, respectively. Low-resolution 3D ULF T1W and T2W input data sets were also independently acquired and then fed to T1W and T2W models, respectively. Figure 5 compares various gray matter, white mater, and CSF structures in superresolution T1W and T2W data sets at two different slices and three orientations from 1 healthy volunteer. Videos S5-S7 present the complete 3D T1W and T2W results in axial, coronal and sagittal orientations, respectively. The 0.055T superresolution results revealed the same structures (e.g., cortical gyrus/sulcus, white and gray matter, and CSF) that were highly correlated spatially among T1W and T2W contrasts. This finding indicated that likely the fine T1W and T2W

F I G U R E 5
Contrast consistency of various brain tissues among T1W and T2W deep learning SR 0.055T images from a healthy volunteer (27-year-old male). Two slices are shown for each orientation. The intensity and boundary of various gray mater, white mater, and CSF structures appear to be highly correlated spatially between the T1W and T2W images, indicating that image details restored by deep learning were not artifactual, but converged among independent deep learning outcomes of two contrasts. image details restored by deep learning were not artifacts or incidental. Instead, they converged independently to the identical structural locations and boundaries. Such complementary consistency in structures among T1W and T2W contrasts further supported the high-fidelity efficacy of the proposed deep learning restoration of 0.055T image quality.
We also performed reproducibility tests by repeatedly acquiring 0.055T low-resolution data sets from the same subjects at different imaging sessions. Superresolution 3D data sets were generated by the models, and then coregistered through simple rigid 3D rotations and translations to correct the head position differences among imaging sessions. High reproducibility was observed in all 5 subjects who underwent these tests for T1W and 5 subjects for T2W. Figure 6 and Figure S7 in Appendix S1 present the typical reproducibility of deep learning-enabled high-resolution 3D T1W and T2W results from 2 volunteers, respectively. By repetitively restoring the same structures throughout the entire 3D brain volume, they demonstrated the high stability of the restored image details by the proposed strategy. Figure 7 presents the original low-resolution and superresolution 0.055T data sets from 2 elderly volunteers exhibiting aged-related brain atrophy. Superresolution images revealed fine details in T1W and T2W images as in other healthy volunteers. Moreover, they also reliably detected brain atrophy (i.e., central ventricle enlargement, expanded CSF space, gray and white matter shrinkage, and bilateral asymmetry). Their spatial extents, boundaries, and contrasts were in excellent agreement with those revealed by 3T MRI, demonstrating the robustness of our proposed deep learning strategy in capturing brain structural aberrations. Figure 8 presents the preliminary experimental 0.055T superresolution results from a patient with chronic ischemic stroke, whereas Figure S8 in Appendix S1 shows the synthetic superresolution results from three high-field T1W and T2W data sets containing various pathologies.

F I G U R E 7
Application of deep learning models to LR 0.055T MRI data sets acquired from 2 elderly subjects: 67-year-old female (A) and 69-year-old male (B). The LR input data sets have 3-mm isotropic resolution. Deep learning SR output data sets have 1.5-mm isotropic resolution. Four axial images are shown for each contrast, together with corresponding HR 3T images with 1.5-mm isotropic resolution. Different extents of brain atrophy, asymmetry, and ventricular enlargement are reliably detected in the 2 subjects, as confirmed by 3T MRI, indicating the reliable restoration of many anatomical details in the presence of brain structural changes due to aging.

F I G U R E 8
Application of deep learning models to low-resolution 0.055T MRI data sets acquired from 1 patient: a 58-year-old female with a small lacunar infarct in the left putamen due to chronic (>11 months) ischemic stroke. (A) Four consecutive axial slices from LR input data sets and deep learning SR output data sets are shown, together with corresponding HR 3T images. (B) Enlarged representative slice for the lesion; location is indicated by the green arrow. All lesion locations and boundaries detected by deep learning SR images were confirmed by 3T MRI.

DISCUSSION AND CONCLUSIONS
In this study, we sought to tackle a major challenge in modern health care: the scarce and limited access to MRI. Our strategy advanced portable ULF MRI quality through deep learning of high-field MRI data. ULF MRI scanners are low-cost to manufacture, maintain, and operate. ULF MRI also holds several inherent advantages to high-field MRI, including open magnet configuration for patient comfort, low acoustic noise levels during scanning, low sensitivity to metallic implants, less image susceptibility artifacts at air/tissue interfaces, and extremely low RF specific absorption rate. 9,10,13,40-42 These features, together with complete RF and magnetic shielding elimination, present ULF MRI as a simpler and more patient-friendly alternative and complement to existing high-field MRI.
At present, the major roadblock to ULF MRI is the significant MR signal loss at ULF, resulting in limited image quality and diagnostic value. This low MR signal challenge at ULF could be partly mitigated through generic adoption of emerging image-reconstruction algorithms such as fingerprinting for fast and quantitative MRI. 24,43 However, such potential remedy is likely insufficient to adequately offset the drastic MR signal loss at ULF. 44 Thus, it is imperative to explore and pursue different rationales and approaches to push the limits of ULF MRI. One untapped approach is to exploit the omnipresent 3D structural features shared across humans for all organs including the brain. They arise from genetically predefined human anatomy, and manifest as a broad range of 3D structures and content characteristics in MR images of various contrasts acquired with different imaging protocols. Such unique commonalities have been mostly neglected in MRI technology development, despite the increasing availability of various large-scale human MRI data sets. Most MRI postprocessing methods for improving image quality, both analytical and deep learning-based ones, are often not focused on isotropic 3D images, which may limit their eventual prowess and efficacy.

Three-dimensional superresolution with noise and artifact suppression for isotropic ULF data
In this study, we devised an end-to-end fully 3D superresolution deep learning strategy specifically for ULF MRI that integrates dual image acquisitions and multiscale high-level feature extraction and attentional fusion. This model effectively recovers fine structural details from two low-resolution data sets while suppressing image noise and artifacts. The model was trained for the two most common neuroimaging protocols (i.e., T1W and T2W) using data sets synthesized from publicly available large-scale 3T human brain MRI data. We implemented the raw data dual-acquisition protocols on a low-cost shielding-free 0.055T MRI head scanner recently developed in our laboratory, and combined them with our proposed superresolution strategy. The results indicated that our models could effectively achieve high-quality brain MRI at 0.055 T with significantly enhanced spatial resolution and extremely low level of noise and artifacts for 3D T1W and T2W imaging (with total scan time under 20 min). These superresolution 0.055T experimental images with synthetic 1.5-mm isotropic resolution could capture many fine 3D brain structures that were consistent among different contrasts, reproducible, and confirmed by 3T MRI. Note that such isotropic 3D imaging eliminates the need for imaging at multiple orientations, which are common in existing MRI procedures at the expense of additional scan time.
Our proposed dual-acquisition approach uses two consecutive 3D image acquisitions, exploiting the nature of multiple acquisitions unique to ULF MRI, where averaging is often needed to gain image SNR. [3][4][5][9][10] Such 3D acquisitions also take advantage of the high acquisition efficiency of 3D encoding sequences at ULF because of the much shorter T 1 values for key brain tissues at ULF. 10,13,45 Meanwhile, leveraging the 3D isotropic nature of anatomical features, the proposed strategy yields dramatic image-quality improvement by restoring numerous anatomical details in ULF images, especially compared with original low-resolution raw images frame by frame (see Video S8 for results from the same volunteer shown in Figures 3 and 4). Such performance of our proposed method also highlights the limitation of human 2D vision, as demonstrated by the difficulty if one attempts to visually relate various 3D structures before and after 3D superresolution. Many local structures are indeed strenuous to perceive visually from the raw multislice 2D images, if not difficult to imagine. On the other hand, such disconnection in human visual perception is not entirely surprising, as our model is designed, trained, and applied in a truly 3D and isotropic manner with deep cross-scale feature extraction and fusion, whereas human vision is inherently 2D to a large extent. Note that this observation also underscores the necessity and power of direct 3D deep learning in future artificial intelligence-based identification of various brain 3D local lesions or global structural abnormalities for medical diagnosis.
Another particular feature of our deep learning model is its ability to suppress artifacts. For example, severe CSF Gibbs ringing artifacts along all three directions caused by the narrow isotropic k-space sampling range during low-resolution ULF MRI data acquisition are completely eliminated (see coronal and sagittal T2W images in Figure 4A, Figure S6 in Appendix S1, and Videos S3 to S8). With such artifact patterns embedded in training input data but not output/target data, deep learning is highly effective in learning and removing the semiglobal, physics-dictated artifacts for high-resolution MRI, as already demonstrated for high-field MRI by relatively simple convolution neural networks. 46,47 It is possible to expand our proposed model to remove other important MRI artifacts, such as partial Fourier image reconstruction artifacts, 48,49 to accelerate ULF MRI data acquisition or improve its image quality for a given scan time.
Our proposed 3D superresolution model not only restores ULF image details through resolution enhancement and effective noise/artifact suppression, but also results in overall image appearance similar to existing high-field MRI. This feature may directly benefit the adoption of ULF MRI in routine clinical practice, as clinicians are often trained through years of experience in reading high-field MRI images. Our approach here may provide a smooth transition for diagnostics using ULF images exhibiting overall appearance or attributes indistinguishable from high-field images. The quality of ULF 3D T1W and T2W images with isotropic 1.5-mm resolution may be sufficient for certain existing MRI applications (such as transcranial magnetic stimulation, deep brain stimulation, and RF ablation).

Other recent 3D superresolution approaches
One very recent study (LF-SynthSR 50 ) explores deep learning superresolution for enhancing ULF MRI, which is based largely on a previously developed method (SynthSR 51 ) from the same group. It shows promising T1W brain image results from commercial 0.065T Hyperfine MRI data. While our work shares some similarities with LF-SynthSR, our approach differs from LF-SynthSR as well as IQT 28 proposed in an earlier preliminary study. First, our proposed method tackles ULF 3D imaging through two repetitive single-contrast acquisitions at 0.055 T with low and isotropic resolution, whereas LF-SynthSR and IQT deals with two highly anisotropic 3D data sets (with two distinct contrasts) or a single multislice 2D data set, respectively. Thus, these three methods entail different deep learning features and architecture design, and are tailored for different application scenarios. Note that LF-SynthSR uses paired ULF T1W and T2W images and an additional segmentation training loss to synthesize high-quality T1W MPRAGE-like images, especially for brain segmentation and volumetric measurement applications. Our dual-acquisition strategy, on the other hand, provides a contrast-specific implementation for enhancing ULF T1W or T2W images with isotropic resolution, which are preferable in typical neuroimaging protocols.

Challenges and limitations
Future challenges for ULF MRI lie in its clinical adoption and integration into health care settings for specific applications in disease theranostics. Application of the proposed method to patients with lesions is promising (Figure 8 and Figure S8 in Appendix S1). However, any eventual clinical application requires further careful and comprehensive optimization and evaluation. First, the performance of our proposed strategy is not yet fully evaluated in this study. It needs to be more quantitatively assessed, such as through more elaborate brain-tissue segmentation/volumetric measurements 50,51 and/or quantitative comparison to experimental high-field MRI as the ground truth. Second, studies should evaluate and optimize the sensitivity and specificity of our proposed deep learning-enabled ULF MRI in detecting various types of brain diseases or injuries with different extents and locations as well as contrasts. These future studies will fine-tune the models for optimal balance among hardware, raw image SNR and resolution, scan time, and contrast (including close contrast matching between ULF and high-field images for both normal and various pathological tissues). For example, our proposed strategy can resolve numerous image details with large cerebrum but not the fine structures in the lower brain regions, especially the cerebellum, when compared with 3T MRI. The detailed gray-matter/white-matter structures within cerebellar hemispheres are blurred in both testing results (Figure 2 and Videos S1 and S2) and experimental ULF results (Figures 4 and 5, and Videos S3-S8), likely due to the much lower image SNR in cerebellum because of low cerebellum signal intensity and diminishing RF coil detection sensitivity in the region. Third, it is imperative to further reduce ULF MRI scan time, or increase image resolution for given scan time, such as by extending deep learning to image formation from highly incomplete or undersampled k-space data. Fourth, it is necessary to extend the proposed strategy to other key neuroimaging protocols, such as fluid-attenuated inversion recovery and DWI. Note that we and others have recently demonstrated the feasibility of fluid-attenuated inversion recovery and DWI at ULF. [6][7][8]10 Finally, ULF-MRI protocols and their contrast-related parameters will need to be further optimized for effective delineation of various type of brain diseases and injuries through extensive clinical investigations. This is necessary because MR relaxation properties of various types of lesions are expected to differ to a certain extent from those at high field. Together with sequence parameters, they may not lead to identical lesion contrasts at ULF. One possibility is to train our model with simulated and/or real pathological data sets with varying lesion appearances and contrasts. Future ULF MRI technology development will also encompass areas unique to ULF, including biophysics. For example, biological tissues are expected to exhibit dramatically shorter T 1 and longer proton T 2 (as well as T 2 *) at ULF. 13,45,[52][53][54] This can lead to more time-efficient data-acquisition schemes, given the significantly fast longitudinal magnetization recovery during repetitive excitations, slow transverse magnetization decay during signal readout, and extremely low-RF specific absorption rate. Another avenue is to increase MR signal detection sensitivity by cryogen-cooled 55 or cryogen-free conduction-cooled RF coils using increasingly available cryocooler technology. Such approach is particularly effective for ULF MRI because, at very low resonance frequency, the noise in the MR signal is dominated by RF receiver coil noise, as the sample noise is negligible. 56 In conclusion, we developed a novel dual-acquisition 3D superresolution method to push the present performance limits of ULF MRI by restoring image details. Our method effectively enhances image resolution and suppress noise/artifacts by learning from large-scale, high-quality, high-field MRI data. The proposed strategy enables brain imaging with high quality at 0.055 T with synthetic 1.5-mm isotropic resolution for two common protocols. This strategy may pave the way for accessible and low-cost portable ULF MRI to its practical adoption in clinical applications, especially in point-of-care scenarios and/or in low-income and mid-income countries.

SUPPORTING INFORMATION
Additional supporting information may be found in the online version of the article at the publisher's website.
Video S1. Deep learning model of T 1 -weighted (T1W) testing results from the same subject as in Figure 2.
Video S2. Deep learning model of T 2 -weighted (T2W) testing results from the same subject as in Figure 2.
Video S3. Applying deep learning models to low-resolution 0.055T ultralow-field (ULF) T1W MRI as in Figure 3.
Video S4. Applying deep learning models to low-resolution 0.055T ULF T2W MRI as in Figure 3.
Video S5. Axial T1W and T2W image-contrast consistency from the same subject as in Figure 5.
Video S6. Coronal T1W and T2W image-contrast consistency from the same subject as in Figure 5.
Video S7. Sagittal T1W and T2W image-contrast consistency from the same subject as in Figure 5.
Video S8. Low-resolution (LR) and superresolution (SR) T1W and T2W results with and without the proposed deep learning from the subject shown in Figure 3.