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

Abstract

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, its performance remains limited by poor image quality. Here, a computational approach is formulated to advance ULF MR brain imaging through deep learning of large-scale publicly available 3T brain data. Methods: A dual-acquisition 3D superresolution model is developed for ULF brain MRI at 0.055 T. It consists of deep cross-scale feature extraction, attentional fusion of two acquisitions, and reconstruction. Models for T 1 -weighted and T 2 -weighted imaging were trained with 3D ULF image data sets synthesized from the high-resolution 3T brain data from the Human Connectome Project.They were applied to 0.055T brain MRI with two repetitions and isotropic 3-mm acquisition resolution in healthy volunteers, young and old, as well as patients. Results:The proposed approach significantly enhanced image spatial resolution and suppressed noise/artifacts. It yielded high 3D image quality at 0.055 T for the two most common neuroimaging protocols with isotropic 1.5-mm synthetic resolution and total scan time under 20 min. Fine anatomical details were restored with intrasubject reproducibility, intercontrast consistency, and confirmed by 3T MRI. Conclusion:The proposed dual-acquisition 3D superresolution approach advances ULF MRI for quality brain imaging through deep learning of high-field brain data. Such strategy can empower ULF MRI for low-cost brain imaging, especially in point-of-care scenarios or/and in low-income and mid-income countries.

Whole-body magnetic resonance imaging at 0.05 Tesla

“…In this study, we have implemented and demonstrated such an image formation method, PF-SR (26), applied to brain, spine, liver, and knee imaging, illustrating the ability of such data-driven image formation in enhancing image resolution while suppressing noise and artifacts. Our previous studies (23,26) and the preliminary brain and spine tests using synthetic datasets in this study have also shown the potential of applying this approach to datasets that contain brain and spine lesions.…”

“…Utilizing deep learning for enhanced image formation at 0.05 Tesla MR signal at 0.05 T is several orders of magnitude weaker than at 3 T, the standard highfield strength, due to its proportionality to field strength squared (B 0 2 ) (32), causing high image noise and poor resolution in ULF MRI. To overcome this challenge, we turned to computing and devised deep learning-based reconstruction methods for ULF MRI image formation that are driven by the large-scale high-field MRI data (23,26). We designed a partial Fourier super-resolution (PF-SR) method that integrates image reconstruction and super-resolution (fig.…”

“…S3. In brief, the model applied multi-scale feature extraction with a residual group (RG) inspired by the residual channel attention network (75) and a modified residual channel attention block for extracting multiscale high-level features (23). Small kernel sizes at the top scale level enabled local image feature extraction, whereas an increased receptive field of 3D convolution layers at middle to bottom scale levels facilitated semi-global image feature learning (76)(77)(78).…”

“…Channel and spatial attentions were utilized to modulate highlevel features based on their inter-channel and inter-spatial relationships (79). The modulated features were then fed into a cascade of RGs, up-sampled to a high-resolution feature space using a 3D sub-pixel convolution layer, and transformed into a high-resolution 3D image residue using a 3D convolution layer (23). The final high-resolution 3D image output was generated by combining the image residue and trilinearly up-sampled model input.…”

“…Concurrently, deep learning advances offer exceptional capabilities for multidimensional feature extraction ( 20 , 21 ), presenting approaches to address the low magnetic resonance (MR) signal-to-noise ratio (SNR) inherent to ULF. For example, deep learning super-resolution strategies have been recently pursued for brain ULF MRI to suppress image noise and boost resolution by leveraging the homogeneous brain structures and image contrasts available in human brain high-field MRI data ( 22 , 23 ). However, these developments have been confined to imaging of the brain ( 9 – 14 ) and extremities ( 24 ).…”