Accelerated motion robust low-field neonatal magnetic resonance imaging system and method

Abstract

An exemplary AI system and method are disclosed for reconstructing high-fidelity motion-corrected MRI scans from low-field, motion-corrupted MRI scans (e.g., neonatal MRI data) using a neural network (NN) model in combination with a physics-based motion model. The exemplary AI system and method can enhance image quality and remove motion artifacts which can allow for scans to be performed with significantly reduced scan time, a particular benefit for neonatal subjects as well as all patients when employing low-field MRI.

Description

Attorney Docket No.10046-657WO1 8531 ACCELERATED MOTION ROBUST LOW-FIELD NEONATAL MAGNETIC RESONANCE IMAGING SYSTEM AND METHOD Related Applications This U.S. patent application claims priority to, and the benefit of, U.S. provisional application, US 63 / 708,608, filed on October 17, 2024, and U.S. provisional application, US 63 / 715,881, each of which is hereby incorporated by reference herein in their entirety. Background

[0001] Low-field magnetic resonance imaging (MRI) systems typically provide diagnostic image quality using a lower field strength (e.g., under 1 Tesla). Low-field MRI systems are safe for assessing potential brain abnormalities of infants during the neonatal period, among other usages. However, because of low field strength, low-field MRI can experience long scan times and motion aliasing artifacts, which can produce non-diagnostic images. While AI systems are being employed to reduce noise and improve the reconstruction of MRI images and other imaging modalities, their use for neonatal scans, particularly for neonatal brains, has been challenging because of the lack of training data for newborns. Motion artifacts are also a particularly challenging issue during the imaging of newborns and neonatal subjects.

[0002] There is a benefit to improving low-field MRI systems. Summary

[0003] An exemplary AI system and method are disclosed for reconstructing high- fidelity motion-corrected MRI scans from low-field, motion-corrupted MRI scans (e.g., neonatal MRI data) using a neural network (NN) model (e.g., a diffusion model) in combination with a physics-based motion model. The exemplary AI system and method can enhance image quality and remove motion artifacts, which can allow for scans to be performed with significantly reduced scan time, a particular benefit for neonatal subjects as well as all patients when employing low-field MRI. The implementation, once developed for low-field MRI, can be readily applied to conventional MRI systems to generate super- resolution images from standard resolution scans. The exemplary AI system and method address a key technical challenge associated with neonatal scans – the scarcity of training data of neonatal subjects. The exemplary AI system and method provide an improvement for computing technology in facilitating the training of an AI system using scarce or undersampled training data where such training data would not overwise be unavailable.Attorney Docket No.10046-657WO1 8531

[0004] As used herein, the term “scarce” or “undersampled” in the context of training data refers to an insufficient amount, i.e., not adequately sampled or surveyed, for use in conventional training methodology to meet the technical needs for the training to produce a statistically and clinically meaningful outcome.

[0005] While motion artifacts in adult MRI scans can be mitigated by reducing scan times through undersampled acquisitions and leveraging a variety of correction techniques and combining parallel imaging, the unique challenge presented by the low-field neonatal MRI setting precludes the direct application of these accelerated MRI techniques as those employed for adults. Many low-field systems measure signal with a single-channel receive array, so parallel imaging cannot be applied. Pre-trained models from adult patients cannot be used because brain structure can vary greatly between neonates and adults. In addition, the images are inherently noisier in lower-field neonatal MRI, and significantly less data has been acquired and publicly shared, making it challenging to train machine learning reconstruction models, from both a data quality and quantity perspective. Recent machine learning algorithms for accelerated MRI reconstruction yield state-of-the-art results by employing a point-wise mapping between undersampled and fully-sampled data; the mapping is highly susceptible to test time shifts in the measurement operator. More recently, generative methods can be developed that can learn a prior over clean images with robustness to test time shifts in the forward operator; however, the technique requires high-quality image data (i.e., high SNR) for training, which is unavailable for neonatal subjects.

[0006] A study was performed that developed an example of the exemplary AI system and method by (i) modifying existing popular diffusion network architectures to support inputs with varying matrix sizes (a common technical issue in MRI, therefore expanding the set of potential training images), (ii) training a single model on all data with class embeddings (rather than stretching the dataset thin by training a separate model for each image contrast and orientation, (iii) employing self-supervised de-noising model to boost the SNR of the dataset before training. The study observed a reduced scan time of single-coil Fast Spin Echo and Spin Echo sequences by an average of 1.5× while reconstructing high- fidelity images from realistically under-sampled measurements. The study also observed substantial motion correction.

[0007] Different from conventional MRI systems, which can have extended acquisition times due to frequent motion artifacts, the exemplary system and method facilitate accelerated imaging and motion correction using multi-channel receive arrays. Different from adult MRI reconstruction methods, which rely on high-SNR datasets and pre-trained modelsAttorney Docket No.10046-657WO1 8531 that do not generalize to neonatal brain anatomy, the exemplary system and method are tailored to neonatal imaging through the above-discussed training technique that addresses both data scarcity and anatomical variability. Different from current machine learning methods, which require large, high-quality datasets, the exemplary system and method uses generative modeling and self-supervised denoising to improve robustness and performance in low-SNR, limited-data environments.

[0008] In some implementations, a NN model (e.g., diffusion generative model) is trained using a scarce neonatal MRI dataset, where training inputs are (i) encoded with class embeddings of motion-corrupted scans at different orientation and contrasts and (ii) employed denoised versions of the same scarce dataset serve as ground truth generated by a second trained AI model configured for noise removal. The approach facilitates the training of an NN model that can used for low-field neonatal MRI scans and generalizable to non- neonatal patients. The exemplary AI training is generalizable across varying image contrasts and orientations while compensating for low signal-to-noise ratio (SNR) and limited data availability. In some implementations, the physics-based motion model performs motion- informed posterior sampling to correct motion artifacts in the output of the trained NN model, resulting in motion-free, high-fidelity MRI scans.

[0009] By integrating the above-discussed training technique to generate a neural network model that can operate with state-of-the-art physics-informed motion correction, the exemplary system and method can improve the accessibility and diagnostic utility of neonatal MRI within the neonatal intensive care unit (NICU). The exemplary system and method can enable faster, higher-quality imaging without the need for sedation or patient transport, thereby reducing risk and expanding clinical access to neonatal populations.

[0010] In an aspect, a method is disclosed comprising: receiving, by a processor, an MRI image or data object of a subject acquired by an MRI scanner; generating, via a trained neural network model and physics-based motion model operating in combination, a higher- fidelity image or data object of the subject that reduces motion and noise from the received MRI image or data object; and outputting the higher-fidelity image or data object, wherein the higher-fidelity image or data object is employed for diagnosis or treatment of the subject, wherein the trained neural network model was generated by (i) training a denoising AI model using scarce training dataset, (ii) generating a denoised scarce training dataset from the scarce training dataset using the trained denoising AI model, and (iii) applying the scarce training dataset as input to the training of the neural network model, wherein the denoised scarce training dataset is employed as a ground truth during the training.Attorney Docket No.10046-657WO1 8531

[0011] In some embodiments, the scarce training dataset includes scans of the subject for a plurality of orientations and contrasts, and the training of the neural network model employed the plurality of orientations and contrasts with class embedding.

[0012] In some embodiments, the scarce training dataset includes low-field MRI scans, and the low-field MRI scans were acquired with reduced acquisition time.

[0013] In some embodiments, the scarce training dataset includes low-field neonatal MRI scans, and the low-field neonatal MRI scans were acquired with a reduced acquisition time of a moving subject during the scan.

[0014] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan.

[0015] In some embodiments, the at least two orientations are selected from the group consisting of axial orientation, sagittal orientation, coronal orientation, and oblique orientation.

[0016] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, wherein class embedding is applied for training of the neural network model for each of the at least two orientations for the given contrast.

[0017] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations, each for at least two contrast scans selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, wherein class embedding is applied for training of the neural network model for each of the orientations and contrasts.

[0018] In some embodiments, the MRI image or data object includes 2D MRI scans or 3D MRI scans.

[0019] In some embodiments, the MRI image or data object is acquired by a low-field MRI scanner.

[0020] In some embodiments, the trained neural network model is a trained diffusion generative model.

[0021] In some embodiments, the denoising AI model was trained using only the scarce training dataset and a statistical model of a noise distribution of the scarce training dataset.

[0022] In another aspect, a method is disclosed comprising: receiving, by a processor, an MRI image or data object of a neonatal subject acquired by a low-field neonatal MRIAttorney Docket No.10046-657WO1 8531 scanner; generating, via a trained neural network model and physics-based motion model operating in combination, a higher-fidelity image or data object of the neonatal subject that reduces motion and noise from the received MRI image or data object; and outputting the higher-fidelity image or data object, wherein the higher-fidelity image or data object is employed for diagnosis or treatment of the neonatal subject, wherein the trained neural network model was generated by (i) training a denoising AI model using scarce neonatal training dataset, (ii) generating a denoised scarce training dataset from the scarce neonatal training dataset using the trained denoising AI model, and (iii) applying the scarce neonatal training dataset as input to the training of the neural network model, wherein the denoised scarce training dataset is employed as a ground truth during the training, wherein the scarce neonatal training dataset includes scans of the neonatal subject for a plurality of orientation and contrast, and wherein the training of the neural network model employed the plurality of orientation and contrast with class embedding.

[0023] In some embodiments, the scarce neonatal training dataset includes a low field MRI scans, the low field MRI scans was acquired with reduced acquisition time of a moving neonatal subject during the scan, and the MRI image or data object of the neonatal subject was acquired a moving neonatal subject during the scan.

[0024] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan, and the at least two orientations are selected from the group consisting of axial orientation, sagittal orientation, coronal orientation, and oblique orientation.

[0025] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, and each class embedding is applied for training of the neural network model for each of the at least two orientations for the given contrast.

[0026] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations, each for at least two contrast scans selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, and each class embedding is applied for training of the neural network model for each of the orientations and contrasts.

[0027] In some embodiments, the MRI image or data object includes 2D MRI scans or 3D MRI scans.Attorney Docket No.10046-657WO1 8531

[0028] In some embodiments, the MRI image or data object is acquired by a low-field MRI scanner.

[0029] In some embodiments, the trained neural network model is a trained diffusion generative model.

[0030] In some embodiments, the denoising AI model was trained using only the scarce training dataset and a statistical model of a noise distribution of the scarce training dataset.

[0031] In yet another aspect, a system is disclosed comprising: a processor; and a memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to: receive an MRI image or data object of a subject acquired by an MRI scanner; generate, via a trained neural network model and physics-based motion model operating in combination, a higher-fidelity image or data object of the subject that reduces motion and noise from the received MRI image or data object; and output, the higher-fidelity image or data object, wherein the higher-fidelity image or data object is employed for diagnosis or treatment of the subject, wherein the trained neural network model was generated by (i) training a denoising AI model using scarce training dataset, (ii) generating a denoised scarce training dataset from the scarce training dataset using the trained denoised AI model, and (iii) applying the scarce training dataset as input to the training of the neural network model, wherein the denoised scarce training dataset is employed as a ground truth during the training.

[0032] In some embodiments, the system is configured by any one of the above- discussed methods.

[0033] In yet another aspect, a non-transitory computer-readable medium having instructions stored thereon is disclosed, where execution of the instructions causes a processor to execute any one of the above-discussed systems or to perform any one of the above-discussed methods.

[0034] In yet another aspect, a method is disclosed of training an AI model comprising: receiving, by a processor, a scarce training dataset comprising an MRI image or data object of a subject acquired by an MRI scanner; training a denoising AI model using the scarce training dataset; generating a denoised scarce training dataset from the scarce training dataset using the trained denoising AI model, and applying the scarce training dataset as input to the training a the neural network model (e.g., diffusion generative model), wherein the denoised scarce training dataset is employed as a ground truth during the training in combination with the input.Attorney Docket No.10046-657WO1 8531 Brief Description of Drawings

[0035] Figs.1A – 1B each shows an example system for reconstructing a low-field or low-resolution magnetic resonance imaging (MRI) scan into a higher-fidelity MRI scan with reduced motion artifacts and noise, in accordance with an illustrative embodiment.

[0036] Fig.2A shows an example operational flow of the exemplary system, in accordance with an illustrative embodiment.

[0037] Fig.2B shows an example method of training a neural network (NN) model of the exemplary system, in accordance with an illustrative embodiment.

[0038] Fig.3A shows an example Motion Informed Posterior Sampling (in the exemplary system) that combines a generative learning model with a motion robust MRI measurement model to reconstruct high-fidelity Neonatal MRI images in the presence of rigid motion.

[0039] Fig.3B shows how the exemplary system can mitigate motion artifacts in a real sagittal Neonatal scan by combining motion modeling, generative models, and data from two acquisitions.

[0040] Figs.4A – 4C shows an example training process for a neural network (NN) model (e.g., generative diffusion model) in the exemplary system and an example inference pipeline for the trained NN model.

[0041] Fig.4D shows examples of denoised training samples and prior samples generated with class embeddings using the exemplary system and its trained NN model.

[0042] Fig.4E shows an example algorithmic implementation of the generation of the high-fidelity MRI image from the low-field MRI image in the inference pipeline.

[0043] Fig.5A shows reconstruction results for each orientation and contrast comparing the baseline L1-wavelet to an experimental system (as described in Figs.1A – 1B) using a generative model trained with and without denoising on the training dataset.

[0044] Fig.5B shows how the experimental system / model estimated motion parameters and yielded cleaner results than the original motion-corrupted data on example axial and coronal slices from the test dataset.

[0045] Fig.6A shows the accelerated MRI reconstructions using (i) experimental models trained on each contrast and orientation separately versus (ii) experimental models trained on all data with and without class embeddings.

[0046] Fig.6B shows the experimental model applied to the task of motion correction on measured, motion-corrupted clinical data.Attorney Docket No.10046-657WO1 8531

[0047] Fig.6C shows the experimental model applied to solving the super-resolution inverse problem on axial, coronal, sagittal FSE, and axial SE slices. Detailed Description

[0048] Some references, which may include various patents, patent applications, and publications, are cited in a reference list and discussed in the disclosure provided herein. The citation and / or discussion of such references is provided merely to clarify the description of the disclosed technology and is not an admission that any such reference is “prior art” to any aspects of the disclosed technology described herein. In terms of notation, “[n]” corresponds to the nth reference in the list. For example, [1] refers to the first reference in the list. All references cited and discussed in this specification are incorporated herein by reference in their entirety and to the same extent as if each reference were individually incorporated by reference.

[0049] Example System

[0050] Figs.1A – 1B each shows an example system 100 (shown as 100a and 100b) for reconstructing a low-field or low-resolution magnetic resonance imaging (MRI) scan 104 (shown as 104a and 104b) into a higher-fidelity MRI scan 106 (shown as 106a and 106b) with reduced motion artifacts and noise. Figs.1A and 1B show the inference phase 108 (shown as 108a – 108b, respectively) for the use of the example system 100 and the training phase 114 (shown as scarce dataset-based training phase 114a and 114b) to generate the trained AI model for use during the inference stage 108. During operation, e.g., in the inference phase 108a, 108b, the system 100 is configured to generate, via a combined operation 102 between a physics-based motion model 110 and a trained neural network (NN) model 111, the high-fidelity MRI scan 106 with reduced motion artifacts (e.g., higher field equivalence, lower noise), from the low-field or low-resolution MRI scan 104. The trained NN model 111 in phase 108 was generated and trained in the scarce dataset-based training phase 114. In some embodiments, the trained NN model 111 is a trained diffusion generative model.

[0051] Scarce Dataset-based Training Phase (114). A scarce training dataset refers to a limited and fragmented data collection, e.g., for newborn / neonatal patients (e.g., within the first 28 days of life), where the volume, variety, and accessibility of data are severely constrained. The scarcity arises from logistical challenges in data collection, the rarity of certain neonatal conditions, and the risk of conventional MRI scans for neonatal / newborn patients. As noted above, pre-trained models from adult patients cannot be used becauseAttorney Docket No.10046-657WO1 8531 brain structure can vary greatly between neonates and adults. Scanned images are inherently noisier in low or lower-field neonatal MRI, and significantly less data has been acquired and publicly shared, making it challenging to train machine learning reconstruction models, from both a data quality and quantity perspective. Newborns are additionally noted to be scanned in their natural state, i.e., without sedation and physical constraints; the resulting images / scans are thus corrupted by motion.

[0052] In the examples shown in Figs.1A – 1B, a scarce training dataset 112 (e.g., neonatal dataset) (shown as 112a and 112b) is employed for the training, as that is all that is available. In Fig.1B, to improve the training, the scarce training dataset 112b includes scans (e.g., 2D or 3D) of subjects (e.g., neonatal subjects) at various orientations and contrasts with class embedding. The scarce training dataset 112 can include (i) low-field and / or low- resolution MRI scans, or (ii) low-field and / or low-resolution neonatal MRI scans, which may be acquired (e.g., by a low-field single-coil MRI scanner) with reduced acquisition time (e.g., at an accelerated pace) of a moving subject during the scan. In some embodiments, the MRI scans of the subject for various orientations include at least two orientations from the group consisting of axial orientation, sagittal orientation, coronal orientation, and oblique orientation, for a contrast orientation. In some embodiments, the contrast configuration for at least two orientations is selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging.

[0053] In Figs.1A and 1B, the scarce training dataset 112, or portions thereof, is unconventionally used and reused for multiple training operations (e.g., 116, 122) of multiple models (e.g., 111, 118). In Figs.1A and 1B, the scarce training dataset 112 (shown as 123), or a substantial portion of it, is first used as input to a denoising AI model training module 116, to generate a denoising AI model 118. The denoising AI model training module 116 may be provided with noise distribution data of the scarce training dataset 124 (e.g., from a statistical model) to generate the denoised AI model 118. The scarce training dataset 112 (shown as 126), or a substantial portion of it, is then reused as an input 126a to the denoising AI model 118 to generate a denoised scarce training dataset 120. A description of the denoising AI model training module 116 may be found in Moran, Nick, et al. "Noisier2noise: Learning to denoise from unpaired noisy data." Proceedings of the IEEE / CVF conference on computer vision and pattern recognition.2020, which is incorporated by reference herein.

[0054] Class embeddings 132, from the class embedding module 130, and the denoised training set 120 are used as an input to a neural network model training module 122 to generate the trained NN model 111 (shown as 111a, 111b). The denoised scarce trainingAttorney Docket No.10046-657WO1 8531 data set 120 is provided as ground truth to the neural network model training module 122 for the training. Indeed, the neural network model 111a is trained by the NN model training 122 using the ground truth 120 and training input 128.

[0055] In Fig.1B, the training 114b employs class embedding for different orientations and contrast for the scarce training dataset 112. In Fig.1B, the scarce training data set 112b includes the MRI scan at different orientations (shown as “Scarce training data set at orientation / contrast 1” … “Scarce training data set at orientation / contrast n”). The different orientations are encoded by class embeddings 132 via a class embedding module 130 to generate the training input 132 as a scarce training dataset with class embedding. The training input 132 (generated from the reused scarce training data set 112b) is then used by the training module 122 in the training of the neural network model 111.

[0056] Indeed, the scarce training dataset is used and reused during the training operation to generate the neural network model 111. Each use / reuse is substantially the same as the prior use, e.g., having substantial overlap in the data that is used. In the preferred embodiment, there is 100% overlap between the data (e.g., 123, 126a, 128 in Fig.1A). In other embodiments, at least 90% of the data is reused between the different uses and reuse.

[0057] In some embodiments, the scarce training dataset includes scans of the subject for a plurality of orientations and contrasts, and wherein the training of the neural network model employed the plurality of orientations and contrasts with class embedding.

[0058] In some embodiments, the scarce training dataset includes low-field MRI scans, wherein the low-field MRI scans were acquired with reduced acquisition time.

[0059] In some embodiments, the scarce training dataset includes low-field neonatal MRI scans, wherein the low-field neonatal MRI scans were acquired with reduced acquisition time of a moving subject during the scan.

[0060] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan. In some embodiments, the orientations are selected from the group consisting of axial orientation, sagittal orientation, coronal orientation, and oblique orientation.

[0061] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, wherein class embedding is applied for training of the neural network model for each of the at least two orientations for the given contrast.Attorney Docket No.10046-657WO1 8531

[0062] In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations, each for at least two contrast scans selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, wherein class embedding is applied for training of the neural network model for each of the orientations and contrasts.

[0063] In some embodiments, the MRI image or data object comprises 2D MRI scans or 3D MRI scans.

[0064] In some embodiments, the MRI image or data object is acquired by a low-field MRI scanner.

[0065] Inference Phase (108). In the examples shown in Figs.1A – 1B, the trained NN model 111, once trained via 114a, 114b, is configured to receive a low-field or low- resolution magnetic resonance imaging (MRI) scan 104 (shown as 104a and 104b) to generate a higher-fidelity MRI scan 106 (shown as 106a and 106b) with reduced motion artifacts and noise. High fidelity can be higher field equivalent or higher resolution. The trained NN model 111 is then configured to generate, in combination 102 with the physics- based motion model 110, the higher-fidelity MRI scan 106 (shown as 106a and 106b) (e.g., image or object) of the subject that reduces motion artifacts and noise from the received low- field neonatal MRI scan 104. The trained NN model 111, in combination with the physics- based motion model 110, is then configured to output the higher-fidelity neonatal MRI scan 106, which can be subsequently employed for diagnosis or treatment of the subject.

[0066] In some embodiments, the generation of the higher-fidelity MRI scan 106 of the subject is a motion-informed posterior sampling operation (see Fig.3A and Equation Set 3) that is (i) configured to generate a high-fidelity MRI scan from a motion-corrupt MRI scan and (ii) performed by the physics-based motion model 110.

[0067] Example Method

[0068] Fig.2A shows an example operational flow 200a for the exemplary system 100 (in an inference phase 108), in accordance with an illustrative embodiment. The method 200a includes receiving (202), by a processor, a magnetic resonance imaging (MRI) image or data object (e.g., 104, Figs.1A – 1B) (e.g., 2D MRI scan or 3D MRI scan) acquired by an MRI scanner (e.g., low-field single-coil MRI scanner). The method 200a includes generating (204), via a trained neural network model (e.g., 111, Figs.1A – 1B) (e.g., diffusion generative model) and a physics-based motion model (e.g., 110, Figs.1A – 1B) operating in combination, a higher-fidelity image or data object of the subject (e.g., 106, Figs.1A – 1B) that reduces motion and noise from the received MRI image or data object (e.g., 104, Figs.Attorney Docket No.10046-657WO1 8531 1A – 1B). The method 200a includes outputting (206) the higher-fidelity image or data object (e.g., 106, Figs.1A – 1B) that can be subsequently employed for diagnosis or treatment of the subject.

[0069] Fig.2B shows an example method of training the neural network (NN) model (e.g., 111, Figs.1A – 1B) of the exemplary system 100, in accordance with an illustrative embodiment. The method 200b includes training (210) a denoising artificial intelligence (AI) model (e.g., 118, Figs.1A – 1B) using a scarce training dataset (e.g., 112, Figs.1A – 1B). The method 200b includes generating a denoised scarce training dataset (e.g., 120, Figs.1A – 1B) from the scarce training dataset (e.g., 112, Figs.1A – 1B) using the trained denoising AI model (e.g., 118, Figs.1A – 1B). The method 200b includes applying (214) the scarce training dataset (e.g., 112, Figs.1A – 1B) as input to the training of the neural network (NN) model (e.g., 111, Figs.1A – 1B), and the denoised scarce training dataset (e.g., 120, Figs.1A – 1B) is employed as a ground truth during the training.

[0070] The scarce training dataset (e.g., 112, Figs.1A – 1B) can include scans of a subject for a plurality of orientations and contrasts. The training of the NN model (e.g., 111, Figs.1A – 1B) can employ the plurality of orientations and contrasts with class embeddings (e.g., 422, Fig.4B). In some embodiments, the scarce training dataset (e.g., 112, Figs.1A – 1B) includes low-field MRI scans, and the low-field MRI scans were acquired with reduced acquisition time (e.g., at accelerated pace). In some embodiments, the scarce training data set (e.g., 112, Figs.1A – 1B) includes low-field neonatal MRI scans acquired with reduced acquisition time (e.g., at accelerated pace) of a moving subject during the scan.

[0071] The scans of the subject for the plurality of orientations can include at least two orientations for a given contrast scan. In some embodiments, the at least two orientations are selected from the group consisting of axial orientation, sagittal orientation, coronal orientation, and oblique orientation. In some embodiments, the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, where class embedding is applied for training of the NN model (e.g., 111, Figs.1A – 1B) for each of the at least two orientations for the given contrast.

[0072] In some embodiments, the denoising AI model (e.g., 118, Figs.1A – 1B) was trained using the scarce training dataset (e.g., 112, Figs.1A – 1B) and a statistical model of a noise distribution of the scarce training dataset (e.g., 124, Figs.1A – 1B).

[0073] Example Motion Robust Magnetic Resonance ImagingAttorney Docket No.10046-657WO1 8531

[0074] Motion robust magnetic resonance imaging (MRI) measurement model. The exemplary system can be applied to various MRI contrasts and acquisitions, including two- dimensional (2D), T2-weighted Fast-Spin-Echo (FSE) Neonatal data measured with two average, among other descrbed herein, including 3D scans. A repetition time (TR) of the FSE sequence may include a train of refocusing pulses with signal readout during the generated echoes, followed by deadtime for signal recovery. Assuming in-plane rigid motion that only occurs during sequence deadtime yields 3 parameters per TR that describe Neonatal motion. Combining all motion states for all TRs into the variable ^, the measurement model of the exemplary system can be defined per Equation 1. ^^ ^^^(Eq.1)

[0075] In Equation 1, ^^is the forward operator consisting of translation, rotation, Fourier Transform, coil sensitivities, and undersampling operators, ^ is acquired data, and ^ is the motion-free image estimate.

[0076] Since the required data y consists of two averages, ^^and ^^, the accelerated, motion-robust reconstruction problem can be defined per Equation 2, which can reconstruct amotion-free image x* and motion estimates ^^^^ ^^^ by jointly exploiting data from bothaverages.(Eq.2)

[0077] Deep learning-based generative reconstruction models. Deep learning-based generative models (e.g., as neural network model 111) can serve as priors when solving inverse problems, as they decouple from the measurement model. This means a generative model can be trained on a clinical database of motion-free neonatal MRI images and then applied as a prior when solving the accelerated, motion-robust reconstruction problem described in Equation 2.

[0078] Let ^ be a generative model trained on over 10,000 neonatal MRI images. Then, motion-informed posterior sampling (MI-PS) can solve Equation 2 with the generative prior D. First, MI-PS can initialize the image estimate to Gaussian noise and motion parameters 0, and then during each sampling iteration ^, the algorithm applies Equation Set 3.#$ ^ ^^^ ^ ^^^!% ^&&&Attorney Docket No.10046-657WO1 8531(Eq. Set 3)

[0079] In Equation Set 3,^represents the time steps of the sampling process (e.g., discretized time steps of the ordinary differential equation), ^^^! represents the pre-defined noise schedule, and ^_^ is a motion informed MRI forward model that incorporates fourier encoding and rigid motion. Posterior sampling can simultaneously estimate a clean image ^ and motion parameters ^ from motion corrupt measured data ^.

[0080] Fig.3A shows an example Motion Informed Posterior Sampling (in the exemplary system) that combines a generative deep learning model with a motion robust MRI measurement model (see Equation Set 3) to reconstruct high-fidelity Neonatal MRI images in the presence of rigid motion.

[0081] Fig.3B shows an example mitigation, by the exemplary system 100, of motion artifacts in a real sagittal Neonatal scan by combining motion modeling, generative models, and data from two acquisitions. The exemplary system can simultaneously estimate a clean image and the associated motion parameters. In Fig.3B, the data are acquired using a T2- weighted turbo-spin-echo acquisition with two scans. Without correction, motion during the acquisition results in artifacts in the images of row 302. The motion correction of the exemplary system can reduce motion artifacts in row 304, with the reconstructions using data from both scans yielding the cleanest image. Finally, the exemplary system yields motion estimated associated with the clean image, indicating the extent and nature of motion during the acquisition.

[0082] Theory. Given single-channel MRI measurements y = NK x + ^ in the neonatal setting of the exemplary system, an image can be generated by solving the inverse problem defined per Equation 4.(Eq.4)

[0083] In Equation 4, x 6 Cnis the vectorized image, NK is the 2D Fourier transform evaluated at coordinates K, ^ 6 Cmis Gaussian random noise, y 6 Cmare the acquired measurements. MRI scans can accelerate acquisitions by acquiring fewer measurements thanAttorney Docket No.10046-657WO1 8531 image pixels (i.e., m < n), but this can result in an ill-posed inverse problem that yields non- diagnostic images without suitable regularization.

[0084] Generative models can solve ill-posed inverse problems by learning the statistical prior, p(x), over clean images to guide the reconstruction towards solutions that both match the data and are statistically likely. Specifically, diffusion models

[0013] ,

[0014] can learn p(x) by training a neural network (NN) D^(x) to learn the score&89.&:;^^;!&of progressively noised distributions. Then, the process of reconstructing an image can be viewed as sampling from the posterior distribution x < p(x|y).

[0085] Following

[0014] , the exemplary system can sample from the posterior distribution by using an Euler solver on the reverse ordinary differential equation (ODE) formulation, defined per Equation 5, with s(t) = 1 and ^(t) = 1. >?^ !#^ ^ => ^ ^ >^ !^^ ! ^?^ !^^ !(^ 89. :(Eq.5)

[0086] Using Bayes' rule, the exemplary system can separate the posterior score into a likelihood and prior score as defined per Equation 6.(Eq.6)

[0087] The analytical expression for the likelihood score can be known at time point t = 0, so the exemplary system can use the approximation ^D^^! = E[x0| x]

[0015] . This formulation can decouple the statistical prior from the likelihood, so for Neonatal MRI, a single prior can be reused to solve inverse problems with different sampling patterns, receive coils, timings, and measurement models.

[0088] While the previous studies assumed a static image, motion in MRI can be modeled as forward model uncertainties for rigid body motion, as shown in Equation 7.(Eq.7)

[0089] In Equation 7, R^ is a rotation matrix for all motion states, P^ is a diagonal matrix implementing a linear phase shift describing the horizontal and vertical translations during each motion state, and GHI5is the Non-uniform Fast Fourier Transform (NUFFT) of x at the coordinates R^K. Let ^ = {^, ^} be a variable that holds all information about the rigid body motion. Then, a clean image and its associated motion parameters from an acquisition in the presence of motion can be estimated by solving Equation 8.Attorney Docket No.10046-657WO1 8531(Eq.8)

[0090] Again, since posterior sampling can decouple the prior and likelihood, the same diffusion model, D^, can be applied to solve this ill-posed inverse problem. In particular, the exemplary system can follow the previous study

[0016] , which assumes an independent, uniform prior on the motion parameters, and samples from the joint distribution p(x, ^|y) by solving the reverse ODE with Euler updates.

[0091] Example Neural Network Model Training Process

[0092] Fig.4A shows an example training process 114 for a neural network (NN) model 111 (e.g., generative diffusion model D) in the exemplary system and an example inference pipeline 108 of the trained NN model 111.

[0093] Neonatal dataset. In Fig.4A, a low-field neonatal dataset 112 (shown as 112a, 112b) is acquired (e.g., with the 1T Embrace system) from 128 neonatal subjects. Each neonatal subject is scanned with Fast Spin Echo (FSE) (e.g., at axial, coronal, and / or sagittal orientations), Spin Echo (SE) (e.g., at axial orientation), gradient echo, and / or Magnetization Prepared – Rapid Gradient Echo (MPRAGE) sequences. In some embodiments, the dataset 112 is randomly split into (i) 108 subjects for training and validation, and (ii) 20 subjects for testing. The training dataset can keep the coronal, sagittal, and axial FSE and axial SE scans and remove the first four and last four slices from each volume, resulting in 8659 FSE and 3224 SE slices. Training slices can be resized to a 200 × 200 matrix size.

[0094] Model Training (114). To adapt training 114 to a neonatal setting, the NN model 111 is configured to take varying matrix size inputs so that Heterogeneous matrix size data can be resized and used for training, and the NN model 111 may handle matrix size discrepancies at inference. The NN model 111 can be trained using a training input 132 and a ground truth 120 (e.g., a denoised MRI dataset).

[0095] The training input 132 is generated from a combination / embedding operation 130 re-using the neonatal dataset 112a with class embeddings. Fig.4B shows an example class embedding operation 130 of the training process 114. In Fig.4B, the embedding operation 130 is configured to (i) one-hot encode (420) the contrast and orientation information in the neonatal dataset 112 with class embeddings 422 (e.g., 422a – 422d) (e.g., encode a vector of all zeros except for a value one in the index corresponding to a specific contrast and orientation in the neonatal dataset 112) and (ii) train (426) a multi-layer perception (MLP) 428 to take the encoding vector 424 (e.g., combined dataset and classAttorney Docket No.10046-657WO1 8531 embeddings) as input and output an embedding vector 132 that the NN model 111 can incorporate into its architecture (as a training input). In this way, the NN model 111 can use image contrast and orientation information during training 114 and inference 108.

[0096] Referring back to Fig.4A, the ground truth 120 is a denoised low-field neonatal MRI dataset generated from a self-supervised denoising operation 116 using the neonatal dataset 112b. Fig.4C shows an example self-supervised denoising operation 118 (also referred to as a denoiser or denoising AI model). In some embodiments, the self- supervised denoising operation 118 is a trained Noiser2Noise model

[0017] applied to the neonatal dataset 112b before the training of the NN model 111.

[0097] Fig.4D shows examples of denoised training samples and prior samples generated with class embeddings from the exemplary system and its trained NN model (e.g., 111, Figs.1A – 1B). In Fig.4D, subpanel (a), the top row shows two training samples from a dataset, and the bottom row shows the corresponding training samples after applying a denoiser trained in a self-supervised fashion. In Fig.4D, subpanels (b) – (e) show prior samples generated by the exemplary system when conditioned on class embeddings of FSE axial, sagittal, coronal, and SE axial. The NN model 111 of the exemplary system uses all available training data to learn a statistical prior over neonatal MR images.

[0098] In some embodiments, the NN model 111 is a generative diffusion model having a UNet backbone with (i) 65 million hyperparameters, 6 upsampling and downsampling blocks, block convolution layers, skip connections, and self-attention layers interspersed throughout. To adapt training to the neonatal setting, the NN model 111 is configured to take varying matrix size input (e.g., by modifying its hyperparameters to be compatible with the neonatal dataset 112). Thus, heterogeneous matrix size data can be resized and used for training, and the NN model 111 may handle matrix size discrepancies at inference.

[0099] Inference Pipeline (108). Referring back to Fig.4A, the trained NN model 111 can reconstruct a high-fidelity MRI image 106 using a low-field neonatal MRI image 104 acquired from accelerated (e.g., reduced-time) motion-corrupted scanning 402. In some embodiments, the reconstruction / generation of the high-fidelity MRI image 106 from the low-field MRI image 104 is a motion-informed posterior sampling operation (see Fig.3A and Equation Set 3).

[0100] Fig.4E shows an example algorithmic implementation 400e of the motion- informed posterior sampling operation that generates the high-fidelity MRI image 106 fromAttorney Docket No.10046-657WO1 8531 the low-field MRI image 104 in the inference pipeline 108. As shown, the algorithm 400e is implemented using Equation Set 3.

[0101] Example Artificial Intelligence (AI) and Machine Learning (ML) Models

[0102] Machine Learning. In addition to the machine learning features described above, the exemplary system can be implemented using one or more artificial intelligence and machine learning operations. The term “artificial intelligence” can include any technique that enables one or more computing devices or computing systems (i.e., a machine) to mimic human intelligence. Artificial intelligence (AI) includes but is not limited to knowledge bases, machine learning, representation learning, and deep learning. The term “machine learning” is defined herein to be a subset of AI that enables a machine to acquire knowledge by extracting patterns from raw data. Machine learning techniques include, but are not limited to, logistic regression, support vector machines (SVMs), decision trees, Naïve Bayes classifiers, and artificial neural networks. The term “representation learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, or classification from raw data. Representation learning techniques include, but are not limited to, autoencoders and embeddings. The term “deep learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, classification, etc., using layers of processing. Deep learning techniques include but are not limited to artificial neural networks or multilayer perceptron (MLP).

[0103] An artificial neural network (ANN) is a computing system including a plurality of interconnected neurons (e.g., also referred to as “nodes”). This disclosure contemplates that the nodes can be implemented using a computing device (e.g., a processing unit and memory as described herein). The nodes can be arranged in a plurality of layers, such as an input layer, an output layer, and optionally one or more hidden layers with different activation functions. An ANN having hidden layers can be referred to as a deep neural network or multilayer perceptron (MLP). Each node is connected to one or more other nodes in the ANN. For example, each layer is made of a plurality of nodes, where each node is connected to all nodes in the previous layer. The nodes in a given layer are not interconnected with one another, i.e., the nodes in a given layer function independently of one another. As used herein, nodes in the input layer receive data from outside of the ANN, nodes in the hidden layer(s) modify the data between the input and output layers, and nodes in the output layer provide the results. Each node is configured to receive an input, implement an activation function (e.g., binary step, linear, sigmoid, tanh, or rectified linear unit (ReLU)Attorney Docket No.10046-657WO1 8531 function), and provide an output in accordance with the activation function. Additionally, each node is associated with a respective weight. ANNs are trained with a dataset to maximize or minimize an objective function. In some implementations, the objective function is a cost function, which is a measure of the ANN’s performance (e.g., error such as L1 or L2 loss) during training, and the training algorithm tunes the node weights and / or bias to minimize the cost function. This disclosure contemplates that any algorithm that finds the maximum or minimum of the objective function can be used for training the ANN. Training algorithms for ANNs include but are not limited to backpropagation. It should be understood that an artificial neural network is provided only as an example machine learning model. This disclosure contemplates that the machine learning model can be any supervised learning model, semi-supervised learning model, or unsupervised learning model. Optionally, the machine learning model is a deep learning model. Machine learning models are known in the art and are therefore not described in further detail herein.

[0104] A convolutional neural network (CNN) is a type of deep neural network that has been applied, for example, to image analysis applications. Unlike traditional neural networks, each layer in a CNN has a plurality of nodes arranged in three dimensions (width, height, depth). CNNs can include different types of layers, e.g., convolutional, pooling, and fully-connected (also referred to herein as “dense”) layers. A convolutional layer includes a set of filters and performs the bulk of the computations. A pooling layer is optionally inserted between convolutional layers to reduce the computational power and / or control overfitting (e.g., by downsampling). A fully-connected layer includes neurons, where each neuron is connected to all of the neurons in the previous layer. The layers are stacked similarly to traditional neural networks. GCNNs are CNNs that have been adapted to work on structured datasets such as graphs.

[0105] Deep Generative Models. A deep generative model (DGM) is a type of deep neural network configured to learn and synthesize complex data distributions. Different from discriminative models, which map input data to labels or predictions, a DGM models the underlying probability distribution of the data to generate new samples that resemble those observed in training. DGMs can include different architectures, such as variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion-based models. In a typical configuration, a DGM includes an encoder or generator network that transforms latent variables drawn from a predefined prior distribution into structured outputs (e.g., images, signals, or text). Training may involve optimizing a loss function that measures the divergence between generated and real data distributions, such as a likelihood objective,Attorney Docket No.10046-657WO1 8531 adversarial loss, or score-matching function. The layers of the network are arranged hierarchically to capture multi-scale dependencies, enabling the system to produce high- dimensional, realistic, and coherent samples across diverse modalities.

[0106] Other Supervised Learning Models. A logistic regression (LR) classifier is a supervised classification model that uses the logistic function to predict the probability of a target, which can be used for classification. LR classifiers are trained with a dataset (also referred to herein as a “dataset”) to maximize or minimize an objective function, for example, a measure of the LR classifier’s performance (e.g., an error such as L1 or L2 loss), during training. This disclosure contemplates that any algorithm that finds the minimum of the cost function can be used. LR classifiers are known in the art and are therefore not described in further detail herein.

[0107] A Naïve Bayes’ (NB) classifier is a supervised classification model that is based on Bayes’ Theorem, which assumes independence among features (i.e., the presence of one feature in a class is unrelated to the presence of any other features). NB classifiers are trained with a dataset by computing the conditional probability distribution of each feature given a label and applying Bayes’ Theorem to compute the conditional probability distribution of a label given an observation. NB classifiers are known in the art and are therefore not described in further detail herein.

[0108] A k-NN classifier is an unsupervised classification model that classifies new data points based on similarity measures (e.g., distance functions). The k-NN classifiers are trained with a dataset (also referred to herein as a “dataset”) to maximize or minimize a measure of the k-NN classifier’s performance during training. This disclosure contemplates any algorithm that finds the maximum or minimum. The k-NN classifiers are known in the art and are therefore not described in further detail herein.

[0109] A majority voting ensemble is a meta-classifier that combines a plurality of machine learning classifiers for classification via majority voting. In other words, the majority voting ensemble’s final prediction (e.g., class label) is the one predicted most frequently by the member classification models. The majority voting ensembles are known in the art and are therefore not described in further detail herein.

[0110] Experimental Results and Additional Examples

[0111] A study was conducted to develop an experimental system that employs deep generative models (also referred to as “motion robust measurement model” or “experimental (generative reconstruction) model”) and physics-based motion modeling to shortenAttorney Docket No.10046-657WO1 8531 acquisition time and reduce motion vulnerability of magnetic resonance imaging processes, as described in relation to Figs.1 – 2.

[0112] Experiment #1

[0113] Accelerated and motion reconstruction experiments. The study performed accelerated MRI reconstruction and motion correction experiments to demonstrate the utility of the generative model of the experimental system (“experimental (generative) model”) on neonatal data across varying measurement models. For accelerated MRI reconstruction, the test FSE and SE data were undersampled by an average acceleration rate of 1.5. To achieve realistic undersampling with respect to signal decay

[0018] , the study undersampled the FSE data by throwing away groups of data associated with each echo train, so FSE acceleration was either 1.4 or 1.6, depending on the echo train length and matrix size. The same model, taking advantage of class embedding, reconstructed all images, and the study compared reconstructions using a baseline L1-wavelet [9], the experimental generative model without denoising, and the experimental model with denoising. For the study’s reconstructions, the study averaged 5 posterior samples generated with different random initializations.

[0114] Next, the study identified two acquisitions with motion corruption in the test dataset and prospectively applied the MI-PS algorithm

[0016] with the experimental generative model without any modifications to reduce motion artifacts. The study compared the original, motion-corrupt clinical images to the experimental model.

[0115] Evaluation results. Table 1 presents quantitative comparisons of the accelerated reconstruction experiments separated across contrast and orientation. Table 1. Accelerated Reconstruction on Test Set

[0116] As shown in Table 1, the experimental generative model with denoising achieved comparable or superior average normalized root mean square error (NRMSE) performance across the test set.

[0117] Fig.5A shows reconstruction results for each orientation and contrast comparing the baseline L1-wavelet to the experimental system using a generative model trained with (shown as “Ours Noisy”) and without (shown as “Ours Denoised”) self- supervised denoising on the training dataset. As shown, the L1-wavelet experienced residualAttorney Docket No.10046-657WO1 8531 aliasing artifacts, and the experimental model with denoising reduced error compared to the experimental model without denoising. The quantitative results were computed with respect to the fully-sampled, non-denoised images, so this could bias the comparisons between the experimental model with and without denoising.

[0118] Fig.5B shows how the experimental model (shown as “Ours”) estimated motion parameters and yielded cleaner results (e.g., fewer artifacts) than the original motion- corrupt data on example axial and coronal slices from the test dataset. The experimental model also estimated the associated motion parameters of that scan.

[0119] Experiment #2

[0120] The study performed another set of accelerated MRI reconstruction, motion correction, and super-resolution experiments to demonstrate the utility of the experimental generative model for neonatal MRI across varying measurement models and inverse problems at inference time.

[0121] Accelerated MRI reconstruction. First, test axial, coronal, sagittal FSE slices and axial SE slices were undersampled by an average rate of 2.0×. The study undersampled by throwing away groups of data associated with each echo train to achieve realistic undersampling with respect to signal decay

[0019] . To analyze the effect of combining data with class embedding, the study compared the reconstruction performance of diffusion models trained on all data with and without class embeddings to diffusion models trained on each contrast and orientation separately. No method employed denoising pre-training in the experiment. Second, the study undersampled the test FSE and SE slices by 1.5×, and compared non-learned L1-wavelet [9] based reconstructions to experimental generative models trained with combined data using class embeddings with and without denoising.

[0122] Motion Correction. The study (i) identified axial and coronal FSE acquisitions in the test dataset with motion artifacts and (ii) solved the inverse problem of Equation 8 to estimate motion-free images and the associated motion parameters using the experimental generative model applied in accelerated MRI reconstruction. This experiment used prospective clinical data; thus, no ground truth existed, so the study visually compared the original, motion-corrupted clinical image to the image by the experimental model.

[0123] Super-Resolution. The study lowered the resolution of example axial, coronal, sagittal FSE, and axial SE slices by a factor of 2.5× by discarding the high-frequency k-space measurements. Super-resolution on this low-resolution data was performed by solving the inverse problem described in Equation 4 using the experimental generative model.Attorney Docket No.10046-657WO1 8531

[0124] Evaluation Results. Table 2 shows normalized root mean squared error (NRMSE ×100) comparisons of the 2× accelerated MRI reconstruction experiments using (i) experimental generative models trained on each contrast and orientation separately (shown as Solo) versus (ii) experimental generative models trained on all data with and without class embeddings. Fig.6A shows the same accelerated MRI reconstructions using (i) experimental generative models trained on each contrast and orientation separately versus (ii) experimental generative models trained on all data with and without class embeddings. The models trained on all data achieved lower NRMSE than those trained separately. Additionally, incorporating embeddings improved performance when training using all data. Table 2

[0125] Table 3 compares the quantitative performance of 1.5× accelerated MRI reconstruction using the non-learned L1-wavelet method with experimental generative models trained on all data using class embeddings with and without denoising pre-training. The learning-based models outperformed L1-wavelet, and the experimental model with denoising achieved comparable or superior average NRMSE compared to the experimental model without denoising pre-training. The quantitative results were complex-valued differences computed with respect to the fully-sampled, non-denoised images

[0020] . Table 3

[0126] Fig.6B shows the experimental generative model (shown as “Ours”) applied to the task of motion correction on measured, motion-corrupt clinical data. The fully sampled, clinical axial and coronal images experienced artifacts induced by patient motion during the scan. The motion correction employed the same generative model used for accelerated MRI reconstruction to estimate a motion-free image and associated motion parameters from motion-corrupt data.

[0127] Qualitative reduction in ringing artifacts, highlighted by the arrows, was observed when using the motion correction.Attorney Docket No.10046-657WO1 8531

[0128] Fig.6C shows the experimental generative model applied to solving the super- resolution inverse problem on axial, coronal, sagittal FSE, and axial SE slices. Due to a 2.5× reduction in the extent of k-space sampling and, therefore, scan time, the images in the low- res column exhibited lower image resolution and Gibbs Ringing artifacts, but applying the experimental generative model to solve the inverse problems results in sharper images.

[0129] Discussion

[0130] Neonatal Magnetic Resonance Imaging (MRI) enables non-invasive assessment of potential brain abnormalities during the critical phase of Neonatal and preterm development [1], [2], [3]. However, the necessity of sedation to reduce motion artifacts and transfer of vulnerable patients from the neonatal intensive care unit (NICU) to the scan room precludes access to MRI for many neonatal patients [4] (e.g., where sedation may be unavailable due to the risk to the patient). Previous studies show that lower-field MRI systems (i.e., below 1.5 Tesla) that function directly in the NICU increase accessibility of Neonatal MRI [5], but these systems still experience frequent motion artifacts and low signal- to-noise ratio (SNR).

[0131] Motion artifacts in adult MRI scans have been shown to be mitigatable by reducing scan times through undersampled acquisitions and leveraging various correction techniques [6], [7]. Clinically routine methods to accelerate MRI combine parallel imaging [8], which exploits the multi-channel signal receive array, with hand-crafted spatial regularization and Compressed Sensing [9]. Recently, machine learning algorithms for accelerated MRI reconstruction have yielded state-of-the-art results

[0010] . End-to-end methods learn a point-wise mapping between undersampled and fully sampled data but are highly susceptible to test time shifts in the measurement operator. More recently, generative methods learn a prior over clean images with robustness to test time shifts in the forward operator

[0011] ,

[0012] . However, both these techniques require high-quality image data (i.e., high SNR) for training.

[0132] The unique challenge presented by the low-field neonatal MRI setting precludes direct application of these accelerated MRI techniques for adults. Many low-field systems measure signal with a single-channel receive array [5], so parallel imaging cannot be applied. Pretrained models from adult patients cannot be used because brain structure can vary greatly between neonates and adults. In addition, the images are inherently noisier in lower-field neonatal MRI, and significantly less data has been acquired and publicly shared, making it challenging to train machine learning reconstruction models, from both a data quality and quantity perspective.Attorney Docket No.10046-657WO1 8531

[0133] The instant study developed a prototype, as an example of the exemplary system and method, that can accelerate and improve motion robustness of low-field Neonatal MRI acquired. The prototype was an in-NICU 1T Embrace System (Aspect Imaging) through diffusion-based generative modeling

[0013] . The instant study observed a reduction in scan time of single-coil Fast Spin Echo and Spin Echo sequences by an average of 1.5× by using the pre-trained prior to reconstruct high-fidelity images from realistically under- sampled measurements. The generative model was applied to the task of motion correction, which was observed to substantiallyk reduce motion artifacts in prospectively acquired data.

[0134] While the study presented an initial qualitative and quantitative analysis, future steps may be required to take the study towards clinical adoption. First, a reader study with board-certified radiologists working with neonatal MR images is needed to evaluate whether the accelerated and motion-corrected images maintain diagnostic utility. Second, accelerated MRI should be evaluated prospectively by directly acquiring undersampled data from the scanner instead of retrospectively throwing away echo trains from fully sampled data. The motion correction experiments in the study were prospective, as the study did not throw away any data before applying the exemplary system and method. Finally, posterior sampling for a single slice took roughly 15 seconds on the study’s H100 GPU. An ODE solver that runs faster and parallel to produce image volumes in clinically acceptable scan times may also be needed.

[0135] Additionally, the exemplary system and method may require longer reconstruction times to create an image from data than baseline techniques. However, reconstruction times can be reduced through a more optimized implementation of our algorithm and other advancements in generative modeling.

[0136] Conclusion

[0137] The construction and arrangement of the systems and methods, as shown in the various implementations, are illustrative only. Although only a few implementations have been described in detail in this disclosure, many modifications are possible (e.g., variations in sizes, dimensions, structures, shapes, and proportions of the various elements, values of parameters, mounting arrangements, use of materials, colors, orientations, etc.). For example, the position of elements may be reversed or otherwise varied, and the nature or number of discrete elements or positions may be altered or varied. Accordingly, all such modifications are intended to be included within the scope of the present disclosure. The order or sequence of any process or method steps may be varied or re-sequenced according to alternative implementations. Other substitutions, modifications, changes, and omissions mayAttorney Docket No.10046-657WO1 8531 be made in the design, operating conditions, and arrangement of the implementations without departing from the scope of the present disclosure.

[0138] The present disclosure contemplates methods, systems, and program products on any machine-readable media for accomplishing various operations. The implementation of the present disclosure may be implemented using existing computer processors, or by a special-purpose computer processor for an appropriate system, incorporated for this or another purpose, or by a hardwired system. Implementations within the scope of the present disclosure include program products, including machine-readable media for carrying or having machine-executable instructions or data structures stored thereon. Such machine- readable media can be any available media that can be accessed by a general-purpose or special-purpose computer or other machine with a processor. By way of example, such machine-readable media can comprise RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to carry or store desired program code in the form of machine- executable instructions or data structures, and which can be accessed by a general purpose or special purpose computer or other machine with a processor.

[0139] When information is transferred or provided over a network or another communications connection (either hardwired, wireless, or a combination of hardwired or wireless) to a machine, the machine properly views the connection as a machine-readable medium. Thus, any such connection is properly termed a machine-readable medium. Combinations of the above are also included within the scope of machine-readable media. Machine-executable instructions include, for example, instructions and data that cause a general-purpose computer, special-purpose computer, or special-purpose processing machines to perform a certain function or group of functions.

[0140] Although the figures show a specific order of method steps, the order of the steps may differ from what is depicted. Also, two or more steps may be performed concurrently or with partial concurrence. Such variation will depend on the software and hardware systems chosen and on designer choice. All such variations are within the scope of the disclosure. Likewise, software implementations could be accomplished with programming techniques with rule-based logic and other logic to accomplish the various connection steps, processing steps, comparison steps, and decision steps.

[0141] Each and every feature described herein, and each and every combination of two or more of such features, is included within the scope of the present invention, provided that the features included in such a combination are not mutually inconsistent.Attorney Docket No.10046-657WO1 8531

[0142] Although example embodiments of the disclosed technology are explained in detail herein, it is to be understood that other embodiments are contemplated. Accordingly, it is not intended that the disclosed technology be limited in its scope to the details of construction and arrangement of components set forth in the following description or illustrated in the drawings. The disclosed technology is capable of other embodiments and of being practiced or carried out in various ways.

[0143] It must also be noted that, as used in the specification and the appended claims, the singular forms “a,” “an” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” or “approximately” one particular value and / or to “about” or “approximately” another particular value. When such a range is expressed, other exemplary embodiments include from the one particular value and / or to the other particular value.

[0144] By “comprising” or “containing” or “including” is meant that at least the named compound, element, particle, or method step is present in the composition or article or method, but does not exclude the presence of other compounds, materials, particles, method steps, even if the other such compounds, material, particles, method steps have the same function as what is named.

[0145] Unless otherwise expressly stated, it is in no way intended that any method set forth herein be construed as requiring that its steps be performed in a specific order. Accordingly, where a method claim does not actually recite an order to be followed by its steps or it is not otherwise specifically stated in the claims or descriptions that the steps are to be limited to a specific order, it is no way intended that an order be inferred, in any respect. This holds for any possible non-express basis for interpretation, including: matters of logic with respect to the arrangement of steps or operational flow; plain meaning derived from grammatical organization or punctuation; the number or type of embodiments described in the specification.

[0146] While the methods and systems have been described in connection with certain embodiments and specific examples, it is not intended that the scope be limited to the particular embodiments set forth, as the embodiments herein are intended in all respects to be illustrative rather than restrictive.

[0147] The following patents, applications, and publications, as listed below and throughout this document, are hereby incorporated by reference in their entirety herein. Reference List #1Attorney Docket No.10046-657WO1 8531 [1] Lianne Woodward, Peter Anderson, Nicola Austin, Kelly Howard, and Terrie Inder, “Neonatal mri to pre-dict neurodevelopmental outcomes in preterm infants,” The New England Journal of Medicine, 2006. [2] H. Kidokoro, JJ. Neil, and T.E. Inder, “New mr imag-ing assessment tool to define brain abnormalities in very preterm infants at term,” American Journal of Neurora-diology, 2013. [3] Shamik Trivedi, Zachary Vesoulis Rakesh Rao, Steve Liao, Joshua Shimony, Robert McKinstry, and Amit Mathur, “A validated clinical mri injury scoring sys-tem in neonatal hypoxic-ischemic encephalopathy,” Pe-diatric Radiology, 2017. [4] Jessica Dubois, Marianne Alison, Serena Counsell, Lu-cie Hertz-Pannier, Petra Huppi, and Manon Benders, “Mri of the neonatal brain: A review of methodologi-cal challenges and neuroscientific advances,” Journal of Magnetic Resonance Imaging, 2020. [5] Kirsten Thiim, Elizabeth Singh, Srinivasan Mukundan, Ellen Grant, Edward Yang, Mohamed El-Dib, and Terrie Inder, “Clinical experience with an in-nicu magnetic resonance imaging system,” Journal Of Perinatology, 2022. [6] Julian Maclaren, Michael Herbst, Oliver Speck, and Maxim Zaitsev, “Prospective motion correction in brain imaging: a review,” Magnetic resonance in medicine, vol.69, no.3, pp.621–636, 2013. [7] Jakob M. Slipsager, Stefan L. Glimberg, Liselotte Højgaard, Rasmus R. Paulsen, Paul Wighton, M. Dy-lan Tisdall, Camilo Jaimes, Borjan A. Gagoski, P. Ellen Grant, Andre´ van der Kouwe, Oline V. Olesen, and Robert Frost, “Comparison of prospective and retro-spective motion correction in 3d-encoded neuroanatom-ical mri,” Magnetic Resonance in Medicine, 2022. [8] Anagha Deshmane, Vikas Gulani, Mark Griswold, and Nicole Seiberlich, “Parallel mr imaging,” Journal of Magnetic Resonance Imaging, 2012. [9] Shreyas Vasanawala, Marcus Alley, Brian Hargreaves, Richard Barth, John Pauly, and Michael Lustig, “Im-proved pediatric mr imaging with compressed sensing,” Radiology, 2010.

[0010] Reinhard Heckel, Mathews Jacob, Akshay Chaudhari, Or Perlman, and Efrat Shimron, “Deep learning for accelerated and robust mri reconstruction,” Magnetic Resonance Materials in Physics, Biology and Medicine, 2024.

[0011] Ajil Jalal, Marius Arvinte, Giannis Daras, Eric Price, Alexandros G Dimakis, and Jonathan I Tamir, “Ro-bust compressed sensing mri with deep generative pri-ors,” Advances in Neural Information Processing Sys-tems, 2021.Attorney Docket No.10046-657WO1 8531

[0012] Hyungjin Chung and Jong Chul Ye, “Score-based diffu-sion models for accelerated mri,” Medical Image Anal-ysis, 2022.

[0013] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole, “Score-based generative modeling through stochastic differential equations,” 2021.

[0014] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine, “Elucidating the design space of diffusion-based generative models,” 2022.

[0015] Hyungjin Chung, Jeongsol Kim, Michael T. Mccann, Marc L. Klasky, and Jong Chul Ye, “Diffusion posterior sampling for general noisy inverse problems,” 2023.

[0016] Brett Levac, Sidharth Kumar, Ajil Jalal, and Jonathan Tamir, “Accelerated motion correction with deep gen-erative diffusion models,” Magnetic Resonance in Medicine, 2024.

[0017] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady, “Noisier2noise: Learning to denoise from un-paired noisy data,” Conference on Computer Vision and Pattern Recognition, 2020.

[0018] Junaid Rajput, Simon Weinmueller, Jonathan Endres, Peter Dawood, Florian Knoll, Andreas Maier, and Moritz Zaiss, “Death by retrospective undersampling -caveats and solutions for learning-based mri reconstruc-tions,” MICCAI, 2024. Reference List #2 [1'] R. Heckel, M. Jacob, A. Chaudhari, O. Perlman, and E. Shimron, ”Deep learning for accelerated and robust MRI reconstruction,” Magn. Reson. Mater. in Phys, Biol. and Med., vol.37, pp.335-368, Jul.2024. [2'] A. Jalal, M. Arvinte, G. Daras, E. Price, A. Dimakis, J. Tamir, ”Robust Compressed Sensing MRI with Deep Generative Priors,” in Proc. of NeurIPS, 2021, vol.34, pp. 14938-14954. [3'] H. Chung and J.C. Ye, ”Score-based diffusion models for accelerated MRI,” Med. Image Anal., vol.80, pp.102479, Aug.2022. [4'] L.J. Woodward, P.J. Anderson, N.C. Austin, K. Howard, and T.E. Inder, ”Neonatal MRI to predict neurodevelopmental outcomes in preterm infants,” N. Engl. J. Med., vol. 17, pp.685-694, Aug.2006. [5'] H. Kidokoro, J.J. Neil, and T.E. Inder, ”New MR imaging assessment tool to define brain abnormalities in very preterm infants at term,” AJNR Am. J. Neuroradiol., vol. 34, pp.2208-2214, Dec.2013.Attorney Docket No.10046-657WO1 8531 [6'] J. Deubois et al., ”MRI of the Neonatal Brain: A Review of Methodological Challenges and Neuroscientific Advances,” J. Magn. Reson. Imag., vol.53, pp.1318-1343, May 2021. [7'] J.M. Slipsager et al., ”Comparison of prospective and retrospective motion correction in 3D-encoded neuroanatomical MRI,” Magn. Reson. Med., vol.87, pp.629-645, Feb. 2022. [8'] A. Deshmane, V. Gulani, M.A. Griswold, and N. Seiberlich, ”Parallel MR imaging,” J. Magn. Reson. Imag., vol. 36, pp.55-72, July 2012. [9'] S.S. Vasanawala, M.T. Alley, B.A. Hargreaves, R.A. Barth, J.M. Pauly, and M. Lustig, ”Improved pediatric MR imaging with compressed sensing,” Radiology, vol.256, pp. 607-616, Aug.2010. [10'] G. McGibney, M.R. Smith, S.T. Nichols, and A. Crawley, ”Quantitative evaluation of several partial Fourier reconstruction algorithms used in MRI,” Magn. Reson. Med., vol.30, pp.51-59, July 1993. [11'] S. Kiryu et al., ”Clinical Impact of Deep Learning Reconstruction in MRI,” Radiographics, vol.43, pp. e220133, June 2023. [12'] K.R. Thiim et al., ”Clinical experience with an in-NICU magnetic resonance imaging system,” J. Perinatol., vol.42, pp.873-879, July 2022. [13'] A.A. Figaji, ”Anatomical and Physiological Differences between Children and Adults Relevant to Traumatic Brain Injury and the Implications for Clinical Assessment and Care,” Front Neurol., vol.8, Dec.2017. [14'] Y. Song, J. Sohl-Dickstein, D.P. Kingma, A. Kumar, S. Ermon, and B. Poole, ”Score- Based Generative Modeling through Stochastic Differential Equations,” in Proc. Int. Conf. on Learn. Representations, 2021. [15'] T. Karras, M. Aittala, T. Aila, and S. Laine, ”Elucidating the design space of diffusion- based generative models,” in Proc. of NeurIPS, 2022, vol.35, pp.26565-26577. [16'] H. Chung, J. Kim, M.T. Mccann, M.L. Klasky, J.C. Ye, ”Diffusion posterior sampling for general noisy inverse problems,” in Proc. Int. Conf. on Learn. Representations, 2023. [17'] B. Levac, S. Kumar, A. Jalal, and J.I. Tamir, ”Accelerated motion correction with deep generative diffusion models,” Magn. Reson. Med., vol.92, pp.853-868, Apr.2024. [18'] N. Moran, D. Schmidt, Y. Zhong, and P. Coady, ”Noisier2noise: Learning to denoise from unpaired noisy data,” in Proc. IEEE Conf. on Comp. Vis. and Pattern Recognit., 2020, pp 12064-12072.Attorney Docket No.10046-657WO1 8531 [19'] J.R. Rajput et al., ”Death by Retrospective Undersampling-Caveats and Solutions for LearningBased MRI Reconstructions,” in Proc. Int. Conf. on Med. Image Comput. and Computer-Assisted Intervention, 2024, pp 233-241. [20'] J. Wang, D. An, and J.P. Haldar, ”The “hidden noise” problem in MR image reconstruction,” Magn. Reson. Med., vol.92, pp.982-996, Sep.2024. [21'] E.J. Hu et al., ”LoRA: Low-Rank Adaptation of Large Language Models,” in Proc. Int. Conf. on Learn. Representations, 2022.

Claims

Attorney Docket No.10046-657WO1 8531 What is claimed:

1. A method comprising: receiving, by a processor, an MRI image or data object of a subject acquired by an MRI scanner; generating, via a trained neural network model and physics-based motion model operating in combination, a higher-fidelity image or data object of the subject that reduces motion and noise from the received MRI image or data object; and outputting the higher-fidelity image or data object, wherein the higher-fidelity image or data object is employed for diagnosis or treatment of the subject, wherein the trained neural network model was generated by (i) training a denoising AI model using scarce training dataset, (ii) generating a denoised scarce training dataset from the scarce training dataset using the trained denoising AI model, and (iii) applying the scarce training dataset as input to the training of the neural network model, wherein the denoised scarce training dataset is employed as a ground truth during the training.

2. The method of claim 1, wherein the scarce training dataset includes scans of the subject for a plurality of orientations and contrasts, and wherein the training of the neural network model employed the plurality of orientations and contrasts with class embedding.

3. The method of claim 1 or 2, wherein the scarce training dataset includes low-field MRI scans, wherein the low-field MRI scans were acquired with reduced acquisition time.

4. The method of claim 1 or 2, wherein the scarce training dataset includes low-field neonatal MRI scans, wherein the low-field neonatal MRI scans were acquired with reduced acquisition time of a moving subject during the scan.

5. The method of any one of claims 2-4, wherein the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan.

6. The method of claim 5, wherein the at least two orientations are selected from the group consisting of axial orientation, sagittal orientation, coronal orientation, and oblique orientation.Attorney Docket No.10046-657WO1 8531 7. The method of any one of claims 2-4, wherein the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, wherein class embedding is applied for training of the neural network model for each of the at least two orientations for the given contrast.

8. The method of any one of claims 2-4, wherein the scans of the subject for the plurality of orientations include at least two orientations, each for at least two contrast scans selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, wherein class embedding is applied for training of the neural network model for each of the orientations and contrasts.

9. The method of any one of claims 1-8, wherein the MRI image or data object comprises 2D MRI scans or 3D MRI scans.

10. The method of any one of claims 1-9, wherein the MRI image or data object is acquired by a low-field MRI scanner.

11. The method of any one of claims 1-10, wherein the trained neural network model is a trained diffusion generative model.

12. The method of any one of claims 1-11, wherein the denoising AI model was trained using only the scarce training dataset and a statistical model of a noise distribution of the scarce training dataset.

13. A method comprising: receiving, by a processor, an MRI image or data object of a neonatal subject acquired by a low-field neonatal MRI scanner; generating, via a trained neural network model and physics-based motion model operating in combination, a higher-fidelity image or data object of the neonatal subject that reduces motion and noise from the received MRI image or data object; and outputting the higher-fidelity image or data object, wherein the higher-fidelity image or data object is employed for diagnosis or treatment of the neonatal subject,Attorney Docket No.10046-657WO1 8531 wherein the trained neural network model was generated by (i) training a denoising AI model using scarce neonatal training dataset, (ii) generating a denoised scarce training dataset from the scarce neonatal training dataset using the trained denoising AI model, and (iii) applying the scarce neonatal training dataset as input to the training of the neural network model, wherein the denoised scarce training dataset is employed as a ground truth during the training, wherein the scarce neonatal training dataset includes scans of the neonatal subject for a plurality of orientation and contrast, and wherein the training of the neural network model employed the plurality of orientation and contrast with class embedding.

14. The method of claim 13, wherein the scarce neonatal training dataset includes a low field MRI scans, wherein the low field MRI scans was acquired with reduced acquisition time of a moving neonatal subject during the scan, and wherein the MRI image or data object of the neonatal subject was acquired a moving neonatal subject during the scan.

15. The method of claim 13 or 14, wherein the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan, wherein the at least two orientations are selected from the group consisting of axial orientation, sagittal orientation, coronal orientation, and oblique orientation.

16. The method of any one of claims 13-15, wherein the scans of the subject for the plurality of orientations include at least two orientations for a given contrast scan selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, and wherein each class embedding is applied for training of the neural network model for each of the at least two orientations for the given contrast.

17. The method of any one of claims 13-15, wherein the scans of the subject for the plurality of orientations include at least two orientations, each for at least two contrast scans selected from the group consisting of spin echo, fast spin echo, gradient echo, diffusion weighted imaging, and susceptibility weighted imaging, wherein each class embedding is applied for training of the neural network model for each of the orientations and contrasts.Attorney Docket No.10046-657WO1 8531 18. The method of any one of claims 13-17, wherein the MRI image or data object comprises 2D MRI scans or 3D MRI scans.

19. The method of any one of claims 13-18, wherein the MRI image or data object is acquired by a low-field MRI scanner.

20. The method of any one of claims 13-19, wherein the trained neural network model is a trained diffusion generative model.

21. The method of any one of claims 13-20, wherein the denoising AI model was trained using only the scarce training dataset and a statistical model of a noise distribution of the scarce training dataset.

22. A method of training an AI model comprising: receiving, by a processor, a scarce training dataset comprising an MRI image or data object of a subject acquired by an MRI scanner; training a denoising AI model using the scarce training dataset; generating a denoised scarce training dataset from the scarce training dataset using the trained denoising AI model, and applying the scarce training dataset as input to the training a the neural network model (e.g., diffusion generative model), wherein the denoised scarce training dataset is employed as a ground truth during the training in combination with the input.

23. A system comprising: a processor; and a memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to: receive an MRI image or data object of a subject acquired by an MRI scanner; generate, via a trained neural network model and physics-based motion model operating in combination, a higher-fidelity image or data object of the subject that reduces motion and noise from the received MRI image or data object; and output, the higher-fidelity image or data object, wherein the higher-fidelity image or data object is employed for diagnosis or treatment of the subject,Attorney Docket No.10046-657WO1 8531 wherein the trained neural network model was generated by (i) training a denoising AI model using scarce training dataset, (ii) generating a denoised scarce training dataset from the scarce training dataset using the trained denoised AI model, and (iii) applying the scarce training dataset as input to the training of the neural network model, wherein the denoised scarce training dataset is employed as a ground truth during the training.

24. The system of claim 23, wherein the system is configured by a method of any one of claims 1-22.

25. A non-transitory computer-readable medium having instructions stored thereon, wherein execution of the instructions by a processor causes the processor to execute the system of any one of claims 23 and 24 or to perform a method of any one of claims 1-22.

Images