Methods and systems for optimizing magnet array configurations

Abstract

Disclosed herein is a method for optimizing a magnet array configuration by a deep learning model. The method comprises providing an initialized magnetic strength into a deep learning model, adjusting parameters of a magnetic array to determine an optimized magnetic field, and outputting a learned magnetization strength according to convergence of a loss function.

Description

Attorney Docket No. 59984-715601METHODS AND SYSTEMS FOR OPTIMIZING MAGNET ARRAY CONFIGURATIONSCROSS-REFERENCE

[0001] This application claims the benefit of U.S. Patent Application No. 63 / 740,816 filed December 31, 2024, and U.S. Patent Application No. 63 / 741,377 filed January 2, 2025, which are incorporated herein in their entirety.BACKGROUND

[0002] MRI devices are often large and complex machines that are expensive and inaccessible to many patients. Low-cost portable MRIs can be provided to increase patient access to MRI devices at a lower cost. Low-cost portable MRIs often have inhomogeneous magnetic fields due to their magnet array configuration, impacting image quality.SUMMARY

[0003] Disclosed herein, is a method for optimizing a magnet array, comprising:inputting an initialized magnetization strength for a plurality of magnets within a magnet array into a convolutional neural network, defining a loss function, at the neural network, wherein the loss function measures a homogeneity of a magnetic field, based at least in part on the initialized magnetic strength, training the neural network by minimizing the loss function to generate a trained neural network, wherein minimizing the loss function is based at least in part on adjusting parameters of the neural network and adjusting a plurality of parameters of the magnet array during the training, and using the trained neural network to output a learned magnetization strength, based at least in part on minimizing the loss function.

[0004] In some embodiments, the initialized magnetization strength is determined by an initial size, position, orientation, or any combination thereof, of the plurality of magnets.

[0005] In some embodiments, the initialized magnetization strength is at least about 70 mT.

[0006] In some embodiments, the plurality of parameters of the magnet array comprises at least one of a size, position, or orientation of the plurality of magnets.

[0007] In some embodiments, the neural network is a convolutional neural network (CNN).

[0008] In some embodiments, the CNN contains a plurality of convolutional layers.

[0009] In some embodiments, the plurality of convolutional layers comprise a rectified linear unit (ReLU) between the plurality of convolutional layers.

[0010] In some embodiments, the plurality of convolutional layers comprise at least four convolutional layers.Attomey Docket No. 59984-715601

[0011] In some embodiments, the plurality of convolutional layers comprise about eight convolutional layers.

[0012] In some embodiments, the plurality of convolutional layers comprise kernels having a kernel size n*n with n>2.

[0013] In some embodiments, n is 3.

[0014] In some embodiments, the initializing the magnetization strength is done at a predetermined region of interest (RO I) of the magnet array.

[0015] In some embodiments, the magnetic field loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in parts per million (ppm).

[0016] In some embodiments, the loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in magnetization magnitude (M).

[0017] In some embodiments, the training the neural network further comprises using an Adaptive Moment Estimation (ADAM) optimization technique for gradient descent.

[0018] In some embodiments, the magnet array is a Halbach array.

[0019] In some embodiments, the Halbach array is configured in a cylindrical Halbach array.

[0020] In some embodiments, disclosed herein is a method for optimizing a magnetic field homogeneity of a magnet array by adjusting parameters comprising size, orientation, angle, and position of a plurality of magnetic blocks within the magnet array by a deep learning model.

[0021] In some embodiments, the present disclosure provides a system comprising one or more computer processors and computer memory coupled thereto. The computer memory comprises machine executable code that, upon execution by the one or more computer processors, implements any of the methods above or elsewhere herein.

[0022] Disclosed herein is a system comprising a magnet array comprising a plurality of magnets; and a processor configured to: input an initialized magnetization strength for the plurality of magnets into a neural network; define a loss function, at the neural network, based at least in part on the initialized magnetic strength, wherein the loss function measures a homogeneity of a magnetic field; train the neural network by minimizing the loss function to generate a trained neural network, wherein minimizing the loss function is based at least in part on adjusting parameters of the neural network and adjusting a plurality of parameters of the magnet array during the training; and use the neural network to output a learned magnetization strength, based at least in system on minimizing the loss function. In some embodiments, the processor is further configured to determine the initialized magnetization strength by an initial size, position, orientation or any combination thereof, of the plurality of magnets. In some embodiments, the initialized magnetization strength is at least about 70 ml. In some embodiments, the plurality of parameters of the magnet array comprises at least one of a size,Attomey Docket No. 59984-715601position, or orientation of the plurality of magnets. In some embodiments, the neural network is a convolutional neural network (CNN). In some embodiments, the CNN contains a plurality of convolutional layers. In some embodiments, the plurality of convolutional layers comprise a rectified linear unit (ReLU) between the plurality of convolutional layers. In some embodiments, the plurality of convolutional layers comprise at least four convolutional layers. In some embodiments, the plurality of convolutional layers comprise about eight convolutional layers. In some embodiments, the plurality of convolutional layers comprise kernels having a size n X n with n > 2. In some embodiments, n is 3. In some embodiments, the initialized magnetization strength is from a predetermined region of interest (RO I) of the magnet array. In some embodiments, the loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in parts per million (ppm). In some embodiments, the loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in magnetization magnitude (M). In some embodiments, the processor is configured to train the neural network using an Adaptive Moment Estimation (ADAM) optimization technique for gradient descent. In some embodiments, the magnet array is a Halbach array. In some embodiments, the Halbach array is configured in a cylindrical Halbach array.

[0023] Disclosed herein is a method for denoising a magnetic resonance imaging (MRI) device image, the method comprising: generating a magnetic resonance image from a primary coil of an MRI device; generating a reconstructed image from the magnetic resonance image using a reconstruction pipeline; generating a plurality of values from a plurality of external noise coils of the MRI device; generating, using an MRI feature extraction network, an image substantially free of electromagnetic interference (EMI), based at least in part on the magnetic resonance image; training a deep learning model to combine the plurality of values into a simulated EMI using an EMI feature extraction network; optimizing a loss function during the training of the deep learning model using the image substantially free of EMI, the simulated EMI, and the reconstructed image, wherein the loss function minimizes a difference between the image substantially free of EMI, the simulated EMI, and the reconstructed image; and producing a denoised image using the deep learning model, based at least in part on the minimized loss function.

[0024] In some embodiments, the MRI feature extraction network comprises a plurality of convolutional layers. In some embodiments, the plurality of convolutional layers comprise kernels of size n X n, with n >2. In some embodiments, n is 3. In some embodiments, the EMI feature extraction network comprises a plurality of convolutional layers. In some embodiments, the plurality of convolutional layers comprise kernels of size n X n, with n >2. In some embodiments, n is 3. In some embodiments, the EMI feature extraction network learns to outputAttomey Docket No. 59984-715601a simulated EMI of the primary coil, based at least in part on the plurality of noise signals. In some embodiments, the loss function minimizes a discrepancy between the substantially EMI-free image and the simulated EMI-free image. In some embodiments, the EMI feature extraction network includes a rectified linear unit (ReLu) activation function configured to perform nonlinear transformation between the plurality of convolutional layers. In some embodiments, the MRI feature extraction network includes a rectified linear unit (ReLu) activation function configured to perform nonlinear transformations between the plurality of convolution layers of the MRI feature extraction network. In some embodiments, the plurality of convolutional layers comprise at least four layers. In some embodiments, the plurality of convolutional layers comprise at most 10 layers. In some embodiments, the plurality of convolutional layers comprise at least four layers. In some embodiments, the plurality of convolutional layers comprise at most 10 layers.

[0025] Disclosed herein is a system comprising: a magnetic resonance imaging (MRI) device comprising a primary coil and a plurality of external noise coils, wherein the primary coil is configured to generate a magnetic resonance image; a computer processor configured to: generate a reconstructed image from the magnetic resonance image using a reconstruction pipeline; generate a plurality of values from the plurality of external noise coils; generate, using an MRI feature extraction network, an image substantially free of electromagnetic interference (EMI), based at least in part on the magnetic resonance image; train a deep learning model to combine the plurality of values into a stimulated EMI using an EMI feature extraction network; optimize a loss function during the training of the deep learning model using the image substantially fee of EMI, the stimulated EMI, and the reconstructed image, wherein the loss function minimizes a difference between the image substantially free of EMI, the stimulated EMI, and the reconstructed image; and produce a denoised image using the deep learning model, based at least in part on the minimized loss function. In some embodiments, the MRI device is a low field MRI device. In some embodiments, the MRI feature extraction network comprises a plurality of convolutional layers. In some embodiments, the plurality of convolutional layers comprise kernels of size n X n, with n >2. In some embodiments, n is 3. In some embodiments, the EMI feature extraction network comprises a plurality of convolutional layers. In some embodiments, the plurality of convolutional layers comprise kernels of size n X n, with n >2. In some embodiments, n is 3. In some embodiments, the EMI feature extraction network learns to output a simulated EMI of the primary coil, based at least in part on the plurality of noise signals. In some embodiments, the loss function minimizes a discrepancy between the substantially EMI-free image and the simulated EMI-free image. In some embodiments, the EMI feature extraction network includes a rectified linear unit (ReLu) activation function configured to performAttomey Docket No. 59984-715601nonlinear transformation between the plurality of convolutional layers. In some embodiments, the MRI feature extraction network includes a rectified linear unit (ReLu) activation function configured to perform nonlinear transformations between the plurality of convolution layers of the MRI feature extraction network. In some embodiments, the plurality of convolutional layers comprise at least four layers. In some embodiments, the plurality of convolutional layers comprise at most 10 layers. In some embodiments, the plurality of convolutional layers comprise at least four layers. In some embodiments, the plurality of convolutional layers comprise at most 10 layers.

[0026] Provided herein is a method for optimizing a magnet array. The method can comprise training a neural network by minimizing a loss function measuring a homogeneity of a magnetic field based at least in part on adjusting parameters of the neural network and adjusting a plurality of parameters of the magnet array during the training. The method can comprise using the trained neural network to output a learned magnetization strength.

[0027] Disclosed herein is a method for denoising a magnetic resonance imaging (MRI) device image. The method can comprise optimizing a loss function during training of a deep learning model using (a) an image substantially free of electromagnetic interference (EMI) generated via an MRI feature extraction network, (b) a simulated EMI, and (c) a reconstructed MRI image. In some cases, the loss function minimizes a difference between the image substantially free of EMI, the simulated EMI, and the reconstructed image. The method can comprise producing a denoised image using the deep learning model.

[0028] Additional aspects and advantages of the present disclosure will become readily apparent to those skilled in this art from the following detailed description, wherein only illustrative embodiments of the present disclosure are shown and described. As will be realized, the present disclosure is capable of other and different embodiments, and its several details are capable of modifications in various obvious respects, all without departing from the disclosure.Accordingly, the drawings and description are to be regarded as illustrative in nature, and not as restrictive.INCORPORATION BY REFERENCE

[0029] All publications, patents, and patent applications mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent, or patent application was specifically and individually indicated to be incorporated by reference. To the extent publications and patents or patent applications incorporated by reference contradict the disclosure contained in the specification, the specification is intended to supersede and / or take precedence over any such contradictory material.Attorney Docket No. 59984-715601BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The novel features of the disclosure are set forth with particularity in the appended claims. A better understanding of the features and advantages of the present disclosure will be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the disclosure are utilized, and the accompanying drawings (also “Figure” and “FIG.” herein), of which:

[0031] FIG. 1 provides a schematic illustration of magnet array design optimization using deep learning in accordance with example embodiments described herein.

[0032] FIG. 2A is an image depicting a predefined ROI located within a Halbach cylinder magnet setup in accordance with example embodiments described herein.

[0033] FIG. 2B provides the evolution of the maximum, minimum and mean magnetization strength that have been learned and optimized during the training of a convolutional neural network (CNN), across the ROI depicted in FIG. 2A in accordance with example embodiments described herein.

[0034] FIG. 2C provides the progression loss function values and parts per million (ppm) deviations across training epochs across the ROI depicted in FIG. 2A in accordance with example embodiments described herein.

[0035] FIG. 3A is an image depicting a predefined ROI located within a Halbach cylinder magnet setup in accordance with example embodiments described herein.

[0036] FIG. 3B provides evolution of a maximum, minimum and mean magnetization strength that have been learned and optimized during the training of the CNN across the ROI depicted in FIG. 3A in accordance with example embodiments described herein.

[0037] FIG. 3C provides progression loss function values and the ppm deviations across training epochs across the ROI depicted in FIG. 3A in accordance with example embodiments described herein.

[0038] FIG. 4A is an image depicting a predefined ROI located within a Halbach cylinder magnet setup in accordance with example embodiments described herein.

[0039] FIG. 4B provides evolution of a maximum, minimum and mean magnetization strength that have been learned and optimized during the training of the CNN across the ROI depicted in FIG. 4A in accordance with example embodiments described herein.

[0040] FIG. 4C provides progression loss function values and the ppm deviations across the training epochs across the ROI depicted in FIG. 4A in accordance with example embodiments described herein.

[0041] FIG. 5 shows a computer system that is programmed or otherwise configured to implement methods provided herein in accordance with example embodiments described herein.Attomey Docket No. 59984-715601

[0042] FIG. 6A illustrates a method for EMI reduction in accordance with example embodiments described herein.

[0043] FIG. 6B illustrates a reconstruction pipeline in accordance with example embodiments described herein.

[0044] FIG. 6C illustrates architecture of an EMI feature extraction network in accordance with example embodiments described herein.

[0045] FIG. 6D illustrates architecture of an MRI feature extraction network in accordance with example embodiments described herein.

[0046] FIG. 7 illustrates noise reduction performance of the method disclosed herein, compared to EMI-contaminated images in a first human brain image, in accordance with example embodiments described herein.

[0047] FIG. 8 illustrates noise reduction performance of the method disclosed herein, compared to EMI-contaminated images in a second human brain image, in accordance with example embodiments described herein.

[0048] FIG. 9 illustrates noise reduction performance of the method disclosed herein, compared to EMI-contaminated images in a third human brain image, in accordance with example embodiments described herein.

[0049] FIG. 10 illustrates noise reduction performance of the method disclosed herein, compared to EMI-contaminated images in a fourth human brain image, in accordance with example embodiments described herein.

[0050] FIG. 11 illustrates noise reduction performance of the method disclosed herein, compared to EMI-contaminated images in a fifth human brain image, in accordance with example embodiments described herein.

[0051] FIG. 12 illustrates noise reduction performance of the method disclosed herein, compared to EMI-contaminated images in a sixth human brain image, in accordance with example embodiments described herein.

[0052] FIG. 13A depicts components of an MRI scanning system including a dome-shaped housing for a magnetic array, the dome-shaped housing surrounding a region of interest therein and further depicting the dome-shaped housing positioned to receive at least a portion of the head of a patient reclined on the table into the region of interest, in accordance with example embodiments described herein.

[0053] FIG. 13B depicts a patient’s head positioned in the region of interest of the MRI scanning system of FIG. 13A, in accordance with example embodiments described herein.Attomey Docket No. 59984-715601

[0054] FIG. 14 is a perspective view of an alternative dome-shaped housing for a magnetic array for use with the MRI scanning system of FIG. 13, wherein access apertures are defined in the dome-shaped housing, in accordance with example embodiments described herein.

[0055] FIG. 15 is a perspective view of an alternative dome-shaped housing for a magnetic array for use with the MRI scanning system of FIG. 13, wherein access apertures and an adjustable gap is defined in the dome-shaped housing, in accordance with example embodiments described herein.

[0056] FIG. 16 depicts a dome-shaped housing for use with an MRI scanning system having an access aperture in the form of a centrally-defined hole, in accordance with example embodiments described herein.

[0057] FIG. 17 is a cross-sectional view of the dome-shaped housing of FIG. 16, in accordance with example embodiments described herein.DETAILED DESCRIPTION

[0058] The present disclosure is related to magnetic resonance imaging (MRI) magnetic array configuration.

[0059] Provided herein is a method for optimizing permanent magnet arrays by deep learning models. In some cases, a method for optimizing a magnet array may comprise initializing a magnetization strength for a plurality of magnets within a magnet array into a neural network; adjusting at least one magnet array parameter to minimize a magnetic field loss function;

[0060] training the neural network based at least in part on the adjusting the at least one magnet array parameter to minimize the magnetic field loss function; and outputting a learned magnetization strength from the neural network. In some cases, the initialized magnetization strength is determined by an initial size, position, orientation, or any combination thereof, of the plurality of magnets. In some cases, the magnet array parameter comprises at least one of a size, position, or orientation of the plurality of magnets. In some cases, the neural network is a convolutional neural network (CNN). In some cases, the initializing the magnetization strength is done for a predetermined region of interest (ROI) of the magnet array. In some cases, the magnetic field loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in parts per million (ppm). In some cases, the loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in magnetization magnitude (M). In some cases, the magnet array is a Halbach array. In some cases, the Halbach array is configured in a cylindrical Halbach array.

[0061] Optimizing permanent magnet arrays by the methods disclosed herein can be used in a low-cost portable MRI. A cylindrical Halbach array magnet configuration is a permanent magnet array configuration that can produce a homogeneous Bo field inside a cylindrical shell and zeroAttomey Docket No. 59984-715601Bo field outside of the shell. Homogeneity of the Bo field inside the cylinder is important for MRI applications, wherein the homogeneity of the Bo field impacts features comprising image distortion, artifacts, and signal-to-noise ratio (SNR). To develop a numerical optimization model for a magnet array involves describing a magnetic field distribution. In some cases, the magnetic field distribution can be optimized by adjusting parameters of a magnetic array comprising shapes, sizes, positions, angular orientations, and material grades for at least one magnet blocks in a plurality of magnetic blocks within an array.

[0062] Provided herein is a deep learning method to address the challenge of optimizing magnet array designs. Optimization is crucial for producing a homogeneous magnetic field, denoted as Bo, which is important for effective functioning of low-field portable MRI scanners. A homogeneous Bo field ensures that a MRI scanner can generate clear and accurate images with less distortions. The magnet array configuration depends on several factors comprising 1) a predefined imaging ROI, 2) a position of at least one magnet block in a plurality of magnet blocks configured in a Halbach array, 3) an initial orientation and angle of the at least one magnet block in the plurality of magnet blocks, 4) a size of the at least one magnet block in the plurality of magnet blocks, and 5) a total number of magnetic blocks in the plurality of magnetic blocks.

[0063] The method disclosed herein can address challenges inherent to magnet array configurations. Optimizing magnet array configurations is important for producing a homogenous magnetic field. Often, low-cost portable MRI devices comprise magnetic arrays that produce an inhomogeneous magnetic field. A homogenous magnetic field ensures that an MRI device can generate clear and accurate images with less distortions.

[0064] In some embodiments, the method disclosed herein provides a method for optimizing a magnet array configuration by using a deep learning model.

[0065] In some embodiments, the magnet array configuration can comprise a Halbach array configuration. In some embodiments, the magnet array configuration can comprise permanent magnets. In some embodiments, the magnet array configuration can comprise permanent magnets in a Halbach array configuration. In some embodiments, the magnet array configuration can comprise permanent magnets in a cylindrical Halbach array configuration. In some embodiments, the magnet array comprises at least one rectangular magnet block in a plurality of magnet blocks. In some embodiments, the magnet array comprises at least one cylindrical magnet block in a plurality of magnet blocks.

[0066] In some embodiments of the method disclosed herein, a magnet array configuration can be adjusted to optimize magnetic field homogeneity using a deep learning model. In some embodiments of the method, a magnet array configuration can be adjusted to maximize magnetic field homogeneity using a deep learning model. In some embodiments, a magnet arrayAttomey Docket No. 59984-715601configuration can be adjusted to minimize magnetic field inhomogeneity using a deep learning model. In some embodiments, the deep learning model comprises a convolutional neural network (CNN), a multilayer perceptron, a neural network, an autoencoder, or any combination thereof. In some embodiments of the method, a magnet array configuration can be adjusted to maximize magnetic field homogeneity using a CNN. In some embodiments of the method, a magnet array configuration can be adjusted to minimize magnetic field inhomogeneity using a CNN. In some embodiments, a magnet array configuration can be adjusted to optimize magnetic field homogeneity by adjusting parameters comprising size, shape, orientation, angle, position, or any combination thereof, of at least one magnetic block in a plurality of magnetic blocks within a magnetic array. In some embodiments, the magnet array configuration can be adjusted to optimize magnetic field homogeneity by adjusting parameters comprising size, shape, orientation, angle, position, or any combination thereof, of the least one magnetic block in the plurality of magnetic blocks within the magnetic array using a CNN.

[0067] Optimizing permanent magnet arrays by the methods disclosed herein can be used in a low-cost portable MRI. A cylindrical Halbach array magnet configuration is a permanent magnet array configuration that can produce a homogeneous BO field inside a cylindrical shell and zero BO field outside of the shell. Homogeneity of the BO field inside the cylinder is important for MRI applications, wherein the homogeneity of the BO field impacts features comprising image distortion, artifacts, and signal-to-noise ratio (SNR). To develop a numerical optimization model for a magnet array involves describing a magnetic field distribution. In some cases, the magnetic field distribution can be optimized by adjusting parameters of a magnetic array comprising shapes, sizes, positions, angular orientations, and material grades for at least one magnet blocks in a plurality of magnetic blocks within an array.

[0068] Provided herein is a deep learning method to address the challenge of optimizing magnet array designs. Optimization is crucial for producing a homogeneous magnetic field, denoted as Bo, which is important for effective functioning of low-field portable MRI scanners. A homogeneous Bo field ensures that a MRI scanner can generate clear and accurate images with less distortions. The magnet array configuration depends on several factors comprising 1) a predefined imaging ROI, 2) a position of at least one magnet block in a plurality of magnet blocks configured in a Halbach array, 3) an initial orientation and angle of the at least one magnet block in the plurality of magnet blocks, 4) a size of the at least one magnet block in the plurality of magnet blocks, and 5) a total number of magnetic blocks in the plurality of magnetic blocks.

[0069] In some embodiments, disclosed herein is a method for optimizing a magnetic field homogeneity of a magnet array by adjusting parameters comprising size, orientation, angle and position of a plurality of magnetic blocks within the magnet array by a deep learning model.Attomey Docket No. 59984-715601Magnet Array Design Optimization

[0070] The present disclosure provides a method for optimizing magnet array configurations. The method provides a deep learning model which can minimize a loss function. The deep learning model can comprise a CNN, a multi-layered perceptron, a recurrent neural network, an autoencoder, or any combination thereof. The method can provide a CNN to minimize the loss function. The loss function can calculate inhomogeneity of a magnetic field Bo across a target ROI. The deep learning model can adjust parameters related to a plurality of magnetic blocks comprising position, angle, orientation, or any combination thereof, of the plurality of magnetic blocks. The deep learning model can adjust parameters related to at least one magnetic block in a plurality of magnetic blocks comprising position, angle, orientation, or any combination thereof, of the at least one magnetic block in the plurality of magnetic blocks. The plurality of magnetic blocks can be configured in a Halbach array configuration. The deep learning model can adjust the configuration of the plurality of magnetic blocks to minimize the inhomogeneity of the magnetic field Bo across the ROI. The deep learning model can adjust the configuration of the plurality of magnetic blocks to maximize the uniformity of the magnetic field Bo across the ROI. The deep learning model can adjust the configuration of the plurality of magnetic blocks configured in a Halbach array to minimize the inhomogeneity of the magnetic field Bo across the ROI. The deep learning model can adjust the configuration of the plurality of magnetic blocks configured in a Halbach array to maximize the uniformity of the magnetic field BO across the ROI.

[0071] The method disclosed herein for optimizing magnet array design can be iterated to adjust parameters of the plurality of magnetic blocks comprising adjusting a size of at least one magnetic block in the plurality of magnetic blocks, adjusting a position of at least one magnetic block in the plurality of magnetic blocks, adjusting an orientation of at least one magnetic block in the plurality of magnetic blocks, or any combination thereof. The adjusting the parameters can be performed based on feedback from a loss of function computation performed by the deep learning model. The adjusting the parameters from feedback from the loss of function computation performed by the deep learning model can optimize the configuration of the plurality of magnetic blocks to minimize inhomogeneity of the magnetic field Bo across the ROI. The adjusting the parameters from the feedback from the loss of function computation performed by the deep learning model can optimize the configuration of the plurality of magnetic blocks in a Halbach array configuration to minimize inhomogeneity of the magnetic field Bo across the ROI. The adjusting the parameters from feedback from the loss of function computation performed by the deep learning model can optimize the configuration of the plurality of magnetic blocks to maximize uniformity of the magnetic field Bo across the ROI. The adjusting theAttorney Docket No. 59984-715601parameters from the feedback from the loss of function computation performed by the deep learning model can optimize the configuration of the plurality of magnetic blocks in a Halbach array configuration to maximize uniformity of the magnetic field Bo across the ROI.

[0072] A magnetic field homogeneity loss function as described herein can be defined as a magnetic field homogeneity parts per million (ppm) of a Bo across a target ROI, according to:

[0073] The magnetic field homogeneity loss function can be minimized comprising adjusting a size of at least one magnetic block in the plurality of magnetic blocks in at least one dimension. The at least one dimension can comprise a length, width, height, or any combination thereof, of the at least one magnetic block in the plurality of magnetic blocks.

[0074] The magnetic field homogeneity loss function can be constrained to ensure that an optimization solution complies with permissible conditions. The permissible conditions can comprise a mean magnitude of the Bo field across a target ROI, and a magnetization magnitude (M) in a form required by a Dipole matrix. The magnetization magnitude may remain below a maximum allowable magnetization magnitude (Mmax). The Mmax can be measured in units of magnetic dipole moment. The mean magnitude of the Bo field across a target ROI can be defined as:Bmean = IIBII{ROI}

[0075] In some cases, the Bmean can be constrained to not surpass about 70 millitesla (mT). In some embodiments, the Bmean can be constrained to not surpass from about 35 to about 50 mT, from about 50 to about 70 mT, from about 70 to about 85 mT, from about 95 to about 105 mT, or from about 35 to about 105 mT. The magnetic field homogeneity loss function can be constrained according to a constraint loss function:^constraint — Bmean — 70) + (M — Mzm)

[0076] The loss function can be minimized by the deep learning model comprising the magnetic field homogeneity loss function and the constraint loss function according to:FossFunction — Homogeneity + Fconstraint

[0077] FIG. 1 provides a schematic illustration of an example method 100 for minimizing the loss function by the deep learning model. For example, provided in FIG. 1, an initialized magnetic strength 101 was input to the deep learning model. In some cases, the initialized magnetic strength can be about 70 mT. In some embodiments, the initialized magnetic strength can be from about 35 to about 105. In some embodiments, the initialized magnetic strength can be from about 35 to about 50, from about 35 to about 70, from about 35 to about 85, from about 35 to about 95, from about 35 to about 105, from about 50 to about 70, from about 50 to aboutAttorney Docket No. 59984-71560185, from about 50 to about 95, from about 50 to about 105, from about 70 to about 85, from about 70 to about 95, from about 70 to about 105, from about 85 to about 95, from about 85 to about 105, or from about 95 to about 105. In some embodiments, the initialized magnetic strength can be about 35, about 50, about 70, about 85, about 95, or about 105. In some embodiments, the initialized magnetic strength can be at least about 35, about 50, about 70, about 85, or about 95. In some embodiments, the initialized magnetic strength can be from at most about 50, about 70, about 85, about 95, or about 105. In some cases, the deep learning model can comprise a CNN 102. In some cases, the CNN can comprise a plurality of convolution layers in one dimension (ID). In some cases, the convolution layers can comprise 1024 kernels. In some embodiments, the convolution layer can comprise about 1024 kernels. In some embodiments, the convolution layers can comprise from about 524 to about 724 kernels, from about 724 to about 1024 kernels, from about 1024 to about 1224 kernels, from about 1224 to aboutl524 kernels, or from about 524 to about 1524 kernels. In some cases, the kernels can have a kernel size of 3x3. In some embodiments, the kernels can have a kernel size of 2x2. In some embodiments, the kernels can have a kernel size of 4x4. In some embodiments, the kernels can have a kernel size of 5x5. In some embodiments, the kernels can have a kernel size of 6x6. In some embodiments, the kernels can have a kernel size of 7x7. In some embodiments, the kernels can have a kernel size of 8x8. In some embodiments, the kernels can have a kernel size of 9x9. In some embodiments, the kernels can have a kernel size of 10x10. In some embodiments, the kernels can have a kernel size of greater thanlOxlO.In some cases, the CNN 102 can employ a rectifier linear unit (ReLu) activation function for non-linear processing between the plurality of convolution layers. In some cases, the deep learning model can comprise a CNN 102 comprising more than about eight convolution layers in one dimension. In some cases, the deep learning model can comprise a CNN 102 comprising at least about four convolutional layers. In some cases, the CNN 102 can output a learned magnetization strength 103. In some cases, the learned magnetization strength 103 can comprise an optimized magnetization strength and a flux density map for at least one magnet block in the plurality of magnet blocks. In some cases, the learned magnetization strength can be achieved when the loss function 104 value converges. In some cases, training the CNN 102 can extend over a total of around 10,000 epochs, and can comprise a learning rate or training step size set at 0.0001. In some embodiments, training the CNN 102 can extend over a total from about 5,000 to 7,500 epochs, from about 7,500 to about 10,000 epochs, from about 10,000 to about 12,500 epochs, from about 12,500 to about 15,000 epochs, or from about 5,000 to about 15,000 epochs. In some embodiments, training the CNN 102 can comprise a learning rate or training step size set at about 0.0001. In some embodiments, training the CNN 102 can comprise a learning rate or training step size set from about 0.00001 to about 0.00005, from about 0.00005Attomey Docket No. 59984-715601to about 0.0001, from about 0.0001 to about 0.00015, from about 0.00015 to about 0.0002, or from about 0.00001 to about 0.0002. In some cases, the learning rate comprises a training step size. In some cases, training the CNN comprises using an Adaptive Moment Estimation (ADAM). In some cases, the ADAM technique optimizes for gradient descent. In some cases, the learned magnetization strength and the flux density map of the at least one magnet block can be validated by a simulation software such as COMSOL.

[0078] Initial measured parameters of the magnetic field can comprise a quantification of size, position, quantity, and orientation of the at least one magnetic block in the plurality of magnetic blocks. In some cases, a pre-designated ROI can be input to the deep learning model to minimize the homogeneity of the magnetic field.

[0079] The ppm of a magnetic field can express the uniformity or inhomogeneity of the magnetic field. The ppm can be used to measure the deviation of the magnetic field from an initial value over a specific volume or ROI. A deviation of the ppm of a magnetic field can be calculated by comparing a local maximum magnetic field strength to a local minimum magnetic field strength, wherein a difference between the local maximum magnetic field strength and the local minimum magnetic field strength are expressed as parts per million (ppm). Inhomogeneity can be quantified using ppm according to:PPM = (max (||B||2) - mm (||B||2)) x 106,{ROI} {ROI}

[0080] wherein 106scales the fraction to parts per million (ppm).

[0081] The method disclosed herein may be implemented to optimize magnetic field strength of a magnet array. In some cases, the magnet array can have an average magnetic field strength that exceeds 70 mT. In some embodiments, the magnet array can have an average magnetic field strength of from about 70 to about 80mT, from about 80 to about 90 mT, from about 90 to about 100 mT, from about 100 to about 110 mT, from about 110 to about 120 mT, from about 120 to about 130 mT, from about 130 to aboutl40 mT, or from about 140 to about 150 mT. In some embodiments, the magnet array can have an average magnetic field strength that exceeds 150 mT. In some cases, the magnetization strength for at least one of the magnetic blocks in the plurality of magnetic blocks can be capped at a predefined maximum value. For example, FIGs.2A-4C illustrate effectiveness of the method disclosed herein to minimize a magnetic field inhomogeneity across different ROIs implementations with constraints comprising an average magnetic field that must exceed 70 mT.

[0082] FIG. 2A provides a predefined ROI within a Halbach cylinder magnet array. The ID CNN as disclosed herein was used to reduce inhomogeneity of the magnetic field within the predefined ROI 202, while adhering to the magnetization strength limits. FIG. 2B provides a magnetic field analysis across training epochs 204, and the evolution of a maximum (maxAttomey Docket No. 59984-715601Bmag), minimum (min Bmag), and mean magnetization (mean Bmag) strengths that were learned and optimized during training of the ID CNN. The magnetic field analysis 204 tracks how changes to the maximum, minimum, and mean magnetization strengths change across the training epochs. A Bmin value represents a lower bound constraint, wherein the lower bound constraint is the minimum average magnetization strength that the model must maintain. The Bmin used during this experiment was 70 ml. FIG. 2B provides a progression of the loss function values and the ppm deviations across training epochs 206. FIG. 2C illustrates the success of the model in progressively reducing errors and deviations in the magnetic field strength across epochs as training progresses 206.

[0083] FIGS. 3A-3C provide magnetic field analysis across epochs 304 and magnetic field PPM analysis across epochs 306, as disclosed herein, for a different ROI 302 implementation. FIGS.4A-4C provide magnetic field analysis across epochs 404, and magnetic field PPM analysis across epochs 406, as disclosed herein, for a different ROI 402 implementation.MRI Image Denoising Using Self-Supervised Deep Learning

[0084] Magnetic resonance imaging (MRI) devices often require shielding from electromagnetic interference (EMI), such as by Faraday shielding. MRI devices are expensive and can be inaccessible to patients due to limitations such as the requirement for shielding from EMI. Point-of-care use of an MRI device requires operation of the MRI device outside of a Faraday shielded room. Deep learning models are increasingly applied in image processing. However, these methods can demand a significant volume of ground truth data for training.

[0085] The present disclosure provides a self-SUPERvised learning approach for Comprehensive Elimination for Electromagnetic Anomalies and Noise (SUPERCLEAN) that reduces noise in low field MRI images without the need for EMI-free images as ground truth references.

[0086] In some embodiments, disclosed herein, is a method for denoising an MRI device image from an MRI device using a self-supervised deep learning framework, without the need for ground truth reference images for training.

[0087] In some embodiments, the MRI device can comprise a primary coil and a plurality of external noise coils (Ac). In some embodiments, the primary coil can generate a magnetic field to enable the MRI device to capture the MRI device image. In some embodiments, the plurality of external noise coils can function as external detectors to reduce electromagnetic noise surrounding the primary coil. In some embodiments, the MRI device can be a low field MRI device.

[0088] Deep learning models require substantial clean, true reference images to train the deep learning model to implement MRI device image denoising. Deep learning models can implementAttorney Docket No. 59984-715601an External Dynamic Interference Estimation and Removal (EDITER) method to reduce EMI-derived noise from a low field MRI device image. The EDITER method comprises a measurement domain and implements a non-local mean noise reduction on an image domain. Due to substantial EMI noise in a primary coil, which correlates with noise in a plurality of coils within a low field MRI device, the EDITER method cannot effectively implement the non-local mean method to provide an EMI-free image.

[0089] In some embodiments, disclosed herein, is a method for denoising an MRI device image from an MRI device comprising the use of the EDITER method and a deep learning framework.

[0090] Also described herein, in certain embodiments, the MRI device can comprise a primary coil and a plurality of external noise coils (Nc). In certain embodiments, the primary coil can generate a magnetic field to enable the MRI device to capture the MRI device image. In certain embodiments, the plurality of external noise coils can function as external detectors to reduce electromagnetic noise surrounding the primary coil. In certain embodiments, the MRI device can be a low field MRI device.

[0091] In some embodiments, the deep learning framework can comprise a self-supervised learning framework. In some embodiments, the self-supervised learning framework can denoise an image from the primary coil to produce an image substantially free of EMI.

[0092] Disclosed herein is an MRI device comprising a primary coil and a plurality of external noise coils (Nc), wherein the plurality of external noise coils may be configured to capture EMI noise surrounding the primary coil. In some cases, a noise signal detected by the plurality of noise coils in a measurement domain can be represented as follows:

[0093] In some cases, measurement data captured by the plurality of noise coils in the measurement domain can be represented as f0G C 'il. In some cases, M xN denote pixel points. In some cases, f0includes both EMI, denoted by e / G C ' / , and an EMI-free k-space signal, denoted by f G CMxN. In some cases, there can exist a linear relationship (equation 1) represented as follows:fo = ef+ f0* (1)

[0094] In some cases, the primary coil can enable the MRI device to capture an image. In some cases, the corresponding image captured by the primary coil can be represented by x0e CXxVand can be reconstructed from the measured data f0by using a reconstruction pipeline to generate a reconstructed image. In some cases, the reconstruction pipeline comprises model based methods or discrete Fourier transform. In some cases, the reconstructed image can be expressed in an image domain, from equation 1, as follows:Attorney Docket No. 59984-715601x0= e + x0(2)

[0095] In some cases, the self-supervised learning model can remove the EMI noise e from the image domain to approximate the EMI eliminated image xa.Image Reconstruction Pipeline

[0096] In some cases, the image reconstruction pipeline is a process that comprises converting a k-space signal to an image. In some cases, converting the k-space signal to the image comprises a measurement domain denoising process, a multi-echo correction process, a zero-padding process, a distortion correction process, and an image domain denoising process. In some cases, the measurement domain denoising process comprises the EDITER. In some cases, the EDITER can estimate the EMI-free data xa. In some cases, the image domain denoising method utilizes a nonlocal mean method.Self-Supervised Deep Learning Model

[0097] In some cases, the self-supervised deep learning model comprises an EMI feature extraction network and an MRI feature extraction network. In some cases, the EMI feature extraction network can utilize the capabilities of the plurality of external noise coils, denoted as {x , i = 1 • • • Nc to generate EMI features ee. In some cases, the EMI features eecan be used for training a loss function £ .

[0098] In some cases, the MRI feature extraction network can generate EMI-free data with an input of the reconstructed image x0. In some cases, xocan comprise noise (e) or an EMI value denoted by e eanc[anEMI-free image signal, denoted by xae C' / x. In some cases, the x0value can provide a linear relationship between the EMI value and the EMI-free image signal value according to:x0= e + XQ = (the EMI value) + (the EMI-free image signal value).

[0099] In some cases, an optimized denoised image can be produced upon the convergence of the loss function. In some cases, the loss function can be represented as:

[0100] In some cases, the EMI-free image can be x0— ee, and the loss function can minimize the discrepancy between the EMI-free image and an estimated learned image xe,.

[0101] Disclosed herein, the self-supervised learning framework as applied to the deep learning model can generate a plurality of optimized, denoised images produced upon the convergence of a loss function. In some cases, the deep learning model disclosed herein can comprise a neural network. In some cases, the neural network can the EMI feature extraction network, as disclosed herein. In some cases, the EMI feature extraction network can process and analyze dataAttomey Docket No. 59984-715601containing a plurality of noise inputs. In some cases, the EMI feature extraction network can engage in a linear process to amalgamate data from a plurality of noise coil channels, as denotedinto a single channel. In some cases, the linear process to amalgamate data from the plurality of noise channels simulates EMI noise in the primary coil, as denoted by ee. In some cases, the EMI feature extraction network can simulate an EMI noise. In some cases, the EMI feature extraction network can be structured around a plurality of convolutional layers. In some cases, the plurality of convolutional layers can comprise at least four layers. In some cases, the plurality of convolutional layers can comprise at most ten layers. In some cases, the plurality of convolutional layers employs kernels. In some cases, the kernels are of size n x n wherein n > 2. In some embodiments, the kernels are of size 3x3, 4x4, 5x5, 6x6, 7x7, 8x8, 9x9, 10x10, or greater. In some cases, the EMI feature extraction network can be structured around five convolutional layers, wherein the convolutional layers comprise 32 kernels of size 3x3. In some cases, the EMI feature extraction network can employ a rectified linear unit (ReLu) activation function. In some cases, the ReLu activation function can perform nonlinear transformation between the plurality of convolutional layers.

[0102] In some cases, the deep learning model as disclosed herein can comprise the MRI feature extraction network. In some cases, the MRI feature extraction network can reconstruct an image or a signal from incomplete or degraded data. In some cases, the MRI feature extraction network can establish a mapping between the primary coil image, the substantially EMI-free image, and xe< . In some cases, the mapping can render xe< to approximate the EMI-free image. In some cases, the MRI feature extraction network can comprise a plurality of convolutional layers. In some cases, the plurality of convolutional layers can comprise kernels. In some cases, the kernels are of size n x n wherein n > 2. In some cases, the plurality of convolutional layers of the MRI feature extraction network can comprise 10 convolution layers, wherein the convolution layers comprise 128 kernels, with a kernel size of 3x3. In some embodiments, the plurality of convolution layers of the MRI feature extraction network comprise from about 1 to about 3 convolution layers, from about 3 to about 5 convolution layers, from about 5 to about 8 convolution layers, from about 8 to about 10 convolution layers, from about 10 to about 13 convolution layers, from about 13 to about 15 convolution layers, from about 15 to about 18 convolution layers, or from about 18 to about 20 convolution layers. In some embodiments, the convolution layers comprise 16 kernels, 32 kernels, 64 kernels, 96 kernels, 128 kernels, 160 kernels, 192 kernels, 224 kernels, or 256 kernels. In some cases, the MRI feature extraction network can employ a ReLu activation function. In some cases, the ReLu activation function employed by the MRI feature extraction network can perform nonlinear transformations between the plurality of convolution layers of the MRI feature extraction network.Attomey Docket No. 59984-715601

[0103] In some cases, the deep learning framework disclosed herein can remain unaffected by contrast variations in the image from the MRI device. In some cases, the deep learning framework can process multi-contrast data into EMI-free images. In some cases, the image from the MRI device can be a series of {x^}, where i denotes an image with one contrast. In some cases, the deep learning framework can address noise in multiple datasets with different contrasts as a single batch.

[0104] In an example of a method 600, illustrated in FIG. 6A, values from a plurality of noise coils 602 can be transformed by a non-uniform fast Fourier transformation (NUFFT) 604 and input into the EMI feature extraction network 606. The EMI feature extraction network 606 can output EMI features eeor e leaned 608. A primary coil value 610 can be input to the MRI feature extraction network 616. The MRI feature extraction network 616 can output Xiearned 614. Xieamed 614 and c / et,raet / 608 and xe, 618 can be input to the Loss Function 620. FIG. 6A illustrates an architecture of the EMI feature extraction network 106 and an architecture of the MRI feature extraction network 116. The combination network 606 comprises a plurality of convolutional layers and a plurality of ReLu activation functions between the plurality of convolutional layers. The MRI feature extraction network 616 comprises a plurality of convolutional layers and a plurality of ReLu activation functions between the plurality of convolutional layers.

[0105] In an example of a method 600, illustrated in FIG. 6A, the values from the plurality of noise coils 602 can be transformed by a non-uniform fast Fourier transformation (NUFFT). In some cases, a plurality of NUFFT-transformed values 604 can be input into the EMI feature extraction network 606 to output ee608.

[0106] In some cases, the primary coil value (f0) 610 can be input to a reconstruction pipeline 612, as illustrated in FIG. 6B. In some cases, the reconstruction pipeline 612 can comprise EDITER. In some cases, the reconstruction pipeline 612 utilizes EDITER to denoise the primary coil value 610, to generate a denoised primary coil value. In some cases, the reconstruction pipeline 612 comprises a multi-echo correction. In some cases, the denoised primary coil value can be corrected by the multi-echo correction to generate a multi-echo corrected primary coil value. In some cases, the multi-echo correction comprises a technique to correct distortions in the primary coil value 610 by acquiring multiple echo times for a plurality of primary coil values 610. In some cases, the multi-echo corrected primary coil value can be zero-padded. In some cases, zero-padding comprises a technique for adding zeros to the end of the multi-echo corrected primary coil value, to generate a zero-padded primary coil value. In some cases, the zero-padded primary coil value can be transformed by a NUFFT. In some cases, the zero-padded primary coil value can increase a number of samples input into the NUFFT and increase the performance of the NUFFT. In some cases, the NUFFT can output a NUFFT-transformed primary coil value. InAttomey Docket No. 59984-715601some cases, the reconstruction pipeline 612 comprises a distortion correction technique to generate a distortion-corrected image. In some cases, the distortion-corrected image can be further denoised using a non-local mean denoising method, to output a reconstructed image x0614

[0107] In some cases, as illustrated by FIG. 6A, the reconstructed image x0614 can be input to the MRI feature extraction network 616 to output xe, 618. In some cases, ee608 can be generated by the EMI feature extraction network 606, and xe, 618 can be generated by the MRI feature extraction network 616 by optimizing a loss function 620 during the training of the deep learning model. In some cases, optimizing the loss function 620 comprises minimizing a difference between the EMI-free image, the simulated EMI, and the reconstructed image.

[0108] FIG. 6C illustrates an architecture of the EMI feature extraction network 606. The EMI feature extraction network 606 can comprise a convolutional neural network. In some cases, the EMI feature extraction network 606 comprises a plurality of convolutional layers. In some cases, the convolutional layers comprise a plurality of ReLu activation functions between the plurality of convolutional layers.

[0109] FIG. 6D an architecture of the MRI feature extraction network 616. In some cases, the MRI feature extraction network 616 comprises a convolutional neural network. In some cases, the MRI feature extraction network comprises a plurality of convolutional layers. In some cases, the MRI feature extraction network 616 comprises a plurality of ReLu activation functions between the plurality of convolutional layers.

[0110] The deep learning model can operate independently of a specific contrast of a plurality of output images. Therefore, the deep learning model can simultaneously process a plurality of contrast images and output a plurality of denoised images from the plurality of contrast images. The deep learning model may not discriminate based on the type of image contrast during a training phase or a testing phase.[oni] To assess the effectiveness of the method disclosed herein to denoise low field MRI images using self-supervised deep learning, the method was tested on MRI images from human brains. The method as disclosed herein assessed three different imaging contrasts comprising Ti weighted, T2 weighted, and T2 fluid-attenuated inversion recovery (T2-FLAIR). The performance of the method disclosed herein to reduce noise is illustrated in FIGs. 2-7. An EMI-contaminated image was an image produced by the primary coil. The EMI-contaminated images were raw scans obtained directly from the primary coil. In FIGs. 2-7, the EMI-contaminated images comprise an intended brain signal and a plurality of environmental noise, without any initial processing to mitigate the plurality of environmental noise. The EMI-contaminated images were processed by the reconstruction pipeline, as disclosed herein, to output a reconstructed image xQ.Attomey Docket No. 59984-715601However, In some cases, the denoising performance of ED ITER and the non-local mean method may not provide an ideal EMI-free image due to a lack of adequate characterization of the EMI signal. Thus, In some cases, the output image of the reconstruction pipeline was further processed using the MRI feature extraction network, as disclosed herein.

[0112] In FIGs. 7-12, the output images x0were processed using training with the loss function as disclosed herein. In FIGs. 7-12, proposed network images were images processed by the deep learning method disclosed herein. Qualitative analysis of these images, comparing different subjects and contrasts, indicates that the method disclosed herein is highly effective. The method disclosed herein not only clears out background noise but also removes unwanted noise patterns within brain signal areas. Noise reduction achieved by the method disclosed herein are consistent across different imaging contrasts, illustrating the methods robustness and versatility.

[0113] In FIG. 7, EMI-contaminated images from Ti weighted imaging contrasts 702, EMI-contaminated images from T2 weighted imaging contrasts 704, and EMI-contaminated images from FLAIR imaging contrasts 706 are presented from a first human brain image. In FIG. 7, output images from Ti weighted imaging contrasts 708, output images from T2 weighted imaging contrasts 710, and output images from FLAIR imaging contrasts 712, are illustrated. In FIG. 7, the proposed network images from TI weighted imaging contrasts 714, the proposed network images from T2 weighted imaging contrasts 716, and the proposed network images from FLAIR imaging contrasts 718, are illustrated.

[0114] In FIG. 8, EMI-contaminated images from Ti weighted imaging contrasts 802, EMI-contaminated images from T2 weighted imaging contrasts 804, and EMI-contaminated images from FLAIR imaging contrasts 806 are presented from a second human brain image. In FIG. 8, output images from Ti weighted imaging contrasts 808, output images from T2 weighted imaging contrasts 810, and output images from FLAIR imaging contrasts 812, are illustrated. In FIG. 8, the proposed network images from TI weighted imaging contrasts 814, the proposed network images from T2 weighted imaging contrasts 816, and the proposed network images from FLAIR imaging contrasts 818, are illustrated.

[0115] In FIG. 9, EMI-contaminated images from Ti weighted imaging contrasts 902, EMI-contaminated images from T2 weighted imaging contrasts 904, and EMI-contaminated images from FLAIR imaging contrasts 906 are presented from a third human brain image. In FIG. 9, output images from Ti weighted imaging contrasts 908, output images from T2 weighted imaging contrasts 910, and output images from FLAIR imaging contrasts 912, are illustrated. In FIG. 9, the proposed network images from TI weighted imaging contrasts 914, the proposed network images from T2 weighted imaging contrasts 916, and the proposed network images from FLAIR imaging contrasts 418, are illustrated.Attomey Docket No. 59984-715601

[0116] Turning to FIG. 10, EMI-contaminated images from Ti weighted imaging contrasts 1002, EMI-contaminated images from T2 weighted imaging contrasts 1004, and EMI-contaminated images from FLAIR imaging contrasts 1006 are presented from a fourth human brain image. Output images from Ti weighted imaging contrasts 1008, output images from T2 weighted imaging contrasts 1010, and output images from FLAIR imaging contrasts 1012, are illustrated. The proposed network images from TI weighted imaging contrasts 1014, the proposed network images from T2 weighted imaging contrasts 1016, and the proposed network images from FLAIR imaging contrasts 1018 are illustrated.

[0117] In FIG. 11, EMI-contaminated images from Ti weighted imaging contrasts 1102, EMI-contaminated images from T2 weighted imaging contrasts 1104, and EMI-contaminated images from FLAIR imaging contrasts 1106 are presented from a fifth human brain image. Output images from Ti weighted imaging contrasts 1108, output images from T2 weighted imaging contrasts 1110, and output images from FLAIR imaging contrasts 1112, are illustrated. The proposed network images from TI weighted imaging contrasts 1114, the proposed network images from T2 weighted imaging contrasts 1116, and the proposed network images from FLAIR imaging contrasts 1118, are illustrated.

[0118] In FIG. 12, EMI-contaminated images from Ti weighted imaging contrasts 1202, EMI-contaminated images from T2 weighted imaging contrasts 1204, and EMI-contaminated images from FLAIR imaging contrasts 1206 are presented from a sixth human brain image. Output images from Ti weighted imaging contrasts 1208, output images from T2 weighted imaging contrasts 1210, and output images from FLAIR imaging contrasts 1212, are illustrated. The proposed network images from TI weighted imaging contrasts 1214, the proposed network images from T2 weighted imaging contrasts 1216, and the proposed network images from FLAIR imaging contrasts 1218, are illustrated.MRI Scanning Systems

[0119] FIG. 13A depicts an MRI scanning system 1300 that includes a dome-shaped housing 1302 configured to receive a patient’s head. The dome-shaped housing 1302 can further include at least one access aperture configured to allow access to the patient’s head to enable a neural intervention. A space within the dome-shaped housing 1302 forms the region of interest for the MRI scanning system 1300. Target tissue in the region of interest is subjected to magnetization fields / pulses, as further described herein, to obtain imaging data representative of the target tissue.

[0120] For example, referring to FIG. 13B, a patient can be positioned such that his / her head is positioned within the region of interest within the dome-shaped housing 1302. The brain can beAttomey Docket No. 59984-715601positioned entirely within the dome-shaped housing 1302. In such instances, to facilitate intracranial interventions (e.g. neurosurgery) in concert with MR imaging, the dome-shaped housing 1302 can include one or more apertures that provide access to the brain. Apertures can be spaced apart around the perimeter of the dome-shaped housing.

[0121] The MRI scanning system 1300 can include an auxiliary cart that houses certain conventional MRI electrical and electronic components, such as a computer, programmable logic controller, power distribution unit, and amplifiers, for example. The MRI scanning system 1300 can also include a magnet cart that holds the dome-shaped housing 1302, gradient coil(s), and / or a transmission coil, as further described herein. Additionally, the magnet cart can be attached to a receive coil in various instances. Referring primarily to FIG. 13A, the dome-shaped housing 1302 can further include RF transmission coils, gradient coils 1304 (depicted on the exterior thereof), and shim magnets 1306 (depicted on the interior thereof). Alternative configurations for the gradient coil(s) 1304 and / or shim magnets 1306 are also contemplated. In various instances, the shim magnets 1306 can be adjustably positioned in a shim tray within the domeshaped housing 1302, which can allow a technician to granularly configure the magnetic flux density of the dome-shaped housing 1302.

[0122] Various structural housings for receiving the patient’s head and enabling neural interventions can be utilized with an MRI scanning system, such as the MRI scanning system 1300. In one aspect, the MRI scanning system 1300 may be outfitted with an alternative housing, such as a dome-shaped housing 1402 (FIG. 14) or a two-part housing 1502 (FIG. 15) configured to form a dome-shape. The dome-shaped housing 1402 defines a plurality of access apertures 1403; the two-part housing 1502 also defines a plurality of access apertures 1503 and further includes an adjustable gap 1505 between the two parts of the housing.

[0123] In various instances, the housings 1402 and 1502 can include a bonding agent 1508, such as an epoxy resin, for example, that holds a plurality of magnetic elements 1510 in fixed positions. The plurality of magnetic elements 1510 can be bonded to a structural housing 1512, such as a plastic substrate, for example. In various aspects, the bonding agent 1508 and structural housing 1512 may be non-conductive or diamagnetic materials. Referring primarily to FIG. 15, the two-part housing 1502 comprises two structural housings 1512. In various aspect, a structural housing for receiving the patient’s head can be formed from more than two sub-parts. The access apertures 1503 in the structural housing 1512 provide a passage directly to the patient’s head and are not obstructed by the structural housing 1512, bonding agent 1508, or magnetic elements 1510. The access apertures 1503 can be positioned in an open space of the housing 1502, for example.Attomey Docket No. 59984-715601

[0124] There are many possible configurations of neural interventional MRI devices that can achieve improved access for surgical intervention. Many configurations build upon two main designs, commonly known as the Halbach cylinder and the Halbach dome.

[0125] In various instances, a dome-shaped housing for an MRI scanning system, such as the system 100, for example, can include a Halbach dome defining a dome shape and configured based on several factors including main magnetic field Bo strength, field size, field homogeneity, device size, device weight, and access to the patient for neural intervention. In various aspects, the Halbach dome comprises an exterior radius and interior radius at the base of the dome. The Halbach dome may comprise an elongated cylindrical portion that extends from the base of the dome. In one aspect, the elongated cylindrical portion comprises the same exterior radius and interior radius as the base of the dome and continues from the base of the dome at a predetermined length, at a constant radius. In another aspect, the elongated cylindrical portion comprises a different exterior radius and interior radius than the base of the dome (see e.g. FIGS.14 and 15). In such instances, the different exterior radius and interior radius of the elongated cylindrical portion can merge with the base radii in a transitional region.

[0126] FIG. 16 illustrates an exemplary Halbach dome 1600 for an MRI scanning system, such as the system 1300, for example, which defines an access aperture in the form of a hole or access aperture 1603, where the dome 1600 is configured to receive a head and brain B of the patient P within the region of interest therein, and the access aperture 1603 is configured to allow access to the patient P to enable neural intervention with a medical instrument and / or robotically-controlled surgical tool, in accordance with at least one aspect of the present disclosure. The Halbach dome 1600 can be built with a single access aperture 1603 at the top side 1618 of the dome 1600, which allows for access to the top of the skull while minimizing the impact to the magnetic field. Additionally or alternatively, the dome 1500 can be configured with multiple access apertures around the structure 1616 of the dome 1600, as shown in FIGS. 14 and 15.

[0127] The diameter Dhole of the access aperture 1603 may be small (e.g. about 2.54 cm) or very large (substantially the exterior rextdiameter of the dome 1600). As the access aperture 1603 becomes larger, the dome 1600 begins to resemble a Halbach cylinder, for example. The access aperture 1603 is not limited to being at the apex of the dome 1600. The access aperture 1603 can be placed anywhere on the surface or structure 1616 of the dome 1600. In various instances, the entire dome 1600 can be rotated so that the access aperture 1603 can be co-located with a desired physical location on the patient P.

[0128] FIG. 17 depicts relative dimensions of the Halbach dome 1600, including a diameter Dhoie of the access aperture 1603, a length L of the dome 1600, and an exterior radius rextand an interior radius nnof the dome 1600. The Halbach dome 1600 comprises a plurality of magneticAttorney Docket No. 59984-715601elements that are configured in a Halbach array and make up a magnetic assembly. The plurality of magnetic elements may be enclosed by the exterior radius rextand interior radius nnin the structure 1616 or housing thereof. In one aspect, example dimensions may be defined as: nn= 19.3 cm; rext = 23.6 cm; L = 38.7 cm; and 2.54 cm < D < 19.3 cm. In some embodiments, the rH1is from about 13 cm to about 14 cm, about 13 cm to about 15 cm, about 13 cm to about 16 cm, about 13 cm to about 17 cm, about 13 cm to about 18 cm, about 13 cm to about 19 cm, about 13 cm to about 20 cm, about 13 cm to about 21 cm, about 13 cm to about 22 cm, about 13 cm to about 23 cm, about 13 cm to about 24 cm, about 14 cm to about 15 cm, about 14 cm to about 16 cm, about 14 cm to about 17 cm, about 14 cm to about 18 cm, about 14 cm to about 19 cm, about 14 cm to about 20 cm, about 14 cm to about 21 cm, about 14 cm to about 22 cm, about 14 cm to about 23 cm, about 14 cm to about 24 cm, about 15 cm to about 16 cm, about 15 cm to about 17 cm, about 15 cm to about 18 cm, about 15 cm to about 19 cm, about 15 cm to about 20 cm, about 15 cm to about 21 cm, about 15 cm to about 22 cm, about 15 cm to about 23 cm, about 15 cm to about 24 cm, about 16 cm to about 17 cm, about 16 cm to about 18 cm, about 16 cm to about 19 cm, about 16 cm to about 20 cm, about 16 cm to about 21 cm, about 16 cm to about 22 cm, about 16 cm to about 23 cm, about 16 cm to about 24 cm, about 17 cm to about 18 cm, about 17 cm to about 19 cm, about 17 cm to about 20 cm, about 17 cm to about 21 cm, about 17 cm to about 22 cm, about 17 cm to about 23 cm, about 17 cm to about 24 cm, about 18 cm to about 19 cm, about 18 cm to about 20 cm, about 18 cm to about 21 cm, about 18 cm to about 22 cm, about 18 cm to about 23 cm, about 18 cm to about 24 cm, about 19 cm to about 20 cm, about 19 cm to about 21 cm, about 19 cm to about 22 cm, about 19 cm to about 23 cm, about 19 cm to about 24 cm, about 20 cm to about 21 cm, about 20 cm to about 22 cm, about 20 cm to about 23 cm, about 20 cm to about 24 cm, about 21 cm to about 22 cm, about 21 cm to about 23 cm, about 21 cm to about 24 cm, about 22 cm to about 23 cm, about 22 cm to about 24 cm, or about 23 cm to about 24 cm. In some embodiments, the nnis from about 13 cm, about 14 cm, about 15 cm, about 16 cm, about 17 cm, about 18 cm, about 19 cm, about 20 cm, about 21 cm, about 22 cm, about 23 cm, or about 24 cm. In some embodiments, the nnis from at least about 13 cm, about 14 cm, about 15 cm, about 16 cm, about 17 cm, about 18 cm, about 19 cm, about 20 cm, about 21 cm, about 22 cm, or about 23 cm. In some embodiments, the rin is from at most about 14 cm, about 15 cm, about 16 cm, about 17 cm, about 18 cm, about 19 cm, about 20 cm, about 21 cm, about 22 cm, about 23 cm, or about 24 cm. In some embodiments, the rextis from about 19 cm to about 30 cm. In some embodiments, the rextis from about 19 cm to about 20 cm, about 19 cm to about 21 cm, about 19 cm to about 22 cm, about 19 cm to about 23 cm, about 19 cm to about 24 cm, about 19 cm to about 25 cm, about 19 cm to about 26 cm, about 19 cm to about 27 cm, about 19 cm to about 28 cm, about 19 cm to about 29 cm, about 19 cm to about 30 cm, about 20 cm to about 21 cm, aboutAttorney Docket No. 59984-71560120 cm to about 22 cm, about 20 cm to about 23 cm, about 20 cm to about 24 cm, about 20 cm to about 25 cm, about 20 cm to about 26 cm, about 20 cm to about 27 cm, about 20 cm to about 28 cm, about 20 cm to about 29 cm, about 20 cm to about 30 cm, about 21 cm to about 22 cm, about 21 cm to about 23 cm, about 21 cm to about 24 cm, about 21 cm to about 25 cm, about 21 cm to about 26 cm, about 21 cm to about 27 cm, about 21 cm to about 28 cm, about 21 cm to about 29 cm, about 21 cm to about 30 cm, about 22 cm to about 23 cm, about 22 cm to about 24 cm, about 22 cm to about 25 cm, about 22 cm to about 26 cm, about 22 cm to about 27 cm, about 22 cm to about 28 cm, about 22 cm to about 29 cm, about 22 cm to about 30 cm, about 23 cm to about 24 cm, about 23 cm to about 25 cm, about 23 cm to about 26 cm, about 23 cm to about 27 cm, about 23 cm to about 28 cm, about 23 cm to about 29 cm, about 23 cm to about 30 cm, about 24 cm to about 25 cm, about 24 cm to about 26 cm, about 24 cm to about 27 cm, about 24 cm to about 28 cm, about 24 cm to about 29 cm, about 24 cm to about 30 cm, about 25 cm to about 26 cm, about 25 cm to about 27 cm, about 25 cm to about 28 cm, about 25 cm to about 29 cm, about 25 cm to about 30 cm, about 26 cm to about 27 cm, about 26 cm to about 28 cm, about 26 cm to about 29 cm, about 26 cm to about 30 cm, about 27 cm to about 28 cm, about 27 cm to about 29 cm, about 27 cm to about 30 cm, about 28 cm to about 29 cm, about 28 cm to about 30 cm, or about 29 cm to about 30 cm. In some embodiments, the rextis from about 19 cm, about 20 cm, about 21 cm, about 22 cm, about 23 cm, about 24 cm, about 25 cm, about 26 cm, about 27 cm, about 28 cm, about 29 cm, or about 30 cm. In some embodiments, the rextis from at least about 19 cm, about 20 cm, about 21 cm, about 22 cm, about 23 cm, about 24 cm, about 25 cm, about 26 cm, about 27 cm, about 28 cm, or about 29 cm. In some embodiments, the rext is from at most about 20 cm, about 21 cm, about 22 cm, about 23 cm, about 24 cm, about 25 cm, about 26 cm, about 27 cm, about 28 cm, about 29 cm, or about 30 cm. In some embodiments, the L is from about 30 cm to about 52 cm. In some embodiments, the L is from about 30 cm to about 32 cm, about 30 cm to about 34 cm, about 30 cm to about 36 cm, about 30 cm to about 38 cm, about 30 cm to about 40 cm, about 30 cm to about 42 cm, about 30 cm to about 44 cm, about 30 cm to about 46 cm, about 30 cm to about 48 cm, about 30 cm to about 50 cm, about 30 cm to about 52 cm, about 32 cm to about 34 cm, about 32 cm to about 36 cm, about 32 cm to about 38 cm, about 32 cm to about 40 cm, about 32 cm to about 42 cm, about 32 cm to about 44 cm, about 32 cm to about 46 cm, about 32 cm to about 48 cm, about 32 cm to about 50 cm, about 32 cm to about 52 cm, about 34 cm to about 36 cm, about 34 cm to about 38 cm, about 34 cm to about 40 cm, about 34 cm to about 42 cm, about 34 cm to about 44 cm, about 34 cm to about 46 cm, about 34 cm to about 48 cm, about 34 cm to about 50 cm, about 34 cm to about 52 cm, about 36 cm to about 38 cm, about 36 cm to about 40 cm, about 36 cm to about 42 cm, about 36 cm to about 44 cm, about 36 cm to about 46 cm, about 36 cm to about 48 cm, about 36 cm to about 50 cm, about 36 cm to about 52 cm, aboutAttorney Docket No. 59984-71560138 cm to about 40 cm, about 38 cm to about 42 cm, about 38 cm to about 44 cm, about 38 cm to about 46 cm, about 38 cm to about 48 cm, about 38 cm to about 50 cm, about 38 cm to about 52 cm, about 40 cm to about 42 cm, about 40 cm to about 44 cm, about 40 cm to about 46 cm, about 40 cm to about 48 cm, about 40 cm to about 50 cm, about 40 cm to about 52 cm, about 42 cm to about 44 cm, about 42 cm to about 46 cm, about 42 cm to about 48 cm, about 42 cm to about 50 cm, about 42 cm to about 52 cm, about 44 cm to about 46 cm, about 44 cm to about 48 cm, about 44 cm to about 50 cm, about 44 cm to about 52 cm, about 46 cm to about 48 cm, about 46 cm to about 50 cm, about 46 cm to about 52 cm, about 48 cm to about 50 cm, about 48 cm to about 52 cm, or about 50 cm to about 52 cm. In some embodiments, the L is from about 30 cm, about 32 cm, about 34 cm, about 36 cm, about 38 cm, about 40 cm, about 42 cm, about 44 cm, about 46 cm, about 48 cm, about 50 cm, or about 52 cm. In some embodiments, the L is from at least about 30 cm, about 32 cm, about 34 cm, about 36 cm, about 38 cm, about 40 cm, about 42 cm, about 44 cm, about 46 cm, about 48 cm, or about 50 cm. In some embodiments, the L is from at most about 32 cm, about 34 cm, about 36 cm, about 38 cm, about 40 cm, about 42 cm, about 44 cm, about 46 cm, about 48 cm, about 50 cm, or about 52 cm. In some embodiments, the D is from about 2 cm to about 24 cm. In some embodiments, the D is from about 2 cm to about 4 cm, about 2 cm to about 6 cm, about 2 cm to about 8 cm, about 2 cm to about 10 cm, about 2 cm to about 12 cm, about 2 cm to about 14 cm, about 2 cm to about 16 cm, about 2 cm to about 18 cm, about 2 cm to about 20 cm, about 2 cm to about 22 cm, about 2 cm to about 24 cm, about 4 cm to about 6 cm, about 4 cm to about 8 cm, about 4 cm to about 10 cm, about 4 cm to about 12 cm, about 4 cm to about 14 cm, about 4 cm to about 16 cm, about 4 cm to about 18 cm, about 4 cm to about 20 cm, about 4 cm to about 22 cm, about 4 cm to about 24 cm, about 6 cm to about 8 cm, about 6 cm to about 10 cm, about 6 cm to about 12 cm, about 6 cm to about 14 cm, about 6 cm to about 16 cm, about 6 cm to about 18 cm, about 6 cm to about 20 cm, about 6 cm to about 22 cm, about 6 cm to about 24 cm, about 8 cm to about 10 cm, about 8 cm to about 12 cm, about 8 cm to about 14 cm, about 8 cm to about 16 cm, about 8 cm to about 18 cm, about 8 cm to about 20 cm, about 8 cm to about 22 cm, about 8 cm to about 24 cm, about 10 cm to about 12 cm, about 10 cm to about 14 cm, about 10 cm to about 16 cm, about 10 cm to about 18 cm, about 10 cm to about 20 cm, about 10 cm to about 22 cm, about 10 cm to about 24 cm, about 12 cm to about 14 cm, about 12 cm to about 16 cm, about 12 cm to about 18 cm, about 12 cm to about 20 cm, about 12 cm to about 22 cm, about 12 cm to about 24 cm, about 14 cm to about 16 cm, about 14 cm to about 18 cm, about 14 cm to about 20 cm, about 14 cm to about 22 cm, about 14 cm to about 24 cm, about 16 cm to about 18 cm, about 16 cm to about 20 cm, about 16 cm to about 22 cm, about 16 cm to about 24 cm, about 18 cm to about 20 cm, about 18 cm to about 22 cm, about 18 cm to about 24 cm, about 20 cm to about 22 cm, about 20 cm to about 24 cm, or about 22 cm to aboutAttomey Docket No. 59984-71560124 cm. In some embodiments, the D is from about 2 cm, about 4 cm, about 6 cm, about 8 cm, about 10 cm, about 12 cm, about 14 cm, about 16 cm, about 18 cm, about 20 cm, about 22 cm, or about 24 cm. In some embodiments, the D is from at least about 2 cm, about 4 cm, about 6 cm, about 8 cm, about 10 cm, about 12 cm, about 14 cm, about 16 cm, about 18 cm, about 20 cm, or about 22 cm. In some embodiments, the D is from at most about 4 cm, about 6 cm, about 8 cm, about 10 cm, about 12 cm, about 14 cm, about 16 cm, about 18 cm, about 20 cm, about 22 cm, or about 24 cm.

[0129] Based on the above example dimensions, a Halbach dome 1600 with an access aperture 1603 may be configured with a magnetic flux density Bo of around 72 ml, and an overall mass of around 35 kg. It will be appreciated that the dimensions may be selected based on particular applications to achieve a desired magnetic flux density Bo, total weight of the Halbach dome 1600 and / or magnet cart, and geometry of the neural intervention access aperture 1603.

[0130] In various aspects, the Halbach dome 1600 may be configured to define multiple access apertures 1603 placed around the structure 1616 of the dome 1600. These multiple access apertures 1603 may be configured to allow for access to the patient’s head and brain B using tools (e.g., surgical tools) and / or a surgical robot.

[0131] In various aspects, the access aperture 1603 may be adjustable. The adjustable configuration may provide the ability for the access aperture 1603 to be adjusted using either a motor, mechanical assist, or a hand powered system with a mechanical iris configuration, for example, to adjust the diameter Dhole of the access aperture 1603. This would allow for configuration of the dome without an access aperture 1603, conducting an imaging scan, and then adjusting the configuration of the dome 1600 and mechanical iris thereof to include the access aperture 1603 and, thus, to enable a surgical intervention therethrough.

[0132] Halbach domes and magnetic arrays thereof for facilitating neural interventions are further described in International Patent Application No. PCT / US2022 / 72143, titled NEURAL INTERVENTIONAL MAGNETIC RESONANCE IMAGING APPARATUS, filed May 5, 2022, which is incorporated by reference herein in its entirety.Computer Systems

[0133] The present disclosure provides computer systems that are programmed to implement methods of the disclosure. FIG. 5 shows a computer system 501 that is programmed or otherwise configured to train and implement the deep learning model. The computer system 501 can regulate various aspects of training and implementing a deep learning model of the present disclosure, such as, for example, implementing a CNN. The computer system 501 can regulate various aspects of implementing a deep learning model of the present disclosure while adjusting at least one of the magnetic blocks in the plurality of magnetic blocks in a Halbach array, such as,Attomey Docket No. 59984-715601for example, calculating Lhomogeneity, Lconstraint, LossFunction, PPM, and performing magnetic field analysis. The computer system 501 can be an electronic device of a user or a computer system that is remotely located with respect to the electronic device. The electronic device can be a mobile electronic device.

[0134] The computer system 501 includes a central processing unit (CPU, also “processor” and “computer processor” herein) 505, which can be a single core or multi core processor, or a plurality of processors for parallel processing. The computer system 501 also includes memory or memory location 510 (e.g., random-access memory, read-only memory, flash memory), electronic storage unit 515 (e.g., hard disk), communication interface 520 (e.g., network adapter) for communicating with one or more other systems, and peripheral devices 525, such as cache, other memory, data storage and / or electronic display adapters. The memory 510, storage unit 515, interface 520 and peripheral devices 525 are in communication with the CPU 505 through a communication bus (solid lines), such as a motherboard. The storage unit 515 can be a data storage unit (or data repository) for storing data. The computer system 501 can be operatively coupled to a computer network (“network”) 530 with the aid of the communication interface 520. The network 530 can be the Internet, an internet and / or extranet, or an intranet and / or extranet that is in communication with the Internet. The network 530 in some cases is a telecommunication and / or data network. The network 530 can include one or more computer servers, which can enable distributed computing, such as cloud computing. The network 530, in some cases with the aid of the computer system 501, can implement a peer-to-peer network, which may enable devices coupled to the computer system 501 to behave as a client or a server.

[0135] The CPU 505 can execute a sequence of machine-readable instructions, which can be embodied in a program or software. The instructions may be stored in a memory location, such as the memory 510. The instructions can be directed to the CPU 505, which can subsequently program or otherwise configure the CPU 505 to implement methods of the present disclosure. Examples of operations performed by the CPU 505 can include fetch, decode, execute, and writeback.

[0136] The CPU 505 can be part of a circuit, such as an integrated circuit. One or more other components of the system 501 can be included in the circuit. In some cases, the circuit is an application specific integrated circuit (ASIC).

[0137] The storage unit 515 can store files, such as drivers, libraries and saved programs. The storage unit 515 can store user data, e.g., user preferences and user programs. The computer system 501 in some cases can include one or more additional data storage units that are external to the computer system 501, such as located on a remote server that is in communication with the computer system 501 through an intranet or the Internet.Attomey Docket No. 59984-715601

[0138] The computer system 501 can communicate with one or more remote computer systems through the network 530. For instance, the computer system 501 can communicate with a remote computer system of a user. Examples of remote computer systems include personal computers (e.g., portable PC), slate or tablet PC’s (e.g., Apple® iPad, Samsung® Galaxy Tab), telephones, Smart phones (e.g., Apple® iPhone, Android-enabled device, Blackberry®), or personal digital assistants. The user can access the computer system 501 via the network 530.

[0139] Methods as described herein can be implemented by way of machine (e.g., computer processor) executable code stored on an electronic storage location of the computer system 501, such as, for example, on the memory 510 or electronic storage unit 515. The machine executable or machine readable code can be provided in the form of software. During use, the code can be executed by the processor 505. In some cases, the code can be retrieved from the storage unit 515 and stored on the memory 510 for ready access by the processor 505. In some situations, the electronic storage unit 515 can be precluded, and machine-executable instructions are stored on memory 510.

[0140] The code can be pre-compiled and configured for use with a machine having a processer adapted to execute the code, or can be compiled during runtime. The code can be supplied in a programming language that can be selected to enable the code to execute in a pre-compiled or as-compiled fashion.

[0141] Aspects of the systems and methods provided herein, such as the computer system 501, can be embodied in programming. Various aspects of the technology may be thought of as “products” or “articles of manufacture” potentially in the form of machine (or processor) executable code and / or associated data that is carried on or embodied in a type of machine readable medium. Machine-executable code can be stored on an electronic storage unit, such as memory (e.g., read-only memory, random-access memory, flash memory) or a hard disk.“Storage” type media can include any or all of the tangible memory of the computers, processors or the like, or associated modules thereof, such as various semiconductor memories, tape drives, disk drives and the like, which may provide non-transitory storage at any time for the software programming. All or portions of the software may at times be communicated through the Internet or various other telecommunication networks. Such communications, for example, may enable loading of the software from one computer or processor into another, for example, from a management server or host computer into the computer platform of an application server. Thus, another type of media that may bear the software elements includes optical, electrical, and electromagnetic waves, such as used across physical interfaces between local devices, through wired and optical landline networks and over various air-links. The physical elements that carry such waves, such as wired or wireless links, optical links or the like, also may be considered asAttomey Docket No. 59984-715601media bearing the software. As used herein, unless restricted to non-transitory, tangible “storage” media, terms such as computer or machine “readable medium” refer to any medium that participates in providing instructions to a processor for execution.

[0142] Hence, a machine-readable medium, such as computer-executable code, may take many forms, including but not limited to, a tangible storage medium, a carrier wave medium or physical transmission medium. Non-volatile storage media include, for example, optical or magnetic disks, such as any of the storage devices in any computer(s) or the like, such as may be used to implement the databases, etc. shown in the drawings. Volatile storage media include dynamic memory, such as main memory of such a computer platform. Tangible transmission media include coaxial cables; copper wire and fiber optics, including the wires that comprise a bus within a computer system. Carrier-wave transmission media may take the form of electric or electromagnetic signals, or acoustic or light waves such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media therefore include for example: a floppy disk, a flexible disk, hard disk, magnetic tape, any other magnetic medium, a CD-ROM, DVD or DVD-ROM, any other optical medium, punch cards paper tape, any other physical storage medium with patterns of holes, a RAM, a ROM, a PROM and EPROM, a FLASH-EPROM, any other memory chip or cartridge, a carrier wave transporting data or instructions, cables or links transporting such a carrier wave, or any other medium from which a computer may read programming code and / or data. Many of these forms of computer readable media may be involved in carrying one or more sequences of one or more instructions to a processor for execution.

[0143] The computer system 501 can include or be in communication with an electronic display 535 that comprises a user interface (UI) 540 for providing, for example, visualization of the ROI, visualization of the magnetic field analysis, and visualization of the magnetic field ppm analysis. Examples of UFs include, without limitation, a graphical user interface (GUI) and web-based user interface.

[0144] Methods and systems of the present disclosure can be implemented by way of one or more algorithms. An algorithm can be implemented by way of software upon execution by the central processing unit 505. The algorithm can, for example, comprise a deep learning model, such as a CNN.Definitions

[0145] Throughout this application, various embodiments may be presented in a range format. It should be understood that the description in range format is merely for convenience and brevity and should not be construed as an inflexible limitation on the scope of the disclosure.Attomey Docket No. 59984-715601Accordingly, the description of a range should be considered to have specifically disclosed all the possible subranges as well as individual numerical values within that range. For example, description of a range such as from 1 to 6 should be considered to have specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6 etc., as well as individual numbers within that range, for example, 1, 2, 3, 4, 5, and 6. This applies regardless of the breadth of the range.

[0146] The ranges disclosed herein also encompass any and all overlap, sub-ranges, and combinations thereof. Language such as “up to,” “at least,” “greater than,” “less than,” “between,” and the like includes the number recited. Numbers preceded by a term such as “approximately”, “about”, and “substantially” as used herein include the recited numbers, and also represent an amount close to the stated amount that still performs a desired function or achieves a desired result. The term “about” or “approximately” may mean within an acceptable error range for the particular value, which will depend in part on how the value is measured or determined, e.g., the limitations of the measurement system. For example, the terms “approximately”, “about”, and “substantially” may refer to an amount that is within less than 10% of, within less than 5% of, within less than 1% of, within less than 0.1% of, and within less than 0.01% of the stated amount. For example, “about” may mean within 1 or more than 1 standard deviation, per the practice in the art. Alternatively, “about” may mean a range of up to 20%, up to 10%, up to 5%, or up to 1% of a given value. As used herein, the term “about” a number refers to that number plus or minus 10% of that number. The term “about” a range refers to that range minus 10% of its lowest value and plus 10% of its greatest value. Where particular values are described in the application and claims, unless otherwise stated the term “about” meaning within an acceptable error range for the particular value may be assumed.

[0147] As used in the specification and claims, the singular forms “a”, “an” and “the” include plural references unless the context clearly dictates otherwise. For example, the term “a sample” includes a plurality of samples, including mixtures thereof.

[0148] Whenever the term “at least,” “greater than,” or “greater than or equal to” precedes the first numerical value in a series of two or more numerical values, the term “at least,” “greater than” or “greater than or equal to” applies to each of the numerical values in that series of numerical values. For example, greater than or equal to 1, 2, or 3 is equivalent to greater than or equal to 1, greater than or equal to 2, or greater than or equal to 3.

[0149] Whenever the term “no more than,” “less than,” or “less than or equal to” precedes the first numerical value in a series of two or more numerical values, the term “no more than,” “less than,” or “less than or equal to” applies to each of the numerical values in that series of numericalAttomey Docket No. 59984-715601values. For example, less than or equal to 3, 2, or 1 is equivalent to less than or equal to 3, less than or equal to 2, or less than or equal to 1.

[0150] As used herein, the term "substantially" in reference to a given parameter, property, or condition means and includes to a degree that one of ordinary skill in the art would understand that the given parameter, property, or condition is met with a degree of variance, such as within acceptable manufacturing tolerances. By way of example, depending on the particular parameter, property, or condition that is substantially met, the parameter, property, or condition may be at least 90.0% met, at least 95.0% met, at least 99.0% met, or even at least 99.9% met.

[0151] The terms “determining,” “measuring,” “evaluating,” “assessing,” “assaying,” and “analyzing” are often used interchangeably herein to refer to forms of measurement. The terms include determining if an element is present or not (for example, detection). These terms can include quantitative, qualitative or quantitative and qualitative determinations. Assessing can be relative or absolute. “Detecting the presence of’ can include determining the amount of something present in addition to determining whether it is present or absent depending on the context.

[0152] The section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described

[0153] While preferred embodiments of the present disclosure have been shown and described herein, it will be obvious to those skilled in the art that such embodiments are provided by way of example only. It is not intended that the disclosure be limited by the specific examples provided within the specification. While the disclosure has been described with reference to the aforementioned specification, the descriptions and illustrations of the embodiments herein are not meant to be construed in a limiting sense. Numerous variations, changes, and substitutions will now occur to those skilled in the art without departing from the disclosure. Furthermore, it shall be understood that all aspects of the disclosure are not limited to the specific depictions, configurations or relative proportions set forth herein which depend upon a variety of conditions and variables. It should be understood that various alternatives to the embodiments of the disclosure described herein may be employed in practicing the disclosure. It is therefore contemplated that the disclosure shall also cover any such alternatives, modifications, variations, or equivalents. It is intended that the following claims define the scope of the disclosure and that methods and structures within the scope of these claims and their equivalents be covered thereby.

Claims

1. Attorney Docket No. 59984-715601CLAIMS WHAT IS CLAIMED IS:

1. A method for optimizing a magnet array, the method comprising:inputting an initialized magnetization strength for a plurality of magnets within a magnet array into a neural network;defining a loss function, at the neural network, wherein the loss function measures a homogeneity of a magnetic field, based at least in part on the initialized magnetic strength;training the neural network by minimizing the loss function to generate a trained neural network, wherein minimizing the loss function is based at least in part on adjusting parameters of the neural network and adjusting a plurality of parameters of the magnet array during the training; andusing the trained neural network to output a learned magnetization strength, based at least in part on minimizing the loss function.

2. The method of claim 1, wherein the initialized magnetization strength is determined by an initial size, position, orientation, or any combination thereof, of the plurality of magnets.

3. The method of claim 2, wherein the initialized magnetization strength is at least about 70 mT.

4. The method of any one of claims 1 to 3, wherein the plurality of parameters of the magnet array comprises at least one of a size, position, or orientation of the plurality of magnets.

5. The method of any one of claims 1 to 4, wherein the neural network is a convolutional neural network (CNN).

6. The method of claim 5, wherein the CNN contains a plurality of convolutional layers.

7. The method of claim 6, wherein the plurality of convolutional layers comprise a rectified linear unit (ReLU) between the plurality of convolutional layers.

8. The method of claims 6 or 7, wherein the plurality of convolutional layers comprise at least four convolutional layers.Attomey Docket No. 59984-7156019. The method of claims 6 or 7, wherein the plurality of convolutional layers comprise about eight convolutional layers.

10. The method of any one of claims 6 to 9, wherein the plurality of convolutional layers comprise kernels having a size n x n with n > 2.

11. The method of claim 10, wherein n is 3.

12. The method of any one of claims 1 to 11, wherein the initialized magnetization strength is from a predetermined region of interest (RO I) of the magnet array.

13. The method of claim 12, wherein the loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in parts per million (ppm).

14. The method of claim 12, wherein the loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in magnetization magnitude (M).

15. The method of any one of claims 1 to 14, wherein training the neural network further comprises using an Adaptive Moment Estimation (ADAM) optimization technique for gradient descent.

16. The method of any one of claims 1 to 15, wherein the magnet array is a Halbach array.

17. The method of claim 16, wherein the Halbach array is configured in a cylindrical Halbach array.

18. A system comprising:a magnet array comprising a plurality of magnets;a processor; anda memory comprising instructions for the processor to:i) input an initialized magnetization strength for the plurality of magnets into a neural network;ii) define a loss function, at the neural network, based at least in part on the initialized magnetic strength, wherein the loss function measures a homogeneity of a magnetic field;Attorney Docket No. 59984-715601iii) train the neural network by minimizing the loss function to generate a trained neural network, wherein minimizing the loss function is based at least in part on adjusting parameters of the neural network and adjusting a plurality of parameters of the magnet array during the training; andiv) use the neural network to output a learned magnetization strength, based at least in system on minimizing the loss function.

19. The system of claim 18, wherein the processor is further configured to determine the initialized magnetization strength by an initial size, position, orientation or any combination thereof, of the plurality of magnets.

20. The system of claim 19, wherein the initialized magnetization strength is at least about 70 mT.

21. The system of any one of claims 18 to 20, wherein the plurality of parameters of the magnet array comprises at least one of a size, position, or orientation of the plurality of magnets.

22. The system of any one of claims 18 to 21, wherein the neural network is a convolutional neural network (CNN).

23. The system of claim 22, wherein the CNN contains a plurality of convolutional layers.

24. The system of claim 23, wherein the plurality of convolutional layers comprise a rectified linear unit (ReLU) between the plurality of convolutional layers.

25. The system of claims 23 or 24, wherein the plurality of convolutional layers comprise at least four convolutional layers.

26. The system of claims 23 or 24, wherein the plurality of convolutional layers comprise about eight convolutional layers.

27. The system of any one of claims 23 to 26, wherein the plurality of convolutional layers comprise kernels having a size n x n with n > 2.

28. The system of claim 27, wherein n is 3.Attomey Docket No. 59984-71560129. The system of any one of claims 18 to 28, wherein the initialized magnetization strength is from a predetermined region of interest (RO I) of the magnet array.

30. The system of claim 29, wherein the loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in parts per million (ppm).

31. The system of claim 29, wherein the loss function calculates a homogeneity of a magnetic field across the predetermined ROI, measured in magnetization magnitude (M).

32. The system of any one of claims 18 to 31, wherein the processor is configured to train the neural network using an Adaptive Moment Estimation (ADAM) optimization technique for gradient descent.

33. The system of any one of claims 18 to 32, wherein the magnet array is a Halbach array.

34. The system of claim 33, wherein the Halbach array is configured in a cylindrical Halbach array.

35. A method for denoising a magnetic resonance imaging (MRI) device image, the method comprising:generating a magnetic resonance image from a primary coil of an MRI device; generating a reconstructed image from the magnetic resonance image using a reconstruction pipeline;generating a plurality of values from a plurality of external noise coils of the MRI device; generating, using an MRI feature extraction network, an image substantially free of electromagnetic interference (EMI), based at least in part on the magnetic resonance image; training a deep learning model to combine the plurality of values into a simulated EMI using an EMI feature extraction network;optimizing a loss function during the training of the deep learning model using the image substantially free of EMI, the simulated EMI, and the reconstructed image, wherein the loss function minimizes a difference between the image substantially free of EMI, the simulated EMI, and the reconstructed image; andproducing a denoised image using the deep learning model, based at least in part on the minimized loss function.

36. The method of claim 35, wherein the MRI device is a low field MRI device.Attomey Docket No. 59984-71560137. The method of claim 35 or 36, wherein the MRI feature extraction network comprises a plurality of convolutional layers.

38. The method of claim 37, wherein the plurality of convolutional layers comprise kernels of size n x n, with n >2.

39. The method of claim 38, wherein n is 3.

40. The method of any one of claims 35 to 39, wherein the EMI feature extraction network comprises a plurality of convolutional layers.

41. The method of claim 40, wherein the plurality of convolutional layers comprise kernels of size n x n, with n >2.

42. The method of claim 41, wherein n is 3.

43. The method of claim 40, wherein the EMI feature extraction network includes a rectified linear unit (ReLu) activation function configured to perform nonlinear transformation between the plurality of convolutional layers.

44. The method of any one of claims 35 to 43, wherein the EMI feature extraction network learns to output a simulated EMI of the primary coil, based at least in part on the plurality of noise signals.

45. The method of any one of claims 35 to 44, wherein the loss function minimizes a discrepancy between the substantially EMI-free image and the simulated EMI-free image.

46. The method of any one of claims 37 to 39, wherein the MRI feature extraction network includes a rectified linear unit (ReLu) activation function configured to perform nonlinear transformations between the plurality of convolution layers of the MRI feature extraction network.

47. The method of claim 37, wherein the plurality of convolutional layers comprise at least four layers.

48. The method of claim 37, wherein the plurality of convolutional layers comprise at most 10 layers.Attorney Docket No. 59984-71560149. The method of claim 40, wherein the plurality of convolutional layers comprise at least four layers.

50. The method of claim 40, wherein the plurality of convolutional layers comprise at most 10 layers.

51. A system comprising:a magnetic resonance imaging (MRI) device comprising:a primary coil configured to generate a magnetic resonance image; and a plurality of external noise coils; anda processor;a memory comprising instructions for the processor to:i) generate a reconstructed image from the magnetic resonance image using a reconstruction pipeline;ii) generate a plurality of values from the plurality of external noise coils;iii) generate, using an MRI feature extraction network, an image substantially free of electromagnetic interference (EMI), based at least in part on the magnetic resonance image;iv) train a deep learning model to combine the plurality of values into a stimulated EMI using an EMI feature extraction network;v) optimize a loss function during the training of the deep learning model using the image substantially fee of EMI, the stimulated EMI, and the reconstructed image, wherein the loss function minimizes a difference between the image substantially free of EMI, the stimulated EMI, and the reconstructed image; and vi) produce a denoised image using the deep learning model, based at least in part on the minimized loss function.

52. The system of claim 51, wherein the MRI device is a low field MRI device.

53. The system of claim 51 or 52, wherein the MRI feature extraction network comprises a plurality of convolutional layers.

54. The system of claim 53, wherein the plurality of convolutional layers comprise kernels of size n x n, with n >2.

55. The system of claim 54, wherein n is 3.Attomey Docket No. 59984-71560156. The system of any one of claims 51 to 55, wherein the EMI feature extraction network comprises a plurality of convolutional layers.

57. The system of claim 56, wherein the plurality of convolutional layers comprise kernels of size n x n, with n >2.

58. The system of claim 57, wherein n is 3.

59. The system of any one of claims 51 to 58, wherein the EMI feature extraction network learns to output a simulated EMI of the primary coil, based at least in part on the plurality of noise signals.

60. The system of any one of claims 51 to 59, wherein the loss function minimizes a discrepancy between the substantially EMI-free image and the simulated EMI-free image.

61. The system of claim 56, wherein the EMI feature extraction network includes a rectified linear unit (ReLu) activation function configured to perform nonlinear transformation between the plurality of convolutional layers.

62. The system of claim 53, wherein the MRI feature extraction network includes a rectified linear unit (ReLu) activation function configured to perform nonlinear transformations between the plurality of convolution layers of the MRI feature extraction network.

63. The system of claim 53, wherein the plurality of convolutional layers comprise at least four layers.

64. The system of claim 53, wherein the plurality of convolutional layers comprise at most 10 layers.

65. The system of claim 56, wherein the plurality of convolutional layers comprise at least four layers.

66. The system of claim 56, wherein the plurality of convolutional layers comprise at most 10 layers.

67. A method for optimizing a magnet array, the method comprising:Attomey Docket No. 59984-715601training a neural network by minimizing a loss function measuring a homogeneity of a magnetic field based at least in part on adjusting parameters of the neural network and adjusting a plurality of parameters of the magnet array during the training; andusing the trained neural network to output a learned magnetization strength.

68. A method for denoising a magnetic resonance imaging (MRI) device image, the method comprising:optimizing a loss function during training of a deep learning model using (a) an image substantially free of electromagnetic interference (EMI) generated via an MRI feature extraction network, (b) a simulated EMI, and (c) a reconstructed MRI image, wherein the loss function minimizes a difference between the image substantially free of EMI, the simulated EMI, and the reconstructed image; andproducing a denoised image using the deep learning model.

Images