Distortion correction of low-field magnetic resonance images using machine learning and leveraging paired pre-operative images

Abstract

Disclosed herein is a system including a database storing preoperative high resolution images of an object of interest and a control circuit. The control circuit includes a processor and a memory. The memory stores instructions executable by the processor to obtain a low field strength magnetic resonance image of the object of interest. The memory stores further instructions executable by the processor to input the low field strength MRI of the object of interest to a generator model of a pre-trained generative adversarial network. The generator model is pre-trained with low field strength MRI and paired high resolution images to correct for image distortion. The memory stores further instructions executable by the processor to output a distortion corrected image of the object of interest from the generator model based on the low field strength MRI and to transmit the distortion corrected image of the object of interest to a user interface.

Description

[0001] Cross-referencing This application claims the benefit of U.S. Patent Application No. 18 / 407,758, filed January 9, 2024, the entire contents of which are incorporated herein by reference. Background Technology

[0002] This disclosure relates to magnetic resonance imaging (MRI), medical imaging, medical interventions, and surgical interventions. MRI systems typically comprise large, complex machines that generate significantly high magnetic fields, posing substantial limitations on the feasibility of certain surgical interventions. These limitations may include restricted physical access to the patient by the surgeon and / or surgical robot and / or limitations on the use of certain electrical and mechanical components near the MRI scanner. Such limitations are inherent in the basic design of many existing systems and are difficult to overcome. Summary of the Invention

[0003] In one general aspect, this disclosure describes a system. The system includes a database storing preoperative high-resolution images of an object of interest, and control circuitry including a processor and a memory. The memory stores instructions executable by the processor to acquire low-field magnetic resonance imaging (MRI) images of the object of interest during surgery. The memory stores further instructions executable by the processor to input the low-field MRI of the object of interest into a pre-trained generative adversarial network (GAN) generator model during surgery, wherein the generator model is pre-trained with the low-field MRI and paired high-resolution images to correct image distortion. The memory stores further instructions executable by the processor to output a distortion-corrected image of the object of interest from the generator model during surgery based on the low-field MRI. The memory stores further instructions executable by the processor to transmit the distortion-corrected image of the object of interest to a user interface during surgery.

[0004] In at least one aspect, the paired high-resolution images are selected from high-field MRI and high-resolution computed tomography images.

[0005] In at least one aspect, the generative adversarial network (GAN) is trained to a desired performance level using at least one set of training images, wherein each set of training images includes a high-resolution training image of a training object of interest and a low-field MRI training image of the training object of interest. In at least one aspect, the system further includes training control circuitry to obtain a first set of training images preoperatively, wherein the first set of training images includes a first low-field MRI training image and a first high-resolution training image of a first training object of interest. In at least one aspect, the training control circuitry is further configured to input the first low-field MRI training image into the generator model of the GAN preoperatively to generate a first distortion-corrected training image. In at least one aspect, the training control circuitry is further configured to input the first distortion-corrected training image and the first high-resolution training image into the discriminator model of the GAN preoperatively to evaluate the first training image. In at least one aspect, the training control circuitry is further configured to update one of the generator model and the discriminator model preoperatively based on the evaluation of the first distortion-corrected training image by the discriminator model.

[0006] In at least one aspect, the discriminator model includes a discriminator neural network, and the generator model includes a generator neural network, wherein the training control circuitry is further configured to classify the first distortion-corrected training image from the generator model as “real” based on the discriminator model, and to adjust at least one weight of at least one layer of the discriminator neural network.

[0007] In at least one aspect, the discriminator model includes a discriminator neural network, and the generator model includes a generator neural network, wherein the training control circuitry is further configured to classify the first distortion-corrected training image as a “fake” based on the discriminator model, and to adjust at least one weight of at least one layer of the generator neural network of the generator model.

[0008] In at least one aspect, the low-field MRI includes distortions of the anatomical structures depicted in the low-field MRI.

[0009] In at least one aspect, the paired high-resolution images have a first resolution, wherein the low-field MRI has a second resolution, wherein the second resolution is lower than the first resolution, and wherein the memory stores further instructions executable by the processor to adjust the first resolution of the paired high-resolution images based on the second resolution of the low-field MRI prior to training the generative adversarial network.

[0010] In at least one aspect, adjusting the first resolution based on the second resolution includes smoothing each paired high-resolution image.

[0011] In at least one aspect, the memory stores further instructions executable by the processor to generate low-field MRI intraoperatively using a low-field magnetic field.

[0012] In at least one aspect, the system further includes a dome-shaped housing configured to house a magnet array, wherein the magnet array is arranged to generate the low-field magnetic field toward the object of interest within the field of view, wherein the low-field magnetic field has a magnetic field strength less than or equal to 1 T. In at least one aspect, the system further includes a radio frequency coil assembly configured to selectively excite magnetization of the object of interest within the field of view.

[0013] In at least one aspect, the memory stores further instructions executable by the processor to transmit the distortion-corrected image of the object of interest to the user interface in real time during surgery.

[0014] In at least one aspect, the object of interest includes the anatomical structure of a particular patient.

[0015] In another general aspect, this disclosure describes a training system for a generative adversarial network (GAN). The training system includes a training processor and a training memory storing multiple sets of training images, each set including a high-resolution training image of a training object of interest (POI) and paired low-field MRI training images of the POI. The memory stores instructions executable by the processor to obtain a first set of training images, wherein the first set of training images includes a first low-field MRI training image and a first high-resolution training image of a first POI; input the first low-field MRI training image into a generator model of the GAN to generate a first distortion-corrected training image; and input the first distortion-corrected training image and the first high-resolution training image into a discriminator model of the GAN. The memory stores instructions executable by the processor to evaluate the first distortion-corrected training image by the discriminator model, thereby identifying the first distortion-corrected training image as either "real" or "fake," and to update the GAN preoperatively based on the evaluation of the first distortion-corrected training image by the discriminator model. Updating the generative adversarial network includes updating the discriminator model if the discriminator model classifies the first distortion-corrected training image from the generator model as "real", and updating the generator model if the discriminator model classifies the first distortion-corrected training image from the generator model as "fake".

[0016] In at least one aspect, the discriminator model includes a discriminator neural network, and the generator model includes a generator neural network. In at least one aspect, the training memory stores instructions executable by the training processor to classify the first distortion-corrected training image from the generator model as "real" based on the discriminator model, adjust at least one weight of at least one layer of the discriminator neural network of the discriminator model, and to classify the first distortion-corrected training image as "fake" based on the discriminator model, adjusting at least one weight of at least one layer of the generator neural network of the generator model.

[0017] In at least one aspect, the system further includes training the generator model to a desired performance level by obtaining at least another set of training images of different objects, and further training the generative adversarial network with the at least another set of training images.

[0018] In another general aspect, this disclosure describes a method comprising training a generative adversarial network (GAN) to a desired performance level preoperatively. Training the GAN includes inputting a first set of training images into the GAN. The first set of training images includes a first low-field MRI training image and a first high-resolution training image of a first object of interest, and wherein the GAN includes a generator model and a discriminator model. Training the GAN further includes having the generator model generate a first distortion-corrected training image based on the first low-field MRI training image, having the discriminator model receive the first distortion-corrected training image and the first high-resolution training image, having the discriminator model classify the first distortion-corrected training image as “real” or “fake,” and updating the GAN based on the classification. Updating the GAN includes updating the discriminator model if the first distortion-corrected training image is classified as “real,” and updating the generator model if the first distortion-corrected training image is classified as “fake.” The method also includes sending a notification to a user interface based on the GAN reaching the desired performance level.

[0019] In at least one aspect, the method further includes obtaining a high-field MRI of the first training object of interest preoperatively.

[0020] In at least one aspect, the method further includes obtaining a high-resolution computed tomography image of the first training object of interest preoperatively.

[0021] In at least one aspect, the method includes generating a distortion-corrected image with minimal distortion after training the generative adversarial network to the desired performance level. Generating the distortion-corrected image includes obtaining a low-field MRI of the object of interest intraoperatively using a low-field magnetic resonance imaging (MRI) system. The low-field MRI includes a dome-shaped housing and a magnet array arranged around the dome-shaped housing. Generating the distortion-corrected image also includes inputting the low-field MRI of the object of interest into the generator model intraoperatively, wherein the generator model is used to generate the distortion-corrected image, and transmitting the distortion-corrected image intraoperatively to a user interface, wherein the generation and transmission of the distortion-corrected image are performed in real time.

[0022] In at least one aspect, the method further includes projecting a low-field magnetic field from the magnet array onto the object of interest located within the field of view, wherein the low-field magnetic field has a magnetic field strength of less than or equal to 1 T; sending a sequence of radio frequency pulses to a radio frequency coil assembly configured to selectively excite magnetization of the object of interest within the field of view; and receiving and recording an output signal from the radio frequency coil assembly. Attached Figure Description

[0023] The various aspects of the organization and operation methods described herein, as well as their further purposes and advantages, can be best understood by referring to the following description in conjunction with the accompanying figures.

[0024] Figure 1 An assembly of an MRI scanning system according to at least one aspect of the present disclosure is depicted, the assembly including a dome-shaped housing of a magnetic array surrounding a region of interest therein, and further depicting the dome-shaped housing positioned for receiving at least a portion of the head of a patient lying on a table into the region of interest.

[0025] Figure 2 It is for use with at least one aspect of this disclosure Figure 1 A perspective view of an alternative dome-shaped housing for a magnetic array used in MRI scanning systems, wherein an access hole is defined in the dome-shaped housing.

[0026] Figure 3 It is for use with at least one aspect of this disclosure Figure 1 A perspective view of an alternative dome-shaped housing for a magnetic array used in MRI scanning systems, wherein access holes and adjustable gaps are defined within the dome-shaped housing.

[0027] Figure 4 A dome-shaped housing for use with an MRI scanning system, according to at least one aspect of the present disclosure, is depicted, the dome-shaped housing having an access hole in the form of a centrally defined hole.

[0028] Figure 5 It is based on at least one aspect of this disclosure Figure 4 A cross-sectional view of the dome-shaped shell.

[0029] Figure 6 A control schematic diagram of an MRI system according to at least one aspect of this disclosure is depicted.

[0030] Figure 7 It is a method for obtaining imaging data from an MRI system according to at least one aspect of this disclosure.

[0031] Figure 8 An MRI scanning system and a robotic system according to at least one aspect of this disclosure are described.

[0032] Figure 9 A method for eliminating distortions from LF-MRI by applying a machine learning model using paired images, according to at least one aspect of this disclosure, is described.

[0033] Figure 10A method for training a generative adversarial network to eliminate distortions from LF-MRI using paired HF-MRI is described according to at least one aspect of this disclosure.

[0034] Figure 11 A method for eliminating distortion from LF-MRI during surgery is described according to at least one aspect of this disclosure.

[0035] Figure 12 This is a block diagram of a system for eliminating distortions from LF-MRI during surgery, according to at least one aspect of this disclosure.

[0036] In the various views, the corresponding reference characters represent the corresponding components. The examples listed in this document illustrate various disclosed implementations, are one form, and such examples should not be construed as limiting its scope in any way. Detailed Implementation

[0037] The applicant of this application also owns the following patent applications, each of which is incorporated herein by reference in its entirety: - International patent application number PCT / US2022 / 72143, filed on May 5, 2022, entitled NEURAL INTERVENTIONAL MAGNETIC RESONANCEIMAGING APPARATUS; - U.S. Patent Application No. 18 / 057,207, filed on November 19, 2022, entitled SYSTEM AND METHOD FOR REMOVING ELECTROMAGNETIC INTERFERENCE FROM LOW-FIELD MAGNETIC RESONANCE IMAGES; - U.S. Patent Application No. 18 / 147,418, filed on December 28, 2022, entitled MODULARIZED MULTI-PURPOSE MAGNETICRESONANCE PHANTOM; - U.S. Patent Application No. 18 / 147,542, filed on December 28, 2022, entitled INTRACRANIAL RADIO FREQUENCY COIL FOR INTRAOPERATIVE MAGNETIC RESONANCE IMAGING; - U.S. Patent Application No. 18 / 147,556, filed on December 28, 2022, entitled DEEP LEARNING SUPER-RESOLUTION TRAINING FOR ULTRA LOW-FIELD MAGNETIC RESONANCE IMAGING; - U.S. Patent Application No. 18 / 153,111, filed on January 11, 2023, entitled "ACCELERATING MAGNETIC RESONANCE IMAGINGUSING PARALLEL IMAGING AND ITERATIVE IMAGE RECONSTRUCTION"; - U.S. Patent Application No. 18 / 153,175, filed on January 11, 2023, entitled FAST T2-WEIGHTED AND DIFFUSION-WEIGHTEDCHIRPED-CPMG SEQUENCES; - U.S. Patent Application No. 18 / 450,010, filed August 15, 2023, entitled "ITERATIVE SHIMMING FOR LOW-FIELD HEAD-OPTIMIZED MRI"; and - U.S. Patent Application No. 18 / 459,712, filed on September 1, 2023, entitled ACTIVE SHIMMING FOR LOW-FIELD MAGNETICRESONANCE IMAGING.

[0038] Before explaining the various aspects of interventional magnetic resonance imaging apparatus in detail, it should be noted that the illustrative examples are not limited in application or use to the details of the construction and arrangement of the components shown in the figures and description. The illustrative examples may be implemented or combined in other aspects, variations, and modifications, and may be practiced or performed in various ways. Furthermore, unless otherwise stated, the terminology and expressions used herein are chosen for the convenience of describing the illustrative examples to the reader, and not for the purpose of limiting the illustrative examples. Moreover, it will be understood that one or more aspects, expressions, and / or examples described below may be combined with any one or more other aspects, expressions, and / or examples described below.

[0039] Various aspects involve neurointerventional magnetic resonance imaging (MRI) devices that allow for the integration of surgical intervention and guidance with MRI. This includes providing physical access to the area surrounding the patient and accessing the patient's head using one or more access ports. Additionally, neurointerventional MRI devices can allow the use of robotic-guided tools and / or conventional surgical instruments. In various cases, neurointerventional MRI can be used during surgery to obtain scans of the patient's head and / or brain during surgical interventions, such as procedures like brain biopsies or neurosurgery.

[0040] Figure 1 An MRI scanning system 100 is depicted, comprising a dome-shaped housing 102 configured to receive a patient's head. The dome-shaped housing 102 may also include at least one access hole configured to allow access to the patient's head for neurological intervention. The space within the dome-shaped housing 102 forms a region of interest (ROI) for the MRI scanning system 100. Target tissue within the RIO is subjected to a magnetizing field / pulse, as further described herein, to obtain imaging data representative of the target tissue.

[0041] For example, a patient can be positioned such that his / her head is located within a region of interest within the dome-shaped housing 102. The brain can be completely positioned within the dome-shaped housing 102. In such cases, to facilitate intracranial interventions (e.g., neurosurgery) in conjunction with MR imaging, the dome-shaped housing 102 may include one or more openings that provide access to the brain. The openings may be spaced apart around the periphery of the dome-shaped housing.

[0042] MRI scanning system 100 may include an auxiliary cart (e.g., see...) Figure 6 The MRI scanning system 100 may also include a magnet cart 540 that houses certain conventional MRI electrical and electronic components, such as a computer, programmable logic controller, power distribution unit, and amplifier. The magnet cart may also house a dome-shaped housing 102, gradient coils, and / or transmit coils, as further described herein. Additionally, in various cases, the magnet cart may be attached to a receive coil. Main Reference Figure 1 The dome-shaped housing 102 may also include an RF transmitting coil, a gradient coil 104 (depicted on its exterior), and a shim magnet 106 (depicted on its interior). Alternative configurations of the gradient coil 104 and / or the shim magnet 106 are also contemplated. In various cases, the shim magnet 106 may be adjustably positioned in a shim tray within the dome-shaped housing 102, which allows a technician to finely configure the magnetic flux density of the dome-shaped housing 102.

[0043] Various structural housings used to receive a patient's head and enable neurointervention can be used with MRI scanning systems, such as MRI scanning system 100. In one aspect, MRI scanning system 100 can be equipped with alternative housings, such as dome-shaped housing 202. Figure 2 ) or configured to form a two-part housing 302 in a dome shape ( Figure 3 The dome-shaped housing 202 defines a plurality of access holes 203; the two-part housing 302 also defines a plurality of access holes 303 and includes an adjustable gap 305 between the two parts of the housing.

[0044] In various cases, housings 202 and 302 may include a binder 308 (such as epoxy resin), for example, binder 308 holding a plurality of magnetic elements 310 in a fixed position. The plurality of magnetic elements 310 may be bonded to structural housing 312, for example, such as a plastic substrate. In various aspects, binder 308 and structural housing 312 may be non-conductive or antimagnetic materials. Main Reference Figure 3 The two-part housing 302 includes two structural housings 312. In various aspects, the structural housing for receiving the patient's head may be formed from more than two sub-parts. An access hole 303 in the structural housing 312 provides direct access to the patient's head and is not obstructed by the structural housing 312, the bonding agent 308, or the magnetic element 310. For example, the access hole 303 may be positioned within an open space of the housing 302.

[0045] Neurointerventional MRI devices have many possible configurations that can improve accessibility to surgical interventions. Many configurations are built on two main designs, often referred to as the Halbach cylinder and Halbach dome as described in the following articles: Cooley et al. (e.g., Cooley, CZ; Haskell, MW; Cauley, SF; Sappo, C.; Lapierre, CD; Ha, CG; Stockmann, JP; and Wald, LL (2018), Design of sparse Halbach magnet arrays for portable MRI using agenetic algorithm, IEEE transactions on magneticsThe article “Design of sparse Halbach magnet arrays for portable MRI using a genetic algorithm” by Cooley et al., published in IEEE Transactions on Magnetics, 54(1), 5100112 in 2018, is incorporated herein by reference in its entirety.

[0046] In various cases, the dome-shaped housing of an MRI scanning system (such as system 100) may, for example, include a Halbach dome, which defines the dome shape and is configured based on several factors, including the strength of the main magnetic field B0, field size, field homogeneity, device size, device weight, and access for neurological intervention on the patient. In various aspects, the Halbach dome includes an outer radius and an inner radius at the base of the dome. The Halbach dome may include an elongated cylindrical portion extending from the base of the dome. In one aspect, the elongated cylindrical portion includes the same outer and inner radii as the base of the dome and extends from the base of the dome by a predetermined length with a constant radius. In another aspect, the elongated cylindrical portion includes an outer and inner radius different from the base of the dome (e.g., see...). Figure 2 and Figure 3 In such cases, the different outer and inner radii of the slender cylindrical portion can be merged with the base radius in the transition region.

[0047] Figure 4 The illustration depicts an exemplary Halbach dome 400 of an MRI scanning system (such as system 100) according to at least one aspect of the present disclosure. For example, the exemplary Halbach dome 400 defines an access aperture in the form of a hole or access aperture 403, wherein the dome 400 is configured to receive the head and brain B of a patient P in a region of interest therein, and the access aperture 403 is configured to allow access to the patient P to perform neurointerventions using medical devices and / or robotically controlled surgical instruments. The Halbach dome 400 may have a single access aperture 403 constructed at the top side 418 of the dome 400, which allows access to the top of the skull while minimizing the influence of magnetic fields. Additionally or alternatively, the dome 300 may be configured with multiple access apertures around a structure 416 of the dome 400, such as... Figure 2 and Figure 3 As shown.

[0048] Approaching the diameter D of hole 403 孔 It can be small (e.g., about 2.54 cm) or very large (basically the outer diameter r of a dome of 400). 外For example, as the access hole 403 becomes larger, the dome 400 begins to resemble a Halbach cylinder. The access hole 403 is not limited to being located at the apex of the dome 400. The access hole 403 can be placed anywhere on the surface of the dome 400 or on the structure 416. In various cases, the entire dome 400 can be rotated such that the access hole 403 can coincide with a desired physical location on the patient P.

[0049] Figure 5 The relative dimensions of Halbach Dome 400 are depicted, including the diameter D near the aperture 403. 孔 The length L of dome 400 and the outer radius r of dome 400 外 and inner radius r 内 The Halbach dome 400 includes multiple magnetic elements arranged in a Halbach array and forming a magnetic assembly. These magnetic elements can be formed by the outer radius r within structure 416 or its housing. 外 and inner radius r 内 Enclosed. In one respect, the example size can be defined as: r 内 =19.3 cm; r 外 =23.6 cm; L=38.7 cm; and 2.54 cm≤D<19.3 cm.

[0050] Based on the example dimensions above, the Halbach dome 400 with proximity aperture 403 can be configured to have a magnetic flux density B0 of approximately 72 mT and a total mass of approximately 35 kg. It will be understood that dimensions can be selected based on specific applications to achieve the desired magnetic flux density B0, the total weight of the Halbach dome 400 and / or the magnet trolley, and the geometry of the neural intervention proximity aperture 403.

[0051] In various aspects, the Halbach dome 400 can be configured to define a plurality of access holes 403 placed around a structure 416 of the dome 400. These plurality of access holes 403 can be configured to allow access to the patient’s head and brain B using tools (e.g., surgical instruments) and / or surgical robots.

[0052] In various respects, the proximity hole 403 can be adjustable. An adjustable configuration provides the ability to adjust the proximity hole 403 using a motor, mechanical aids, or a manual system with a mechanical iris configuration, for example, to adjust the diameter D of the proximity hole 403. 孔 This will allow the dome to be configured for imaging scans without access hole 403, and then the configuration of the dome 400 and its mechanical iris to include access hole 403, thus enabling surgical intervention.

[0053] The Halbach dome and its magnetic array for facilitating neural intervention are further described in international patent application PCT / US2022 / 72143, filed May 5, 2022, entitled NEURAL INTERVENTIONAL MAGNETIC RESONANCE IMAGING APPARATUS (which is incorporated herein by reference in its entirety).

[0054] Now for reference Figure 6 A schematic diagram of an MRI system 500 is shown. For example, an MRI scanning system 100 ( Figure 1 Various dome-shaped housings and their magnetic arrays, as further described herein, can be incorporated into the MRI system 500. For example, the MRI system 500 includes a housing 502, which may resemble the dome-shaped housing 102 in many respects. Figure 1 ), 202 Figure 2 ) and / or 302 ( Figure 3 The housing 502 is dome-shaped and configured to form a region of interest or field of view 552 therein. For example, in various aspects of this disclosure, the housing 502 may be configured to receive a patient's head.

[0055] The housing 502 includes a magnet assembly 548 having a plurality of magnets (e.g., a Halbach magnet array) disposed therein. In various aspects, a main magnetic field B0 generated by the magnet assembly 548 extends into a field of view 552, which contains an object (e.g., a patient's head) imaged by the MRI system 500.

[0056] The MRI system 500 also includes an RF transmit / receive coil 550. The RF transmit / receive coil 550 is combined into an integrated transmit-receive (Tx / Rx) coil. In other cases, the RF transmit coil may be separate from the RF receive coil. For example, the RF transmit coil may be integrated into a housing 502, and the RF receive coil may be positioned within the housing 502 to obtain imaging data.

[0057] The housing 502 also includes one or more gradient coils 504 configured to generate a gradient field to facilitate imaging of an object in a field of view 552 generated by the magnet assembly 548, for example, which is enclosed by a dome-shaped housing and an array of dome-shaped magnetic elements therein. A shimming tray adapted to receive a shimming magnet 506 may also be incorporated into the housing 502.

[0058] During imaging, the main magnetic field B0 extends into the field of view 552. The direction of the effective magnetic field (B1) changes in response to RF pulses emitted by the RF transmit / receive coil 550 and the associated electromagnetic field. For example, the RF transmit / receive coil 550 can be configured to selectively transmit RF signals or pulses to objects in the field of view 552, such as tissue from a patient's brain. These RF pulses can alter the effective magnetic field experienced by the spins in the sample tissue.

[0059] Housing 502 communicates with auxiliary trolley 530, which is configured to supply power to housing 502 and send / receive control signals to / from housing 502. Auxiliary trolley 530 includes a power distribution unit 532, a computer 542, a spectrometer 544, a transmit / receive switch 545, an RF amplifier 546, and a gradient amplifier 558. In various cases, housing 502 may communicate with multiple auxiliary trolleys, and each trolley may support one or more of the power distribution unit 532, computer 542, spectrometer 544, transmit / receive switch 545, RF amplifier 546, and / or gradient amplifier 558.

[0060] Computer 542 communicates with spectrometer 544 via signals, and computer 542 is configured to send and receive signals between computer 542 and spectrometer 544. When an object in field of view 552 is excited by an RF pulse from RF transmit / receive coil 550, the precession of the object generates an induced current or MR current, which is detected by RF transmit / receive coil 550 and sent to RF preamplifier 556. RF preamplifier 556 is configured to amplify or amplify the excitation data signals and send them to spectrometer 544. Spectrometer 544 is configured to send excitation data to computer 542 for storage, analysis, and image construction. For example, computer 542 is configured to combine multiple stored excitation data signals to create an image. In various cases, computer 542 communicates with at least one database 562 of stored reconstruction algorithm 564 and / or pulse sequence 566. Computer 542 is configured to use the reconstruction algorithm to generate MR image 568.

[0061] The signal can also be relayed from the spectrometer 544 via the RF power amplifier 546 and the transmit / receive switch 545 positioned between the spectrometer 544 and the RF power amplifier 546 to the RF transmit / receive coil 550 in the housing 502. The signal can also be relayed from the spectrometer 544 via the gradient power amplifier 558 to the gradient coil 560 in the housing 502. For example, the RF power amplifier 546 is configured to amplify the signal and send it to the RF transmit coil 560, and the gradient power amplifier 558 is configured to amplify the gradient coil signal and send it to the gradient coil 560.

[0062] In various cases, the MRI system 500 may include a noise cancellation coil 554. For example, the auxiliary cart 530 and / or the computer 542 may communicate signalically with the noise cancellation coil 554. In other cases, the noise cancellation coil 554 may be optional. For example, some MRI systems disclosed herein may not include supplemental / auxiliary RF coils for detecting and eliminating electromagnetic interference (i.e., noise).

[0063] Figure 7 A flowchart depicting the process 570 for acquiring MRI images is shown. This flowchart, for example, can be implemented by an MRI system 500. In various cases, at box 572, the target object (e.g., a portion of a patient's anatomy) is positioned within a main magnetic field B0 of a region of interest (e.g., region of interest 552), such as within the dome-shaped housing (e.g., magnet assembly 548) of various MRI scanners further described herein. The main magnetic field B0 is configured to magnetically polarize the hydrogen protons (¹H protons) of the target object (e.g., all organs and tissues) and is referred to as the net longitudinal magnetization M0. It is proportional to the proton density (PD) of the tissue and develops exponentially over time, with a time constant referred to as the longitudinal relaxation time T1 of the tissue. For example, the T1 value of a single tissue depends on many factors including its microstructure, its water and / or lipid content, and the strength of the polarizing magnetic field. For these reasons, the T1 value of a given tissue sample depends on age and health status.

[0064] At box 574, a time-varying oscillating magnetic field B1, i.e., an excitation pulse, is applied to the magnetically polarized target object using an RF coil (e.g., RF transmit / receive coil 550). The carrier frequency of the pulse B1 field is set to the resonant frequency of the 1H protons, which causes the longitudinal magnetization to flip from its equilibrium longitudinal direction, resulting in a rotating magnetization vector. Depending on the flip angle used, this vector magnetization can typically have both transverse and longitudinal magnetization components. Common B1 pulses include inversion pulses, or 180-degree pulses and 90-degree pulses. A 180-degree pulse reverses the magnetization direction of the 1H protons along the longitudinal axis. A 90-degree pulse rotates the magnetization of the 1H protons by 90 degrees, so that the magnetization lies in the transverse plane. The MR signal is proportional to the transverse component of the magnetization and is a time-varying current detected by a suitable RF coil. These MR signals decay exponentially over time, and their time constant is called the transverse relaxation time T2, which, for example, also depends on the microstructure, water / lipid content, and the strength of the magnetic field used.

[0065] At box 576, the MR signal is spatially encoded by exposing the target object to an additional magnetic field (referred to as a gradient field) generated by a gradient coil (e.g., gradient coil 560). The gradient field, which varies linearly in space, is applied in pulses over short time intervals and varies spatially in each direction. The end result is the generation of multiple spatially encoded MR signals, which are detected at box 577 and can be reconstructed to form an MR image depicting a slice of the examined object. An RF receiving coil (e.g., RF transmit / receive coil 550) can be configured to detect the spatially encoded RF signals. The slice can be oriented in a transverse, sagittal, coronal, or any oblique plane.

[0066] At box 578, the spatially encoded signal of each slice of the scanned area is digitized and mathematically spatially decoded using a computer reconstruction program (e.g., computer 542) to generate an image depicting the internal anatomical structure of the examined object. In various cases, the reconstruction program can inversely transform the spatially encoded data (k-space data) into geometrically decoded data using an (inverse) Fourier transform.

[0067] Figure 8 A graphical illustration is provided of a robotic system 680 that can be used for neurointerventions utilizing an MRI scanning system 600. The robotic system 680 includes a computer system 696 and a surgical robot 682. The MRI scanning system 600 may be similar to the MRI system 500 and may include a dome-shaped housing with access holes and a magnetic array, as further described herein. For example, the MRI system 500 may include one or more access holes defined in a Halbach magnet array within a permanent magnet assembly to provide access to one or more anatomical sites of a patient being imaged during a medical procedure. In various configurations, the robotic arm and / or tools of the surgical robot 682 are configured to extend through the access holes in the permanent magnet assembly to reach the patient or target site. Each access hole may provide access to the patient and / or surgical site. For example, in the case of multiple access holes, the multiple access holes may allow access from different directions and / or proximal positions.

[0068] According to various implementation schemes, the robotic system 680 is configured to be placed outside the MRI system 600. For example... Figure 8 As shown, the robot system 680 may include a robot arm 684 configured for movement with one or more degrees of freedom. According to various embodiments, the robot arm 684 includes one or more robot arm sections, including a hollow shaft 686 and an end effector 688. The hollow shaft 686 and the end effector 688 are configured to move, rotate, and / or thrash within a variety of motion ranges via one or more motion controllers 690. Figure 8The double-headed curved arrows in the diagram represent exemplary rotational movements generated by the motion controller 690 at various joints in the robot arm 684.

[0069] According to various embodiments, the robotic arm 684 of the robotic system 682 is configured to approach various anatomical sites of interest via or around the MRI scanning system 600. According to various embodiments, the access aperture is designed to accommodate the size of the robotic arm 684. For example, the access aperture defines a circumference configured to accommodate the robotic arm 684, the hollow shaft 686, and the end effector 688 passing through it. In various cases, the robotic arm 684 is configured to approach various anatomical sites of the patient from around one side of the magnetic imaging device 600. For example, the hollow shaft 686 and / or the end effector 688 may be adapted to receive a robotic tool 692, such as a biopsy needle with a cutting edge 694, used to collect biopsy samples from the patient.

[0070] The reader will understand that the robot system 682 can be used in conjunction with various dome-shaped and / or cylindrical magnetic housings further described herein. Furthermore, Figure 8 The robotic system 682 and robotic tool 692 described herein are exemplary. Alternative robotic systems can be used in conjunction with the various MRI systems disclosed herein. Furthermore, handheld surgical instruments and / or additional imaging devices (e.g., endoscopes) and / or systems can also be used in conjunction with the various MRI systems disclosed herein.

[0071] In various aspects of this disclosure, the MRI systems described herein may include low-field-strength MRI (LF-MRI) systems. In such cases, for example, the main magnetic field B0 generated by the permanent magnet assembly may be between 0.1 T and 1.0 T. In other cases, the MRI systems described herein may include ultra-low-field-strength MRI (ULF-MRI) systems. In such cases, for example, the main magnetic field B0 generated by the permanent magnet assembly may be between 0.03 T and 0.1 T.

[0072] For example, high magnetic fields (such as those above 1.0 T) can impede the use of certain electrical and mechanical components near an MRI scanner. For instance, the presence of surgical instruments and / or surgical robotic components, including those made of metal (especially ferrous metals), can be hazardous near high magnetic fields, as such tools may be pulled toward the magnetization source. Furthermore, high magnetic fields typically require specially designed rooms and additional precautions and shielding measures to limit magnetic interference. Despite the limitations of high-field MRI (HF-MRI) systems, low-field and ultra-low-field MRI systems present various challenges in acquiring high-quality images with sufficient resolution and quality to achieve desired imaging goals.

[0073] Compared to higher-field MRI systems, LF-MRI and ULF-MRI systems typically define relatively poor overall magnetic field homogeneity. For example, as further described herein, the dome-shaped housing for the magnet array may include a Halbach permanent magnet array, which, in various aspects of this disclosure, generates a magnetic field B0 with homogeneity between 1,000 ppm and 10,000 ppm in the region of interest.

[0074] Under various conditions, the inherent MRI signal is proportional to the field strength. Therefore, the signal-to-noise ratio (SNR) associated with LF-MRI and ULF-MRI systems can theoretically be as low as less than one-twentieth of the SNR of a high-field MRI system. For example, the SNR of a 70 mT MRI system can be approximately 5% of the SNR of a 1.5T MRI system.

[0075] When using a non-uniform main magnetic field to acquire LF-MRI, LF-MRI exhibits a high degree of spatial distortion. The main magnetic field of LF-MRI can range from 0.05T to 1.0T. In some cases, the non-uniform main magnetic field is generated by a small array of permanent magnets (e.g., a Halbach array). Within the available imaging volume, the homogeneity of the main magnetic field can fluctuate from a few parts per thousand (ppt) to up to one percent. During imaging of pulse sequences, these fluctuations can cause equally large phase and / or frequency shifts, resulting in distortion due to spatial errors (translation) (errors reaching several centimeters) and, for example, unwanted rotation and / or shearing. On the other hand, clinical HF-MRI is typically performed within a superconducting main magnetic field between 1T and 3T, for example, where the imaging volume homogeneity is better than five parts per million (ppm) within a spherical volume with a diameter of 30 cm. Clinical HF-MRI scans typically show extremely low spatial distortion, generally at the pixel or subpixel level, corresponding to errors on the order of millimeters.

[0076] In various aspects, machine learning methods can be applied to LF-MRI for distortion correction, resulting in higher-quality images. LF-MRI systems offer numerous benefits, such as enabling invasive and image-guided neuroMRI procedures, for example, brain biopsies and tumor resection / ablation. For instance, intraoperative MRI, i.e., MR imaging acquired during surgery, can be used to guide surgical tools appearing in the images. In various cases, invasive and image-guided neuroMRI procedures typically rely on one or more prior HF-MRI or high-resolution computed tomography (CT) scans for procedure planning. Such prior HF-MRI or high-resolution CT images often have very low spatial distortion and can be paired with higher-distortion LF-MRI scans available at the time of surgery. These two images can be used to train a machine learning model (e.g., a generative adversarial network (GAN)) to synthesize distortion-corrected images. In various cases, multiple HF-MRI and / or high-resolution CT images can be paired with LF-MRI. During surgery, the LF-MRI can be fed into a machine learning model to synthesize distortion-corrected images.

[0077] Machine learning methods can be trained on pairs of high-resolution images (e.g., HF-MRI and / or high-resolution CT images) and LF-MRI from one or more subjects until the distortion-corrected images output from the machine learning model are deemed suitable for interventional or image-guided procedures. In various cases, senior neuroradiologists, neurologists, and / or neurosurgeons can determine when distortion-corrected images are appropriate. High-resolution images are paired with LF-MRI because these images belong to the same object of interest from the same subject (e.g., the same organ in the same patient). In various cases, paired images can be obtained within a predetermined time frame (e.g., 3 days, 24 hours, or 8 hours). In some aspects, the target image used in the machine learning model with LF-MRI is HF-MRI. In an alternative aspect, the target image may be a high-resolution CT image of an object of interest with key features (e.g., the limbic system, ventricles, and skull location). In yet another alternative aspect, the target image is a combination of both HF-MRI and high-resolution CT images (e.g., using both HF-MRI and high-resolution CT images for the target image).

[0078] In some respects, it is beneficial to modify high-resolution images to make them more similar to the generated output image (which is based on LF-MRI) for use with machine learning models. In one respect, one or more intermediate image targets can be used with the machine learning model; for example, the intermediate target could be a modified HF-MRI or a modified high-resolution CT image, wherein the HF-MRI or high-resolution CT image is modified to reduce its spatial resolution in two or three dimensions (i.e., smoothed), thereby making the modified HF-MRI or modified high-resolution CT image more similar to the spatial resolution of the LF-MRI and the generated distortion-corrected output image. The paired targets (e.g., HF-MRI or high-resolution CT images) can initially be made as similar to the LF-MRI as possible. For example, the image contrast of the paired images can be matched using FLAIR, T2-weighted, T1-weighted, etc.

[0079] In some respects, training a machine learning model is a multi-stage process. For example, the model might initially be trained using LF-MRI and intermediate target images (e.g., modified HF-MRI or modified high-resolution CT images). As performance improves, this process can be modified to replace the target image (a lower-resolution image) with a high-resolution HF-MRI. In other words, as the machine learning model is continuously trained and improved over time, it can adapt to larger differences, or at least larger potential differences, between the target considered "distortion-free" and the generated output image.

[0080] Once the improved image from the object-based LF-MRI image is output from the machine learning model, it can be further optimized to produce the final distortion-corrected LF-MRI for the patient. For example, the optimization step may include upsampling the output, synthesized distortion-corrected image to match the resolution of the template image for registration. Alternatively or additionally, the optimization step may include applying intensity correction to produce more uniform image contrast. In some aspects, one or more modified or additional HF-MRI or high-resolution CT scans, along with the LF-MRI, can be acquired prior to surgery and used to optimize the machine learning model, thereby maximizing the accuracy and fidelity of the distortion-corrected LF-MRI.

[0081] Figure 9A method 700 for eliminating distortions from LF-MRI using a pre-trained machine learning model is described, wherein the machine learning model is trained on at least one LF-MRI and paired high-resolution images (e.g., HF-MRI or high-resolution CT images). The LF-MRI can be acquired during surgery using an LF-MRI or ULF-MRI system (e.g., MRI scanning system 100, which includes any of various dome-shaped housings and magnetic arrays) and / or MRI system 500. In various aspects, method 700 and / or portions thereof may be powered by a computing device (e.g., computer 542). Figure 6 (Implemented by one or more control circuits).

[0082] Method 700 includes, during surgery (i.e., intraoperatively), controlling circuitry to obtain LF-MRI of the object of interest (704). The LF-MRI is generated using a low-intensity magnetic field. In various aspects, the LF-MRI is generated by, for example, an MRI scanning system 100 (which includes any of various dome-shaped housings and magnetic arrays) or an MRI system 500. In at least one aspect, the LF-MRI includes at least one distortion. The distortion may be caused by inhomogeneities of the low-intensity magnetic field, nonlinearities of the gradient coil magnetic field, and / or eddy currents associated with switching of the gradient coil. In at least one aspect, the distortion is a distortion of the anatomical structure or object of interest depicted in the LF-MRI. Some examples of distortion are spatial errors (translation) (which can be several centimeters), such as undesirable rotation, shearing, and signal convergence to the same pixel or voxel in the image.

[0083] Method 700 also includes, during the surgical procedure, the control circuit inputting the LF-MRI of the object of interest into a machine learning model (708). In at least one aspect, the machine learning model includes a generator model of a GAN. In all aspects, the machine learning model operates on the control circuit and in an unsupervised manner.

[0084] The machine learning model, based on LF-MRI, outputs a distortion-corrected image of the object of interest in real time (e.g., during surgery). In at least one aspect, the machine learning model minimizes any distortions in the LF-MRI to generate the output image. Method 700 also includes, during surgery, control circuitry transmitting the distortion-corrected image to a user interface (710), such as the display of computer 542, or any other device that allows the user to view the image.

[0085] As an example, a GAN can be used to train a generator model to generate images intraoperatively based on LF-MRI of an object of interest. For instance, a GAN can be used to train a generator model in Method 700 for generating distortion-corrected images. A GAN consists of two competing models: a generator model and a discriminator model. The generator model includes a first neural network that can be used to generate images. The discriminator model includes a second neural network that evaluates images to determine whether a given image is a “real” image or a “synthetic” (i.e., “fake”) image. The generator model generates distortion-corrected synthetic or fake LF-MRI images from real LF-MRI, while the discriminator determines whether the input image is a “real” or “synthetic” image, i.e., whether it was generated by the generator. In other words, the discriminator model determines whether the image generated by the generator model belongs to the same image set or the same image type as the real image. The two models work together such that the generator model’s goal is to make the discriminator model classify the generated image as “real,” and the discriminator model’s goal is to correctly classify the generated image as “synthetic.”

[0086] Both models are updated to optimize their respective objectives. For example, the generator neural network can be penalized when the discriminator neural network determines that a generated image is fake, and rewarded when the discriminator determines that the synthesized image is real. Similarly, the discriminator neural network can be penalized if it determines that a generated image is real or identifies a real image as fake. Furthermore, the discriminator neural network can be rewarded when it correctly determines that a generated image is synthetic and a real image is real. In all cases, as the generator and discriminator models continuously optimize—that is, as the generator model generates higher-quality images and the discriminator model becomes increasingly unable to distinguish between synthetic and real images—the process utilized by GANs can improve the overall quality of images generated by the generator model.

[0087] A GAN can be trained using pairs of high-resolution images and LF-MRI from multiple test subjects, where each pair consists of a high-resolution image and an LF-MRI from the same test subject. Partially paired images can be between a high-resolution CT image and an LF-MRI from the same subject, and partially paired images can be between an HF-MRI and an LF-MRI from the same subject. For example, LF-MRI and high-resolution images can be input into the GAN, where the generator model receives the LF-MRI, and the discriminator model receives the high-resolution image and a distortion-corrected image output by the generator model. The generator model can generate images based on the LF-MRI. High-resolution images (e.g., HF-MRI or high-resolution CT images) and images from the generator model can be input into the discriminator model. The generator and discriminator models can be trained on multiple pairs of LF-MRI and high-resolution images until the image generated by the generator model (i.e., distortion-corrected LF-MRI) is considered compliant with the requirements of interventional and / or image-guided procedures. This process refers to... Figure 10 A more detailed description was provided.

[0088] Figure 10 A method 800 for training a GAN using paired HF-MRI to eliminate distortions from LF-MRI is described. In various aspects, LF-MRI is acquired before surgery (i.e., preoperatively) using an LF-MRI or ULF-MRI system (e.g., MRI scanning system 100, which includes any of various dome-shaped housings and magnetic arrays) and / or MRI system 500). In various aspects, method 800 or parts thereof may be powered by a computing device (e.g., computer 542). Figure 6 (Implemented by one or more control circuits).

[0089] Method 800 includes control circuitry to acquire a high-resolution image (802) of a first object of interest (e.g., the patient's brain). In at least one aspect, as previously described herein, the high-resolution image is an HF-MRI, and the HF-MRI is generated using a high-field magnetic field. Alternatively, the high-resolution image is a high-resolution CT image. In all aspects, the high-resolution image is acquired prior to surgery (i.e., preoperatively). The high-resolution image can be either an HF-MRI or a high-resolution CT image, and the subsequent procedures of method 800 remain unchanged.

[0090] Method 800 also includes, preoperatively, controlling the circuitry to acquire an LF-MRI of the object of interest (804). The LF-MRI is generated using a low-field magnetic field. The high-resolution image and the low-field image are paired images of the same object acquired within a relatively short time period. In various aspects, the LF-MRI is generated by an MRI scanning system 100 (which includes any of various dome-shaped housings and magnetic arrays) or an MRI system 500. In at least one aspect, the LF-MRI includes at least one distortion.

[0091] In some aspects, method 800 also includes control circuitry to modify the high-resolution image (806). The modification may transform the high-resolution image (e.g., an HF-MRI or high-resolution CT image) into an image that more closely matches the LF-MRI. In at least one aspect, the spatial resolution of the high-resolution image is higher than that of the LF-MRI. In one aspect, the spatial resolution of the high-resolution image is adjusted to match the spatial resolution of the LF-MRI. For example, the high-resolution image may be smoothed to adjust its spatial resolution to match that of the LF-MRI. The modified high-resolution image can be used as an intermediate target during the training process of a machine learning model. The modified high-resolution image provides the benefit of a closer match to the LF-MRI compared to the original high-resolution image. In some alternative aspects, modifying the high-resolution image (806) may not be desirable and / or necessary, and therefore, the control circuitry may not modify the high-resolution image before inputting it into the machine learning model.

[0092] In at least one aspect, Method 800 includes prior training of the GAN using multiple paired high-resolution images and LF-MRI obtained from healthy volunteers and patients during an early imaging timeframe. During training, the discriminator model is trained directly on both real and generated images and is responsible for classifying images as “real” or “fake,” while the generator model is trained using feedback from the discriminator model, as further described herein. The training phase seeks to minimize a suitable loss function for both the generator and discriminator models when applied together to the training dataset. In one aspect, the discriminator model seeks to maximize the probability that it correctly assigns the “real” label to real HF images and the “fake” label to the synthetic LF images it receives from the generator model. Simultaneously, the generator model seeks to maximize the probability that its synthetic LF-MRI images are classified as “real” by the discriminator model.

[0093] HF-MRI and LF-MRI are input into GAN 820. Method 800 also includes control circuitry to input LF-MRI and a high-resolution image (or a modified high-resolution image) into GAN 820. Method 800 also includes control circuitry to input LF-MRI into the generator model of the GAN (808). Method 800 also includes, via control circuitry, an image generated by the generator model based on the LF-MRI output (810). Distortion in the LF-MRI is minimized to produce the generated image. Method 800 also includes control circuitry to input the generated image and the high-resolution image (or a modified high-resolution image) into a discriminator model (812). Method 800 also includes control circuitry to determine whether the generated image is classified as “real” or “fake” based on the output of the discriminator model (814). The discriminator model seeks to maximize the correct probability assigned to “real” and “fake” images, while the generator model seeks to minimize the probability that its image is predicted as “fake” or maximize the probability that its image is predicted as “real”.

[0094] If the discriminator model determines that the generated image is a "forgery," method 800 continues to update the generator model by the control circuit (816). The control circuit can update the generator model by adjusting the weights and / or biases of the neural network of the generator model. In at least one aspect, adjustments are made using an optimization method, wherein the generator model seeks to minimize the probability that its image is predicted as a "forgery" or to maximize the probability that its image is predicted as "real." Updating the generator model is used to improve the generated images produced by the generator model. For example, as the generator model is updated, the generated images will increasingly resemble high-resolution images.

[0095] If the discriminator model determines that the generated image is "real," method 800 continues to update the discriminator model by the control circuit (818). The control circuit updates the discriminator model by adjusting the weights and / or biases of the neural network of the discriminator model. In at least one aspect, adjustments are made using an optimization method, where the discriminator model seeks to maximize the correct probabilities assigned to "real" and "fake" images. For example, updating the discriminator model is used to improve the classification of generated images as "fake" and high-resolution images as "real." Both the discriminator model and the generator model can be improved during training, so that the improvement of one model does not significantly exceed that of the other.

[0096] During GAN training, the competing generator and discriminator models are improved simultaneously. Once either the generator or discriminator model is updated, Method 800 is repeated to continuously update the GAN. This update process enables the generator model to produce higher-quality images based on LF-MRI images. For example, as the generator model is optimized, it is trained to produce higher-quality images, and as the discriminator model is optimized, it is trained to improve its ability to detect forged images. With the improved quality of the generated images, more distortion can be removed, and the generated images appear more like high-resolution images.

[0097] The GAN is trained to a desired performance level. For example, the GAN can be trained until the distortion-corrected image output from the generator model is considered suitable for invasive or image-guided procedures. In various situations, senior neuroradiologists, neurologists, and / or neurosurgeons can determine whether the distortion-corrected image is appropriate. Method 800 also includes a control circuit sending a notification to the user interface based on the GAN reaching the desired performance level. For example, once the control circuit determines that the GAN has reached the desired performance level, a professional can review the distortion-corrected image to determine whether it is suitable for invasive or image-guided procedures.

[0098] In at least one aspect, different pairs of high-resolution images (e.g., HF-MRI or high-resolution CT images) and LF-MRI can be used during GAN training. For example, multiple pairs of high-resolution images and LF-MRI can be used during training. In each aspect, a new pair of high-resolution images and LF-MRI can be used in every iteration of Method 800, or the same pair of high-resolution images and LF-MRI can be used multiple times before changing to a new pair of high-resolution images and LF-MRI. Paired high-resolution images and LF-MRI are images of the same object of interest from the same patient. If the high-resolution image is HF-MRI, one MRI is acquired with a high-field magnetic field and the other with a low-field magnetic field. Using more than one pair of high-resolution images and LF-MRI can provide GAN with different images and objects for training and can improve the quality of the machine learning tool in various situations. As an example, GAN can be trained with hundreds of pairs of matched LF-MRI and HF-MRI from various patient types. In addition to individuals such as those who have suffered a stroke, have undergone tumor resection, have undergone craniotomy, and / or have significant midline shift due to compensatory hydrocephalus ex vacuo, patients may also include healthy individuals. In some cases, training the model on multiple anatomical structures, including highly abnormal structures, such as cases where skull defects may be present, is advantageous.

[0099] In at least one aspect, the generator model is trained without updating the discriminator model, and then the discriminator model is trained without updating the generator model, or vice versa. The advantage of this approach is that one model can be trained without changing the other, thus improving one model relative to the other. Since the model is trained on something invariant, this process also simplifies optimization work that improves the model being trained. This approach can also be beneficial if one model performs worse than the other. For a GAN to function properly, the two models need to have similar levels of result quality. For example, a tool can be considered of acceptable quality when the discriminator model detects only about half of the total images as "fakes," and its generator and discriminator models can be considered sufficiently similar in quality.

[0100] Figure 11 The text describes a generator model trained by a generative adversarial network for use during surgery (e.g., regarding...). Figure 10Method 900 (discussed in method 800) eliminates distortion from LF-MRI. Similar to method 700, LF-MRI can be acquired during surgery using an LF-MRI or ULF-MRI system (e.g., MRI scanning system 100 (which includes any of various dome-shaped housings and magnetic arrays) or MRI system 500). In various aspects, method 900 and / or parts thereof can be computed by a computing device (e.g., computer 542). Figure 6 (Implemented by one or more control circuits).

[0101] Method 900 includes, during surgery (i.e., intraoperatively), control circuitry acquiring an LF-MRI of the object of interest (904). In various aspects, the LF-MRI is generated by, for example, an MRI scanning system 100 (which includes any of various dome-shaped housings and magnetic arrays) or an MRI system 500. In at least one aspect, the LF-MRI includes at least one distortion. Method 900 also includes control circuitry inputting the LF-MRI into a generator model of the GAN (908), and the generator model generating an image based on the LF-MRI output, wherein the distortion in the LF-MRI is minimized to produce the generated image. Subsequently, method 900 also includes control circuitry transmitting the generated image to a user interface (910), such as the display of a computer 542, or any other device that allows the user to view the image.

[0102] Figure 12 A system 1000 is described for eliminating distortions from an LF-MRI 1004 during surgery using a machine learning model. System 1000 includes a computing device 1010 having a memory 1014, a processor 1012, and a machine learning model 1016, which is a generative adversarial network. The computing device 1010 can communicate with one or more other computing devices via a network 1040. The computing device 1010 can be implemented as a server, desktop computer, laptop computer, and / or mobile device (e.g., tablet device or mobile phone device). In various cases, the computing device 1010 may represent multiple communicating computing devices, such as multiple communicating servers.

[0103] Processor 1012 may represent two or more processors on computing device 1010 that execute in parallel and utilize corresponding instructions stored in memory 1014. Memory 1014 represents a non-transitory computer-readable storage medium. Memory 1014 may represent one or more different types of memory, such as that utilized by computing device 1010. In addition to storing instructions that enable processor 1012 to implement machine learning model 1016 and computer-readable instructions stored in memory 1014, memory 1014 may be used to store data, such as imaging data, pre-trained models (such as machine learning models), algorithms, and / or their subroutines.

[0104] During the training of the machine learning model, a user can input an LF-MRI 1004 and a paired high-resolution image 1002 (e.g., an HF-MRI or high-resolution CT image) into the machine learning model 1016 of the computing device 1010. Alternatively, the machine learning model 1016 can run on the computing device 1010 as a component of a cloud network in which a user inputs the LF-MRI 1004 and the paired high-resolution image 1002 from another computing device via a network (such as network 1040) during training. After training the machine learning model 1016 to the desired performance level, the machine learning model 1016 can generate distortion-corrected images based on the LF-MRI (without requiring a high-resolution image). The LF-MRI can be acquired using an MRI scanning system. For example, the MRI scanning system could be an MRI scanning system 100 (which includes any of various dome-shaped housings and magnetic arrays) or an MRI system 500. The machine learning model 1016 is capable of real-time distortion correction based on the paired LF-MRI 1004 and the high-resolution image 1002.

[0105] Figure 12 The machine learning model 1016 shown is a generative adversarial network (GAN), which includes a generator model and a discriminator model. The generator model is a generator neural network 1018, which includes multiple layers such as 1020a and 1020b. The discriminator model is a discriminator neural network 1022, which includes multiple layers such as 1024a and 1024b. Both the generator neural network 1018 and the discriminator neural network 1022 can have any number of layers. Although... Figure 12 Both the generator neural network 1018 and the discriminator neural network 1022 have at least 3 layers, but both neural networks can have only 1 or 2 layers. Furthermore, while the generator neural network 1018 and the discriminator neural network 1022 can have the same number of layers, they do not necessarily have to have the same number of layers.

[0106] Such as about Figure 10As discussed, during the training of the machine learning model or GAN (1016), LF-MRI 1004 and high-resolution image 1002 are input into GAN 1016. LF-MRI 1004 is input into generator neural network 1018, and high-resolution image 1002 is input into discriminator neural network 1022. Discriminator neural network 1022 receives the generated image from generator neural network 1018 and high-resolution image 1002. (See also...) Figure 10 The discriminator neural network 1022 and the generator neural network 1018 are trained simultaneously. Similar to method 800, the high-resolution image 1002 can be an HF-MRI or high-resolution CT image. The generator neural network 1018 and the discriminator neural network 1022 are trained preoperatively. (See also...) Figure 11 The method described in 900 uses a trained generator neural network 1018 in the procedure.

[0107] In various aspects, upon receiving the paired high-resolution images 1002 and LF-MRI 1004, the processor 1012 of the computing device 1010 can execute the training of a machine learning model 1016 to train a generative adversarial network based on the paired high-resolution images 1002 and LF-MRI 1004. For example, machine-readable instructions stored in memory 1014 can be used by the processor 1012 to run the machine learning model 1016 according to, for example, method 800 (… Figure 10 The generator neural network 1018 and the discriminator neural network 1022 are trained. In at least one aspect, both the LF-MRI 1004 and the high-resolution image 1002 are images of the same object of interest from the same patient.

[0108] During the procedure, the LF-MRI 1004 can be acquired using an MRI scanning system. For example, the MRI scanning system could be an MRI scanning system 100 (which includes any of various dome-shaped housings and magnetic arrays) or an MRI system 500. The LF-MRI 1004 is input into a generator neural network 1018. The generator neural network 1018 generates a generated image that minimizes any distortion of the LF-MRI. Based on the output of the generator neural network 1018, the generated image is output to a display 1030.

[0109] In various aspects, upon receiving the LF-MRI 1004, the processor 1012 of the computing device 1010 can execute a machine learning model 1016 to apply a generator neural network 1018 to the LF-MRI 1004. For example, machine-readable instructions stored in memory 1014 can be used by the processor 1012 to run the generator neural network 1018 according to, for example, method 700 (… Figure 9 ) and Method 900 ( Figure 11 Based on LF-MRI of the object of interest from the patient, images with minimal distortion of the object of interest are generated in real time.

[0110] Although the various aspects disclosed herein relate to brain imaging and / or neurointervention, the reader will understand that the various systems and methods disclosed herein can be used in a variety of situations to image other parts and / or different structures of a patient's anatomy.

[0111] While several forms have been described and illustrated, the applicant does not intend to limit or restrict the scope of the appended claims to such details. Many modifications, variations, alterations, substitutions, combinations, and equivalents of these forms can be implemented and will now be apparent to those skilled in the art without departing from the scope of this disclosure. Furthermore, the structure of each element associated with a described form can alternatively be described as means for providing the function performed by that element. Moreover, where the materials of some components are disclosed, other materials may be used. Therefore, it should be understood that the foregoing description and the appended claims are intended to cover all such modifications, combinations, and variations falling within the scope of the disclosed forms.

[0112] The foregoing detailed description has illustrated various forms of apparatus and / or processes using block diagrams, flowcharts, and / or examples. As long as such block diagrams, flowcharts, and / or examples contain one or more functions and / or operations, those skilled in the art will understand that each function and / or operation in such block diagrams, flowcharts, and / or examples can be implemented individually and / or collectively by various hardware, software, firmware, or virtually any combination thereof. Those skilled in the art will recognize that some aspects of the forms disclosed herein can be implemented, wholly or partially equivalently, in integrated circuits, as one or more computer programs running on one or more computers (e.g., as one or more programs running on one or more computer systems), as one or more programs running on one or more processors (e.g., as one or more programs running on one or more microprocessors), as firmware, or as virtually any combination thereof, and those skilled in the art will recognize that designing circuits and / or writing software and / or firmware code in accordance with this disclosure will be entirely within the skill of those skilled in the art. Furthermore, those skilled in the art will understand that the mechanisms of the subject matter described herein are capable of being distributed in various forms as one or more program products, and that the illustrative form of the subject matter described herein applies regardless of the specific type of signal-bearing medium actually used to perform the distribution.

[0113] Instructions for programming logic to execute various public aspects can be stored in the system's memory, such as dynamic random access memory (DRAM), cache, flash memory, or other storage devices. Furthermore, instructions can be distributed via a network or through other computer-readable media. Therefore, machine-readable media can include any mechanism for storing or transmitting information in a machine-readable (e.g., computer-readable) form, but is not limited to floppy disks, optical disks, read-only optical disks (CD-ROMs) and magneto-optical disks, read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic cards or optical cards, flash memory, or tangible machine-readable storage devices for transmitting information over the Internet via electrical signals, optical signals, acoustic signals, or other forms of propagation signals (e.g., carrier waves, infrared signals, digital signals, etc.). Therefore, non-transitory computer-readable media includes any type of tangible machine-readable medium suitable for storing or transmitting electronic instructions or information in a machine-readable (e.g., computer-readable) form.

[0114] As used in any part of this document, the term "control circuitry" can refer to, for example, hardwired circuitry, programmable circuitry (e.g., a computer processor, processing unit, processor, microcontroller, microcontroller unit, controller, digital signal processor (DSP), programmable logic device (PLD), programmable logic array (PLA), or field-programmable gate array (FPGA) including one or more individual instruction processing cores), state machine circuitry, firmware storing instructions executed by the programmable circuitry, and any combination thereof. Control circuitry can be embodied collectively or individually as part of a larger system, such as integrated circuits (ICs), application-specific integrated circuits (ASICs), system-on-a-chip (SoCs), desktop computers, laptop computers, tablet computers, servers, smartphones, etc. Therefore, as used herein, "control circuitry" includes, but is not limited to, circuitry having at least one discrete circuit, circuitry having at least one integrated circuit, circuitry having at least one application-specific integrated circuit, circuitry forming a general-purpose computing device configured by a computer program (e.g., a general-purpose computer configured by a computer program that at least partially executes the processes and / or devices described herein, or a microprocessor configured by a computer program that at least partially executes the processes and / or devices described herein), circuitry forming a storage device (e.g., in the form of random access memory), and / or circuitry forming a communication device (e.g., a modem, a communication switch, or an optoelectronic device). Those skilled in the art will recognize that the subject matter described herein can be implemented in analog or digital modes or some combination thereof.

[0115] As used in any aspect herein, the term "logic" can mean an application, software, firmware, and / or circuit configured to perform any of the operations described above. Software can be embodied in software packages, code, instructions, instruction sets, and / or data recorded on a non-transitory computer-readable storage medium. Firmware can be embodied in hard-coded (e.g., non-volatile) code, instructions, or instruction sets and / or data in a storage device.

[0116] As used in any part of this document, the terms “component,” “system,” “module,” etc., may refer to an entity related to the control circuitry of a computer, which may be hardware, a combination of hardware and software, software, or software in execution.

[0117] As used anywhere in this document, "algorithm" refers to a self-consistent sequence of steps that leads to a desired result, where "step" refers to the manipulation of physical quantities and / or logical states, which, although not necessarily required, can take the form of electrical or magnetic signals capable of being stored, transmitted, combined, compared, and otherwise manipulated. These signals are typically referred to as bits, values, elements, symbols, characters, terms, numbers, etc. These terms and similar terms may be associated with appropriate physical quantities and are merely convenient labels applied to these quantities and / or states.

[0118] The network may include a packet-switched network. Communication devices may be able to communicate with each other using the selected packet-switched network communication protocol. An example communication protocol may include an Ethernet communication protocol that allows communication using Transmission Control Protocol / Internet Protocol (TCP / IP). The Ethernet protocol may conform to or be compatible with the Ethernet standard entitled "IEEE 802.3 Standard" published by the Institute of Electrical and Electronics Engineers (IEEE) in December 2008 and / or subsequent versions of that standard. Alternatively or additionally, communication devices may be able to communicate with each other using the X.25 communication protocol. The X.25 communication protocol may conform to or be compatible with standards issued by the International Telecommunication Union-Telecommunication Standardization Sector (ITU-T). Alternatively or additionally, communication devices may be able to communicate with each other using the Frame Relay communication protocol. The Frame Relay communication protocol may conform to or be compatible with standards issued by the International Telegraph and Telephone Consultative Committee (CCITT) and / or the American National Standards Institute (ANSI). Alternatively or additionally, transceivers may be able to communicate with each other using the Asynchronous Transfer Mode (ATM) communication protocol. The ATM communication protocol may conform to or be compatible with the ATM standard entitled "ATM-MPLS Network Interoperability 2.0" and / or subsequent versions of that standard, published by the ATM Forum in August 2001. Of course, this document also anticipates different and / or subsequently developed connection-oriented network communication protocols.

[0119] As is evident from the foregoing disclosure, unless otherwise expressly stated, it is understood that throughout the foregoing disclosure, discussions using terms such as “processing,” “calculating,” “operating,” “determining,” and “displaying” refer to the actions and processes of a computer system or similar electronic computing device that manipulates data represented as physical (electronic) quantities in computer system registers and memories and converts that data into other data similarly represented as physical quantities in computer system memories or registers or other such information storage, transmission, or display devices.

[0120] One or more components may be referred to herein as “configured to,” “configurable to,” “operable / working for,” “adaptable / fittable,” “capable of,” “suitable / compliant,” etc. Those skilled in the art will recognize that, unless the context otherwise requires, “configured to” generally includes active state components and / or inactive state components and / or standby state components.

[0121] Those skilled in the art will recognize that, in general, the terminology used herein, and especially in the appended claims (e.g., the body of the appended claims), is typically intended as “open-ended” terms (e.g., the term “including” should be interpreted as “including but not limited to”, the term “having” should be interpreted as “having at least”, the term “includes” should be interpreted as “including but not limited to”, etc.). Those skilled in the art will further understand that if a particular number of introductory claim statements are intentional, such intention will be expressly stated in the claims, and without such statements, such intention does not exist. For example, to aid understanding, the appended claims may contain the use of introductory phrases “at least one” and “one or more” to introduce claim statements. However, the use of such phrases should not be construed as implying that a claim statement introduced by the indefinite article “a / a(a)” or “a / a(an)” limits any particular claim containing such an introductory claim statement to a claim containing only one such statement, even when the same claim includes the introductory phrase “one or more” or “at least one” and indefinite articles such as “a / a(a)” or “a / a(an)” (e.g., “a / a(a)” and / or “a / a(an)” should generally be interpreted as meaning “at least one” or “one or more”); the same applies to the use of definite articles used to introduce claim statements.

[0122] Furthermore, even when a specific number of statements in the introductory claims is explicitly stated, those skilled in the art will recognize that such statements should generally be interpreted as referring to at least the number stated (e.g., in the absence of other modifiers, the naked statement "two statements" generally means at least two statements, or two or more statements). Moreover, when using conventions such as "at least one of A, B, and C," generally, such constructions are intended to be understood by those skilled in the art (e.g., "a system having at least one of A, B, and C" will include, but is not limited to, systems having only A, only B, only C, both A and B, both A and C, both B and C, and / or both A, B, and C). When using conventions such as "at least one of A, B, or C," generally, such constructions are intended to be understood by those skilled in the art (e.g., "a system having at least one of A, B, or C" will include, but is not limited to, systems having only A, only B, only C, both A and B, both A and C, both B and C, and / or both A, B, and C). Those skilled in the art will further understand that, generally, whether in the specification, claims, or drawings, extractive terms and / or phrases presenting two or more alternative terms should be understood to contemplate the possibility of including one, any, or both of the terms, unless the context otherwise specifies. For example, the phrase “A or B” will generally be understood to include the possibility of including “A” or “B” or “A and B”.

[0123] Regarding the appended claims, those skilled in the art will understand that the operations described herein can generally be performed in any order. Furthermore, although the various operation flowcharts are presented sequentially, it should be understood that the various operations can be performed in any order other than those shown, or can be performed simultaneously. Unless the context otherwise specifies, examples of such alternative orders may include overlapping, interlacing, interrupted, reordered, ascending, preparatory, supplementary, simultaneous, reversed, or other variations of the order. Furthermore, unless the context otherwise specifies, terms such as “in response to,” “related,” or other past participle adjectives are generally not intended to exclude such variations.

[0124] It is important to note that any reference to "one aspect," "in one respect," "example," "an example," etc., implies that the specific feature, structure, or characteristic described in connection with that aspect is included in at least one aspect. Therefore, the phrases "in one aspect," "in one respect," "in an example," and "in an example" appearing in various places throughout the specification do not necessarily refer to the same aspect. Furthermore, specific features, structures, or characteristics can be combined in any suitable manner with one or more aspects.

[0125] Any patent application, patent, non-patent publication, or other disclosure mentioned in this specification and / or listed in any application data sheet is incorporated herein by reference, provided that the incorporated material is consistent with this document. Thus, and to the extent necessary, the disclosure expressly set forth herein supersedes any conflicting material incorporated herein by reference. Any material or portion thereof that is said to be incorporated herein by reference but conflicts with existing definitions, statements, or other disclosures set forth herein will be incorporated only if there is no conflict between the incorporated material and existing disclosures. In summary, the many benefits of adopting the concepts described herein have been described. For illustrative and descriptive purposes, one or more forms of the above description have been presented. It is not intended to be exhaustive or to limit the precise forms disclosed. Modifications or variations may be made in accordance with the above teachings. One or more forms were chosen and described to illustrate principles and practical applications, thereby enabling those skilled in the art to utilize the various forms and make various modifications to suit a particular intended use. The claims filed herein are intended to define the overall scope.

Claims

1. A system comprising: A database that stores high-resolution preoperative images of objects of interest; as well as Control circuitry including a processor and a memory, wherein the memory stores instructions executable by the processor to: Low-field magnetic resonance imaging (MRI) of the object of interest was obtained during the procedure. During the procedure, the low-field MRI of the object of interest is input into a pre-trained generative adversarial network generator model, wherein the generator model is pre-trained with the low-field MRI and paired high-resolution images to correct image distortion. During the procedure, based on the low-field MRI, a distortion-corrected image of the object of interest is output from the generator model; as well as During the procedure, the distortion-corrected image of the object of interest is transmitted to the user interface.

2. The system of claim 1, wherein the paired high-resolution images are selected from high-field MRI and high-resolution computed tomography images.

3. The system of claim 1, wherein the generative adversarial network is trained to a desired performance level using at least one set of training images, wherein each set of training images includes a high-resolution training image of an object of interest and a low-field MRI training image of the object of interest, and wherein the system further includes training control circuitry to: A first set of training images is obtained preoperatively, wherein the first set of training images includes a first low-field MRI training image and a first high-resolution training image of a first training object of interest. Before surgery, the first low-field MRI training image is input into the generator model of the generative adversarial network to generate the first distortion-corrected training image. Before surgery, the first distortion-corrected training image and the first high-resolution training image are input into the discriminator model of the generative adversarial network to evaluate the first training image. as well as Based on the evaluation of the first distortion-corrected training image by the discriminator model before surgery, one of the generator model and the discriminator model is updated.

4. The system of claim 3, wherein the discriminator model comprises a discriminator neural network, and the generator model comprises a generator neural network, and wherein the training control circuit is further configured to classify the first distortion-corrected training image from the generator model as "real" based on the discriminator model, and to adjust at least one weight of at least one layer of the discriminator neural network.

5. The system of claim 3, wherein the discriminator model comprises a discriminator neural network, and the generator model comprises a generator neural network, and wherein the training control circuit is further configured to classify the first distortion-corrected training image as "forgery" based on the discriminator model, and to adjust at least one weight of at least one layer of the generator neural network of the generator model.

6. The system of claim 1, wherein the low-field MRI includes distortions of the anatomical structures depicted in the low-field MRI.

7. The system of claim 1, wherein the paired high-resolution images have a first resolution, wherein the low-field MRI has a second resolution, wherein the second resolution is lower than the first resolution, and wherein the memory stores further instructions executable by the processor to adjust the first resolution of the paired high-resolution images based on the second resolution of the low-field MRI prior to training the generative adversarial network.

8. The system of claim 1, wherein adjusting the first resolution based on the second resolution includes smoothing each paired high-resolution image.

9. The system of claim 1, wherein the memory stores further instructions executable by the processor to generate low-field MRI intraoperatively using a low-field magnetic field.

10. The system according to claim 9, further comprising: A dome-shaped housing configured to house an array of magnets arranged to generate the low-intensity magnetic field toward the object of interest within the field of view, wherein the low-intensity magnetic field has a magnetic field strength of less than or equal to 1 T. as well as A radio frequency coil assembly configured to selectively excite the magnetization of the object of interest within the field of view.

11. The system of claim 1, wherein the memory stores further instructions executable by the processor to transmit the distortion-corrected image of the object of interest to the user interface in real time during surgery.

12. The system of claim 1, wherein the object of interest comprises the anatomical structure of a particular patient.

13. A training system for generative adversarial networks, the training system comprising: Training processor; as well as A training memory stores multiple sets of training images, each set including a high-resolution training image of a training object of interest and a paired low-field MRI training image of the training object of interest, and wherein the memory stores instructions executable by the processor to: Obtain a first set of training images, wherein the first set of training images includes a first low-field MRI training image and a first high-resolution training image of a first object of interest; The first low-field-strength MRI training image is input into the generator model of the generative adversarial network to generate the first distortion-corrected training image. The first distortion-corrected training image and the first high-resolution training image are input into the discriminator model of the generative adversarial network; The first distortion-corrected training image is evaluated by the discriminator model to identify it as either "real" or "fake". as well as Based on the evaluation of the first distortion-corrected training image performed by the discriminator model, the generative adversarial network is updated preoperatively, wherein updating the generative adversarial network includes: If the discriminator model classifies the first distortion-corrected training image from the generator model as "real", then the discriminator model is updated; as well as If the discriminator model classifies the first distortion-corrected training image from the generator model as "forgery", then the generator model is updated.

14. The training system of claim 13, wherein the discriminator model comprises a discriminator neural network, and the generator model comprises a generator neural network, and wherein the training memory stores instructions executable by the training processor to: Based on the discriminator model, the first distortion-corrected training image from the generator model is classified as "real," and at least one weight of at least one layer of the discriminator neural network of the discriminator model is adjusted; and Based on the discriminator model, the first distortion-corrected training image is classified as "fake", and at least one weight of at least one layer of the generator neural network of the generator model is adjusted.

15. The training system of claim 13, further comprising training the generator model to the desired performance level by: Obtain at least another set of training images of different objects; and The generative adversarial network is further trained using at least another set of training images.

16. A method comprising: The generative adversarial network is trained to the desired performance level before surgery, wherein training the generative adversarial network includes: The first set of training images is input into the generative adversarial network, wherein the first set of training images includes a first low-field MRI training image and a first high-resolution training image of a first object of interest, and wherein the generative adversarial network includes a generator model and a discriminator model. The generator model generates a first distortion-corrected training image based on the first low-field MRI training image; The discriminator model receives the first distortion-corrected training image and the first high-resolution training image; The discriminator model classifies the first distortion-corrected training image as "real" or "fake"; and The generative adversarial network is updated based on the classification, wherein updating the generative adversarial network includes: If the first distortion-corrected training image is classified as "real", the discriminator model is updated; and if the first distortion-corrected training image is classified as "fake", the generator model is updated; the method further includes: A notification is sent to the user interface once the generative adversarial network achieves the desired performance level.

17. The method of claim 16, further comprising obtaining a high-field MRI of the first training object of interest preoperatively.

18. The method of claim 16, further comprising obtaining a high-resolution computed tomography image of the first training object of interest preoperatively.

19. The method of claim 16, further comprising, after training the generative adversarial network to the desired performance level, generating a distortion-corrected image with minimal distortion, wherein generating the distortion-corrected image comprises: Intraoperatively, a low-field MRI of the object of interest is obtained using a low-field magnetic resonance imaging system, wherein the low-field MRI includes a dome-shaped housing and a magnet array arranged around the dome-shaped housing. During the procedure, the low-field MRI of the object of interest is input into the generator model, wherein the generator model is used to generate the distortion-corrected image; as well as The distortion-corrected image is transmitted to the user interface during the procedure, wherein the generation and transmission of the distortion-corrected image are performed in real time.

20. The method of claim 19, further comprising: A low-intensity magnetic field is projected from the magnet array toward the object of interest located within the field of view, wherein the low-intensity magnetic field has a magnetic field strength of less than or equal to 1T; A sequence of radio frequency pulses is sent to a radio frequency coil assembly, the radio frequency coil assembly being configured to selectively excite the magnetization of the object of interest within the field of view; as well as Receive and record the output signal from the radio frequency coil assembly.

Images