System and method for generating robust phase images in magnetic resonance images

Abstract

The invention is titled "System and method for generating robust phase images in magnetic resonance images". The present disclosure provides a computer-implemented method of correcting phase and reducing noise in a magnetic resonance (MR) phase image. The method includes executing a neural network model for analyzing MR images, wherein the neural network model is trained with a pair of an original image and a corrupted image, wherein the corrupted image includes corrupted phase information, the original image is the corrupted image with reduced corrupted phase information, and a target output image of the neural network model is the original image. The method further includes receiving a MR image including corrupted phase information, and analyzing the received MR image using the neural network model. The method further includes deriving an original phase image of the received MR image based on the analyzing, wherein the derived original phase image includes reduced corrupted phase information compared to the received MR image; and outputting the MR image based on the derived original phase image.

Description

BACKGROUND

[0001] The field of the present disclosure relates generally to systems and methods of correcting phase and reducing noise in phase images, and more particularly, to systems and methods of using a neural network model to reduce noise in phase images in medical images.

[0002] Magnetic resonance imaging (MRI) has proven useful for the diagnosis of many diseases. MRI provides detailed images of soft tissue, abnormal tissue such as tumors, and other structures that cannot be easily imaged by other imaging modalities such as computed tomography (CT) imaging. Moreover, MRI operates without exposing the patient to ionizing radiation experienced in modalities such as CT and X-ray.

[0003] Furthermore, MRI is unique compared to other imaging modalities in that the MRI signal is represented by a complex number, rather than by a scalar or real number. Thus, the image value of each image pixel includes a magnitude and a phase. The phase of MR images is typically discarded, although it carries important information and can be used in many applications, particularly in applications for estimating parameters that cannot be estimated with magnitude alone. Magnitude operations also change the noise distribution and affect image quality and parameter estimation using MR images. SUMMARY

[0004] In one aspect, a computer-implemented method of correcting phase and reducing noise in magnetic resonance (MR) phase images is provided. The method includes executing a neural network model for analyzing MR images, wherein the neural network model is trained with a pair of an original image and a corrupted image, wherein the corrupted image includes corrupted phase information, the original image is the corrupted image with reduced corrupted phase information, and a target output image of the neural network model is the original image. The method further includes receiving a MR image including corrupted phase information, and analyzing the received MR image using the neural network model. The method further includes deriving an original phase image of the received MR image based on the analyzing, wherein the derived original phase image includes reduced corrupted phase information compared to the received MR image; and outputting the MR image based on the derived original phase image.

[0005] In another aspect, a phase image correction and noise reduction system is provided, the phase image correction and noise reduction system comprising a phase image correction and noise reduction computing device comprising at least one processor in communication with at least one memory device. The at least one processor is programmed to execute a neural network model for analyzing MR images, wherein the neural network model is trained with a pair of an original image and a corrupted image, wherein the corrupted image comprises corrupted phase information, the original image is the corrupted image with reduced corrupted phase information, and a target output image of the neural network model is the original image. The at least one processor is further programmed to receive an MR image comprising corrupted phase information, and analyze the received MR image using the neural network model. The at least one processor is further programmed to derive an original phase image of the received MR image based on the analysis, wherein the original phase image comprises reduced corrupted phase information compared to the received MR image, and output the MR image based on the original phase image. BRIEF DESCRIPTION OF DRAWINGS

[0006] FIG. 1 is a schematic illustration of an example magnetic resonance imaging (MRI) system.

[0007] FIG. 2A is an example phase image correction and noise reduction system.

[0008] FIG. 2B is a flowchart of an example method of reducing noise in a phase image.

[0009] FIG. 3 is an example neural network model for FIG. 2A the system shown in

[0010] FIG. 4A is a noise-free phase image of a phantom.

[0011] FIG. 4B is a phase image contaminated with noise FIG. 4A shown in

[0012] FIG. 4C is an original phase image of the phase image shown in FIG. 4B denoised using deep learning.

[0013] FIG. 4D is a phase error image of the original phase image shown in FIG. 4C

[0014] is an original phase image of the phase image shown in FIG. 4E FIG. 4B is an original phase image of the phase image shown in

[0015] FIG. 4F FIG. 4E ​​The phase error image of the original phase image is shown.

[0016] FIG. 5A is the magnitude image of another phantom.

[0017] FIG. 5B yes FIG. 5A A magnitude image of the phantom contaminated by noise.

[0018] FIG. 5C yes FIG. 5A Phase image of the phantom.

[0019] FIG. 5D yes FIG. 5B Phase image of the phantom.

[0020] FIG. 5E It uses deep learning to denoise FIG. 5B The original phase image of the phantom.

[0021] FIG. 5F yes FIG. 5E The deep learning denoising results are comparable to FIG. 5A The difference between the real phase of the image.

[0022] FIG. 5G Low-pass filtering is used FIG. 5B The original phase image of the phantom.

[0023] FIG. 5H yes FIG. 5G The low-pass filtering result is FIG. 5A The difference between the real phase of the image.

[0024] FIG. 6A is the magnitude image of the phantom.

[0025] FIG. 6B yes FIG. 6A Phase image of the phantom, where the acquisition number (NEX) is ten.

[0026] FIG. 6C are the true phase image, the original phase image denoised using deep learning, and the original phase image low-pass filtered using ten acquisitions.

[0027] FIG. 6D After phase correcting the initial image using deep learning denoising results FIG. 6A A virtual image of the phantom.

[0028] FIG. 6E After phase correcting the initial image using the low-pass filtering result FIG. 6A A virtual image of the phantom.

[0029] FIG. 7Ais a real image of a diffusion weighted image that has been denoised using deep learning.

[0030] FIG. 7B is a real image of a diffusion weighted image that has been low pass filtered. FIG. 7A

[0031] FIG. 7C is a real image of a diffusion weighted image that has been denoised using deep learning. FIG. 7A

[0032] FIG. 7D is a real image of a diffusion weighted image that has been low pass filtered. FIG. 7B

[0033] FIG. 7E is an imaginary image of a diffusion weighted image that has been denoised using deep learning and then phase corrected with an initial image.

[0034] FIG. 7F is an imaginary image of a diffusion weighted image that has been low pass filtered and then phase corrected with an initial image.

[0035] FIG. 8A is a schematic diagram of a neural network model.

[0036] FIG. 8B is a schematic diagram of a neuron in the neural network model shown in FIG. 8A

[0037] FIG. 9 is a schematic diagram of an exemplary convolutional neural network.

[0038] FIG. 10 is a block diagram of an exemplary computing device. DETAILED DESCRIPTION

[0039] The present disclosure includes systems and methods of correcting and reducing noise in magnetic resonance (MR) phase images of a subject. As used herein, corrupted phase information includes noise, undesirable artifacts, background phase, or any combination thereof. Undesirable artifacts can be caused by eddy currents, physiological noise, B0 drift, or other system imperfections. A corrupted image is an image that includes corrupted phase information. An original image is an image in which the corrupted phase information is reduced. Denoising includes reducing or removing the corrupted phase information, and can include correcting the phase and / or reducing or removing noise in the phase image. Method aspects will be partially apparent and partially discussed explicitly in the following description.

[0040] ​​​​In magnetic resonance imaging (MRI), a subject is placed in a magnet. While the subject is in the magnetic field generated by the magnet, the magnetic moments of nuclei such as protons attempt to align with the magnetic field but precess about the magnetic field in a random order at the Larmor frequency of the nuclei. The magnetic field of the magnet is referred to as B0 and extends in the longitudinal or z-direction. During the acquisition of an MRI image, a magnetic field (referred to as the excitation field Bi) in the x-y plane and close to the Larmor frequency is generated by a radio frequency (RF) coil and can be used to rotate or "tip" the net magnetic moment Mz of the nuclei from the z-direction toward the transverse or x-y plane. After the excitation signal Bi is terminated, the nuclei emit a signal referred to as an MR signal. In order to use the MR signal to generate an image of the subject, magnetic field gradient pulses (Gx, Gy, and Gz) are used. The gradient pulses are used to scan through k-space, the spatial or distance frequency space. There is a Fourier relationship between the acquired MR signal and the image of the subject, so the image of the subject can be derived by reconstructing the MR signal.

[0041] FIG. 1 A schematic diagram of an exemplary MRI system 10 is shown. In an exemplary embodiment, the MRI system 10 includes a workstation 12 having a display 14 and a keyboard 16. The workstation 12 includes a processor 18, such as a commercially available programmable machine running a commercially available operating system. The workstation 12 provides an operator interface that allows scan definitions to be entered into the MRI system 10. The workstation 12 is coupled to a pulse sequence server 20, a data acquisition server 22, a data processing server 24, and a data store server 26. The workstation 12 and each of the servers 20, 22, 24, and 26 are in communication with each other.

[0042] In an exemplary embodiment, the pulse sequence server 20 is responsive to instructions downloaded from the workstation 12 to operate a gradient system 28 and a radio frequency ("RF") system 30. The instructions are used to produce gradient waveforms and RF waveforms in an MR pulse sequence. The RF coil 38 and the gradient coil assembly 32 are used to perform the defined MR pulse sequence. The RF coil 38 is shown as a whole body RF coil. The RF coil 38 can also be a local coil that can be placed near the anatomy to be imaged, or a coil array that includes multiple coils.

[0043] In an exemplary embodiment, the gradient waveforms used to perform the defined scan are produced and applied to the gradient system 28, which excites the gradient coils in the gradient coil assembly 32 to produce the magnetic field gradients Gx, Gy, and Gz used for position encoding of the MR signal. The gradient coil assembly 32 forms part of a magnet assembly 34 that also includes a polarizing magnet 36 and the RF coil 38.

[0044] In an example implementation, the RF system 30 includes an RF transmitter for generating RF pulses used in the MR pulse sequence. The RF transmitter is responsive to scan definition and direction from the pulse sequence server 20 to generate RF pulses having the required frequency, phase, and pulse amplitude waveforms. The generated RF pulses can be applied by the RF system 30 to the RF coil 38. The responsive MR signals detected by the RF coil 38 are received by the RF system 30, amplified, demodulated, filtered, and digitized under the direction of commands generated by the pulse sequence server 20. The RF coil 38 is described as a transmitter and receiver coil such that the RF coil 38 transmits RF pulses and detects MR signals. In one implementation, the MRI system 10 can include a transmitter RF coil that transmits RF pulses and a separate receiver coil that detects MR signals. The transmission channel of the RF system 30 can be connected to the RF transmission coil and the receiver channel can be connected to the separate RF receiver coil. Typically, the transmission channel is connected to the entire body RF coil 38 and each receiver segment is connected to a separate local RF coil.

[0045] In an example implementation, the RF system 30 also includes one or more RF receiver channels. Each RF receiver channel includes an RF amplifier that amplifies the MR signals received by the RF coil 38 to which the channel is connected and a detector that detects and digitizes the I and Q quadrature components of the received MR signals. The magnitude of the received MR signals can then be determined as the square root of the sum of the squares of the I and Q components as shown in Equation (1) below:

[0046]

[0047] And the phase of the received MR signals can also be determined as shown in Equation (2) below:

[0048]

[0049] In an exemplary embodiment, the digitized MR signal samples generated by the RF system 30 are received by the data acquisition server 22. The data acquisition server 22 can operate in response to instructions downloaded from the workstation 12 to receive real-time MR data and provide buffer memory so that no data is lost due to data overflow. In some scans, the data acquisition server 22 only passes the acquired MR data to the data processing server 24. However, in scans where information derived from the acquired MR data is needed to control further execution of the scan, the data acquisition server 22 is programmed to generate the required information and transmit it to the pulse sequence server 20. For example, during a pre-scan, MR data is acquired and used to calibrate the pulse sequence executed by the pulse sequence server 20. In addition, navigator signals can be acquired during the scan and used to adjust operating parameters of the RF system 30 or gradient system 28, or to control the order in which views are sampled in k-space.

[0050] In an exemplary embodiment, the data processing server 24 receives MR data from the data acquisition server 22 and processes the MR data according to instructions downloaded from the workstation 12. Such processing may include, for example, performing Fourier transforms on the raw k-space MR data to produce two-dimensional or three-dimensional images, applying filters to reconstructed images, performing backprojection image reconstruction on the acquired MR data, generating functional MR images, and computing motion or flow images.

[0051] In an exemplary embodiment, the image reconstructed by the data processing server 24 is transmitted back to the workstation 12 and stored there. In some embodiments, the real-time image is stored in a database memory cache ( FIG. 1 (not shown), real-time images can be output from the database memory cache to the operator display 14 or a display 46 located near the magnet assembly 34 for use by the attending physician. Batch mode images or selected real-time images can be stored on a disk storage device 48 or in a host database on the cloud. When such images have been reconstructed and transferred to the storage device, the data processing server 24 notifies the data storage server 26. The operator can use the workstation 12 to archive images, generate films, or send images to other facilities via the network.

[0052] A composite MR image can be reconstructed based on the I quadrature MR signal and the Q quadrature MR signal using a process such as Fourier transform. is an MR image in which each pixel is represented by a complex number with a real component and imaginary component where (x,y) is the position of the pixel in the image. The magnitude of the composite MR image (called the magnitude image) is generated as A phase image, i.e. the phase of the complex MR image, can also be generated and used, where the phase A real image is an MR image where the value at each pixel is taken as the real component A real image is an MR image where the value at each pixel is taken as the real component

[0053] In contrast to the magnitude image, the phase image is typically not used, because the complex MR signal is usually contaminated by unwanted static and dynamic background phase variations caused by gradient imperfections, physiological noise such as cardiac and respiratory motion, and the intrinsic magnetic susceptibility of the subject, which is particularly exacerbated near the interface between air and tissue. Therefore, the phase image is typically discarded and only the magnitude image is used. However, the magnitude operation changes the noise distribution from a central Gaussian distribution with zero expectation value to a non-central chi-squared noise distribution with a positive expectation value. Thus, the magnitude operation causes an overestimation of the signal and introduces a significant bias in the derived quantitative maps, especially in applications with insufficient signal-to-noise ratio (SNR). Furthermore, the magnitude operation irreversibly reduces the dynamic range of the image in low SNR applications, thereby reducing the efficiency of improving the SNR by averaging multiple magnitude images. Averaging in the complex domain is not always feasible, because an incoherent background phase can lead to signal cancellation or destructive interference.

[0054] Furthermore, the image phase can encode certain types of information. As discussed above, the magnetic moment of a nucleus precesses around the magnetic field at the Larmor frequency of the nucleus. The Larmor frequency depends on the nucleus and the local magnetic field experienced by the nucleus. The phase information is derived from the time evolution of the Larmor frequency shift of the nucleus. The local magnetic field shift can be caused by different chemical species, such as fat and water, or by local / non-local susceptibility. The local magnetic field can be affected by motion at the macroscopic or microscopic level, such as diffusion, gradients, and susceptibility. The phase image represents this frequency information of the voxels in the subject. Therefore, in applications for measuring phase-related parameters, such as phase-contrast flow assessment and susceptibility quantification, the phase image is saved and used to fit a signal model and derive maps. In these applications, the goodness of fit depends on the quality of the phase estimation.

[0055] Therefore, a robust and general phase estimation is needed prior to processing or correcting the phase information. Known phase correction methods include filter-based phase correction methods or other spatial regularization techniques such as least-squares fitting of the phase to a low-order polynomial, and low-pass filtering of the complex signal or total variation. This method results in feature loss due to spatial smoothing. Furthermore, known methods require a priori knowledge of the phase information or a separately acquired low-resolution phase image, which is time-consuming to acquire and not always practical. More importantly, all known methods rely heavily on the assumption that the background phase is spatially smooth and the noise is stationary in the image space. These assumptions are often invalid due to contamination from physiological noise, motion encoding in the signal phase, strong susceptibility gradients, or system imperfections. Therefore, known methods can result in loss of signal in practical value operations.

[0056] In contrast, the systems and methods disclosed herein use a deep neural network to denoise the phase image and do not require fine-tuning and selection of filters for specific conditions of MR data acquisition, which can vary with the size of the subject and the level of SNR. As used herein, denoising the phase image means reducing the background phase and / or noise in the phase image. Furthermore, the systems and methods disclosed herein provide a robust and general phase estimation without assuming the background phase or the noise.

[0057] FIG. 2A is a schematic diagram of an example phase image correction and noise reduction system 200. In an example embodiment, the system 200 includes a phase image correction and noise reduction computing device 202 configured to reduce noise in a phase image. The computing device 202 also includes a neural network model 204. The system 200 can include a second phase image correction and noise reduction computing device 203. The second phase image correction and noise reduction computing device 203 can be used to train the neural network model 204, and the phase image correction and noise reduction computing device 202 can then use the trained neural network model 204. The second phase image correction and noise reduction computing device 203 can be the same computing device as the phase image correction and noise reduction computing device 202, such that the training and use of the neural network model 204 are on one computing device. Alternatively, the second phase image correction and noise reduction computing device 203 can be a separate computing device from the phase image correction and noise reduction computing device 202, such that the training and use of the neural network model 204 are performed on different computing devices. The phase image correction and noise reduction computing device 202 can be included in the workstation 12 of the MRI system 10, or can be included on a separate computing device in communication with the workstation 12.

[0058] FIG. 2Bis a flowchart of an example method 250. The method 250 can be implemented on the phase image correction and noise reduction system 200. In an example embodiment, the method includes executing 252 a neural network model for analyzing MR images. The neural network model is trained with training images. The training images can be a pair of an original image and a corrupted image, and the target output image of the neural network model is the original image. The corrupted image includes corrupted phase information. The original image is the corrupted image with reduced corrupted phase information. The training images include phase information and are represented in any format that includes phase information, such as images in complex MR images, phase images, real and imaginary pairs, or phasor representations. In some embodiments, the corrupted image is a simulated phase image that is contaminated with undesirable phase artifacts, background phase, and / or noise, and the original image is the simulated phase image without contamination. The training images can be images that are contaminated with simulated noise from a highly heterogeneous dataset that can be from different applications, and / or acquired with different pulse sequences and / or different pulse sequence parameters.

[0059] The method 250 also includes receiving 254 an MR image that includes corrupted phase information. The received MR image can be in any format that includes phase information, such as images in complex MR images, real and imaginary pairs, or vector representations. A phase image can be generated based on the complex MR image by calculating the phase based on the real and imaginary components. The received MR image can be a phase image. The phase image received or generated based on the received MR image is a corrupted phase image. As used herein, a corrupted phase image is a phase image that includes corrupted phase information. The method also includes analyzing 256 the received MR image using the neural network model. The received MR image can be input directly into the neural network, where the neural network model 204 can include one or more layers of neurons configured to generate a phase image based on the received image. In some embodiments, the received MR image is processed and a phase image is generated and input into the neural network model 204. The neural network model 204 outputs an MR image with reduced corrupted phase information. The output image from the neural network model 204 includes phase information and can be in any format that includes phase information, such as images in complex images, phase images, real and imaginary pairs, or vector representations. Further, the method 250 includes deriving 258 an original phase image of the received MR image based on the analysis. In some embodiments, the output phase image is calculated based on the output image from the neural network model 204. In other embodiments, the original phase image is output directly from the neural network model 204. The derived original phase image includes reduced corrupted phase information compared to the corrupted phase image of the received MR image, i.e., with reduced corrupted phase information. The method 250 also includes outputting 260 an image based on the original phase image.

[0060] The original phase image can be used to improve image quality. For example, a de-noised complex image can be generated by multiplying the initial magnitude image with the original phase image. A de-noised magnitude image can be generated based on the de-noised complex image. The de-noised complex image and / or the de-noised magnitude image can be input into downstream imaging processing and data processing.

[0061] In some embodiments, the original phase image is used for diffusion weighted imaging. For example, a de-noised complex image is derived by multiplying the initial complex image with the conjugate of the original phase image. Artifacts caused by the corrupted phase information in the de-noised complex image are reduced or removed. Furthermore, in diffusion weighted imaging, the SNR can be low, and the background phase prevents efficient averaging of multiple number of acquisitions (NEX). The original phase image can be used to further increase image quality by averaging de-noised complex images of multiple NEX to increase SNR.

[0062] In one embodiment, the original phase image is used for multi-pulse diffusion weighted imaging, where k-space of an imaging slice is read out with multiple RF excitation pulses. MR signals from multiple shots are combined, and a MR image is reconstructed based on the combined MR signals. The MR image is then input into the neural network model 204 and the original phase image is derived. The original phase image is used to generate de-noised complex images and / or magnitude images, which are then used to generate diffusion weighted images with improved image quality.

[0063] In another embodiment, the original phase image is used for imaging using a fast spin echo sequence such as PROPELLER (Periodically Rotated Overlapping Parallel Lines with Enhanced Reconstruction) where k-space is sampled in a radial fashion. In PROPELLER, each "blade" can have slightly different phase corruption due to system defects such as eddy currents. Data or images acquired by the PROPELLER sequence are received, and an original phase image for each blade is derived. De-noised complex images and / or magnitude images are generated based on the original phase images to reduce corrupted phase information before combining the blade images acquired with PROPELLER, thereby increasing image quality.

[0064] Diffusion weighted imaging and PROPELLER are used herein as examples only. The derived original phase image can be used to improve image quality of images acquired with overlapping acquisition by providing a phase image with reduced corrupted phase information before or after combining data acquired from overlapping acquisition. For example, the derived original phase image can be used for view sharing in cardiac cine imaging, where the derived original phase image is used to improve image quality of images based on shared views from adjacent segments.

[0065] In some embodiments, the raw phase image is used for phase sensitive inversion recovery (IR) imaging. In IR imaging, an IR RF pulse is applied, which is a 180° RF pulse and tilts the magnetization to the -z axis. The IR RF pulse can also be referred to as an IR preparation pulse. Then, the magnetization recovers through Tl relaxation between the application of the IR pulse and the excitation RF pulse. IR imaging can be used to suppress fat or dilute cerebrospinal fluid (CSF) by choosing the time between the IR pulse and the excitation RF pulse to be the time when the magnetization from fat or CSF is close to or equal to zero. IR imaging can also be used to increase tissue contrast when dynamic range doubling. However, the contrast between positive and negative magnetization is lost in magnitude images. Phase sensitive IR imaging uses the phase image to recover the phase difference between those voxels. The raw phase image generated using the methods and systems disclosed herein increases the accuracy of phase sensitive IR imaging by providing a phase image in which the background phase and / or noise is reduced.

[0066] In one embodiment, the raw phase image is used for Dixon chemical shift imaging (CSI). In Dixon CSI, a first image and a second image are acquired. The first MR image is based on a first MR signal that includes a first echo of a first signal and a second signal, and the second MR image is based on a second MR signal that includes a second echo of the first signal and the second signal. The first echo has a first phase between the first signal and the second signal, the second echo has a second phase between the first signal and the second signal, and the first phase and the second phase are different from each other. In some embodiments, an in-phase image and an out-of-phase image are acquired. That is, the first phase is zero and the second phase is 180°. In the in-phase image, the MR signal from water and fat is in-phase or in-phase echoes with the fat signal and the water signal. In the out-of-phase image, the MR signal from water and fat is out-of-phase or anti-phase echoes with the fat signal and the water signal. Then, a fat-only image or a water-only image is generated based on the sum and subtraction of the in-phase image and the out-of-phase image. However, the accuracy of the generated fat-only image or water-only image is deteriorated due to the local susceptibility effect. The raw phase image is used to correct the local susceptibility effect and correct the fat-only image and the water-only image.

[0067] In some embodiments, the raw phase images are used to generate phase sensitive maps, such as MR flow assessment, MR susceptibility imaging, tissue mechanical property maps in MR elastography, and temperature maps in MR thermometry, which can be used in MR guided focused ultrasound (MRgFUS). In MR flow assessment, because the phase shift is proportional to the velocity when a bipolar encoding gradient is applied, the phase information contained in the phase images is used to measure the flow. The raw phase images increase the accuracy of the flow assessment by providing phase images in which the background phase and / or noise is reduced. In MR elastography, shear waves are introduced into the body and a phase sensitive MR sequence is used to measure the mechanical properties of the tissue in response to the compression or vibration introduced by the shear waves. As in MR flow assessment, the phase shift accumulates in moving tissue and does not accumulate in stationary tissue. Thus, the phase images are used to measure the displacement and reconstruct the mechanical properties of the tissue. Similarly, the accuracy of the measurements is improved by using the raw phase images.

[0068] In MR susceptibility imaging, a quantitative susceptibility map is generated based at least in part on the phase images. Similarly, the generated susceptibility map is improved using the improved phase images of the raw phase images.

[0069] In MR thermometry, because temperature changes cause the resonance frequency of the nuclei to change, the phase shift from heating is used to measure the temperature or temperature changes in the tissue. Again, the raw phase images are used to improve the measured temperature or temperature changes by providing phase images in which the noise and / or background phase is reduced. The generated temperature maps can be used to guide the focused ultrasound treatment in MRgFUS.

[0070] In summary, the raw phase images denoised by deep learning denoising (DL denoising) can be used to increase the accuracy of the measurements or quantification of parameters that are sensitive to phase changes by using phase sensitive MR imaging. The raw phase images can also be used to improve the image quality in applications where inaccurate phase estimation affects the image quality and the assessment.

[0071] In some embodiments, the systems and methods are applied to multi-channel and / or multi-coil systems. For example, the phase images of each channel or each coil are denoised using DL denoising and the raw phase images are used to improve the image quality of the images and / or increase the accuracy of the measurements or quantification. The raw phase images can be used before or after combining the data from the multi-channel or multi-coil.

[0072] FIG. 3is a schematic diagram of an exemplary neural network model 204. The neural network model 204 can include a convolutional neural network 302. The neural network 302 accepts a MR phase image 304 that is contaminated with corrupted phase information, and returns an original phase image 306 in which the background phase 308 has been substantially de-noised. The input and output of the neural network 302 can also be complex images, magnitude and phase image pairs, or real or imaginary image pairs. The neural network 302 can be trained with a loss function that measures the inference error of the neural network 302. The loss function can be expressed as min(f1(input) - f2(output)), where f1and f2are functions of the input and output of the neural network 302, respectively. In some embodiments, the loss function includes a constraint based on prior knowledge of the phase information. For example, the prior knowledge is that phase wrapping or phase aliasing can be caused by the computational method of the phase. Based on the phase formula The computed phase is limited to the range of [-π, π]. However, in real life, the phase can be any real value. Therefore, adding multiple 2π phases appears as the same phase value, resulting in phase aliasing. In phase sensitive imaging applications where the phase is used to encode physical phenomena such as flow, motion, or temperature, the prior knowledge includes physical laws. Phases that exhibit violation of physical laws are penalized by the loss function. For example, in flow imaging or quantitative susceptibility mapping (QSM), the flow is divergence free. In thermal imaging, the thermal phase is expected to dissipate relatively uniformly. For displacement imaging such as elastography, the displacement is expected to obey physical laws of motion.

[0073] FIG. 4A through FIG. 4F A comparison of phase correction using DL de-noising versus de-noising by filtering radial phase body phantom is provided. FIG. 4A is a phase image 402 with a radial pattern. The phase image 402 corresponds to an imaging sequence of PROPELLER. FIG. 4B is a phase image 404 of the phase image 402 that is contaminated with corrupted phase information. FIG. 4C through FIG. 4D is an original phase image 406 using deep learning or neural network based systems and methods as disclosed herein, and a phase error 408 between the original phase image 406 and the true phase image 402. FIG. 4E through FIG. 4F are an original phase image 416, and a phase error 418 between the original phase image 416 and the true phase image 402, respectively. Comparing the original phase images 406 and 416, the neural network based de-noising demonstrates better high spatial frequency response and lower phase error in the low SNR background region 422 of the phantom than low pass filtering. High spatial frequency phase information 424 is lost or smoothed by filtering.

[0074] FIG. 5A through FIG. 5H Simulation results of a phase contrast detail phantom are shown. FIG. 5Ais the magnitude image 501 of the phantom without contamination. FIG. 5B is the noise magnitude image 503 of the phantom. FIG. 5C is the phase image 505 of the phantom without contamination. FIG. 5D is the corrupted phase image 507 of the phantom. As FIG. 5A and FIG. 5C indicate, the pixel SNR of the phantom increases from left to right ( FIG. 5A ) and the spatial frequency of the phase increases from top to bottom ( FIG. 5C ). FIG. 5E and FIG. 5F are the original phase image 509 and the phase error image 511 of the phantom using DL denoising. FIG. 5G and FIG. 5H are the original phase image 513 and the phase error image 515 of the phantom using low-pass filtering. Again, DL denoising demonstrates better performance in the phantom regions 517 with lower SNR and the phantom regions 519 with higher spatial frequency phase content.

[0075] FIG. 6A through FIG. 6E shows simulation results of a phantom with multiple repetitions where the background phase fluctuates randomly between repetitions, for example in diffusion weighted imaging using echo planar imaging (WI-EPI). Ten repetitions or 10 NEX are used as an example herein. FIG. 6A is the noise magnitude image with one acquisition. FIG. 6B shows the corrupted phase image with repetitions. FIG. 6C the ground truth phase image 601 (top row), the original phase image 603 using DL denoising (middle row), and the original phase image 605 using low-pass filtering (bottom row) are compared. FIG. 6D and FIG. 6E show the virtual image 609 after denoising using DL denoising ( FIG. 6D ) and the virtual image 611 after low-pass filtering ( FIG. 6E ). As FIG. 6C indicates, DL denoising has better performance than denoising by low-pass filtering, where the original phase image 603 by DL denoising is closer to the ground truth phase image 601 than the original phase image 605 by low-pass filtering, especially in the background region 607. These accurate phase images 603 are also reflected by the noisy virtual image 609 after correcting the original image with the DL denoising result phase and averaging the complex results.

[0076] FIG. 7A through FIG. 7F shows brain diffusion weighted images acquired with b = 300 s / mm 2 and NEX = 10. FIG. 7A is the real image 702 with DL denoising, FIG. 7B is the real image 704 with low-pass filtering.FIG. 7C and FIG. 7D are FIG. 7A and FIG. 7B The real images 702, 704 are shown with a reduced window level 706 to show the background signal. FIG. 7E and FIG. 7F The virtual images 708, 710 are shown after denoising with DL denoising FIG. 7E ) or low pass filtering FIG. 7F ) denoising. In the real images 702, a darker background is observed, indicating better phase correction for low SNR regions with DL denoising than with low pass filtering. In addition, the minimum signal in the virtual image 708 shows overall better phase correction with DL denoising than with low pass filtering.

[0077] FIG. 8A An exemplary artificial neural network model 204 is shown. The exemplary neural network model 204 includes neuron layers 502, 504-1 to 504-n, and 506, including an input layer 502, one or more hidden layers 504-1 to 504-n, and an output layer 506. Each layer can include any number of neurons, i.e., FIG. 8A q, r, and n in FIG. 8A It should be understood that the methods and systems described herein can be implemented using neural networks having different structures and configurations than those depicted in

[0078] In exemplary embodiments, the input layer 502 can receive different input data. For example, the input layer 502 includes a first input ai representing a training image, a second input a2 representing a pattern identified in the training image, a third input a3 representing an edge of the training image, and the like. The input layer 502 can include thousands or more inputs. In some embodiments, the number of elements used by the neural network model 204 changes during the training process, and if, for example, some neurons are determined to have less relevance during execution of the neural network, they are bypassed or ignored.

[0079] In exemplary embodiments, each neuron in the hidden layers 504-1 to 504-n processes one or more inputs from the input layer 502 and / or one or more outputs from neurons in one of the previous hidden layers to generate a decision or output. The output layer 506 includes one or more outputs, each output indicating a label, a confidence factor, a weight describing the input, and / or an output image. However, in some embodiments, the output of the neural network model 204 is obtained from the hidden layers 504-1 to 504-n in addition to or instead of the output from the output layer 506.

[0080] In some embodiments, each layer has a discrete, identifiable function with respect to the input data. For example, if n is equal to 3, the first layer analyzes the first dimension of the input, the second layer analyzes the second dimension of the input, and the last layer analyzes the third dimension of the input. The dimensions may correspond to aspects that are considered to be highly deterministic, then to those that are considered to be of medium importance, and finally to those that are considered to be less relevant.

[0081] In other embodiments, the layers are not clearly delineated in terms of the functionality they perform. For example, two or more of the hidden layers 504-1 through 504-n may share decisions related to labels, with no single layer making independent decisions regarding labels.

[0082] FIG. 8B Depicted is a diagram corresponding to the embodiment of FIG. 8A Example neuron 550 of a neuron labeled "1,1" in hidden layer 504-1 of FIG. For each input to neuron 550 (e.g., FIG. 8A The input layer 502 in the input) is weighted so that the input a1 to a p corresponds to weights w1 to w2 determined during the training process of the neural network model 204 p .

[0083] In some embodiments, some inputs lack explicit weights, or have weights below a threshold. The weights are applied to a function α (labeled by reference numeral 510), which may be a summation and may produce a value z1 that is input to a function labeled f 1,1 (z1) function 520. Function 520 is any suitable linear or nonlinear function. FIG. 5B As depicted, function 520 produces a plurality of outputs that may be provided to neurons in subsequent layers or used as outputs of neural network model 204. For example, the outputs may correspond to index values ​​into a list of labels, or may be calculated values ​​used as inputs to subsequent functions.

[0084] It should be understood that the structure and function of the depicted neural network model 204 and neurons 550 are for illustrative purposes only, and that other suitable configurations exist. For example, the output of any given neuron may depend not only on the values ​​determined by past neurons, but also on future neurons.

[0085] The neural network model 204 can include a convolutional neural network (CNN), a deep learning neural network, a reinforcement or enhanced learning module or program, or a combined learning module or program that learns in two or more areas or aspects of interest. Supervised and unsupervised machine learning techniques can be used. In supervised machine learning, a processing element can be provided with example inputs and their associated outputs, and can attempt to discover general rules that map the inputs to the outputs, such that when a subsequent new type of input is provided, the processing element can accurately predict the correct output based on the discovered rules. The neural network model 204 can be trained using unsupervised machine learning programs. In unsupervised machine learning, a processing element can need to find its own structure in unlabeled example inputs. Machine learning can involve recognizing and identifying patterns in existing data in order to facilitate predictions on subsequent data. Models can be created based on example inputs in order to make effective and reliable predictions on new inputs.

[0086] Additionally or alternatively, a machine learning program can be trained by inputting a sample data set or certain data, such as images, object statistics and information into the program. The machine learning program can use a deep learning algorithm, which can primarily focus on pattern recognition, and can be trained after processing a number of examples. The machine learning program can include, individually or in combination, Bayesian program learning (BPL), speech recognition and synthesis, image or object recognition, optical character recognition, and / or natural language processing. The machine learning program can also include natural language processing, semantic analysis, automated reasoning, and / or machine learning.

[0087] Based on these analyses, the neural network model 204 can learn how to recognize and then can apply to analyze characteristics and patterns of image data, model data, and / or other data. For example, the model 204 can learn to recognize features in a series of data points.

[0088] FIG. 9 is a block diagram of an example CNN 600 that can be included in the neural network model 204. The CNN 600 includes a convolutional layer 608. In a convolutional layer, convolution is used instead of general matrix multiplication in a neural network model. In one example, a 1x1 convolution is used to reduce the number of channels in the neural network 600. The neural network 600 includes one or more convolutional layer blocks 602, a fully connected layer 604 in which the neurons in the layer are connected to every neuron in the previous layer, and an output layer 606 that provides an output.

[0089] In an example implementation, the convolutional layer block 602 includes a convolutional layer 608 and a pooling layer 610. Each convolutional layer 608 is flexible in its depth, such as the number of convolutional filters and the size of the convolutional filters. The pooling layer 610 is used to simplify the underlying computations and reduce the dimensionality of the data by combining the outputs of a cluster of neurons at the previous layer into a single neuron in the pooling layer 610. The convolutional layer block 602 can also include a normalization layer 612 between the convolutional layer 608 and the pooling layer 610. The normalization layer 612 is used to normalize the distribution within a batch of training images and update the weights in the layer after normalization. The number of convolutional layer blocks 602 in the neural network 600 can depend on the image quality of the training images and the level of detail in the extracted features.

[0090] In operation, in training, a training image and other data, such as extracted features of the training image, are input into one or more convolutional layer blocks 602. A mask corresponding to the training image is provided as output of the output layer 606. The neural network 600 is adjusted during training. Once the neural network 600 is trained, an input image is provided to the one or more convolutional layer blocks 602, and the output layer 606 provides an output including a mask associated with the input image.

[0091] The workstations 12 described herein, as well as the phase image correction and noise reduction computing devices 202, 203, can be any suitable computing device 800 and software implemented therein. FIG. 10 is a block diagram of an example computing device 800. In an example implementation, the computing device 800 includes a user interface 804 that receives at least one input from a user. The user interface 804 can include a keyboard 806 that enables the user to input relevant information. The user interface 804 can also include, for example, a pointing device, a mouse, a stylus, a touch-sensitive panel (e.g., a touchpad and a touch screen), a gyroscope, an accelerometer, a position detector, and / or an audio input interface (e.g., including a microphone).

[0092] Further, in an example implementation, the computing device 800 includes a display interface 817 that presents information, such as input events and / or validation results, to a user. The display interface 817 can also include a display adapter 808 coupled to at least one display device 810. More particularly, in an example implementation, the display device 810 can be a visual display device, such as a cathode ray tube (CRT), a liquid crystal display (LCD), a light-emitting diode (LED) display, and / or an “electronic ink” display. Alternatively, the display interface 817 can include an audio output device (e.g., an audio adapter and / or a speaker) and / or a printer.

[0093] The computing device 800 also includes a processor 814 and a memory device 818. The processor 814 is coupled to the user interface 804, the display interface 817, and the memory device 818 via a system bus 820. In an example embodiment, the processor 814 communicates with a user, such as by prompting the user via the display interface 817 and / or by receiving user input via the user interface 804. The term “processor” generally refers to any programmable system including systems and microcontrollers, reduced instruction set computers (RISC), complex instruction set computers (CISC), application specific integrated circuits (ASICs), programmable logic circuits (PLCs), and any other circuit or processor that can execute the functions described herein. The above examples are exemplary only, and are thus not intended to limit in any way the definition and / or meaning of the term “processor.”

[0094] In this example embodiment, the memory device 818 includes one or more devices that enable information, such as executable instructions and / or other data, to be stored and retrieved. Additionally, the memory device 818 includes one or more computer-readable media, such as but not limited to dynamic random access memory (DRAM), static random access memory (SRAM), solid state disk, and / or a hard disk. In an example embodiment, the memory device 818 stores, without limitation, application source code, application object code, configuration data, additional input events, application state, assertion statements, validation results, and / or any other type of data. In an example embodiment, the computing device 800 can also include a communication interface 830 coupled to the processor 814 via the system bus 820. Additionally, the communication interface 830 is communicatively coupled to the data acquisition device.

[0095] In an example embodiment, the processor 814 can be programmed by encoding operations using one or more executable instructions and providing the executable instructions in the memory device 818. In an example embodiment, the processor 814 is programmed to select a plurality of measurements received from the data acquisition device.

[0096] In operation, a computer executes computer-executable instructions embodied within one or more computer-readable media to implement aspects of the application described and / or illustrated herein. The order of execution or performance of the operations in the embodiments of the application illustrated and described herein is not essential, unless otherwise specified. That is, it is contemplated that executing or performing a particular operation before, contemporaneously with, or after another operation is within the scope of aspects of the application. For example, it is contemplated that a particular operation within an embodiment of the application includes, without limitation, executable instructions in the computer-readable media executing consecutively, concurrently, or at different times.

[0097] At least one technical effect of the systems and methods described herein includes (a) reducing corrupted phase information in a phase image; (b) generating robust phase images; and (c) universally generating phase images without assuming phase and noise.

[0098] Exemplary embodiments of systems and methods of phase correction and noise reduction in a phase image are described in detail above. These systems and methods are not limited to the specific embodiments described herein, but rather, components of the systems and / or operations of the methods can be used independently and separately from other components and / or operations described herein. Further, the described components and / or operations can also be defined in, or used in, other systems, methods, and / or apparatus beyond the scope of the description. For example, the described systems and methods can be used in conjunction with other systems and methods described herein, and / or in other systems and methods not specifically described herein.

[0099] Although specific features of various embodiments of the application can be shown in some drawings and not in other drawings, this is for convenience only. In accordance with the principles of the application, any of the features of a drawing can be referenced and / or claimed in conjunction with any of the features of any other drawing.

[0100] This written description uses examples to disclose the application, including the best mode, and also to enable any person skilled in the art to practice the application, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the application is defined by the claims, and can include other examples that occur to those skilled in the art. Such other examples are intended to fall within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent

Claims

1. A computer-implemented method of correcting phase and reducing noise in a magnetic resonance (MR) phase image, comprising: executing a neural network model for analyzing MR images, wherein the neural network model is trained with a pair of a raw image and a corrupted image, wherein the corrupted image includes corrupted phase information, the raw image is the corrupted image with reduced corrupted phase information, and a target output image of the neural network model is the raw image; receiving an MR image, the MR image including corrupted phase information; analyzing the received MR image using the neural network model; deriving a raw phase image of the received MR image based on the analysis, wherein the derived raw phase image includes reduced corrupted phase information compared to the received MR image; and outputting an MR image based on the derived raw phase image; wherein: receiving an MR image further comprises receiving a first MR image and a second MR image, wherein the first MR image is based on a first MR signal including a first echo of a first signal and a second signal, and the second MR image is based on a second MR signal including a second echo of the first signal and the second signal, the first echo having a first phase between the first signal and the second signal, the second echo having a second phase between the first signal and the second signal, and the first phase and the second phase are different from each other; deriving a raw phase image further comprises deriving a first raw phase image of the first MR image and a second raw phase image of the second MR image using the neural network model; and the method further comprises generating an image of only the first signal and an image of only the second signal based at least in part on the first raw phase image and the second raw phase image.

2. The method of claim 1, wherein the corrupted phase information includes noise in a phase image, and the derived raw phase image includes reduced noise in a phase image compared to the received MR image.

3. The method of claim 1, wherein the neural network model is trained with a loss function, the loss function including a constraint based on prior knowledge of phase information.

4. The method of claim 1, the method further comprising: generating a phase sensitive map based at least in part on the raw phase image.

5. The method of claim 4, wherein generating a phase sensitive map further comprises: generating a quantitative susceptibility map based at least in part on the raw phase image.

6. The method of claim 4, wherein generating a phase sensitive map further comprises: deriving a flow velocity based on the raw phase image.

7. The method of claim 4, wherein generating a phase sensitive map further comprises: generating a map of tissue mechanical properties based at least in part on the raw phase image.

8. The method of claim 4, wherein generating a phase sensitive map further comprises: generating a temperature map based at least in part on the raw phase image.

9. The method of claim 1, wherein the first echo is an in-phase echo of the first signal and the second signal, and the second echo is an anti-phase echo of the first signal and the second signal.

10. The method of claim 1, wherein the first signal comprises a water signal, and the second signal comprises a fat signal.

11. The method of claim 1, wherein the received MR image comprises an image acquired with overlapping acquisition.

12. The method of claim 11, wherein the received MR image comprises an image acquired with multiple shots.

13. The method of claim 1, wherein the received MR image is a diffusion weighted image.

14. The method of claim 1, wherein the received MR image comprises an image acquired with multiple shots.

15. The method of claim 1, wherein the received MR image comprises an image acquired with inversion recovery preparation pulses.

16. A phase image correction and noise reduction system comprising a phase image correction and noise reduction computing device comprising at least one processor in communication with at least one memory device, and the at least one processor is programmed to: execute a neural network model for analyzing MR images, wherein the neural network model is trained with a pair of an original image and a corrupted image, wherein the corrupted image comprises corrupted phase information, the original image is the corrupted image with the corrupted phase information reduced, and a target output image of the neural network model is the original image; receive an MR image, the MR image comprising corrupted phase information; analyze the received MR image using the neural network model; derive an original phase image of the received MR image based on the analysis, wherein the original phase image comprises reduced corrupted phase information compared to the received MR image; and output an MR image based on the original phase image; wherein: receiving an MR image further comprises receiving a first MR image and a second MR image, wherein the first MR image is based on a first MR signal comprising a first echo of a first signal and a second signal, and the second MR image is based on a second MR signal comprising a second echo of the first signal and the second signal, the first echo having a first phase between the first signal and the second signal, the second echo having a second phase between the first signal and the second signal, and the first phase and the second phase are different from each other; deriving an original phase image further comprises deriving a first original phase image of the first MR image and a second original phase image of the second MR image using the neural network model; and the at least one processor is further programmed to generate an image of only the first signal and an image of only the second signal based at least in part on the first original phase image and the second original phase image.

17. The system of claim 16, the at least one processor is further programmed to: generating a phase sensitive map based at least in part on the original phase image.

18. The system of claim 16, wherein the first echo is an in-phase echo of the first signal and the second signal, the second echo is an anti-phase echo of the first signal and the second signal, the first signal comprises a water signal, and the second signal comprises a fat signal.

Images