Systems and methods for noise reduction in magnetic resonance images

Abstract

A computer-implemented method for reducing noise in magnetic resonance (MR) images is provided. The method includes performing a neural network model to analyze the MR images, wherein the neural network model is trained with a pair of original images and damaged images. The original image is the noise-reduced damaged image, and the target output image of the neural network model is the original image. The method further includes: receiving a first MR signal and a second MR signal; reconstructing the first MR image and the second MR image based on the first MR signal and the second MR signal; and analyzing the first MR image and the second MR image using the neural network model. The method further includes: deriving a denoised MR image based on the analysis, wherein the denoised MR image is a combined image based on the first MR image and the second MR image; and outputting the denoised MR image.

Description

Background Technology

[0001] The field of this disclosure relates generally to systems and methods for image processing, and more particularly to systems and methods for noise reduction in medical images using neural network models.

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

[0003] Noise and artifacts in MR images affect image quality and can therefore interfere with diagnosis. Known methods are disadvantageous in some respects and require improvement. Summary of the Invention

[0004] In one aspect, a computer-implemented method for reducing noise in magnetic resonance (MR) images is provided. The method includes performing a neural network model to analyze the MR images, wherein the neural network model is trained with a pair of original images and damaged images, wherein the original images are the noise-reduced damaged images, and the target output image of the neural network model is the original images. The method further includes: receiving a first MR signal and a second MR signal; reconstructing the first MR image and the second MR image based on the first MR signal and the second MR signal; and analyzing the first MR image and the second MR image using the neural network model. The method further includes: deriving a denoised MR image based on the analysis, wherein the denoised MR image is a combined image based on the first MR image and the second MR image; and outputting the denoised MR image.

[0005] In another aspect, an MR noise reduction system is provided. The system includes a noise reduction computing device comprising at least one processor communicating 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 original images and damaged images, wherein the original images are the noise-reduced damaged images, and the target output image of the neural network model is the original images. The at least one processor is also programmed to receive a first MR signal and a second MR signal, reconstruct the first MR image and the second MR image based on the first MR signal and the second MR signal, and analyze the first MR image and the second MR image using the neural network model. The at least one processor is further programmed to derive a denoised MR image based on the analysis, wherein the denoised MR image is a combined image based on the first MR image and the second MR image, and output the denoised MR image. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of an exemplary magnetic resonance imaging (MRI) system.

[0007] Figure 2A This is an example noise reduction system.

[0008] Figure 2B This is a flowchart of an exemplary method for reducing noise.

[0009] Figure 3A Images are not created using the systems and methods described herein.

[0010] Figure 3B Is with Figure 3A Images of the same anatomical structures located in the same locations, wherein the images are generated using the exemplary methods described herein.

[0011] Figure 4A This is a schematic diagram of a neural network model.

[0012] Figure 4B yes Figure 4A A schematic diagram of neurons in the neural network model shown.

[0013] Figure 5 This is a block diagram of an exemplary computing device.

[0014] Figure 6 This is a block diagram of an exemplary server computing device. Detailed Implementation

[0015] This disclosure includes systems and methods for reducing noise in magnetic resonance (MR) images of a subject. As used herein, the subject is a human, animal, or phantom, or part of a human, animal, or phantom, such as an organ or tissue. Noise, artifacts, other unwanted signals, or any combination thereof are collectively referred to as noise. Artifacts may be caused by eddies, physiological noise, B0 drift, B0 and / or B1 inhomogeneities, or other system defects. The original image is a noise-reduced image compared to the damaged image. Reducing or removing noise is collectively referred to as noise reduction or denoising. A denoised image is a noise-reduced image. Method aspects will be apparent in part and will be explicitly discussed in part in the following description.

[0016] In magnetic resonance imaging (MRI), the subject is placed in a magnet. When the subject is in a magnetic field generated by the magnet, the magnetic moments of nuclei, such as protons, attempt to align with the magnetic field, but precess around the field in a random order at the Larmor frequency of the nucleus. The magnetic field of the magnet is called B0 and extends longitudinally, or in the z-direction. During the acquisition of MRI images, a magnetic field in the xy-plane and close to the Larmor frequency (called the excitation field B1) is generated by a radio frequency (RF) coil and can be used to rotate or "tilt" the net magnetic moment Mz of the nucleus from the z-direction toward the transverse or xy-plane. After the excitation signal B1 terminates, the nucleus emits a signal, which is called the MR signal. To generate an image of the subject using the MR signal, magnetic field gradient pulses (Gx, Gy, and Gz) are used. The gradient pulses are used to scan the reverse direction of space or distance through k-space, spatial frequency, and 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.

[0017] Figure 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 scanning protocols to be input 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 storage server 26. The workstation 12 and each of the servers 20, 22, 24, and 26 communicate with each other.

[0018] In an exemplary embodiment, the pulse sequence server 20 operates the gradient system 28 and the radio frequency (“RF”) system 30 in response to instructions downloaded from workstation 12. The instructions are used to generate gradient waveforms and RF waveforms in the MR pulse sequence. The RF coil 38 and the gradient coil assembly 32 are used to execute the prescribed 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 anatomical structure to be imaged, or a coil array comprising multiple coils.

[0019] In an exemplary embodiment, a gradient waveform for performing a delimited scan is generated and applied to a gradient system 28, which excites gradient coils in gradient coil assembly 32 to generate a magnetic field gradient G for position encoding of the MR signal. x G y and G z The gradient coil assembly 32 forms part of the magnet assembly 34, which also includes a polarized magnet 36 and an RF coil 38.

[0020] In an exemplary embodiment, RF system 30 includes an RF transmitter for generating RF pulses used in an MR pulse sequence. The RF transmitter responds to a scan scheme and orientation from pulse sequence server 20 to generate RF pulses with desired frequency, phase, and pulse amplitude waveforms. The generated RF pulses can be applied to RF coil 38 via RF system 30. The responsive MR signal detected by RF coil 38 is received by RF system 30 and amplified, demodulated, filtered, and digitized under the direction of commands generated by pulse sequence server 20. RF coil 38 is described as both a transmitter and a receiver coil, such that RF coil 38 transmits RF pulses and detects MR signals. In one embodiment, MRI system 10 may include a transmitter RF coil for transmitting RF pulses and a separate receiver coil for detecting MR signals. A transmission channel of RF system 30 may be connected to the RF transmitting coil, and a receiver channel may be connected to a separate RF receiver coil. Typically, the transmission channel is connected to the whole-body RF coil 38, and each receiver segment is connected to a separate local RF coil.

[0021] In an exemplary embodiment, the RF system 30 further includes one or more RF receiver channels. Each RF receiver channel includes an RF amplifier that amplifies the MR signal received by the RF coil 38 to which the channel is connected; and a detector that detects and digitizes the I quadrature components and Q quadrature components of the received MR signal. The magnitude of the received MR signal 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:

[0022]

[0023] Furthermore, the phase of the received MR signal can also be determined as shown in equation (2) below:

[0024]

[0025] In an exemplary embodiment, digitized MR signal samples generated by RF system 30 are received by data acquisition server 22. Data acquisition server 22 can operate in response to instructions downloaded from workstation 12 to receive real-time MR data and provide buffer storage so that no data is lost due to data overflow. In some scans, data acquisition server 22 simply transmits the acquired MR data to data processing server 24. However, in scans where information derived from the acquired MR data is needed to control further execution of the scan, data acquisition server 22 is programmed to generate the required information and transmit it to pulse sequence server 20. For example, during a pre-scan, MR data is acquired and used to calibrate a pulse sequence performed by pulse sequence server 20. Additionally, navigator signals can be acquired during the scan and used to adjust operating parameters of RF system 30 or gradient system 28, or to control the view sequence for sampling k-space.

[0026] 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 a Fourier transform on the raw k-space MR data to generate a two-dimensional or three-dimensional image, applying filters to the reconstructed image, performing back-projection image reconstruction on the acquired MR data, generating a functional MR image, and calculating a motion or flow image.

[0027] In an exemplary embodiment, the image reconstructed by the data processing server 24 is transmitted back to workstation 12 and stored there. In some embodiments, the real-time image is stored in a database storage cache. Figure 1 (Not shown in the image) Real-time images can be output from the database storage cache to the operator's display 14 or a display 46 located near the magnet assembly 34 for use by the attending physician. Batch-processed images or selected real-time images can be stored on disk storage 48 or a host database in 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 workstation 12 to archive images, generate films, or send images to other facilities via a network.

[0028] In known methods using neural network models, the neural network model is typically trained to perform a target operation or produce a target result. For example, for Dixon fat suppression, two or more images are used to generate a water-only image where the fat signal is suppressed. In other known methods using neural network models, during training, two or more images from the training dataset are provided as input to the neural network model, and the corresponding water-only image is provided as the output of the neural network model. That is, the neural network model is specifically trained for a specific task involving combining two or more images. In other known methods, the trained neural network is typically used to analyze the same original images as those in the training dataset, rather than to analyze derived images of the original images. For example, a neural network model is trained using noisy MR images and noise-reduced MR images, and the trained neural network model is used to analyze the original MR images acquired from an MR scanner to reduce noise in the images, rather than to analyze derived images of the original MR images.

[0029] In contrast, in the systems and methods disclosed herein, the neural network model can be a pre-existing, trained model and is not specifically trained using the dataset used or the operations performed during analysis using the neural network model. As used herein, the dataset includes images, MR signals, or any combination thereof. For example, the neural network model is trained to reduce noise in an image. The training dataset can differ from the dataset used for inference. The input image during inference can be a derived image or a processed image of the original image. The original images can have different noise distributions. Furthermore, the neural network model is not trained to perform operations for generating the derived image. Additionally, the neural network model can take complex signals or images as input, thereby improving the accuracy of inference.

[0030] Figure 2AThis is a schematic diagram of an exemplary noise reduction system 200. In an exemplary embodiment, system 200 includes a noise reduction computing device 202-1 configured to reduce noise in MR images. Computing device 202-1 also includes a neural network model 204. System 200 may include a second noise reduction computing device 202-2. The second noise reduction computing device 202-2 can be used to train the neural network model 204, and then the noise reduction computing device 202-1 can use the trained neural network model 204. The second noise reduction computing device 202-2 may be the same computing device as noise reduction computing device 202-1, such that the training and use of the neural network model 204 are performed on a single computing device. Alternatively, the second noise reduction computing device 202-2 may be a separate computing device from noise reduction computing device 202-1, such that the training and use of the neural network model 204 are performed on a separate computing device. Noise reduction computing device 202 may be included in workstation 12 of MRI system 10, or may be included on a separate computing device in communication with workstation 12. In one example, the noise reduction computing device 202 is a server computing device and may be cloud-based.

[0031] Figure 2B This is a flowchart of an exemplary method 250. Method 250 may be implemented on a noise reduction system 200. In an exemplary embodiment, method 250 includes performing (252) analysis of MR images by a neural network model. The neural network model is trained with a training dataset. The training dataset may be training images, which may be a pair of original images and damaged images, and the target output image of the neural network model is the original image. The original image is a noise-reduced damaged image. The training images are represented in any format, such as composite MR images, phase images, real and imaginary pairs, images in phasor representation, or magnitude images. In some embodiments, the damaged image is a simulated image contaminated with artifacts and / or noise, and the original image is a simulated phase image without contamination. The training image may be an image contaminated with simulated noise from a heterogeneous dataset, which may come from different applications and / or be acquired with different pulse sequences and / or different pulse sequence parameters. The trained neural network model 204 is configured to reduce noise in the input image and output a noise-reduced image.

[0032] Method 250 further includes receiving (254) a first MR signal and a second MR signal. A first MR image and a second MR image are reconstructed (256) based on the first MR signal and the second MR signal. A neural network model is used to analyze (258) the first MR image and the second MR image. A denoised MR image is derived (260) based on the analysis. The denoised MR image is a combined image based on the first MR image and the second MR image. The denoised MR image is output (262).

[0033] In an exemplary implementation, method 250 is used for Dixon chemical shift imaging (CSI). In some applications, fat suppression is required, where signals from fat are suppressed. Fat suppression can be achieved by applying a fat saturation pulse before applying an excitation pulse. Fat suppression using a fat saturation pulse may fail due to B0 inhomogeneity. Using an inversion recovery pulse is another method of fat suppression, where an inversion recovery pulse is applied and an image is acquired when the fat signal is zero during inversion recovery. Compared to the fat saturation method, the inversion recovery method is relatively insensitive to B0 inhomogeneity but sensitive to B1 inhomogeneity, resulting in a low signal-to-noise ratio (SNR) and potentially undesirable contrast variations. In regions with high B0 inhomogeneity, such as brachial plexus or eccentric slices, the fat saturation method or even the inversion recovery method may fail. Dixon CSI is advantageous in suppressing fat signals because the MR signal is acquired without saturation or being zeroed, and therefore the method is insensitive to either B0 or ​​B1 inhomogeneity and has a relatively high SNR.

[0034] In Dixon CSI, a first image and a second image are acquired. The first MR image is based on a first MR signal comprising a first echo of a first chemical substance and a second chemical substance, and the second MR image is based on a second MR signal comprising a second echo of a first chemical substance and a second chemical substance. In some embodiments, the first chemical substance is water, and the second chemical substance is fat. The first and second echoes have different echo times from each other. Echo time is the time between the excitation RF pulse and the echo during which the MR signal is acquired. In an MR system, the protons of molecules have different resonance frequencies, depending on the chemical structure of the molecules. Different echo times provide different phases for signals from different molecules, such as water and fat. Therefore, the first echo has a first phase difference between the first and second chemical substances, and the second echo has a second phase difference between the first and second chemical substances. The first phase difference and the second phase difference are different from each other due to the different echo times. In some embodiments, in-phase and out-of-phase images are acquired. That is, the first phase difference is zero, while the second phase difference is 180°. For example, in an in-phase image, the MR signals from water and fat are in-phase, or the acquired MR signals are in-phase echoes of fat and water. In heterophase images, the MR signals from water and fat are out of phase, or the acquired MR signals are out-of-phase echoes of fat and water. Then, a fat-only image or a water-only image is generated based on the summation and subtraction of the in-phase and heterophase images. In some implementations, more than two MR signal acquisitions are used. For example, more than two sets of MR signals are acquired, each acquired at a different echo time.

[0035] In some implementations, the system and method are applied to multi-channel and / or multi-coil systems. For example, the image of each channel or each coil is denoised using deep learning (DL) denoising, and the denoised image is used to improve the image quality and / or increase the accuracy of measurement or quantization. The neural network model 204 can be used before or after combining data from the multi-channel or multi-coil system.

[0036] In some implementations, the neural network model 204 may be applied to the first image and the second image before combination. This application can be separate for the first and second images. For example, the first image is input into the neural network model 204 to derive a denoised first image. The second image is input into the neural network model 204 to derive a denoised second image. Alternatively, the application can be combined for the first and second images. For example, the first and second images are jointly input into the neural network model 204, and the denoised first and second images are output from the neural network model. Joint input into the neural network model 204 can increase the accuracy of inference because more information is provided to the neural network model 204. After analysis by the neural network model 204, the denoised first image and the denoised second image are combined to derive one or more denoised final images.

[0037] A combined image can be generated via a linear combination of a first image and a second image. Alternatively, a combined image can be generated via a nonlinear combination of a first image and a second image. For example, in diffusion-weighted imaging, the first image and the second image are diffusion-weighted images with different diffusion weights, and the combined image can be a graph of the apparent diffusion coefficient or diffusion tensor, which is a nonlinear combination of the diffusion-weighted images. In another example, in multi-point Dixon CSI, images with different phase differences between a first chemical substance and a second chemical substance are acquired, and the combined image is a nonlinear combination of the acquired images.

[0038] In other embodiments, the neural network model 204 can be applied to a combined image of the first and second images. The combined image is input into the neural network model 204. If more than one combined image is generated after combination, the combined images can be input into the neural network model individually or jointly. For example, the first and second combined images are input into the neural network model 204 separately, and the neural network model 204 outputs a denoised first combined image and a denoised second combined image, respectively. Alternatively, the first and second combined images are input into the neural network model 204 jointly, and the neural network model 204 outputs a denoised first combined image and a denoised second combined image. Joint input can increase the accuracy of inference.

[0039] Figures 3A to 3BImages 302-noDL and 302-DL are shown with and without method 250 applied. Image 302-DL is an image processed using the system and method described herein, while image 302-noDL is not processed using the system and method described herein. Image 302 is a sagittal image of the spinal cord. Specifically, image 302 is a water image using Dixon CSI, where the fat signal is suppressed. To generate the water image, two or more sets of MR signals are acquired, each acquired at a different echo time (TE) such that the signals from fat and water are in different phases from each other. For example, two sets of MR signals are acquired, the first MR signal being in phase with each other from water and fat, and the second MR signal being out of phase with each other or having a 180° phase difference from fat and water. The in-phase image is reconstructed from the first MR signal. The out-of-phase image is reconstructed from the second MR signal. Because the signals from water and fat are in phase with each other in the in-phase image and have a 180° phase difference in the out-of-phase image, water-only and / or fat-only images can be derived by a linear combination such as the addition or subtraction of in-phase and out-of-phase images. The water-only image can be referred to as the water image. The fat-only image can be referred to as the fat image. Image 302-noDL is derived by adding the in-phase and out-of-phase images. Image 302-DL is derived by applying neural network model 204. In deriving image 302-DL, neural network model 204 is applied to both the in-phase and out-of-phase images. Neural network model 204 outputs a denoised in-phase image and a denoised out-of-phase image. The water image 302-DL is derived by adding the denoised in-phase and out-of-phase images. A denoised fat-only image (not shown) can be generated by subtracting the denoised in-phase image from the denoised out-of-phase image.

[0040] MR is unique compared to other image modalities because MR signals and MR images are represented by complex numbers at data points or pixels. In an exemplary embodiment, the input and output of the neural network model are composite images. Composite MR images can be reconstructed based on I-orthogonal MR signals and Q-orthogonal MR signals using procedures such as Fourier transform. Composite MR images are used herein. It is an MR image in which each pixel is represented by a complex number, and the pixel has a real component. and imaginary components Where (x, y) are the pixel positions in the image. The magnitudes of the composite MR image (called the magnitude image) are generated as follows: It can also generate and use phase images and the phase of composite MR images, where the phase... A real image is a representation of the value at each pixel as a real component. The MR image. The virtual image is where the value at each pixel is used as a virtual component. MR images.

[0041] In an exemplary embodiment, in order to analyze a water image using a neural network model 204, the water image alone includes phase information and is provided as a composite MR image, a pair of magnitude and phase images, a real pair and a virtual pair, or an image in phasor representation. Although the final displayed image is a magnitude image, the phase of the water image alone is added back and provided to the neural network model 204.

[0042] The application of the system and method described herein to two-point Dixon CSI is described as an example for illustrative purposes only. The system and method can be applied to multi-point Dixon CSI. In three- or more-point Dixon CSI, three or more sets of MR signals are acquired, each with a different TE, such that the signals from water and fat have different phase differences for each set. Water-only images and fat-only images are generated based on the multiple sets of MR signals. The neural network model 204 can be applied before or after the combination.

[0043] In Dixon CSI, in-phase and out-of-phase images exhibit different noise distributions. For example, the SNR of the fat region differs between in-phase and out-of-phase images. Compared to out-of-phase images, the fat signal is relatively high in in-phase images because fat and water are in phase, resulting in a higher SNR in the fat region and a lower SNR in the spinal cord region. Neural network model 204 is configured to remove noise while preserving structure. As a result, after analysis of the in-phase image by neural network model 204, more noise is removed in the spinal cord region compared to the fat region, and the denoised in-phase image retains the noise in the fat region. On the other hand, in the out-of-phase image, because fat is out-of-phase with water, the fat region also has a low SNR similar to that of the spinal cord region, thus reducing the noise in the fat region in the denoised out-of-phase image after analysis by neural network model 204. Therefore, in the water image 302-DL, which is a combination of a denoised in-phase image and a denoised out-of-phase image, noise is preserved in the fat region, leading to artifacts.

[0044] Artifacts can be reduced by combining the first and second images before analysis by the neural network model 204. For example, acquired in-phase and out-of-phase images are combined to generate a water-only image and a fat-only image, and the water-only image and fat-only image are analyzed individually or jointly by the neural network model 204 to derive a denoised water-only image and / or a denoised fat-only image. The neural network model 204 performs non-linear operations on the input images, and noise may not propagate uniformly across the images during the analysis by the neural network model 204. Therefore, when the original images are combined before analysis by the neural network model 204, the neural network model will not affect the noise in the original images differently, thereby reducing artifacts.

[0045] The 2D images are used for illustrative purposes only. The systems and methods described herein can be applied to 3DMR datasets. A 2D neural network model 204 can be used to analyze 3D images. For example, a 2D image is provided as input to the 2D neural network model 204, and the model outputs a 2D image. To process 3D images using the 2D neural network model 204, multiple 2D images can be organized in the z or kz direction, and these 2D images are input into the 2D neural network model 204 for analysis. The output is organized into a 3D image. Alternatively, a 3D neural network model 204 that takes a 3D image as input and outputs a 3D image can be used to analyze 3D images.

[0046] As described above, MR images can be reconstructed using Fourier transform. Because Fourier transform is a linear operation, in some implementations, the input to neural network model 204 and / or the output from the neural network model are MR signals, MR images, or any combination thereof.

[0047] Figure 4A An exemplary artificial neural network model 204 is illustrated. 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 may include any number of neurons, i.e., Figure 4A In this context, q, r, and n can be any positive integer. It should be understood that they can be used with... Figure 4A The structures and configurations shown in the paper are different from those of neural networks to implement the methods and systems described herein.

[0048] In an exemplary embodiment, input layer 502 may receive different input data. For example, input layer 502 may include a first input a1 representing a training image, a second input a2 representing a pattern recognized in the training image, a third input a3 representing the edges of the training image, and so on. Input layer 502 may include thousands or more inputs. In some embodiments, the number of elements used by neural network model 204 changes during the training process, and if, for example, some neurons are determined to be less relevant during the execution of the neural network, these neurons are bypassed or ignored.

[0049] In an exemplary embodiment, each neuron in hidden layers 504-1 to 504-n processes one or more inputs from input layer 502 and / or one or more outputs from neurons in one of the previous hidden layers to generate a decision or output. Output layer 506 includes one or more outputs, each output indicating a label, confidence factor, weights describing the input, and / or an output image. However, in some embodiments, the outputs of neural network model 204 are obtained from hidden layers 504-1 to 504-n in addition to or in lieu of the outputs from output layer 506.

[0050] In some implementations, each layer has a discrete, identifiable function relative to the input data. For example, if n equals 3, the first layer analyzes the first dimension of the input, the second layer analyzes the second dimension, and the last layer analyzes the third dimension. Dimensions may correspond to aspects considered strongly deterministic, then to those considered moderately important, and finally to those considered less relevant.

[0051] In other implementations, these layers are not clearly described in terms of their functionality. For example, two or more hidden layers 504-1 to 504-n may share decisions related to the tag, where no single layer makes an independent decision about the tag.

[0052] Figure 4B The corresponding implementation according to one embodiment is shown in Figure 4A An exemplary neuron 550, labeled "1,1", is shown in hidden layer 504-1. For each input to neuron 550 (e.g., ...), Figure 4A The inputs in the input layer 502 are weighted so that the inputs a1 to a... p The weights w1 to w2, which correspond to the weights determined during the training process of neural network model 204, are... p .

[0053] In some implementations, some inputs lack explicit weights or have weights below a threshold. Weights are applied to a function α (labeled by reference numeral 510 in the attached diagram), which can be a summation function and produces a value z1, which is input to a function labeled f. 1,1 The function 520 of (z1). Function 520 can be any suitable linear or nonlinear function. Figure 4B As shown, function 520 produces multiple outputs, which can be provided to neurons in subsequent layers or used as outputs of neural network model 204. For example, the outputs can correspond to index values ​​of a list of labels, or they can be computed values ​​used as inputs to subsequent functions.

[0054] It should be understood that the structure and function of the depicted neural network model 204 and neuron 550 are for illustrative purposes only and 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.

[0055] The neural network model 204 may include a convolutional neural network (CNN), a deep learning neural network, a reinforcement or enhancement learning module or procedure, or a combined learning module or procedure that learns in two or more domains or aspects of interest. Supervised and unsupervised machine learning techniques can be used. In supervised machine learning, the processing element may be given exemplary inputs and their associated outputs, and may attempt to discover general rules that map the inputs to the outputs, such that when subsequent novel inputs are 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 procedures. In unsupervised machine learning, the processing element may need to find its own structure in unlabeled exemplary inputs. Machine learning may involve identifying and recognizing patterns in existing data to facilitate prediction of subsequent data. Models can be created based on exemplary inputs to make effective and reliable predictions on novel inputs.

[0056] Alternatively or in addition to this, the machine learning program can be trained by inputting a sample dataset or some data such as images, object statistics, and information into the program. The machine learning program may use deep learning algorithms, which may focus primarily on pattern recognition, and may be trained after processing multiple examples. The machine learning program may 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 may also include natural language processing, semantic analysis, automated reasoning, and / or machine learning.

[0057] Based on these analyses, neural network model 204 can learn how to identify characteristics and patterns that can then be applied to analyze image data, model data, and / or other data. For example, model 204 can learn to identify features in a series of data points.

[0058] The workstation 12 and noise-reducing computing device 202 described herein can be any suitable computing device 800 and the software implemented therein. Figure 5This is a block diagram of an exemplary computing device 800. In this exemplary embodiment, the computing device 800 includes a user interface 804 that receives at least one input from a user. The user interface 804 may include a keyboard 806 that enables the user to input relevant information. The user interface 804 may also include, for example, a pointing device, a mouse, a stylus, a touch-sensitive panel (e.g., a touchpad and a touchscreen), a gyroscope, an accelerometer, a position detector, and / or an audio input interface (e.g., including a microphone).

[0059] Furthermore, in an exemplary embodiment, computing device 800 includes a presentation interface 817 for presenting information (such as input events and / or verification results) to a user. Presentation interface 817 may also include a display adapter 808 coupled to at least one display device 810. More specifically, in this exemplary embodiment, display device 810 may be a visual display device, such as a cathode ray tube (CRT), liquid crystal display (LCD), light-emitting diode (LED) display, and / or "electronic ink" display. Alternatively, presentation interface 817 may include an audio output device (e.g., an audio adapter and / or a speaker) and / or a printer.

[0060] The computing device 800 also includes a processor 814 and a memory device 818. The processor 814 is coupled to a user interface 804, a presentation interface 817, and the memory device 818 via a system bus 820. In this exemplary embodiment, the processor 814 communicates with a user, such as by prompting the user via the presentation 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 (RISCs), complex instruction set computers (CISCs), application-specific integrated circuits (ASICs), programmable logic circuits (PLCs), and any other circuitry or processor capable of performing the functions described herein. The examples above are merely exemplary and are therefore not intended to limit the definition and / or meaning of the term "processor" in any way.

[0061] In this exemplary embodiment, memory device 818 includes one or more devices that enable information such as executable instructions and / or other data to be stored and retrieved. Furthermore, 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 drives, and / or hard disks. In this exemplary embodiment, memory device 818 stores, but is not limited to, application source code, application object code, configuration data, additional input events, application state, assertion statements, verification results, and / or any other type of data. In this exemplary embodiment, computing device 800 may also include a communication interface 830 coupled to processor 814 via system bus 820. Furthermore, communication interface 830 is communicatively coupled to a data acquisition device.

[0062] In this exemplary embodiment, the processor 814 can be programmed by encoding operations using one or more executable instructions and by providing the executable instructions in the memory device 818. In this exemplary embodiment, the processor 814 is programmed to select multiple measurements received from the data acquisition device.

[0063] In operation, a computer executes computer-executable instructions embodied in one or more computer-executable components stored on one or more computer-readable media to implement aspects of the invention described and / or illustrated herein. Unless otherwise specified, the order of execution or implementation of the operations in the embodiments of the invention shown and described herein is not required. That is, unless otherwise specified, these operations can be performed in any order, and embodiments of the invention may include more or fewer operations than those disclosed herein. For example, it is contemplated that a particular operation may be performed or implemented before, simultaneously with, or after another operation within the scope of various aspects of the invention.

[0064] Figure 6 An exemplary configuration of server computer device 1001, such as computing device 202, is shown. Server computer device 1001 also includes a processor 1005 for executing instructions. For example, instructions may be stored in memory region 1030. Processor 1005 may include one or more processing units (e.g., in a multi-core configuration).

[0065] The processor 1005 is operatively coupled to the communication interface 1015, enabling the server computer device 1001 to communicate with a remote device or another server computer device 1001. For example, the communication interface 1015 can receive data from the workstation 12 via the Internet.

[0066] Processor 1005 may also be operatively coupled to storage device 1034. Storage device 1034 is hardware suitable for any computer operation of storing and / or retrieving data (such as, but not limited to, wavelength variations, temperature, and strain). In some embodiments, storage device 1034 is integrated into server computer device 1001. For example, server computer device 1001 may include one or more hard disk drives as storage device 1034. In other embodiments, storage device 1034 is external to server computer device 1001 and is accessible by multiple server computer devices 1001. For example, storage device 1034 may include multiple storage units, such as hard disks and / or solid-state drives in a Redundant Array of Inexpensive Disks (RAID) configuration. Storage device 1034 may include storage area network (SAN) and / or network attached storage (NAS) systems.

[0067] In some implementations, processor 1005 is operatively coupled to storage device 1034 via storage interface 1020. Storage interface 1020 is any component capable of providing processor 1005 with access to storage device 1034. Storage interface 1020 may include, for example, an Advanced Technology Attachment (ATA) adapter, a Serial ATA (SATA) adapter, a Small Computer System Interface (SCSI) adapter, a RAID controller, a SAN adapter, a network adapter, and / or any component that provides processor 1005 with access to storage device 1034.

[0068] At least one technical effect of the system and method described herein includes: (a) reducing noise in the derived MR images; (b) using datasets of neural network models with different datasets during training and inference; and (c) reducing noise in fat-suppressed Dixon images.

[0069] Exemplary embodiments of noise reduction systems and methods have been described in detail above. These systems and methods are not limited to the specific embodiments described herein, but rather the components of the systems and / or the operation of the methods can be used independently and separately from other components and / or operations described herein. Furthermore, the described components and / or operations may also be defined in other systems, methods, and / or devices, or used in combination with other systems, methods, and / or devices, and are not limited to practice using only the systems described herein.

[0070] Although certain features of various embodiments of the invention may be shown in some figures but not others, this is only for convenience. Any feature of the figures may be referenced and / or claimed in conjunction with any feature of any other figure according to the principles of the invention.

[0071] This written description uses examples to disclose the invention, including the best mode, and also enables those skilled in the art to practice the invention, including making and using any device or system and performing any included methods. The scope of the invention is defined by the claims and may include other examples that would 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 are not indistinguishable from the literal language of the claims, or if they include equivalent structural elements that have minor differences from the literal language of the claims.

Claims

1. A computer-implemented method for reducing noise in magnetic resonance (MR) images, the method comprising: A neural network model for analyzing magnetic resonance images is executed, wherein the neural network model is trained with a pair of original images and damaged images, wherein the original images are the damaged images with noise reduced, and the target output image of the neural network model is the original images; The system receives a first magnetic resonance signal from a portion of the subject and a second magnetic resonance signal from the same portion of the subject, wherein the first and second magnetic resonance signals are acquired using a pulse sequence of an MR system. A first magnetic resonance image and a second magnetic resonance image are reconstructed based on the first magnetic resonance signal and the second magnetic resonance signal, wherein the first magnetic resonance image and the second magnetic resonance image are magnetic resonance images of the part of the subject. The neural network model was used to analyze the first magnetic resonance image and the second magnetic resonance image; The analysis derives a denoised magnetic resonance image, wherein the denoised magnetic resonance image is a combined image based on the first magnetic resonance image and the second magnetic resonance image; and Output the denoised magnetic resonance image.

2. The method of claim 1, wherein executing the neural network model further comprises executing the neural network model, wherein the input of the neural network model includes a composite image.

3. The method according to claim 1, wherein the first magnetic resonance image has a different noise distribution than the second magnetic resonance image.

4. The method according to claim 1, wherein receiving the first magnetic resonance signal and the second magnetic resonance signal further comprises: The system receives a first magnetic resonance signal and a second magnetic resonance signal, wherein the first magnetic resonance signal is a signal of a first echo of a first chemical substance and a second chemical substance, and the second magnetic resonance signal is a signal of a second echo of the first chemical substance and the second chemical substance, wherein the first echo and the second echo have different echo times.

5. The method of claim 4, wherein the first echo is an in-phase echo of the first chemical substance and the second chemical substance, and the second echo is an out-of-phase echo of the first chemical substance and the second chemical substance.

6. The method of claim 4, wherein the first chemical substance comprises water, and the second chemical substance comprises fat.

7. The method according to claim 1, wherein: Reconstructing the first and second magnetic resonance images also includes: The first magnetic resonance image is reconstructed based on the first magnetic resonance signal; and The second magnetic resonance image is reconstructed based on the second magnetic resonance signal; The analysis of the first magnetic resonance image and the second magnetic resonance image further includes: The first magnetic resonance image and the second magnetic resonance image are combined into a third image; and The neural network model is used to analyze the third image; and Exporting denoised magnetic resonance images also includes: The denoised magnetic resonance image is derived based on the analysis of the third image.

8. The method according to claim 1, wherein: Reconstructing the first and second magnetic resonance images also includes: The first magnetic resonance image is reconstructed based on the first magnetic resonance signal; and The second magnetic resonance image is reconstructed based on the second magnetic resonance signal; The analysis of the first magnetic resonance image and the second magnetic resonance image further includes: The neural network model is used to analyze the first magnetic resonance image to derive a first denoised magnetic resonance image; and The neural network model is used to analyze the second magnetic resonance image to derive a second denoised magnetic resonance image; and Exporting denoised magnetic resonance images also includes: The noise-reduced magnetic resonance image is derived by combining the first noise-reduced magnetic resonance image and the second noise-reduced magnetic resonance image.

9. The method according to claim 1, wherein: Reconstructing the first and second magnetic resonance images also includes: The first magnetic resonance image is reconstructed based on the first magnetic resonance signal; and The second magnetic resonance image is reconstructed based on the second magnetic resonance signal; The analysis of the first magnetic resonance image and the second magnetic resonance image further includes: The neural network model is used to jointly analyze the first magnetic resonance image and the second magnetic resonance image; and Exporting denoised magnetic resonance images also includes: The denoised magnetic resonance image is derived based on the joint analysis of the first magnetic resonance image and the second magnetic resonance image.

10. The method of claim 1, wherein receiving the first magnetic resonance signal and the second magnetic resonance signal further includes receiving the first magnetic resonance signal and the second magnetic resonance signal acquired using a multi-channel magnetic resonance system.

11. The method of claim 1, wherein the neural network model is configured to analyze a three-dimensional magnetic resonance dataset.

12. A magnetic resonance noise reduction system, the magnetic resonance noise reduction system comprising a noise reduction computing device, the noise reduction computing device comprising at least one processor, the at least one processor communicating with at least one memory device, and the at least one processor being programmed to perform the method according to any one of claims 1 to 11.

Images