Magnetic resonance imaging device, noise removal method and image processing device

Abstract

The present invention relates to a magnetic resonance imaging device, a noise removal method and an image processing device. Appropriate noise removal processing is also performed on images whose signal levels and noise levels are different depending on different imaging conditions and reconstruction conditions. The magnetic resonance imaging device of the present invention comprises: a measuring unit, which receives the nuclear magnetic resonance signal generated by the subject by a receiving coil; an image reconstruction unit, which processes the nuclear magnetic resonance signal received by the receiving coil to reconstruct the image of the subject; an SNR spatial distribution calculation unit, which uses the noise level of the image and the spatial distribution of the signal to calculate the spatial distribution of the signal-to-noise ratio of the image; and a noise removal unit, which removes the noise of the image based on the spatial distribution of the signal-to-noise ratio.

Description

Technical Field

[0001] The present invention relates to an image processing technology for images acquired by a medical imaging device, and in particular to a noise reduction technology for images acquired by a magnetic resonance imaging device (hereinafter referred to as an MRI device). Background Art

[0002] Images acquired by medical imaging devices contain various noises depending on the accuracy of the device itself, imaging conditions, factors on the inspection subject side, etc. Many techniques for reducing these noises have been developed in accordance with the types and causes of these noises.

[0003] For example, for Gaussian noise, spike noise, etc., noise can be reduced by replacing the pixel value with its average value, central value, or weighted average value in each local area composed of a small number of pixels. These noise reduction processes are specifically implemented through various filters such as Box filter, median filter, Gaussian filter, bilateral filter, non-local mean (NLM) filter, etc. The noise reduction effect and edge preservation effect vary depending on the type of filter.

[0004] In addition, there are noise removal methods that utilize sparse transformations such as wavelet transformation and curvelet transformation, and in recent years, there are noise removal methods that utilize machine learning (noise removal models).

[0005] In MRI, diagnostic images are usually calculated by reconstructing images obtained from multiple receiving coils. Therefore, the noise carried in the MRI image varies according to the configuration of the receiving coils and the image reconstruction method in addition to the imaging conditions, and cannot be properly removed by existing noise removal methods. In response to this, the following method is proposed: the noise of the MRI image is analyzed, and the noise is effectively removed based on the analysis results (for example, Patent Document 1, Patent Document 2, etc.). The technology described in Patent Document 1 proposes that for images reconstructed by the parallel imaging method, based on the fact that there is noise correlation between the image in which the spatially repeated signals are not separated (the image before separation) and the image after the signals are separated (the image after separation), it is added to the constraints of the noise removal processing.

[0006] The method disclosed in Patent Document 2 describes that, in images acquired by multiple receiving coils, the degree of noise removal is first adjusted based on the noise index value of the image of each coil, and the noise is removed for each image of each coil. After that, the noise-removed images of each coil are synthesized to calculate a diagnostic image. As a noise index value, a signal-to-noise ratio (SNR) is derived by performing multiple imaging with different RF conditions, and the SNR is used as the noise index value.

[0007] Prior art literature

[0008] Patent Literature

[0009] Patent Document 1: JP Patent No. 6762284

[0010] Patent Document 2: Japanese Patent Application Publication No. 2020-103890

[0011] Non-patent literature

[0012] Non-patent document 1: Pruessmann KP et al., "SENSE: Sensitivity Encoding for FastMRI", Magnetic Resonance in Medicine, 1999, Vol. 42, pp. 952-962

[0013] When the parallel imaging method (non-patent document 1) is used to shorten the imaging time by using multiple receiving coils, it is known that the SNR is inversely proportional to an index called the g-factor (Geometry Factor). Therefore, during parallel imaging reconstruction, since the noise changes spatially according to the g-factor, even if the technology described in patent document 2 is used to remove the noise of each coil image before parallel imaging reconstruction, the SNR will deteriorate according to the g-factor after parallel imaging reconstruction, and noise removal cannot be performed appropriately.

[0014] Furthermore, in the technology described in Patent Document 2, since SNR is calculated and an index is derived by performing multiple imaging with different RF conditions, there is also a problem that multiple imaging is required for SNR calculation. Summary of the invention

[0015] The present invention aims to effectively remove noise from images having various noise levels due to different imaging conditions, and to obtain a noise-removed image even from an image synthesized from data from a plurality of receiving coils, thereby reducing imaging and calculation for noise removal.

[0016] The present invention solves the above-mentioned problem by using the spatial distribution of the noise level of an image and the spatial distribution of the signal value to obtain the spatial distribution of the SNR and performing noise removal reflecting the spatial distribution of the SNR.

[0017] That is, the magnetic resonance imaging apparatus of the present invention comprises: a measuring unit, which receives the nuclear magnetic resonance signal generated by the object by a receiving coil; an image reconstruction unit, which processes the nuclear magnetic resonance signal received by the receiving coil and reconstructs the image of the object; an SNR spatial distribution calculation unit, which calculates the spatial distribution of the signal-to-noise ratio of the image using the noise level of the image and the spatial distribution of the signal; and a noise removal unit, which removes the noise of the image based on the spatial distribution of the signal-to-noise ratio.

[0018] In addition, the noise removal method of the present invention removes noise from an image acquired by a magnetic resonance imaging device, comprising: a step of calculating the spatial distribution of the signal-to-noise ratio; and a step of removing the noise from the image based on the spatial distribution of the signal-to-noise ratio. The step of calculating the spatial distribution of the signal-to-noise ratio comprises: a step of calculating the noise level of the image; a step of calculating the spatial distribution of the noise of the image; and a step of calculating the spatial distribution of the signal of the image, and the spatial distribution of the signal-to-noise ratio is calculated using the noise level, the spatial distribution of the noise, and the spatial distribution of the signal.

[0019] Furthermore, the image processing device of the present invention removes noise from an image obtained by a medical imaging device, and comprises: an SNR spatial distribution calculation unit that calculates the spatial distribution of the signal-to-noise ratio from the image of the noise processing object; and a noise removal unit that removes noise from the image based on the spatial distribution of the signal-to-noise ratio.

[0020] Effects of the Invention

[0021] According to the present invention, by performing noise removal processing on an image based on the spatial distribution of the signal-to-noise ratio calculated from the image, it is possible to effectively remove noise from an image whose signal level and noise level vary depending on the imaging conditions. In addition, according to the present invention, in an imaging method using a plurality of receiving coils, by applying noise removal to the synthesized image, it is also possible to remove noise that may be generated during synthesis. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is an overall block diagram showing one embodiment of the MRI apparatus of the present invention.

[0023] Figure 2 This is a functional block diagram of the image processing device of the present invention.

[0024] Figure 3 Yes means Figure 1 FIG. 1 is a diagram showing an overview of the processing of an MRI apparatus.

[0025] Figure 4 This is a functional block diagram of the noise processing unit according to the first embodiment.

[0026] Figure 5 This is a diagram showing the flow of noise processing in the first embodiment.

[0027] Figure 6 This is a diagram showing an example of a histogram used for noise level calculation.

[0028] Figure 7 Yes means Figure 5 FIG. 5 is a diagram showing details of the noise level calculation process.

[0029] Figure 8This is a diagram showing the flow of noise processing in the second embodiment.

[0030] Fig. 9 This is a diagram showing the effect of the first embodiment.

[0031] Fig.10 This is a diagram showing the effects of the first and second embodiments.

[0032] Fig.11 It is a figure which shows a display example.

[0033] Description of Reference Numerals

[0034] 10: MRI device, 102: static magnetic field coil, 103: gradient magnetic field coil, 104: shim coil, 105: transmitting coil, 106: receiving coil, 107: transmitter, 108: receiver, 112: gradient magnetic field power supply, 113: shim power supply, 200: computer, 201: display, 203: external storage device, 205: input device, 210: measurement control unit, 230: image reconstruction unit, 240: noise processing unit, 250: SNR distribution calculation unit, 251: noise distribution calculation unit, 252: noise level calculation unit, 253: signal distribution calculation unit, 260: noise removal unit, 270: display control unit, 30: image processing device DETAILED DESCRIPTION

[0035] Embodiments of the MRI apparatus of the present invention will be described below.

[0036] [Overview of MRI Device]

[0037] The MRI apparatus of this embodiment is as follows Figure 1 As shown, generally speaking, it comprises: a measuring unit 100, which measures the nuclear magnetic resonance signals generated from the subject 101; and a computer 200, which controls the measuring unit 100 and uses the nuclear magnetic resonance signals measured by the measuring unit 100 to perform image reconstruction, correction, and other operations.

[0038] The measuring unit 100 comprises: a static magnetic field coil 102, which generates a static magnetic field in a space where a subject 101 is placed; a transmitting unit (105, 107), which transmits high-frequency magnetic field pulses to the subject 101 arranged in the static magnetic field; a receiving unit (106, 108), which receives the nuclear magnetic resonance signal generated by the subject; and a gradient magnetic field coil 103, which gives a magnetic field gradient to the static magnetic field generated by the static magnetic field coil 102 in order to give position information to the nuclear magnetic resonance signal.

[0039] The static magnetic field coil 102 is composed of a normal conductive or superconductive static magnetic field coil, a static magnetic field generating magnet, etc., and there are vertical magnetic field methods, horizontal magnetic field methods, etc. according to the direction of the generated static magnetic field. The shape of the coil and the appearance of the entire device vary according to the method. This embodiment can be used regardless of the type of MRI device.

[0040] The transmitting unit includes: a high-frequency coil 105 for transmission (hereinafter simply referred to as a transmission coil), which transmits a high-frequency magnetic field to the measurement area of ​​the subject 101; and a transmitter 107, which includes a high-frequency oscillator, an amplifier, etc. The receiving unit includes: a high-frequency coil 106 for reception (hereinafter simply referred to as a reception coil), which receives the nuclear magnetic resonance signal generated from the subject 101; and a receiver 108, which includes an orthogonal detection circuit, an A / D converter, etc. In this embodiment, the reception coil is composed of a plurality of channels (small reception coils), each of which is connected to an orthogonal detection circuit and an A / D converter constituting the receiver 108. The nuclear magnetic resonance signal received by the receiver 108 is handed over to the computer 200 as a complex digital signal. The reception coil 106 is composed of a reception coil composed of a plurality of small coils, and the nuclear magnetic resonance signals received by each reception coil are handed over to the computer 200 and synthesized during image reconstruction.

[0041] The gradient magnetic field coil 103 has three sets of gradient magnetic field coils for applying gradient magnetic fields in the x, y, and z directions, respectively, and each is connected to a gradient magnetic field power supply unit 112. Furthermore, the MRI apparatus may also include a shim coil 104 for adjusting static magnetic field distribution and a shim power supply unit 113 for driving the shim coil.

[0042] Furthermore, the measuring unit 100 includes a sequence control device 114 that controls the operation of the measuring unit 100. The sequence control device 114 controls the operation of the gradient magnetic field power supply unit 112, the transmitter 107, and the receiver 108, and controls the timing of applying the gradient magnetic field and the high-frequency magnetic field and receiving the nuclear magnetic resonance signal. The controlled timing diagram is called a pulse sequence, which is pre-set according to the measurement and stored in a storage device or the like included in the computer 200 described later.

[0043] The computer 200 controls the overall operation of the MRI apparatus 10 and performs various calculations on the received nuclear magnetic resonance signals. The computer 200 is an information processing device including a CPU, a memory, a storage device, etc., and is connected to a display 201, an external storage device 203, an input device 205, etc.

[0044] The display 201 is an interface for displaying the results obtained in the calculation process to the operator. The input device 205 is an interface for the operator to input the conditions and parameters required for the measurement and calculation process implemented in this embodiment. The user can input measurement parameters such as the speed multiple in the PI method through the input device 205. The external storage device 203 stores data used in various calculation processes performed by the computer 200, data obtained through the calculation process, input conditions, parameters, etc. together with the storage device inside the computer 200.

[0045] In this embodiment, the computer 200 performs image synthesis and noise removal processing using the sensitivity distribution of the receiving coil as a function of image processing. Figure 1 As shown, the computer 200 includes functional units such as a measurement control unit 210, an image reconstruction unit 230, a noise processing unit 240, and a display control unit 270. The noise processing unit 240 has a function of performing a process for noise removal on the image generated by the image reconstruction unit 230, and includes an SNR distribution calculation unit 250 and a noise removal unit 260.

[0046] The functions of these parts can be realized as software installed in the computer 200, and the CPU loads the program (software) stored in the storage device into the memory and executes it. Various data used in the processing of each function and various data generated during the processing are stored in the storage device or the external storage device 203. In addition, some of the various functions realized by the computer 200 can be realized by hardware such as ASIC (Application Specific Integrated Circuit) and FPGA (field-programmable gate array).

[0047] Furthermore, among the functions implemented by computers, the functions involved in image processing are as follows Figure 2 As shown, this can be realized by an image processing device 30 that is independent of the MRI device 10 and can transmit and receive data with the MRI device 10 .

[0048] The image processing device 30 includes an image receiving unit 310 that receives images from the medical imaging device 10A. Figure 1The computer 200 of the MRI apparatus 10 shown in the figure similarly has a noise processing unit 340 including an SNR distribution calculation unit 350 and a noise removal unit 360. The medical imaging apparatus 10A is not limited to an MRI apparatus, but may be a CT apparatus, an ultrasonic imaging apparatus, or a database storing various medical images. In addition, although not necessarily required, an image reconstruction unit 330 may be provided, which reconstructs an image based on an image or measurement data received from the medical imaging apparatus 10A.

[0049] Next, the operation of the MRI apparatus (mainly the computer 200) and the image processing apparatus of this embodiment will be briefly described. The functions of the various components of the image processing apparatus 30 are basically the same as the image processing functions of the MRI apparatus, and the following will refer to the various components of the MRI apparatus. Figure 3 Shows the flow of actions.

[0050] [Measurement: S301]

[0051] First, the user's imaging sequence and imaging condition settings are received through the input device 205. The imaging sequence is not particularly limited, and for example, the PI method using a high-speed pulse sequence such as a 2D-RSSG (RF-spoiled-Steady-state Acquisition with Rewound Gradient-Echo) sequence can be selected and set. The imaging conditions include the parameters of the imaging sequence (repetition time TR, echo time TE), and the speed factor when performing k-space thinning measurement. In addition, the setting of the number of slices is included in the case of simultaneous excitation of multiple slices. In addition, when these imaging conditions are set as an inspection protocol, the conditions set for the inspection protocol are read in.

[0052] The measurement control unit 210 operates the sequence control device 114 according to the pulse sequence set based on the parameters input by the user, and measures the nuclear magnetic resonance signal (echo signal) of the predetermined condition. The sequence control device 114 controls each part of the MRI device 10 according to the instruction from the measurement control unit 210, and receives the nuclear magnetic resonance signal through the receiving coil. In the case where the receiving coil is a multi-channel coil composed of a plurality of receiving coils, k-space data is collected for each channel.

[0053] [Image reconstruction: S302]

[0054] The image reconstruction unit 230 performs Fourier transform and other operations on the k-space data to reconstruct the image. In the case of the parallel imaging (PI) method, the k-space data of each receiving coil and the sensitivity distribution of multiple receiving coils are used to separate the signals that overlap in space, and images at different positions in space are created and synthesized. In addition, in the case of the compressed sensing (CS) method, repeated operations including L1 norm minimization are performed on the randomly sampled k-space data to reconstruct the image.

[0055] [S303, S304]

[0056] Next, the noise processing unit 240 performs processing for removing noise included in the reconstructed image.

[0057] Noise removal is achieved by performing repeated operations under given constraints to minimize noise, similar to the well-known nonlinear filters including total variation regularization and sparse regularization. In this embodiment, the spatial distribution of the SNR of the image is reflected in the constraints to perform repeated operations. Therefore, the spatial distribution of the signal level of the image, the noise level and its spatial distribution are calculated for the image that is the object of noise removal, and these quantities are used to calculate the SNR spatial distribution (S303). After that, noise removal using the SNR spatial distribution is performed (S304).

[0058] The image after noise removal is stored in the storage device 203 or displayed on the display 201 as needed ( S305 ).

[0059] According to this embodiment, by performing noise removal based on the spatial distribution of SNR calculated from the signal level and noise level of the image, even when the signal level and noise level of the image are different depending on the shooting conditions, noise removal based on the spatial distribution of SNR can be performed with good accuracy. In addition, noise removal with good accuracy can also be performed on an image reconstructed from a spatially repeated signal or a synthesized image using the PI method or the like.

[0060] The following describes the specific processing implementation of the noise processing unit. Figure 1 The outline of the MRI apparatus shown and Figure 3 The outline of the processing shown is common to the following embodiments and can be referred to as appropriate.

[0061] <Implementation Method 1>

[0062] In this embodiment, noise is removed from an image obtained by high-speed imaging using a plurality of receiving coils. If the imaging method is the SENSE method, the target is an image obtained by unwinding and synthesizing the folds generated in the image of each receiving coil by matrix calculation using the image reconstructed for each receiving coil and the sensitivity information of each receiving coil.

[0063] exist Figure 4 The configuration of the noise processing unit 240 of this embodiment is shown. As shown in the figure, the noise processing unit 240 includes an SNR distribution calculation unit 250 and a noise removal unit 260. The SNR distribution calculation unit 250 includes a noise distribution calculation unit 251, a noise level calculation unit 252, and a signal distribution calculation unit 253.

[0064] In such a structure, in the noise processing ( Figure 3 :S303), such as Figure 5 As shown in FIG. 1 , the spatial distribution of noise is calculated from the g factor of the receiving coil (S51), the noise level is calculated from the histogram of the image (S52), and the spatial distribution of the signal (excluding noise) is calculated using the information on the spatial distribution of the noise level (S53). The spatial distribution of SNR is calculated from the spatial distribution of noise and the spatial distribution of the signal (S54) and used in the noise removal operation ( Figure 3 :S304).

[0065] The details of each process are described below.

[0066] [Calculation of noise distribution: S51]

[0067] The noise distribution calculation unit 251 calculates the spatial distribution of noise by obtaining the inverse of the g factor of the receiving coil. The g factor is a value that changes depending on the number and arrangement of the small receiving coils constituting the receiving coil, the multiplication rate of the PI method, etc., and can be obtained in advance using the sensitivity distribution of the receiving coil and the correlation matrix of the noise between the receiving coils.

[0068] Here, the g factor can be calculated using the formula described in Non-Patent Document 1. If the receiving coil sensitivity matrix is ​​set to S and the complex transposed matrix of the receiving coil sensitivity matrix is ​​set to S H , let the noise correlation matrix be Ψ, and let diag[A] be the operator for extracting the diagonal element vector of the matrix A, then the g factor is represented by the following formula (1).

[0069]

Mathematical formula 1

[0070]

[0071] [Calculation of noise level: S52]

[0072] The calculation of the noise level uses the fact that the histogram of the noise of the absolute value image follows the Rayleigh distribution, and calculates the noise SD as the noise level.

[0073] exist Figure 6 (A) shows an example of a histogram of an absolute value image of a human head image (T2WI). As shown in the figure, the original image signal shows a Gaussian distribution in the histogram, but the noise has a peak in the low signal value area. When the noise SD of the real image and the virtual image is set to σ, the standard deviation (SD) of the noise follows the Rayleigh distribution with parameter σ. The Rayleigh distribution can be represented by the following formula (2), and its mode is consistent with σ.

[0074]

Mathematical formula 2

[0075]

[0076] If the maximum likelihood estimate of the mode σ is set to σhat, σhat can be calculated by the following formula (3).

[0077]

Mathematical formula 3

[0078]

[0079] Therefore, σhat is obtained at which the difference (root mean square error: RMSE) between the histogram of the absolute value image and the Rayleigh distribution calculated from the maximum likelihood estimation value σhat becomes minimum, and the SD of the complex noise can be calculated by multiplying it by √2.

[0080] exist Figure 7 The specific processing flow performed by the noise level calculation unit 252 is shown in FIG. 2. First, the complex image is converted into an absolute value image, and the absolute value image is multiplied by the inverse of the g factor map calculated in S51 to create a histogram (S521). Here, this is to correct the spatial noise distribution caused by parallel imaging by multiplying by the inverse of the g factor map and to obtain the standard deviation (noise level) of uniform noise. As an example of the histogram created in this way, Figure 6 (B) shows a histogram H of an absolute value image of a Shepp-logan phantom (SNR=5, noise SD=0.2). When the number of bins of the histogram is Nbin, the histogram H is denoted as H(Xnbin) (Xnbin is the signal value of the nth bin).

[0081] Next, the threshold (initial threshold) Thre of the histogram is calculated by discriminant analysis (S522). Next, the number n of pixels below the threshold Thre and greater than 0 is calculated (S523). Based on the pixel values ​​of the pixels (n pixels) below the threshold Thre and greater than 0, the maximum likelihood estimate σhat is calculated using the aforementioned formula (2) (S524).

[0082] Next, the probability density function p(nbin) of the Rayleigh distribution is calculated using equation (4) ( S525 ).

[0083]

Mathematical formula 4

[0084]

[0085] The error between the calculated probability density function p(nbin) and the histogram H(Xnbin) is calculated by equation (5) ( S526 ), and until the error E is minimized ( S527 ), S523 to S526 are repeated while updating the threshold Thre ( S528 ).

[0086]

Mathematical formula 5

[0087]

[0088] The maximum likelihood estimation value σhat when the error E is minimized is set as the mode σ of the Rayleigh distribution. When the error E is not a minimum value, the threshold Thre is reduced by equation (6).

[0089]

Mathematical formula 6

[0090]

[0091] In formula (6), the coefficient α is the ratio of gradually reducing the threshold value, for example, α=0.1, and is set so that the updated Thre is about 90% of the Thre before the update. If the error becomes smaller through the update, the update is repeated, and when the error becomes larger, the σhat calculated with the Thre before the update is set as the maximum likelihood estimate.

[0092] When the maximum likelihood estimation value σhat that minimizes the final error is determined, it is set to the mode σ of the Rayleigh distribution and multiplied by √2 ( S529 ). That is, the noise level of the complex noise (SD of the noise) is calculated.

[0093] By multiplying the calculated noise level and the g-factor map (ie, relative noise SD), the spatial distribution of the noise level with respect to the image of the processing object is obtained.

[0094] [Calculation of signal distribution: S53]

[0095] The signal distribution calculation unit 253 obtains the spatial distribution of the signal from the intensity distribution itself calculated from the absolute value of the complex image to be processed.

[0096] [Calculation of SNR distribution: S54]

[0097] The SNR distribution calculation unit 250 calculates the spatial distribution of the SNR using the signal spatial distribution calculated in S53 and the spatial distribution of the noise level calculated in S52 .

[0098] Through the above processing S51~S54, the Figure 3 The SNR distribution calculation process S303.

[0099] [Noise Removal: S304]

[0100] The noise removal unit 260 performs noise removal processing by repeated calculation (iterative processing). In this noise removal processing, in addition to the constraint that the image before noise removal and the image after noise removal are approximately equal (hereinafter referred to as the image constraint before and after noise removal), that is, in addition to the constraint for not deviating too much from the original image, the constraint that the noise of the image after the image is mapped to the sparse space is approximately equal to zero (hereinafter referred to as the sparse space constraint), the constraint that the noise of the spatial differential value image of the separated image used in the noise removal of the TV (total variation) method is approximately equal to zero (spatial differential value constraint), and other constraints are added to perform minimization processing with regularization terms, and these constraints are repeatedly calculated using the spatial distribution of SNR calculated by the SNR distribution calculation unit 250 as weights.

[0101] For example, a minimization process represented by equation (7) is performed.

[0102]

Mathematical formula 7

[0103]

[0104] In formula (7), the first term represents the image constraint before and after noise removal, the second term represents the sparse space constraint, y is the image before noise removal, x is the image after noise removal, λ1 and λ2 are regularization parameters of the first term and the second term, respectively, and Ψ is a wavelet transform operator. In addition, Wcons and Wrglr are weight images of the first term and the second term. In this embodiment, as the weight image Wrglr of the second term, for example, the inverse of the SNR spatial distribution "1 / SNR" obtained in S54 is used.

[0105] By adding SNR as a weight to constraints other than the image constraints before and after noise removal (here is the sparse space constraint), the weights of these constraints are increased for images with low SNR and reduced for images with high SNR, thereby preventing excessive noise and improving the accuracy of noise processing.

[0106] In addition, in formula (7), sparse space constraints are exemplified as constraints other than the constraints of the images before and after noise removal, but spatial differential value constraints may be used instead of or together with them. Furthermore, regarding the noise correlation constraints disclosed in Patent Document 1 (Patent No. 6762284), i.e., the image before synthesis of the PI method (hereinafter referred to as the image before separation), a constraint condition that sets the existence of noise correlation between the image before separation and the image after separation as a constraint may be used, and as weights for these constraints, SNR (its inverse) may be used.

[0107] As described above, according to the present embodiment, the noise level is estimated (calculated) based on the image to be processed, and the spatial distribution of the SNR is calculated using the spatial distribution of the noise level and the signal distribution calculated based on the noise level, which is then used to adjust the constraints in the noise removal process. Thus, appropriate noise processing can be performed on images whose noise levels and signal levels differ depending on the imaging conditions.

[0108] In addition, according to this embodiment, by applying this embodiment to the synthesized image in the PI method, the difficulty of noise removal in the PI method, that is, the amount of noise changes spatially before and after synthesis, thereby changing the SNR, can be solved. However, this embodiment can also be applied to the image before synthesis. In this case, the noise removal process can be omitted for the synthesized image, but the noise removal process can also be further performed on the synthesized image.

[0109] Furthermore, since the signal changes spatially (masking) when multiple coils are receiving, it is difficult to remove noise without signal loss. However, since noise removal processing is performed based on the SNR that reflects the spatial distribution of the signal, noise removal can be performed without loss even in areas with weak signal levels.

[0110] In addition, in the present embodiment, the noise removal unit 260 performs minimization processing with a regularization term to perform noise removal, but a noise removal model can also be used, wherein the noise removal model is learned in the following manner, that is, deep learning such as CNN (Covolutional Neural Network) is used to calculate an image with noise removed from images with different noise levels.

[0111] In this case, for example, the network that has been learned on an image with low SNR and the network with high SNR can be spatially switched using the SNR distribution to remove noise.

[0112] <Implementation Method 2>

[0113] The basic processing flow of this embodiment and Figure 5 as well as Figure 7 The process shown is the same as that shown in FIG. 1 , but in this embodiment, the feature is that Figure 5 In the noise level calculation process (S52), the noise level is calculated by excluding the background area with a low signal value, thereby improving the accuracy of the noise level calculation process.

[0114] The following references Figure 8 The processing of this embodiment is described.

[0115] First, with Figure 7 Similarly to S521, the absolute value image is multiplied by the inverse of the g factor to create a histogram (S5201). Figure 7 Similarly to S522 to S528, the maximum likelihood estimate σhat is calculated based on the Rayleigh distribution estimation (S5202). Next, a mask is calculated that sets the area larger than the threshold Thre when calculating σhat to 0 and the area below the threshold Thre to 1 (S5203), and the mask is used to calculate the noise area (the area in the image space) (S5204).

[0116] Next, the spatial differential of the noise region is calculated on the complex image by the following equation (8) (S5205). Here, if the pixel value of coordinate xn is set to I(xn) and the coordinate of the adjacent pixel is set to xn+1, the spatial differential ΔI(xn) is expressed by the following equation.

[0117] (Mathematical formula 8)

[0118] ΔI(xn)=I(xn+1)-I(xn) (8)

[0119] By performing spatial differentiation, only noise is extracted as the difference between adjacent pixels. Then, the mask is multiplied by the spatially differentiated data (S5206). Thus, an image with mostly only noise is obtained. The histogram of the absolute value image is further calculated for the image, and the maximum likelihood estimate σhat is calculated in the same way as S5202 (S5207). This is set to √2 times to obtain the noise level of the complex noise (SD of the noise).

[0120] Then, the noise SD is used to calculate the spatial distribution of the noise level, calculate the signal distribution, and obtain the spatial distribution of the SNR ( Figure 5 :S54), and further use the spatial distribution of SNR to perform noise removal processing ( Figure 3 :S304), which are the same as implementation mode 1.

[0121] According to this embodiment, by using a mask, noise in the background area of ​​the image can be excluded from the object of removal processing. Since only noise is extracted by spatial differentiation of the area where noise is to be processed and the noise level is calculated, the noise level can be calculated with higher accuracy.

[0122] In addition, Figure 8 In the process, the noise area roughly calculated from the absolute value image is spatially differentiated to calculate the noise level (noise SD), but as the absolute value image used in creating the histogram, the absolute value image of the spatially differentiated complex image can also be used.

[0123] <Implementation Examples of Implementation Methods 1 and 2>

[0124] The noise removal method (noise level estimation method) of Embodiment 1 and Embodiment 2 was evaluated using the Shepp-Logan numerical phantom. Fig. 9 as well as Fig.10 The results are shown. Fig. 9 The upper figure is a histogram of all pixels, and the lower figure is a figure showing the consistency between the Rayleigh distribution estimated by the method of implementation mode 1 and the histogram, (a), (b), and (c) are the cases of SNR=5 (noise SD is set to 0.20000), SNR=3.3 (noise SD is set to 0.30000), and SNR=2.5 (noise SD is set to 0.40000), respectively.

[0125] The noise SD estimated from the Rayleigh distribution estimated by the method of Embodiment 1 is 0.20005, 0.29803, and 0.40050 in (a) to (c), respectively, which is confirmed to be very consistent with the set noise SD. However, it is confirmed that the error between the estimated Rayleigh distribution and the histogram increases as the SNR decreases.

[0126] Fig.10 This is a diagram showing the relationship between the estimated noise SD and the set noise SD when different noise estimation methods are used. In the case (a) of the method of the first embodiment (calculated from the absolute value image), it is confirmed that if the noise SD is 0.33 or less (that is, SNR=3 or more), the error is suppressed to a small level. In addition, it is confirmed that the accuracy is improved compared to the case (b) and (a) of calculating using the absolute value image that performs spatial differentiation of the complex image (only in the X direction). In the case (c) of the method of the second embodiment (calculating the noise SD from the absolute value image and then performing spatial differentiation to calculate the noise SD), it is confirmed that the accuracy is improved compared to (a).

[0127] Furthermore, the same results were obtained by estimating the noise levels of two types of human head images captured by different camera methods using the methods of implementation modes 1 and 2. In particular, the method of implementation mode 2 was able to obtain a result with a high degree of consistency with the noise SD estimated based on the noise SD of the image synthesized by MAC even in an image with a low SNR.

[0128] <Display Implementation Method>

[0129] The noise processing of the first and second embodiments can be set as a default process in the image processing unit, but it is also possible to accept user selection of a noise removal method when receiving the imaging conditions via the input device 205. In this case, for example, the display control unit 270 displays the image on the display 201. Fig.11 In the example shown in the figure, the noise removal method of this embodiment can be selected together with well-known noise removal methods such as linear filters such as average filters and Gaussian filters, edge preservation filters such as bilateral filters, nonlinear filters including total variation regularization and sparse regularization, etc. As a result, the user has a higher degree of freedom in selecting the method.

Claims

1. A magnetic resonance imaging device, characterized in that: have: The measuring part receives the nuclear magnetic resonance signal generated by the detected object by the receiving coil; An image reconstruction unit processes the nuclear magnetic resonance signal received by the receiving coil to reconstruct an image of the subject; An SNR spatial distribution calculation unit that calculates a spatial distribution of a signal-to-noise ratio of the image using a noise level of the image and a spatial distribution of a signal; and a noise removal unit that removes noise from the image based on the spatial distribution of the signal-to-noise ratio, The noise removal unit removes noise by performing an error minimization calculation using a plurality of constraints, and uses the spatial distribution of the signal-to-noise ratio as a weight of the constraints.

2. The magnetic resonance imaging apparatus according to claim 1, characterized in that: The measuring unit includes a plurality of receiving coils. The image reconstruction unit reconstructs an image by synthesizing the nuclear magnetic resonance signals respectively received by the plurality of receiving coils or images reconstructed based on the nuclear magnetic resonance signals. The SNR spatial distribution calculation unit calculates the spatial distribution of the signal noise for the image synthesized by the image reconstruction unit.

3. The magnetic resonance imaging apparatus according to claim 2, characterized in that: The image reconstruction unit synthesizes the nuclear magnetic resonance signals received by each receiving coil or the image reconstructed according to the nuclear magnetic resonance signals through parallel imaging calculation.

4. The magnetic resonance imaging apparatus according to claim 1, characterized in that: The SNR spatial distribution calculation unit comprises: a noise level calculation unit, which calculates the noise level of the image; a noise distribution calculation unit that calculates the spatial distribution of noise of the image; and a signal distribution calculation unit, which calculates the spatial distribution of the signal of the image, The spatial distribution of the signal-to-noise ratio is calculated using the noise level, the spatial distribution of the noise, and the spatial distribution of the signal.

5. The magnetic resonance imaging apparatus according to claim 4, characterized in that: The noise level calculation unit calculates a noise level based on a histogram of absolute values ​​of the image.

6. The magnetic resonance imaging apparatus according to claim 5, characterized in that: The noise level calculation unit further calculates the noise level based on a histogram of absolute values ​​of spatial differentials of the image.

7. The magnetic resonance imaging apparatus according to claim 4, characterized in that: The measuring unit includes a plurality of receiving coils. The noise distribution calculation unit calculates the noise distribution using the g-factor of each receiving coil.

8. The magnetic resonance imaging apparatus according to claim 1, characterized in that: The multiple constraints include: The image constraint condition before and after noise removal is that the image before and after noise removal is approximately equal; and A sparse space constraint condition that the noise of the image obtained by mapping the image after noise removal to a sparse space is approximately zero and / or a spatial differential value constraint condition that the noise of the spatial differential value image of the image before noise removal is approximately zero, The noise removing unit uses the spatial distribution of the signal-to-noise ratio as a weight of constraints other than the image constraints before and after the noise removal.

9. The magnetic resonance imaging apparatus according to claim 1, characterized in that: The noise removal unit includes a machine learning model that is learned so as to output an image from which noise has been removed from images with different noise levels, and adjusts coefficients of the machine learning model based on the spatial distribution of the signal-to-noise ratio.

10. A method for removing noise from an image acquired by a medical imaging device, characterized in that: Include: The step of calculating the spatial distribution of the signal-to-noise ratio of the image; and a step of removing noise from the image based on the spatial distribution of the signal-to-noise ratio, The step of calculating the spatial distribution of the signal-to-noise ratio comprises: The step of calculating the noise level of the image; The step of calculating the spatial distribution of noise of the image; and the step of calculating the spatial distribution of the signal of the image, calculating the spatial distribution of the signal-to-noise ratio using the noise level, the spatial distribution of the noise, and the spatial distribution of the signal, Noise is removed by an error minimization operation using a plurality of constraints, and the spatial distribution of the signal-to-noise ratio is used as a weight of the constraints.

11. The noise removal method according to claim 10, characterized in that: The image is an image obtained by synthesizing nuclear magnetic resonance signals respectively received by a plurality of receiving coils of an MRI device or images reconstructed based on the nuclear magnetic resonance signals. The step of calculating the spatial distribution of the noise calculates the spatial distribution of the noise using g factors of the plurality of receiving coils.

12. The noise removal method according to claim 10, characterized in that: The step of calculating the noise level comprises: A step of calculating a noise level based on a histogram of an absolute value image calculated from the image.

13. The noise removal method according to claim 10, characterized in that: The step of calculating the noise level comprises: a step of calculating the spatial differential of the image and extracting the noise region, The noise removal method calculates a noise level from the noise area.

14. An image processing device for removing noise from an image acquired by a medical imaging device, the image processing device comprising: An SNR spatial distribution calculation unit calculates a spatial distribution of a signal-to-noise ratio from an image to be processed for noise; and a noise removal unit that removes noise from the image based on the spatial distribution of the signal-to-noise ratio, The noise removal unit removes noise by performing an error minimization calculation using a plurality of constraints, and uses the spatial distribution of the signal-to-noise ratio as a weight of the constraints.

Images