Magnetic Resonance Imaging Apparatus, Image Processing Apparatus, and Image Processing Method

Abstract

To provide a method performing highly accurate reconstruction and noise removal to various thinning-out patterns including equal interval thinning-out.SOLUTION: An image processing unit for processing measurement data acquired by an MRI device performs image reconstruction by using measurement data of each channel measured in a prescribed thinning-out pattern and the sensitivity distribution of each reception coil. At this time, de-noising of the reconstructed image and arithmetic for holding the consistency between the measurement data of each channel created from an image after de-noising and original measurement data are successively processed. With this, accurate image restoration and de-noising can be performed without depending on the thinning-out pattern.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a magnetic resonance imaging (MRI) apparatus, and more particularly to a technique for reconstructing and removing noise from measurement data obtained by thinning-out measurement. [Background technology]

[0002] One method for shortening imaging time in MRI examinations is to thin out the signals in the spatial frequency domain (k-space) and restore the unmeasured areas through signal processing (k-space undersampling reconstruction method).Among these, the parallel imaging (PI) method, which receives signals simultaneously using multiple receiving coils (channels) and restores the signals by utilizing differences in sensitivity distribution, is widely used.

[0003] Known data restoration methods for the PI method include the SENSE method, which restores data from evenly spaced thinning, as well as the POCSENSE and SPIRiT methods, which use iterative reconstruction to accommodate arbitrary thinning patterns (e.g., Non-Patent Documents 1 and 2). The POCSENSE and SPIRiT methods restore unthinned images by solving an inverse problem that expresses the relationship between measurement data, sensitivity maps, and reconstructed images through iterative calculations. However, the PI method generally has the problem that increasing the thinning rate reduces imaging time but increases noise.

[0004] On the other hand, various methods using convolutional neural networks (CNNs) have been proposed to reduce the increase in noise due to thinning. For example, Patent Document 1 proposes a reconstruction method using a trained deep NN that combines a CNN that removes noise with a layer (DC layer) that performs processing to maintain consistency between the denoised data and the acquired data. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] U.S. Patent No. 10,712,416 [Non-patent literature]

[0006] [Non-Patent Document 1] ”POCSENSE: POCS-based reconstruction for sensitivity encoded magnetic resonance imaging”, Magnetic Resonance in Medicine, 2004 December, 52(6) PP. 1397-1406

Non-Patent Document 2

Disclosure of the Invention

Problems to be Solved by the Invention

[0007] Generally, reconstruction using machine learning learns the entire network including iterative processing with a pair of decimated images and non-decimated images.

[0008] However, there are various decimation patterns such as equidistant and random. For decimation patterns not included in the training data, the accuracy of reconstruction decreases and artifacts and noise remain. Although it is possible to increase the training data, it is difficult to handle innumerable patterns, and furthermore, the structure of the corresponding CNN becomes complex and the calculation time increases.

[0009] The present invention has been made in view of the above circumstances, and an object thereof is to provide a method for performing highly accurate reconstruction and noise removal for various decimation patterns including equidistant decimation.

Means for Solving the Problems

[0010] To solve the above problems, the present invention adds noise removal processing to the iterative calculation for maintaining consistency with measurement data. The noise removal processing is based on the statistical properties of the image and the noise.

[0011] That is, a first aspect of the present invention is an MRI apparatus, which, for example, includes a measurement unit that includes a plurality of reception coils and collects nuclear magnetic resonance signals of a subject for each reception coil, a control unit that controls a collection pattern of the nuclear magnetic resonance signals, and an image processing unit that creates an image of the subject using measurement data composed of nuclear magnetic resonance signals collected for each reception coil. The image processing unit includes a sequential reconstruction unit that reconstructs an image by iterative calculation using the measurement data for each reception coil.

[0012] The sequential reconstruction unit includes an initial value setting unit that sets an initial value of iterative calculation based on the measurement data for each reception coil, a synthesis unit that synthesizes images for each reception coil, a noise removal unit that removes noise from the image based on the statistical properties of the image and the noise included in the image before or after synthesis, and a data consistency maintenance unit that creates denoised k-space data for each reception coil using the denoised image and creates estimated measurement data by integrating the denoised k-space data and the measurement data, and updates the initial value based on the estimated measurement data to perform sequential processing.

[0013] Typically, the measurement data is measurement data undersampled in a predetermined decimation pattern, the sequential reconstruction unit receives a sensitivity map of each reception coil together with the measurement data, and the synthesis unit reconstructs an image using the measurement data and the sensitivity map.

[0014] A second aspect of the present invention is an image processing apparatus that processes MRI images, and includes a sequential reconstruction unit that performs iterative calculation using measurement data acquired for each reception coil of an MRI apparatus to reconstruct an image. The sequential reconstruction unit has the same configuration and function as the sequential reconstruction unit of the above-described MRI apparatus.

[0015] Furthermore, a third aspect of the present invention is MRI DeviceAn image processing method for processing measurement data collected by a plurality of receiving coils and creating a denoised reconstructed image, including a sequential reconstruction process to maintain the consistency between the k-space data after denoising and the measurement data. The sequential reconstruction process includes steps of setting an initial value of the process based on the measurement data for each receiving coil, synthesizing images for each receiving coil, removing noise from the image based on the respective statistical properties of the image before or after synthesis and the noise included in the image data, creating denoised k-space data for each receiving coil using the denoised image, and integrating the denoised k-space data and the measurement data to create estimated measurement data. Based on the estimated measurement data, the initial value is updated and the process is repeated.

Advantages of the Invention

[0016] According to the present invention, by performing denoising processing on the image data obtained within the repeated operation of maintaining consistency, since the denoising process is not affected by the aliasing depending on the decimation pattern, noise can be removed accurately regardless of the decimation pattern. Also, since aliasing is removed by the sequential process including the synthesis process using the coil sensitivity distribution, it can also handle equidistant decimation.

[0017] Furthermore, by combining the denoising process and the data consistency maintaining process, it is possible to prevent excessive smoothing due to the denoising process and remove noise accurately. As a result, high-precision reconstruction and noise removal can be performed for various decimation patterns including equidistant decimation. Also, since a deep learning model such as a CNN can be limited to a part of the processing of the repeated operation, it is possible to expand the universality of the processing target while not requiring a large learning model.

Brief Description of the Drawings

[0018]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Embodiments for Carrying Out the Invention

[0019] Hereinafter, embodiments of the MRI device and the image processing device of the present invention will be described with reference to the drawings. First, an embodiment of the MRI device to which the present invention is applied will be described

[0020] As shown in FIG. 1, the MRI device 10 of the present embodiment mainly includes a measurement unit 100 that measures nuclear magnetic resonance signals generated from a subject 101, and a computer (calculation unit) 200 that controls the measurement unit 100 and performs image reconstruction, correction, and other calculations using the nuclear magnetic resonance signals measured by the measurement unit 100

[0021] The measurement unit 100 includes a static magnetic field coil 102 that generates a static magnetic field in the space where the subject 101 is placed, a transmission unit (105, 107) that transmits a high-frequency magnetic field pulse to the subject 101 disposed in the static magnetic field, a reception unit (106, 108) that receives the nuclear magnetic resonance signal generated by the subject, and a gradient magnetic field coil 103 that applies a magnetic field gradient to the static magnetic field generated by the static magnetic field coil 102 in order to impart position information to the nuclear magnetic resonance signal.

[0022] The static magnetic field coil 102 is composed of a resistive or superconducting static magnetic field coil, a static magnetic field generating magnet, etc. Depending on the direction of the generated static magnetic field, there are a vertical magnetic field method, a horizontal magnetic field method, etc. The shape of the coil and the appearance of the entire apparatus differ depending on the method. This embodiment is applicable to any MRI apparatus of any method.

[0023] The transmission unit includes a transmission high-frequency coil 105 (hereinafter simply referred to as a transmission coil) that transmits a high-frequency magnetic field to the measurement region of the subject 101, and a transmitter 107 including a high-frequency oscillator, an amplifier, etc. The reception unit includes a reception high-frequency coil 106 (hereinafter simply referred to as a reception coil) that receives the nuclear magnetic resonance signal generated from the subject 101, and a receiver 108 including a quadrature detection circuit, an A / D converter, etc. In this embodiment, the reception coil is composed of a plurality of channels (small reception coils), and the quadrature detection circuit and A / D converter that constitute the receiver 108 are connected to each of them. The nuclear magnetic resonance signal for each channel (i.e., for each reception coil) received by the receiver 108 is passed to the computer 200 as a complex digital signal.

[0024] The gradient magnetic field coil 103 has three sets of gradient magnetic field coils that apply gradient magnetic fields in the x direction, y direction, and z direction respectively, and are each connected to a gradient magnetic field power supply unit 112. Furthermore, the MRI apparatus may include a shim coil 104 that adjusts the static magnetic field distribution and a shim power supply unit 113 that drives it.

[0025] Furthermore, the measurement unit 100 includes a sequence control device 114 that controls the operation of the measurement unit 100. The sequence control device 114 controls the operations 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, the high-frequency magnetic field, and receiving the nuclear magnetic resonance signal. The nuclear magnetic resonance signal can be encoded by the gradient magnetic field, and the position and pattern of the nuclear magnetic resonance signal in the k-space where it is arranged as a digital signal are determined by the number of encodings (the number of encoding steps) and the amount of each encoding.

[0026] The time chart of the control by the sequence control device is called a pulse sequence, which is preset according to the measurement and stored in a storage device or the like provided in the computer 200.

[0027] The computer 200 controls the operation of the entire MRI apparatus 100 and performs various arithmetic processes on the received nuclear magnetic resonance signal. For this reason, as shown in FIG. 1, the computer 200 includes functional units such as a measurement control unit 210, an image processing unit 230, and a display control unit 250. In the present embodiment, the image processing unit (computer) 230 performs processes such as synthesis of an image using the sensitivity distribution of the reception coil, noise removal processing, and data consistency retention processing between the measurement data and the restored image.

[0028] The computer 200 is an information processing device including a CPU, a memory, a storage device, etc., and a display 201, an external storage device 203, an input device 205, etc. are connected to the computer 200.

[0029] The display 201 is an interface for displaying the results obtained by the arithmetic processing and the like to the operator. The input device 205 is an interface for the operator to input the conditions, parameters, etc. necessary for the measurement and arithmetic processing performed in the present embodiment. The user can input measurement parameters such as the doubling factor (also referred to as the decimation rate) in the PI method via the input device 205. The external storage device 203, together with the storage device inside the computer 200, holds data used for various arithmetic processes executed by the computer 200, data obtained by the arithmetic processes, input conditions, parameters, etc.

[0030] The functions of each part of the computer 200 can be realized as software incorporated in the computer 200, and are realized by the CPU loading and executing the program (software) held in the storage device into the memory. Various data used for the processing of each function and various data generated during the processing are stored in the storage device or the external storage device 203. Among the various functions realized by the computer 200, some functions may also be realized by hardware such as an ASIC (Application Specific Integrated Circuit) or an FPGA (field-programmable gate array).

[0031] A part of the functions realized by the computer 200, for example, the functions of the image processing unit 230, can also be realized by an independent computer or another image processing device capable of transmitting and receiving data with the MRI device 10, and such an image processing device is also included in the present invention.

[0032] The image processing unit 230 of the present embodiment performs sequential reconstruction including synthesis of an image using the sensitivity distribution of the reception coil, noise removal, and data consistency maintenance processing on the measurement data for each reception coil obtained by the measurement unit 100 of the MRI device 10 performing measurement with a predetermined decimation pattern and decimation rate. Also, in these processes, conversion from k-space data to real-space data (inverse Fourier transform), conversion from real-space data to k-space data (Fourier transform), conversion to a sparse space, etc. are performed.

[0033] The operation of the MRI device 10 with the above configuration is the same as the operation of a known MRI device. The measurement unit 100 receives the imaging conditions and pulse sequence selected according to the imaging method and imaging object or set in advance by the inspection protocol, and sends the imaging conditions and pulse sequence to the sequence control device 114. The sequence control device 114 controls the transmitter 107, the gradient magnetic field power supply unit 112, the receiver 108, etc., and collects measurement data composed of nuclear magnetic resonance signals. Measurement The control unit 210 receives and sends the imaging conditions and pulse sequence to the sequence control device 114, and the sequence control device 114 controls the transmitter 107, the gradient magnetic field power supply unit 112, the receiver 108, etc. to collect measurement data composed of nuclear magnetic resonance signals.

[0034] At this time, when the thinning pattern and the thinning rate are set as imaging conditions, the number of encodings and the amount of encoding applied to the nuclear magnetic resonance signal are controlled so as to be the set thinning pattern and thinning rate.

[0035] Measurement data is collected for each of a plurality of small reception coils, and the measurement data for each channel corresponding to the small reception coil is respectively arranged in a k-space with a predetermined matrix size and used for image reconstruction. That is, in the image processing unit 230 of the computer 200, image reconstruction using the measurement data is performed. At this time, the image processing unit 230 repeatedly performs various operations such as conversion from measurement data to real-space data (image data), synthesis of images for each channel, removal of noise included in the image, and consistency retention processing. Each process included in such sequential reconstruction is implemented using a predetermined algorithm.

[0036] Hereinafter, specific embodiments of the configuration and processing of the image processing unit 230 will be described.

[0037] <First Embodiment> In this embodiment, image reconstruction is performed using the measurement data of each channel measured with a predetermined thinning pattern and the sensitivity distribution of each reception coil. At that time, the denoising of the reconstructed image and the operation for maintaining the consistency between the measurement data of each channel created from the image after denoising and the original measurement data are sequentially processed. As a result, accurate image restoration and denoising are possible without depending on the thinning pattern. In addition, a learning model corresponding to various thinning patterns is not required, and the burden of preparing learning data can be reduced.

[0038] The configuration of the image processing unit 230 of this embodiment will be described with reference to FIG. 2. As shown in the figure, the image processing unit 230 includes a data reception unit 231 and a sequential reconstruction unit 233. The sequential reconstruction unit 233 includes an initial value setting unit 235, a synthesis unit 236, a noise removal unit 237, and a data consistency retention unit 238.

[0039] The data receiving unit 231 receives measurement data (k-space data) of each channel. The initial value setting unit 235 sets an initial value for the iterative calculation by the sequential reconstruction unit 233 based on the measurement data received by the data receiving unit 231. The initial value for the iterative calculation is data to be processed in any of the steps of the iterative calculation, and although there are several modes depending on the step, the following describes a case where the image data input to the synthesis unit 236 is used as the initial value.

[0040] The synthesis unit 236 synthesizes the image data of each channel, which are the initial values set by the initial value setting unit 235, to generate one image data.

[0041] The noise removal unit 237 removes noise (denoising) from the combined image data. There are various denoising methods, but in this embodiment, the denoising process is performed based on the statistical properties of the image and the statistical properties of the noise contained in the image. The details of the method will be described later.

[0042] The data consistency maintaining unit 238 converts the denoised image into k-space data for each channel, integrates the k-space data with the original measurement data, and then converts it back into real space data to update the initial value, thereby maintaining consistency with the measurement data.

[0043] The processing of the image processing unit 230 of this embodiment will be described in detail below with reference to Fig. 3 and Fig. 4. Fig. 3 is a diagram showing the processing flow of the image processing unit 230, and Fig. 4 is a diagram showing the processing diagrammatically.

[0044] [Data reception] When the measurement unit 100 collects measurement data by capturing images with a predetermined thinning pattern and thinning rate, the data receiving unit 231 receives the measurement data for each channel.

[0045] The thinning rate (speed multiplication number) is the reciprocal of the ratio of the number of measurement data to the number of data at each grid point in the k space (if the ratio is 1 / 2, the speed multiplication number is 2).

[0046] The decimation pattern includes an equally spaced decimation pattern (a) in which k-space data is decimated at equal intervals as shown in FIG. 5, a non-equally spaced decimation pattern (b) randomly sampled with a uniform or non-uniform probability density, a pattern (c) decimated at equal or non-equal intervals in which measurement is interrupted halfway and data in some regions is removed, a pattern (d) in which lines under predetermined conditions are removed based on an equal or non-equal interval pattern, etc. The predetermined conditions for the pattern (d) include, for example, the influence of body movement. In this case, there is a method of acquiring a Navier echo corresponding to a measurement point or a measurement line and removing the measurement point or the measurement line whose error from the reference Navier echo is equal to or greater than a certain value. In addition, various body movement detection methods can be used.

[0047] Although not shown in FIG. 5, there is also a pattern in which only the central region is fully scanned to obtain the sensitivity map of each coil from the measurement data, and only the periphery thereof is decimated at equal intervals. Furthermore, there is a case of non-Cartesian scanning in which k-space is scanned radially or helically.

[0048] When the decimation rate and the decimation pattern are preset, the image processing unit 230 performs processing using the set information. Also, information indicating measurement points and non-measurement points such as a mask indicating measurement points may be received by the data reception unit 231 together with the measurement data.

[0049] [Step S301] The initial value setting unit 235 sets the initial value I0 of the iterative calculation of the sequential reconstruction unit 233 based on the input measurement data. In this embodiment, since the iterative calculation is performed based on the real space data, the initial value setting unit 235 converts the measurement data of each receiving coil (hereinafter also referred to as the measurement data of each channel) into real space data and uses it as the initial value. Since the measurement data is data in which a part of k-space is decimated, unmeasured data is estimated and converted into real space data.

[0050] For example, unmeasured points in k-space are zero-filled and subjected to inverse Fourier transform to obtain real-space data. Also, after creating one image data by parallel imaging operations (referred to as parallel reconstruction) such as the known SENSE method, the sensitivity map of each receiving coil is multiplied by the image data, and image data for each channel is created, which may be used as the initial value of each channel image. In the case of non-uniform decimation, parallel reconstruction is performed only for the equally spaced portions.

[0051] In addition, from the studies by the present inventors, when an image created by zero-filling is used as the initial value, it is necessary to increase the number of iterations of the iterative operation in order to remove aliasing. On the other hand, when a parallel reconstruction image is used as the initial value, it has been found that the convergence of the iterative operation is fast because aliasing has been removed.

[0052] The initial value setting unit 235 may further perform filter processing such as Gaussian filtering or bilateral filtering on the above-described image. At this time, the filter strength may be changed according to a G-factor map representing the noise amplification factor of parallel imaging. The noise reduction process performed by the initial value setting unit 235 is a different noise reduction process from the process performed by the noise removal unit 237 described later. By performing parallel reconstruction and filter processing at the time of initial value setting, aliasing and noise are reduced in advance before starting the iterative process, so the number of iterations can be reduced.

[0053] [Steps S302 to S305] The sequential reconstruction unit 233 performs the following iterative operation using the initial value I0 set by the initial value setting unit 235 as the initial value of the image data for each channel.

[0054] [Step S302] Synthesis First, the synthesis unit 236 synthesizes the image data of the receiving coils using the sensitivity map of the receiving coils. As shown in FIG. 6, the synthesis method is to multiply the image data I n of each channel by the coefficient corresponding to the sensitivity map, add them between channels, and obtain one image data M nPerform so-called MAC (Multi Array Coil) synthesis. Note that I n and M n The subscript n of indicates the nth time of the iterative operation (the same applies hereinafter). The initial I n uses the initial value I0. In addition to MAC synthesis, the synthesis of image data may also be an addition process in the k-space where a kernel corresponding to the sensitivity map is convolved with the k-space data of each channel and added between channels.

[0055] [Step S303] Denoising After that, noise Removal section 237 performs denoising processing on the synthesized image data M n in the real space. The denoising process is based on statistical properties. The statistical properties of the image and noise are typically properties reflected in the distribution of pixel values and the distribution of noise. Here, a predetermined algorithm that utilizes the difference between the distribution of a vector composed of all or part of the pixel values of the reconstructed image or a vector obtained by transforming it and the noise distribution is used for denoising processing. Alternatively, since the statistical properties are inherent in the past data, the parameters of a predetermined algorithm for separating noise from the reconstructed image may be adjusted based on the past data.

[0056] Specifically, a combination of wavelet transform and soft thresholding (Method 1), reconstruction using a dictionary representing local structures such as sparse modeling or sparse coding (Method 2), a deep learning model such as a CNN (Convolutional Neural Network) trained to remove Gaussian noise (Method 3), etc. can be mentioned.

[0057] In Method 1, the reconstructed image is sparse, that is, there are many zero elements in the wavelet transform, while the noise is not sparse but has a smaller value compared to the elements of the reconstructed image. The difference in the statistical properties of the two is utilized. Therefore, as shown in the following formula (1), first, the image data (including noise) M is wavelet-transformed (Ψ) and separated into a plurality of elements. On the other hand, soft thresholding processing (S λ) and remove the small value elements. Then, perform the inverse wavelet transform (Ψ -1 ), and image data M'.

[0058]

number

[0059] An example of Method 3 uses a CNN, which is trained to remove noise from a combination of previously captured images with low noise (ground truth image) and high noise at the same position. More specifically, examples of noisy images include an image captured at the same position as the ground truth image with a high thinning rate, or an image created by artificially adding complex Gaussian noise to the ground truth image (complex image or absolute image). Figure 7 shows an example of absolute image noise processing using a CNN. In this example, an absolute image and a phase image are calculated from image M (S701, S702), and each pixel value of the absolute image is divided by the maximum value, resulting in the input to the CNN (S703, S704). The output of the CNN, i.e., the denoised image, is multiplied by the maximum value (S705), and then multiplied by the phase image calculated in S702 (S706) to obtain a denoised image (complex image).

[0060] In the example shown in Figure 7, denoising using CNN is performed only on an absolute value image, but it is also possible to perform denoising on complex images. When processing complex images, a CNN may be configured with two input and output channels, one for the real part and one for the imaginary part, or a known network that handles complex numbers may be used. Processing as a complex image allows denoising including the phase, improving the accuracy of noise removal.

[0061] The structure and number of layers of the CNN are not particularly limited. For example, a CNN with a general structure in which a convolutional layer and an activation layer are sequentially connected may be used. However, preferably, as shown in FIG. 8, a residual learning format in which inputs are added and output is used, which allows the network to learn the difference from the correct image, thereby improving the noise removal accuracy.

[0062] Although higher accuracy can be expected by increasing the number of CNN layers or using a complex network such as U-Net in which layers with multiple resolutions are connected, in this embodiment, the number of CNN layers may be smaller, for example, a CNN with three to five layers, because errors in noise removal are also reduced in the data consistency maintaining unit 238 described below. Since the noise processing in this embodiment is repeated as part of the iterative reconstruction processing, the smaller number of layers allows for lighter processing not only for noise processing but also for the entire iterative reconstruction processing.

[0063] Method 3 (using CNN) can mainly remove Gaussian noise, but it can also be trained to remove aliasing artifacts, ringing artifacts, and motion artifacts contained in the image, treating them as noise. This reduces artifacts along with noise.

[0064] [Step S304] FT Next, the data consistency maintaining unit 238 performs a process to restore the denoised image data M' to k-space data of each channel. Specifically, the image data M' is multiplied by the sensitivity map of each receiving coil (S3041), and then Fourier transformed to obtain k-space data K' of each channel (S3042).

[0065] The data consistency maintenance unit 238 integrates the k-space data K' and the measurement data K0 received by the data reception unit 231 in order to maintain the consistency between the two data (S3043). For example, as shown in FIG. 4, the data of the unmeasured points (estimated data) in the k-space data K' and the data of the actually measured points in the measurement data K0 (the measurement data with the unmeasured points zero-filled to the matrix size of the k-space) are integrated. That is, the measured points in the k-space data K' of each channel are replaced with the measurement data at the same position (Method 1).

[0066] Alternatively, the data of the measured points in the k-space data of each channel and the data at the same position in the measurement data K0 may be weighted and added (Method 2). That is, the integration is performed according to the following formula (2). Method 1 is when λ DC =0, and Method 2 is when λ DC >0.

Equation

[0067] Method 1 is fast in processing, but the noise is removed only from the data with the unmeasured part complemented. In contrast, Method 2 has the advantage that the noise in the measurement data itself is also removed.

[0068] The data consistency maintenance unit 238 performs an inverse Fourier transform on the k-space data Kn of each channel after integration to obtain real-space data.

[0069] [Steps S305, S306] The successive reconstruction unit 233 updates the initial value with the real-space data of each channel integrally created by the data consistency maintenance unit 238 (S306), and repeats the above steps S302 to S304. At this time, the difference between the initial values before and after the update is taken, and when the difference is less than a preset error, the iterative calculation is terminated (S305). With the end of the calculation, the reconstructed image Mn synthesized from the real-space data In before the update becomes the output of the successive reconstruction unit 233.

[0070] As described above, the MRI apparatus (image processing unit) of the present embodiment performs repeated calculations of noise processing and data consistency maintenance processing, thereby enabling successive reconstruction that maintains consistency with measurement data regardless of the decimation pattern during measurement, and also preventing excessive noise. Further, since the folding of the equidistant decimation is removed by the successive reconstruction processing including the synthesis processing using the coil sensitivity distribution and the data consistency maintenance processing, it is also possible to handle equidistant decimation.

[0071] <Modification Example 1 of the First Embodiment> In the first embodiment, the image data of each channel used by the synthesis unit 236 for MAC synthesis is set as the initial value. However, the initial value may be either the real-space data or the k-space data used in each step of the iterative calculation shown in FIG. 4.

[0072] FIG. 9 shows the steps where the initial value can be set. FIG. 9 corresponds to the processing of FIG. 4, and the steps of initial value setting that replace the initial value setting in FIG. 4(a) are shown as (a1), (a2), and (a3). The initial value setting unit 232 sets the initial value to be used in any of these steps.

[0073] In the case of (a1), denoising processing S303 is started with an image reconstructed using the measurement data received by the data reception unit 231 as an initial value. The reconstruction from the measurement data may be MAC synthesis or reconstruction using a parallel imaging operation. (a2) and (a3) use k-space data as an initial value, which is either the initial value before or after integration. In the case of (a2), after performing denoising processing on the reconstructed image Mn or without performing denoising processing, it is converted into k-space data (estimated k-space data in which unmeasured points are estimated) for each channel, and this is used as the initial value K'0 of the k-space data K' before integration. In the case of (a3), after integrating the k-space data K'0 for each channel in (a2) and the measurement data K0, it is used as the initial value K'0 of the estimated k-space data K'.

[0074] In any case, only the starting point of the iterative operation is different, and the processing is the same, and the same effects as those of the first embodiment can be obtained. In the case of the first embodiment or (a1), since real-space data is used as an initial value, it can be configured to receive data after performing various known corrections such as distortion correction in real space, or it can be easily added to perform a filtering process for noise removal and then use it as an initial value. In the case of (a2) and (a3), since the k-space data measured from the device is used as the initial value as it is, the system can be simplified.

[0075] <Modification Example 2 of the First Embodiment> In the first embodiment, the images of each channel were synthesized and denoising processing was performed on the synthesized image. However, in this embodiment, denoising is performed on the image for each channel.

[0076] The processing of this embodiment is shown in FIG. 10. In FIG. 10, the same processing as in FIG. 4 is denoted by the same reference numerals, and the detailed description thereof is omitted. Also in this embodiment, a case is shown where image data for each channel created based on measurement data is set as an initial value, but the same modifications as in the above Modification Example 1 are possible.

[0077] In this embodiment, the synthesis unit 236 creates a reconstructed image (an image in which the foldover is unfolded) for each channel (per-channel reconstruction). While the process of the synthesis unit 236 in the first embodiment generates one image from per-channel data, the process in this embodiment is different in that it creates a reconstructed image for each channel using the information of the sensitivity map. As a method for creating a reconstructed image for each channel, for example, a technique such as convolving the kernel calculated from the sensitivity map with the k-space data is used, like the SPIRiT method.

[0078] The noise removal unit 237 performs noise reduction processing on each of the per-channel reconstructed images created by the synthesis unit 236 (S303). The noise reduction processing is the same as in the first embodiment and is performed using a predetermined algorithm (including CNN) based on the statistical properties of the image and the noise. The image data after noise reduction is returned to the k-space data to obtain the per-channel k-space data K' (S3042). The subsequent integration (S3043), inverse Fourier transform (S3044), and update of the initial value (S305) in the data consistency maintenance unit 238 are the same as in the first embodiment.

[0079] The sequential reconstruction unit 233 takes the difference between the initial values before and after the update, and if the difference is less than a preset error, the iterative calculation is terminated. Then, the synthesis unit 236 performs MAC synthesis on the reconstructed images Mn of each channel immediately before and ends the process.

[0080] In this modification example, the image quality is improved by using a reconstruction method such as the SPIRiT method that is less likely to produce foldover artifacts.

[0081] <Second Embodiment> This embodiment is characterized in that a function for estimating the noise amount of an image is added to the configuration of the first embodiment, and various coefficients and weights (collectively referred to as parameters) used in the iterative calculation can be adjusted based on the estimated noise amount.

[0082] The configuration of the image processing unit 230 according to this embodiment is shown in FIG. 11. In FIG. 11, elements having the same functions as those in FIG. 2 are denoted by the same reference numerals, and redundant descriptions are omitted. As shown in FIG. 11, the image processing unit 230 according to this embodiment includes a noise amount estimation unit 234.

[0083] As shown in FIG. 12, the processing flow in this embodiment is the same as that in the first embodiment and its modified example, except that the estimation of the noise amount and the adjustment based on the noise amount (S310) are added. Therefore, the following description will focus on the different processing.

[0084] The noise amount estimation unit 234 estimates the noise amount using the parallel-reconstructed image. As the parallel-reconstructed image, the reconstructed image generated by the synthesis unit 236 in the first iteration operation, or when the initial value setting unit 235 sets the image data of each channel as the initial value after parallel reconstruction, the reconstructed image used for creating the channel image data can be used.

[0085] As a method for estimating the noise amount from an image, a method of calculating the standard deviation or the mode of the noise as the noise amount σ from the distribution (histogram) of the signal intensity of the image can be adopted. As shown in the histogram of FIG. 13, noise appears at a higher frequency than the image signal in the region where the signal intensity is small. Therefore, the noise amount can be calculated by differentiating from the signal in this distribution.

[0086] Alternatively, instead of using the distribution of the signal intensity, information excluding the subject part may be extracted from the image reconstructed in the same manner as above, and the noise amount may be calculated. For example, the background part of the image may be extracted, and the standard deviation or the mode of the signal intensity of the background part may be used as the noise amount. Also, by creating a spatial differential image of the reconstructed image, an image in which the signal component derived from the subject, whose signal change is gentle compared to the noise, is removed can be obtained, and the noise amount may be calculated from the signal value.

[0087] When calculating the noise amount from the reconstructed image, the image may be corrected by the G-factor or R-factor (decimation rate) of parallel reconstruction. Since the G-factor is information on a map representing the spatial distribution of the noise amplification factor in the parallel reconstruction process, by dividing each pixel value of the image by the G-factor value at the corresponding position, it is possible to calculate the noise amount that does not depend on the G-factor. Although the decimation rate is a single value, generally the noise amount increases in proportion to the square root of the decimation rate. Therefore, by dividing by the square root of the decimation rate, it is possible to calculate the noise amount that does not depend on it. By these corrections, the dependence on the reconstruction process can be reduced and the parameters can be adjusted according to the noise amount of the measurement data itself, so the denoising accuracy becomes stable.

[0088] As a method for estimating the noise amount, in addition to the above-described method, various methods for calculating the noise amount based on the reconstructed image, the measurement data, or an image calculated from them can be used.

[0089] The noise amount calculated by the noise amount estimation unit 234 is used for adjustment such as functions, algorithms, coefficients and thresholds of CNN, etc. used in each step of the iterative operation in the sequential reconstruction unit 233 (mainly the noise removal unit 237), and also for switching of CNN.

[0090] For example, when the noise removal unit 237 performs wavelet transform and soft threshold processing, the threshold λ of the soft threshold processing is adjusted as λ = W·σ or the like. W is a constant and may be set by default or may be adjusted by the user. Also, λ may be determined by a polynomial of the noise amount σ or the like.

[0091] Noise removal unit 237 If it is a CNN, for example, as the CNN, a plurality of CNNs such as a CNN trained with teacher data having a large noise amount and a CNN trained with teacher data having a small noise amount may be prepared, and the CNN may be switched according to the noise amount σ. Specifically, the configuration of the network itself such as the total number of CNNs and the weights of each neuron are switched.

[0092] In the first embodiment, two methods were described as the processing of the data consistency maintaining unit 238. However, the value of "λ" in the common formula (2) of these methods may be varied depending on the noise amount σ, and the method may be switched or the weighting of the data at the time of integration may be adjusted. As an example, when the noise amount is large, integration is performed by adjusting so that the weight of the denoised k-space data becomes large. DC The sequential reconstruction unit 233 performs iterative calculations under conditions adjusted based on the noise amount estimated by the noise amount estimation unit 234. The processing of the sequential reconstruction unit 233 is the same as that of the first embodiment and its modification examples, and the description thereof is omitted.

[0093]

[0094] According to the present embodiment, the intensity of denoising and the like can be adjusted according to the noise amount included in the image, the accuracy of denoising can be improved, and moreover, excessive denoising can be prevented by repeating the denoising process together with the data consistency maintaining process.

[0095] <Modification Example of the Second Embodiment> In the processing of the second embodiment shown in FIG. 12, the noise amount estimation process S310 is performed prior to the iterative calculation. However, as shown in FIG. 14, the noise amount estimation and adjustment may be performed as processes inside the iterative calculation. In this case, in the configuration diagram of FIG. 11, the noise amount estimation unit 234 is changed to be included in the sequential reconstruction unit 233.

[0096] In this modification example, after the initial value is updated after the processing by the data consistency maintaining unit 238, the noise amount estimation and the adjustment of parameters and the like are performed again for the updated image data, and then the subsequent processing (denoising, integration, etc.) is performed. In FIG. 14, step S310 is shown as the process before the synthesis process S302, but it may be the process after the synthesis process S302.

[0097] According to this modification example, although the amount of calculation of the iterative calculation increases, denoising can be performed more appropriately.

[0098] ​As described above, embodiments of the MRI apparatus and the processing method performed by its image processing unit have been explained. However, the processes (techniques) that can be adopted in each step described in these embodiments can be appropriately combined, and thereby it is also possible to improve the efficiency of iterative calculations.

[0099] In addition, the image processing apparatus of the present invention has the functions (for example, the elements shown in FIG. 2) provided by the image processing unit 230 of the above-described MRI apparatus, and can be constructed on a general-purpose computer or workstation. The processing content is the same, and redundant explanations are omitted. Needless to say, such an image processing apparatus may have general image processing functions in addition to the above-described functions, and may also have the attached devices (input / output devices and storage devices) of the computer 200 shown in FIG. 1.

Explanation of Reference Numerals

[0100] 10: MRI apparatus, 100: measurement unit, 200: image processing unit (computer), 210: measurement control unit, 230: image processing unit, 231: data reception unit, 233: sequential reconstruction unit, 234: noise amount estimation unit, 235: initial value setting unit, 236: synthesis unit, 237: noise removal unit, 238: data consistency maintenance unit.

Claims

1. A measurement unit including a plurality of reception coils for collecting nuclear magnetic resonance signals of a subject for each reception coil, a control unit for controlling a collection pattern of the nuclear magnetic resonance signals, and an image processing unit for creating an image of the subject using measurement data composed of the nuclear magnetic resonance signals collected for each reception coil. The image processing unit has a sequential reconstruction unit for reconstructing an image by iterative calculation using the measurement data for each reception coil. The sequential reconstruction unit includes an initial value setting unit for setting an initial value of iterative calculation based on the measurement data for each reception coil, a synthesis unit for synthesizing images for each reception coil into one image, a noise removal unit for removing noise of the one image based on respective statistical properties of the synthesized image and the noise included in the image, a data coherence maintaining unit for multiplying the denoised image by a sensitivity map of each reception coil to create denoised k-space data for each reception coil, and integrating the denoised k-space data and the measurement data to create estimated measurement data, and performing sequential processing of updating the initial value based on the estimated measurement data and repeating synthesis processing, denoising processing, and data coherence processing. A nuclear magnetic resonance imaging apparatus characterized by this.

2. The nuclear magnetic resonance imaging apparatus according to Claim 1, wherein the measurement data is measurement data undersampled in a predetermined decimation pattern, and the sequential reconstruction unit receives a sensitivity map of each reception coil together with the measurement data, and the synthesis unit reconstructs an image using the measurement data and the sensitivity map. A nuclear magnetic resonance imaging apparatus characterized by this.

3. The nuclear magnetic resonance imaging apparatus according to Claim 1, wherein the initial value setting unit sets an estimated image of each reception coil obtained by converting the measurement data for each reception coil into real space data by inverse Fourier transform as an initial value of the image for each reception coil used by the synthesis unit. A nuclear magnetic resonance imaging apparatus characterized by this.

4. The nuclear magnetic resonance imaging apparatus according to Claim 2, wherein the initial value setting unit sets an estimated image of each reception coil created using an image reconstructed from the measurement data for each reception coil and the sensitivity map of the reception coil as an initial value of the image for each reception coil used by the synthesis unit. A nuclear magnetic resonance imaging apparatus characterized by this.

5. The magnetic resonance imaging apparatus according to claim 1, wherein the initial value setting unit sets, as an initial value of the synthesized image used by the noise removal unit, an image reconstructed from the measurement data for each reception coil. The magnetic resonance imaging apparatus is characterized by this.

6. The magnetic resonance imaging apparatus according to claim 1, wherein the initial value setting unit sets, as an initial value, the noise k-space data for each reception coil or the estimated k-space data obtained by integrating the noise k-space data and the measurement data. The magnetic resonance imaging apparatus is characterized by this.

7. The magnetic resonance imaging apparatus according to claim 1, wherein the noise removal unit performs noise removal by utilizing the difference between a vector composed of at least a part of the pixel values of the image to be processed or the distribution of the vector and the noise distribution. The magnetic resonance imaging apparatus is characterized by this.

8. The magnetic resonance imaging apparatus according to claim 7, wherein the noise removal unit performs a sparse conversion that converts the image to be processed into a sparse space, and performs noise removal by performing soft threshold processing on the data after the sparse conversion. The magnetic resonance imaging apparatus is characterized by this.

9. The magnetic resonance imaging apparatus according to claim 1, wherein the noise removal unit includes a neural network trained using past data as learning data, and performs noise removal based on the past data. The magnetic resonance imaging apparatus is characterized by this.

10. The magnetic resonance imaging apparatus according to claim 9, wherein the neural network is trained to remove Gaussian noise. The magnetic resonance imaging apparatus is characterized by this.

11. The magnetic resonance imaging apparatus according to claim 9, wherein the neural network is further trained to remove at least one of wrap-around artifacts, ringing artifacts, and motion artifacts. The magnetic resonance imaging apparatus is characterized by this.

12. The magnetic resonance imaging apparatus according to claim 1, wherein the data consistency maintaining unit creates the estimated measurement data by replacing the data at the measured position (measurement point) among the noise k-space data with the data at the same position in the measurement data. The magnetic resonance imaging apparatus is characterized by this.

13. The magnetic resonance imaging apparatus according to claim 1, wherein a data consistency maintaining unit that performs weighted addition of data of the measured measurement points in the denoised k-space data and data of the measurement points at the same positions in the measurement data to create the estimated measurement data.

14. 2. The magnetic resonance imaging apparatus according to claim 1, The magnetic resonance imaging apparatus further comprises a noise amount estimating unit that estimates the amount of noise in the image or k-space data before denoising by the noise removing unit.

15. 15. The magnetic resonance imaging apparatus according to claim 14, The magnetic resonance imaging apparatus according to claim 1, wherein the noise amount estimation unit estimates the amount of noise using a signal intensity distribution of the image before denoising or an image obtained by removing the object from the image.

16. 15. The magnetic resonance imaging apparatus according to claim 14, a noise amount estimation unit that adjusts, based on the estimated noise amount, at least one of a denoising threshold used by the noise removal unit, a coefficient of a calculation formula used by the data consistency maintenance unit to calculate the estimated measurement data, and, if the noise removal unit includes a neural network, a configuration or weight of the neural network, said magnetic resonance imaging apparatus.

17. 15. The magnetic resonance imaging apparatus according to claim 14, The magnetic resonance imaging apparatus, wherein the noise amount estimating unit is included in the iterative reconstruction unit, and repeats noise amount estimation every time the initial value is updated.

18. An image processing device that processes measurement data collected by a plurality of receiver coils of a magnetic resonance imaging device and reconstructs a denoised image, the image processing device comprising: a sequential reconstruction unit that reconstructs an image by repeated calculations using the measurement data for each of the receiver coils; The successive reconstruction unit includes an initial value setting unit that sets an initial value of iterative calculation based on the measurement data for each reception coil, a synthesis unit that synthesizes the images for each reception coil into one image, a noise removal unit that removes the noise of the one image based on the respective statistical properties of the synthesized image and the noise included in the image, and a data consistency maintaining unit that multiplies the denoised image by the sensitivity map of each reception coil to create denoised k-space data for each reception coil, and integrates the denoised k-space data and the measurement data to create estimated measurement data, and performs a successive process of updating the initial value based on the estimated measurement data and repeating the synthesis process, the denoising process, and the data consistency process. The image processing apparatus is characterized by this.

19. An image processing method for processing measurement data collected by a plurality of reception coils of a magnetic resonance imaging apparatus to create a denoised reconstructed image, including a successive process for maintaining the consistency between the denoised k-space data and the measurement data, wherein the successive process includes a step of setting an initial value of processing based on the measurement data for each reception coil, a step of synthesizing the images for each reception coil into one image, a step of removing the noise of the one image based on the respective statistical properties of the synthesized image and the noise included in the image, a step of multiplying the denoised image by the sensitivity map of each reception coil to create denoised k-space data for each reception coil, and integrating the denoised k-space data and the measurement data to create estimated measurement data, and is characterized by updating the initial value based on the estimated measurement data and repeating the synthesis process, the denoising process, and the data consistency process.

Images