Information processing device and method

Abstract

To utilize a basis with a relatively low contribution rate.SOLUTION: According to one embodiment, an information processing apparatus includes a basis conversion unit, a selection unit, and a basis utilization unit. The basis conversion unit calculates a plurality of bases by performing basis conversion on image data. The selection unit selects one or more specific bases with contribution rates lower than a reference from among the plurality of bases. The basis utilization unit executes data processing by utilizing the specific bases.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] Embodiments disclosed herein and in the drawings relate to information processing devices and methods. [Background technology]

[0002] A technique is known for applying KL (Karhunen-Loeve) expansion to time-series MRI data to obtain major basis pairs (basis pairs with relatively high contribution rates), and then using these major basis pairs to extract correlations between time-series MRI data. [Prior art documents] [Patent Documents]

[0003] [Non-Patent Document 1] Sajan Goud Lingala, Yue Hu, Edward DiBella, Mathews Jacob, “Accelerated dynamic MRI exploiting sparsity and low-rank structure: kt SLR”, IEEE Trans Med Imaging, 2011 May, 30(5), 1042-1054. [Non-Patent Document 2] Hemant K. Aggarwal, Merry P. Mani, Mathews Jacob, “MoDL: Model Based Deep Learning Architecture for Inverse Problems”, arXiv: 1712.02862v4 [cs.CV] 5 Jun 2019. [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] One of the problems that the embodiments disclosed herein and in the drawings aim to solve is the utilization of a base with a relatively low contribution rate. However, the problems that the embodiments disclosed herein and in the drawings aim to solve are not limited to the above problem. Problems corresponding to the effects of each configuration shown in the embodiments described later can also be positioned as other problems. [Means for solving the problem]

[0005] The information processing device according to this embodiment includes a basis conversion unit, a selection unit, and a basis utilization unit. The basis conversion unit calculates a plurality of basis sets by applying a basis conversion to image data. The selection unit selects one or more specific basis sets from the plurality of basis sets whose contribution rate is lower than a standard. The basis utilization unit performs data processing using the specific basis sets. [Brief explanation of the drawing]

[0006] [Figure 1] Figure 1 shows an example of the configuration of the information processing device according to this embodiment. [Figure 2] Figure 2 is a diagram showing the flow of a series of processes related to null space constraint reconstruction by the information processing device according to this embodiment. [Figure 3] Figure 3 is a schematic diagram showing an example of processing including null space constraint reconstruction by the information processing device according to Example 1. [Figure 4] Figure 4 is a schematic diagram of the sparse sampling + ACS line k-space data used in Example 1. [Figure 5] Figure 5 is a schematic diagram showing the process of steps SA2-SA4 in Figure 3. [Figure 6] Figure 6 shows the output images for Example 1(A), Comparative Example (B), and Comparative Example (C). [Figure 7] Figure 7 is a schematic diagram showing an example of processing including null space constraint reconstruction by the information processing device according to Example 2. [Figure 8] Figure 8 is a schematic diagram showing the processes of steps SB2 and SB3 in Figure 7. [Figure 9]FIG. 9 is a diagram schematically showing an example of a process including null space constraint reconstruction by the information processing apparatus according to Example 3. [Figure 10] FIG. 10 is a diagram schematically showing an example of a process including null space constraint reconstruction by the information processing apparatus according to Example 4. [Figure 11] FIG. 11 is a diagram schematically showing a base conversion regarding the parameter axis direction according to Example 5. [Figure 12] FIG. 12 is a diagram schematically showing a base conversion regarding the contrast axis direction according to Example 6. [Figure 13] FIG. 13 is a diagram showing a configuration example of the integrated neural network according to Example 9. [Figure 14] FIG. 14 is a diagram showing a configuration example of the integrated neural network according to Example 10. [Figure 15] FIG. 15 is a diagram showing a series of flows related to null space constraint segmentation by the information processing apparatus according to Example 11. [Figure 16] FIG. 16 is a diagram showing a network configuration example of the deep neural network according to Example 12. [Figure 17] FIG. 17 is a diagram showing a network configuration example of another deep neural network according to Example 12.

MODE FOR CARRYING OUT THE INVENTION

[0007] Hereinafter, embodiments of the information processing apparatus and method will be described in detail with reference to the drawings.

[0008] The information processing device according to this embodiment is a computer that processes various digitized information. The information to be processed is image data collected by medical devices and inspection devices. The medical device is a medical imaging diagnostic device that generates image data for medical purposes. For example, as a medical device, a single modality device such as an X-ray computed tomography (X-ray CT) device, a magnetic resonance imaging (MRI) device, an X-ray diagnostic device, a PET (Positron Emission Tomography) device, a SPECT (Single Photon Emission CT) device, an ultrasound diagnostic device, an optical coherence tomography (optical coherence tomography) device (fundus camera), and an optical ultrasound diagnostic device may be used, or a composite modality device such as a PET / CT device, a SPECT / CT device, a PET / MRI device, or a SPECT / MRI device may be used. Alternatively, as a medical device, an optical camera device used auxiliaryly with a medical imaging diagnostic device or an optical camera device attached to a catheter may be used. The inspection device is an image generation device that generates image data for inspection purposes that are not for medical purposes. For example, X-ray inspection equipment, optical camera equipment, and radar measurement equipment may be used as inspection devices.

[0009] When a medical imaging diagnostic device is an X-ray CT scanner, the stand of the X-ray CT scanner rotates the X-ray tube and X-ray detector around the subject, irradiating the subject with X-rays from the X-ray tube, and detecting the X-rays that pass through the subject with the X-ray detector. The X-ray detector generates an electrical signal with a pulse height corresponding to the detected X-ray dose. This electrical signal is subjected to signal processing such as A / D conversion by the data acquisition circuit. The electrical signal after A / D conversion is called projection data or sinogram data. The console generates a CT image based on the projection data or sinogram data. Projection data and sinogram data are types of raw data. Projection data, sinogram data, and CT images are types of image data.

[0010] When the medical imaging diagnostic device is an MRI device, the MRI device's stand repeatedly applies a gradient magnetic field via a gradient magnetic field coil and an RF pulse via a transmitting coil under the application of a static magnetic field via a static magnetic field magnet. The application of the RF pulse causes the subject to emit an MR signal. The emitted MR signal is received via a receiving coil. The received MR signal is subjected to signal processing such as A / D conversion by the receiving circuit. The MR signal after A / D conversion is called k-space data. The console generates an MR image based on the k-space data. k-space data is a type of raw data. k-space data and MR images are types of image data.

[0011] When a medical imaging diagnostic device is an ultrasound diagnostic device, the ultrasound probe of the ultrasound diagnostic device transmits an ultrasound beam into the subject's body from multiple ultrasound transducers and receives the ultrasound reflected from the subject's body via the ultrasound transducers. The ultrasound transducers generate an electrical signal with a peak value corresponding to the sound pressure of the received ultrasound. This electrical signal is converted using A / D conversion by an A / D converter installed in the ultrasound probe, etc. The electrical signal after A / D conversion is called echo data. The ultrasound diagnostic device itself generates an ultrasound image based on the echo data. Echo data is a type of raw data. Echo data and ultrasound images are types of image data.

[0012] When a medical imaging diagnostic device is a PET scanner, the PET scanner's stand simultaneously measures a pair of 511 keV gamma rays generated by the annihilation of positrons from radionuclides accumulated in the subject with electrons surrounding those radionuclides, using a simultaneous measurement circuit. This generates digital data containing digital values ​​related to the energy values ​​and detection positions of the pair of gamma rays. This digital data is called coincidence data or sinogram data. The console generates a PET image based on the coincidence data or sinogram data. Coincidence data and sinogram data are types of raw data. Coincidence data, sinogram data, and PET images are types of image data.

[0013] When a medical imaging diagnostic device is an X-ray diagnostic device, it generates X-rays from an X-ray tube mounted on a C-arm. An X-ray detector, such as an FPD (Flat Panel Display), mounted on or independently of the C-arm, receives the X-rays generated from the X-ray tube and transmitted through the subject. The X-ray detector generates an electrical signal with a pulse height corresponding to the detected X-ray dose, and this electrical signal is subjected to signal processing such as A / D conversion. The electrical signal after A / D conversion is called projection data or an X-ray image. In the case of cone-beam CT, projection data or X-ray images are also used as raw data. Projection data and X-ray images are types of image data.

[0014] The raw data according to this embodiment is a general term for data that can be converted into an image. In this embodiment, raw data is not limited to original raw data collected by a medical imaging diagnostic device. For example, raw data may be computational raw data generated by applying an inverse transformation process to a medical image. If the raw data is collected by an X-ray CT device, the inverse transformation process is, for example, a forward projection process, and if it is collected by an MRI device, the inverse transformation process is, for example, a Fourier transform process. Furthermore, the raw data according to this embodiment may be hybrid data generated by applying an image transformation to all but one of the dimensions defining the raw data. For example, in the case of three-dimensional raw data, there is hybrid data in which only one or two axes have been image transformed. As an example, hybrid data generated by applying a Fourier transform only with respect to the readout direction of three-dimensional k-space data in magnetic resonance imaging is known. Such hybrid data also falls within the category of raw data according to this embodiment.

[0015] Figure 1 shows an example configuration of the information processing device 1 according to this embodiment. The information processing device 1 may be a computer included in a medical device or an inspection device, or it may be a computer separate from the medical device and the inspection device.

[0016] As shown in Figure 1, the information processing device 1 includes a processing circuit 11, a communication interface 12, a display device 13, an input interface 14, and a storage device 15.

[0017] The processing circuit 11 has a processor such as a CPU (Central Processing Unit). By starting various programs installed in the storage device 15, the processor realizes the acquisition function 111, image conversion function 112, base conversion function 113, selection function 114, base utilization function 115, image processing function 116, data conversion function 117, and output control function 118. Each function 111-118 is not limited to being realized by a single processing circuit. A processing circuit may be configured by combining multiple independent processors, and each processor may realize each function 111-118 by executing a program.

[0018] By implementing the acquisition function 111, the processing circuit 11 acquires image data relating to the subject. The processing circuit 11 may acquire image data from a medical device or inspection device via the communication interface 12, or it may acquire image data from the storage device 15. As described above, image data is a general term for raw data and images based on said raw data. The processing circuit 11 may acquire raw data or images. The acquisition function 111 is an example of an acquisition unit.

[0019] With the implementation of the image conversion function 112, the processing circuit 11 generates an image by applying image conversion to the raw data. The image conversion performed by the image conversion function 112 is an image conversion that does not utilize the basis selected by the selection function 114, which will be described later. As image conversion methods from raw data to images, for example, analytical reconstruction, iterative reconstruction, and machine learning reconstruction can be used. For example, analytical reconstructions related to MR image reconstruction include the Fourier transform or inverse Fourier transform. Analytical reconstructions related to CT image reconstruction include the FBP (filtered back projection) method, the CBP (convolutional back projection) method, or applications thereof. Iterative reconstructions include the EM (expectation maximization) method, the ART (algebraic reconstruction technique) method, or applications thereof. Machine learning reconstruction is a reconstruction method in which all or part of analytical reconstruction or iterative reconstruction is replaced by a neural network. For example, iterative reconstruction incorporating a machine learning model that performs noise reduction is known as a machine learning reconstruction method. Another example of a machine learning reconstruction method is a machine learning model that infers an image from raw data. Furthermore, reconstruction methods such as compressed sensing and parallel imaging reconstruction in magnetic resonance imaging may be used as image conversion techniques. Hereinafter, images generated by the image conversion function 112 will be referred to as converted images. The image conversion function 112 is an example of an image conversion unit.

[0020] With the implementation of the basis transformation function 113, the processing circuit 11 applies a basis transformation to the image data and calculates multiple basis sets. The processing circuit 11 applies a basis transformation to the image data and calculates a number of basis sets corresponding to the number of dimensions of the image data. As the basis transformation in this embodiment, for example, a basis transformation used for dimensionality reduction or dimensionality compression may be used. As such basis transformation methods, for example, principal component analysis (PCA), singular value decomposition, independent component analysis, and encoders of encoder-decoder networks can be used. Principal component analysis is also called KL (Karhunen-Loeve) expansion. The basis transformation function 113 is an example of a basis transformation unit.

[0021] The implementation of the selection function 114 allows the processing circuit 11 to select one or more specific bases from among the multiple bases calculated by the basis conversion function 113, where the contribution rate is lower than the standard. In other words, the specific bases are those among the multiple bases calculated by the basis conversion function 113 that do not have a contribution rate higher than the standard (hereinafter referred to as the primary base). In typical dimensionality reduction, the primary base is used as the base that approximates the data to be compressed, and the specific bases are evaluated as worthless in dimensionality reduction and are rejected. Hereinafter, the specific bases will be referred to as null bases. The number of null bases must be less than the number of bases calculated by the basis conversion function 113, and at least one. The selection function 114 is an example of a selection unit.

[0022] With the implementation of the basis utilization function 115, the processing circuit 11 performs data processing using the null basis selected by the selection function 114. Any data processing that utilizes the null basis is acceptable, such as image transformation or image recognition. When performing image transformation, the processing circuit 11 uses the space spanned by the null basis (hereinafter referred to as the null space) as a constraint condition to apply the image transformation to the raw data and generate an image. In other words, the processing circuit 11 generates the image from the raw data in such a way that it satisfies the constraint condition regarding the projection amount to the null basis. When performing image recognition, the processing circuit 11 uses the null space as a constraint condition to, for example, apply the image transformation to the raw data to internally calculate a time-series feature sequence, and then converts the calculated time-series feature sequence into a recognition result. The feature sequence has the same size as the time-series image (for example, equivalent to 1 channel in the case of a grayscale image) and is data composed of multiple channels (for example, 32 channels). The recognition result is a classification value such as the probability of each of several classes defined separately. The feature sequence is an example of image data. Hereafter, the image generated by the base utilization function 115 will be referred to as the output image.

[0023] With the implementation of the image processing function 116, the processing circuit 11 performs image processing on the converted image and output image. Image processing performed by the image processing function 116 may include, for example, image processing to improve image quality, such as noise reduction and smoothing. However, the image processing performed by the image processing function 116 is not limited to these; conversion processes to display images, such as volume rendering, surface volume rendering, pixel value projection, MPR (Multi-Planer Reconstruction) processing, and CPR (Curved MPR) processing, may also be performed. Furthermore, image processing may include display processes such as gradation processing and scaling, as well as image analysis processing. The image processing function 116 is an example of an image processing unit.

[0024] With the implementation of the data conversion function 117, the processing circuit 11 generates raw data by applying data conversion to various images. Data conversion is a process that performs the reverse process of image conversion, which is carried out in the image conversion function 112.

[0025] With the implementation of the output control function 118, the processing circuit 11 outputs various types of information. For example, the processing circuit 11 displays various types of information via the display device 13 or transmits it to other computers, etc., via the communication interface 12.

[0026] The communication interface 12 is an interface that connects to medical equipment, testing equipment, workstations, PACS (Picture Archiving and Communication System), HIS (Hospital Information System), RIS (Radiology Information System), etc., via a LAN (Local Area Network), etc. The communication interface 12 sends and receives various types of information to and from the connected device. The communication interface 12 is an example of a communication unit.

[0027] The display device 13 displays various information according to the output control function 118 of the processing circuit 11. The display device 13 can be a liquid crystal display (LCD), a CRT (Cathode Ray Tube) display, an organic electroluminescent display (OELD), a plasma display, or any other suitable display. Alternatively, the display device 13 may be a projector. The display device 13 is an example of an output unit.

[0028] The input interface 14 receives various input operations from the user, converts the received input operations into electrical signals, and outputs them to the processing circuit 11. Specifically, the input interface 14 is connected to input devices such as a mouse, keyboard, trackball, switch, button, joystick, touchpad, and touch panel display. The input interface 14 outputs electrical signals to the processing circuit 11 corresponding to the input operations to the input device. The input device connected to the input interface 14 may also be an input device provided on another computer connected via a network or the like. The input interface 14 may also be a voice recognition device that converts audio signals collected by a microphone into instruction signals. The input interface 14 is an example of an input unit.

[0029] The storage device 15 is a storage device that stores various types of information, such as ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), SSD (Solid State Drive), or integrated circuit memory. For example, the storage device 15 stores image data and various programs. In addition to the above storage devices, the storage device 15 may also be a drive device that reads and writes various types of information to and from portable storage media such as CDs (Compact Discs), DVDs (Digital Versatile Discs), flash memory, or semiconductor memory elements. The storage device 15 may be located in another computer connected to the information processing device 1 via a network. The storage device 15 is an example of a storage unit.

[0030] The following describes a series of processing steps related to the reconstruction of an image using null space as a constraint condition (hereinafter referred to as null space constraint reconstruction) by the information processing device 1 according to this embodiment.

[0031] Figure 2 is a diagram showing the flow of a series of processes related to null space constraint reconstruction by the information processing device 1 according to this embodiment. As shown in Figure 2, first, the processing circuit 11 acquires raw data (step S1). In step S1, the processing circuit 11 acquires raw data generated by imaging a subject with a medical device or inspection device.

[0032] Once step S1 is performed, the processing circuit 11 generates a converted image from the raw data acquired in step S1 by implementing the image conversion function 112 (step S2). In step S2, the processing circuit 11 can generate a converted image from the raw data using any of the aforementioned image conversion methods. The converted image may be a 2D image, a 3D image, or a higher-order image.

[0033] When step S2 is performed, the processing circuit 11 implements the basis transformation function 113 to perform a basis transformation on the transformed image generated in step S2 and calculate multiple bases (step S3). When step S3 is performed, the processing circuit 11 implements the selection function 114 to select a null base from the multiple bases calculated in step S3 (step S4).

[0034] Here, we will explain the processing in steps S3 and S4 in detail. While various methods such as principal component analysis, singular value decomposition, independent component analysis, and encoders from encoder-decoder networks can be used as basis transformations for dimensionality reduction and compression, we will use principal component analysis (PCA) as the basis transformation for the purpose of this explanation.

[0035] First, in step S3, the processing circuit 11 calculates a number of eigenvalues ​​and eigenvectors corresponding to the number of dimensions of the transformed image, respectively. Specifically, the processing circuit 11 calculates the covariance matrix of the transformed image and solves the eigenvalue equation of the covariance matrix to calculate a number of eigenvalues ​​and eigenvectors corresponding to the number of dimensions of the transformed image. Eigenvalues ​​correspond to the variance of the transformed image. Eigenvectors are an example of a basis. In this embodiment, the basis calculated by principal component analysis is sometimes called a PCA basis. The PCA basis is calculated as a linear basis.

[0036] Next, in step S4, the processing circuit 11 selects an eigenvector with a contribution rate lower than the criterion from among the multiple eigenvectors (PCA basis) calculated in step S3 as a null basis. Specifically, the processing circuit 11 first calculates the contribution rate of each of the multiple eigenvectors calculated in step S3. The contribution rate is an index that evaluates the degree to which the eigenvector represents all the information contained in the transformed image in an identifiable way. The contribution rate of each eigenvector is defined, for example, by the ratio of the eigenvalue of that eigenvector to the sum of all eigenvalues ​​of all eigenvectors. Next, the processing circuit 11 selects an eigenvector with a contribution rate smaller than the criterion as a null basis. The criterion may be a threshold of the contribution rate, a ratio, or other criteria. For example, the processing circuit 11 may select an eigenvector with a contribution rate below the threshold as a null basis, or it may select an eigenvector that does not meet the specified rank from the top of the contribution rates as a null basis, or it may select an eigenvector that is within the specified rank from the bottom of the contribution rates as a null basis. A null basis may consist of one eigenvector or multiple eigenvectors. The space spanned by one or more null bases is a null space. In this embodiment, principal bases whose contribution ratio is greater than the criterion are rejected.

[0037] Note that the method for selecting a null basis is not limited to the method described above. For example, the processing circuit 11 may consider the eigenvalues ​​as the contribution rates of the eigenvectors and select eigenvectors with eigenvalues ​​lower than the reference as a null basis. Also, the basis transformation is not limited to the method described above. For example, the processing circuit 11 may use the Fourier transform or discrete cosine transform, which are used for data compression of the transformed image, as the basis transformation.

[0038] When step S4 is performed, the processing circuit 11, by realizing the basis utilization function 115, generates an output image from the raw data using the null basis selected in step S4 (step S5). In step S5, the processing circuit 11 generates an output image by applying an image transformation to the raw data acquired in step S1, using the null space spanned by the null basis as a constraint condition. More specifically, the processing circuit 11 generates an output image from the raw data by performing multiple update operations to relatively reduce the error function, which includes a data consistency term based on the difference between the calculated raw data and the raw data, and the projection amount of the updated image onto each null basis. The term relating to the projection amount of the updated image onto each null basis is called the constraint term. The projection amount of the updated image onto the null basis means the projection amount of the updated image onto the null space. In this embodiment, since principal component analysis is performed on the transformed image, the null space is applied in image space. Here, the processing circuit 11 minimizes the error function under the constraint condition that the constraint term is zero or equal to or less than a predetermined value. In other words, the processing circuit 11 generates an output image from raw data under the constraint that the projection amount of the updated image onto the null space is zero or equal to noise.

[0039] Specifically, the processing circuit 11 generates the output image by minimizing the error function EI(x) shown in equation (1) below.

[0040]

number

[0041] As shown in equation (1), the error function EI(x) is defined by the data consistency term ||FSx-y||22 and the constraint term λ||Rx||2 2 It is defined by the sum of the following: Data consistency term ||FSx-y||2 2 This is defined by the norm of the difference between the computed raw data FSx and the raw data y. The computed raw data FSx is defined by the inverse image transformation (data transformation) F between the updated image x and the sensitivity S. Constraint term λ||Rx||² 2 This is defined by the product of the norm of the projection amount Rx of the updated image x onto null space R and the eigenvalue λ. As an example, the L2 norm is used. However, the norm is not limited to the L2 norm; other norms such as the L1 norm may also be used.

[0042] The processing circuit 11 sequentially changes the pixel value of each pixel in the updated image, while applying the constraint term λ||Rx||2 2 We determine an updated image that satisfies both the constraint condition and the condition for minimizing the error function EI(x). The initial image of the updated image may be a transformed image generated by applying an image transformation to the raw data, or it may be an image with an arbitrary pixel value distribution. The constraint condition is a condition to restrict the projection amount Rx of the updated image x onto the null space R to be zero. Since the null space R is spanned by a null basis with a relatively low contribution rate, it is guaranteed that the projection amount Rx of a plausible updated image is zero or noise. Specifically, the constraint condition is the constraint term λ||Rx||2 2 The constraint is set to be zero or less than or equal to a threshold. The threshold is set to a small value close to zero. The minimization condition may be set to the error function EI(x) being zero or less than or equal to a threshold, or it may be set to the number of updates of the updated image. The updated image that satisfies both the constraint and the minimization condition is output as the output image.

[0043] The constraints may be relaxed. In this case, the error function EI(x) is set as shown in equation (2) below.

[0044]

number

[0045] As shown in equation (2), the eigenvalue λ is replaced with the weight matrix W. It is preferable that each matrix element of the weight matrix W is set to a smaller value as the allowable error is larger. For example, each matrix element of the weight matrix W may be set to the reciprocal of the square root of the eigenvalue and the like.

[0046] When step S5 is performed, the processing circuit 11 outputs the output image generated in step S5 by realizing the output control function 118 (step S6). In step S6, the processing circuit 11, for example, displays the output image on the display device 13, stores it in the storage device 15, or transmits it to another computer via the communication interface 12.

[0047] As described above, a series of processes related to null space constraint reconstruction ends.

[0048] Note that the processing flow shown in FIG. 2 is an example, and the present embodiment is not limited thereto. For example, in step S3, the processing circuit 11 may perform a basis transformation on the raw data. In this case, in null space constraint reconstruction, the null space may be applied in the raw data space instead of the image space. For example, the error function ED(x) is expressed as shown in the following equation (3).

[0049]

Equation

[0050] As shown in equation (3), the error function ED(x) is the sum of the data consistency term ||FSx - y||2 2 and the constraint term λ||RFx||2 2 and is defined by the sum of. The data consistency term ||FSx - y||2 2 is defined by the norm of the difference between the calculated raw data FSx and the raw data y, similar to equation (1). The constraint term λ||RFx||2 2The constraint term λ||RFx|| is defined by the product of the norm of the projection amount Rx of the computational raw data Fx based on the updated image x onto the null space R and the eigenvalue λ. As an example, the L2 norm is used. The computational raw data Fx is calculated by applying an inverse image transformation (data transformation) to the updated image x. In this case, the processing circuit 11 sequentially changes the pixel value of each pixel in the updated image while applying the constraint term λ||RFx||2 2 The goal is to determine an updated image that satisfies both the constraint and the condition for minimizing the error function ED(x). In this embodiment as well, the diagonal weight matrix W shown in (2) may be used instead of λ in the error function ED(x).

[0051] According to the above description, the processing circuit 11 calculates multiple bases by applying a basis transformation to image data through the implementation of the basis transformation function 113. However, the data subject to the basis transformation is not limited to image data. The processing circuit 11 calculates multiple bases by applying a basis transformation to signal data. Signal data refers to data that can be processed by a computer, such as image data. Examples of signal data other than image data include waveform data, which is a series of signal intensity values. Waveform data may be time-series data of signal intensity values, frequency-series data, or data of other physical quantity series. Furthermore, according to the above description, data processing by the basis utilization function 115 is assumed to be image transformation or image recognition. However, data processing is not limited to these, and may also include, for example, noise reduction (denoicing) of signal data.

[0052] According to the above embodiment, the information processing device 1 has a processing circuit 11. The processing circuit 11 calculates a plurality of bases by applying a basis transformation to the signal data. The processing circuit 11 selects a null basis from the plurality of bases whose contribution rate is lower than a reference. The processing circuit 11 performs data processing using the null basis.

[0053] The above configuration makes it possible to provide a new data processing method that utilizes null bases, which were previously rejected in principal component analysis and low-rank approximations based on singular value decomposition.

[0054] Next, an embodiment of the information processing device 1 according to this embodiment will be described. In the following embodiment, unless otherwise specified, the information processing device 1 processes image data generated by a magnetic resonance imaging device. Furthermore, principal component analysis (PCA) will be used as the basis transformation method. In addition, the data processing by the basis utilization function 115 will be a null space-constrained reconstruction using the null space as a constraint condition. Furthermore, the various images may be 2D images, 3D images, or higher-order images, but for illustrative purposes, they will be 2D images.

[0055] (Example 1) The processing circuit 11 in Example 1 applies a basis transformation to the image data with respect to the time axis in the basis transformation function 113. The image data subject to the basis transformation is assumed to be a time-series image. Time-series images are known to have strong correlation with respect to the time axis. Decimal reconstruction is performed by utilizing this characteristic. Even if individual images have a low SNR (signal to noise ratio), it is possible to achieve reconstruction with a high SNR.

[0056] Figure 3 is a schematic diagram showing an example of processing including null space constraint reconstruction by the information processing device 1 according to Embodiment 1. As shown in Figure 3, first, the processing circuit 11 acquires time-series sparse sampling + ACS line k-space data by realizing the acquisition function 111. Time-series sparse sampling + ACS line k-space data is an example of time-series k-space data. Time-series sparse sampling + ACS line k-space data is collected by performing sparse sampling acquisition with ACS line acquisition using a magnetic resonance imaging device and transmitted from the magnetic resonance imaging device to the information processing device 1. Sparse sampling acquisition with ACS line acquisition is a data acquisition method that includes sparse sampling acquisition, which collects data by downsampling k-space lines according to a set acceleration rate, and data acquisition of k-space lines (ACS lines) in the central part of the phase encoding direction of k-space. Sparse sampling acquisition with ACS line acquisition is an application of the k-space method, which is a parallel imaging technique, and for example, GRAPPA (generalized auto-calibrating partially parallel acquisition) is known.

[0057] Figure 4 is a schematic diagram of sparse sampling + ACS line k-space data 21. In Figure 4, the solid and dashed arrows represent k-space lines collected by sparse sampling with ACS line collection, and the dotted lines represent k-space lines that were not collected and were therefore thinned out. The dashed arrows represent k-space lines (ACS lines) collected by ACS line collection, and the solid lines represent k-space lines collected by sparse sampling collection. As shown in Figure 4, sparse sampling + ACS line k-space data 21 includes k-space data of k-space lines collected by sparse sampling collection and k-space data of ACS lines. Here, the k-space line in the central part of the phase encoding direction among the ACS lines and k-space lines collected by sparse sampling collection is called the central line, and the k-space data of the central line is called central line data 22. The range of the central part is not particularly limited and can be set to any number of k-space lines. Furthermore, as central line data 22, only the k-space data of the ADS line in the central part may be extracted, or only the k-space data of the k-space lines collected by sparse sampling may be extracted. Note that the ACS line and k-space lines shown in Figure 4 are examples, and the number of k-space lines and the decimation rate are not particularly limited in this embodiment.

[0058] As shown in Figure 3, once time-series sparse sampling + ACS line k spatial data is acquired, the processing circuit 11 extracts the time-series central line data from the time-series sparse sampling + ACS line k spatial data by realizing the acquisition function 111 (step SA1).

[0059] When step SA1 is performed, the processing circuit 11 implements the image conversion function 112 to perform image reconstruction on the central line data of the time series extracted in step SA1 to generate a central image of the time series (step SA2). When step SA2 is performed, the processing circuit 11 implements the basis conversion function 113 to perform principal component analysis in the time axis direction on the central image of the time series generated in step SA2 to calculate multiple PCA basis sets (step SA3). When step SA3 is performed, the processing circuit 11 implements the selection function 114 to select a null basis from the multiple PCA basis sets calculated in step SA3 (step SA4). The details of the processing in steps SA3 and SA4 will now be explained.

[0060] Figure 5 schematically shows the processing in steps SA2-SA4. As shown in Figure 5, first, in step SA2, the processing circuit 11 performs image reconstruction on the time-series central line data extracted in step SA1 to generate the time-series central image 23. In the case of Figure 5, as an example, it is assumed that 16 frames (time points) of central image 23 are generated. The matrix size of each central image 23 is not particularly limited. The number of pixels in each central image 23 is assumed to be K (where K is an integer). The time-series central images 23 are arranged along the time axis in a three-dimensional space (hereinafter referred to as xy-t space) composed of a two-dimensional image axis and a one-dimensional time axis. The two-dimensional image axis is a coordinate axis that defines the position of each pixel in the image. The time axis is a coordinate axis that defines the collection time of the k-space data corresponding to the central image 23. The time-series central images 23 have a strong correlation with respect to the time axis direction.

[0061] As shown in Figure 5, once the central image 23 of the time series is generated, the processing circuit 11 performs principal component analysis on the central image 23 of the time series, which is arranged in xy-t space. Specifically, the processing circuit 11 first records the change in pixel value along the time axis for all pixels k (where k is the pixel index; 1 ≤ k ≤ K) of the central image 23. For example, for each pixel k, a vector Vk(t) is obtained that shows the change in pixel value along the time axis. The number of elements in the vector Vk(t) corresponds to the number of frames, which in the case of Figure 5 is equal to the number of frames "16". In this case, the vector Vk(t) represents a point in 16-dimensional space.

[0062] As shown in Figure 5, once a vector Vk(t) representing the time evolution of pixel values ​​is obtained for all K pixels, the processing circuit 11 performs principal component analysis on the vector Vk(t) of the K pixels and calculates the same number of PCA basis bases Uj (where j is the index of the PCA basis, 1 ≤ j ≤ K) as the number of frames. As shown in Figure 5, if the number of frames is "16", 16 PCA basis bases are calculated. Each PCA basis base Uj has the same number of elements uj,k as the number of frames. For convenience, the PCA basis indices are assumed to be assigned in order from the largest eigenvalue or variance.

[0063] As shown in Figure 5, once the PCA basis Uj is calculated, the processing circuit 11 selects the PCA basis Uj with a smaller contribution rate than the criterion as the null basis Uj. For example, if the criterion is the top 3 variance values ​​of the eigenvalues, PCA basis U1-U3 are set as the primary basis and rejected, and PCA basis U4-U16 are set as the null basis and selected. As mentioned above, the criterion is not limited to the top 3. The processing circuit 11 assigns a flag to the primary basis and stores it in the storage device 15, and assigns a flag to the null basis and stores it in the storage device 15.

[0064] When step SA4 is performed, the processing circuit 11 generates a time-series output image based on the time-series sparse sampling + ACS line k space data and the null basis selected in step SA4 by realizing the basis utilization function 115 (step SA5). Specifically, in step SA5, the processing circuit 11 uses the null space spanned by the null basis selected in step SA4 as a constraint condition and applies an image transformation to the time-series sparse sampling + ACS line k space data to generate an output image. Specifically, the processing circuit 11 minimizes the error function EI(x) exemplified in equation (1) or (2) under the constraint condition that the constraint term is zero or equal to or less than a predetermined value. More specifically, the processing circuit 11 sequentially changes the pixel value of each pixel in the updated image while applying the constraint term λ||Rx||2 2 An updated image is determined that satisfies both the constraint condition and the minimization condition of the error function EI(x). The constraint condition and the minimization condition can be set in the same way as described above. The initial image of the updated image may be a transformed image generated by applying an image transformation to the raw data, or it may be an image with an arbitrary pixel value distribution. In the case of Example 1, which uses k-space data obtained by sparse sampling acquisition, for example, the non-uniform discrete fourier transform (NUDFT) is used as the function F in equation (1) or (2).

[0065] When step SA5 is performed, the processing circuit 11 performs noise reduction processing on the time-series output image generated in step SA5 by realizing the image processing function 116 (step SA6). The output image after noise reduction processing in step SA6 will be called the noise-reduced image. In step SA6, the processing circuit 11 may use a noise reduction filter, a neural network such as a convolutional neural network for noise reduction, or any other method for noise reduction.

[0066] When step SA6 is performed, the processing circuit 11 determines whether or not to terminate the null space constraint reconstruction (step SA7). In step SA7, the processing circuit 11 determines whether or not the termination condition is met. The termination condition may be, for example, that the number of iterations has reached a predetermined number, or that the image quality evaluation index of the SNR of the noise reduction image is below a threshold, or any other arbitrary condition may be set. If the termination condition is not met, the processing circuit 11 determines not to terminate the null space constraint reconstruction, and if the termination condition is met, it determines to terminate the null space constraint reconstruction.

[0067] If it is determined that null space constraint reconstruction should not be terminated (SA7: NO), the processing circuit 11 performs data conversion on the time-series noise reduction image generated in step SA6 by implementing the data conversion function 117 to generate time-series computational k-space data (step SA8). When step SA8 is performed, the processing circuit 11 generates a new time-series output image based on the time-series computational k-space data generated in step SA8 and the null basis selected in step SA4 (step SA5), and performs noise reduction processing on the new time-series output image generated in step SA5 (step SA6). Then the processing circuit 11 again determines whether or not to terminate null space constraint reconstruction (step SA7). In this way, the processing circuit 11 repeats steps SA5-SA8 until it is determined in step SA7 that the termination condition is met.

[0068] If it is determined that null space constraint reconstruction is complete (SA7:YES), the processing circuit 11 implements the output control function 118 to display the time-series noise reduction image generated in step SA6 on the display device 13, save it to the storage device 15, or transmit it to another computer via the communication interface 12.

[0069] With the above steps, the process including null space constraint reconstruction according to Example 1 is completed.

[0070] Figure 6 shows the output images for Example 1(A), Comparative Example (B), and Comparative Example (C). The data conditions for Example 1 are: number of frames = 16, number of phase encodings = 168, number of readouts = 192, number of coils = 1 (single coil), sampling pattern = combination of k-t4x and ACS lines = 32 (acceleration rate = 168 / 66 = 2.5). For Example 1(A), null space-constrained reconstruction was used as the reconstruction condition, with the top 3 eigenvectors rejected and the remaining 13 eigenvectors selected as the null basis. The output image for Example 1(A) is the average image of the time series output images obtained by null space-constrained reconstruction of the sparsely sampled + ACS line k-space data of the time series obtained under the above conditions.

[0071] The data conditions for Comparative Example (B) are the same as those for Example 1 (A). The output image of Comparative Example (B) is the average image of the time-series output images obtained by zero-filling the sparsely sampled + ACS line k-space data of the time series obtained under the above conditions to generate filled k-space data of the time series, and then performing image transformation (non-null space constraint reconstruction) on the filled k-space data of the time series. The data conditions for Comparative Example (C) are: number of frames = 16, number of phase encodes = 168, number of readouts = 192, number of coils = 1 (single coil), sampling pattern = full sampling. The output image of Comparative Example (C) is the average image of the time-series output images obtained by performing image transformation (non-null space constraint reconstruction) on the fully sampled k-space data of the time series obtained under the above conditions.

[0072] As shown in Figure 6, the output image of Example 1(A) can be confirmed to have equivalent image quality to the output image of Comparative Example (C). Therefore, it can be confirmed that the null space-constrained reconstruction using the null basis obtained by principal component analysis in the time axis direction according to Example 1 is useful.

[0073] It should be noted that the process including null space constraint reconstruction according to Example 1 described above is just one example, and various modifications are possible. For example, in the above example, principal component analysis in the time axis direction was applied to the central image of the time series based on the central line data of a time series with a relatively small matrix size in order to reduce the computational load of principal component analysis. However, principal component analysis in the time axis direction may be applied to any transformed image of a time series. For example, it may be applied to a transformed image of a time series based on k-space data of a time series collected by parallel imaging of a real space method such as SENSE (sensitivity encoding) without ACS line acquisition. Furthermore, although principal component analysis in the time axis direction was applied to a transformed image of a time series, it is not necessary for multiple transformed images from multiple time points to be arranged in frame number order along the time axis, and multiple transformed images from multiple time points may be arranged in an order other than frame number order. For example, multiple transformed images may be arranged randomly or pseudo-randomly with respect to frame number.

[0074] As described above, the processing circuit 11 according to Embodiment 1 calculates multiple basis vectors by applying a basis transformation in the time axis direction to a time-series image based on time-series k-space data, selects a null basis with a relatively low contribution from among the calculated multiple basis vectors, and applies a null space constraint reconstruction to the time-series k-space data with the null space spanned by the null basis as a constraint condition to generate a time-series output image.

[0075] According to the above configuration, since the null space spanned by the null basis derived from the basis transformation in the time axis direction is used as a constraint, during the reconstruction process, information with weak temporal correlation contained in the time series image is actively attenuated, and relatively, information with strong temporal correlation is passively emphasized. Thus, according to Example 1, reconstruction using information in the time axis direction becomes possible. Furthermore, since the projection amount of the time series image into the null space is strongly estimated to be zero or noise, the null space-constrained reconstruction according to Example 1 requires less computational load compared to low-rank approximation model reconstruction using the main basis.

[0076] (Example 2) The processing circuit 11 in Example 2 applies a basis transformation to the image data with respect to the image spatial axis direction in the basis transformation function 113. The image data subject to the basis transformation is assumed to be a single image.

[0077] Figure 7 is a schematic diagram showing an example of processing including null space constraint reconstruction by the information processing device 1 according to Embodiment 2. As shown in Figure 7, first, the processing circuit 11 acquires fully sampled k-space data by realizing the acquisition function 111. The fully sampled k-space data is collected by performing full sampling acquisition with a magnetic resonance imaging device and transmitted from the magnetic resonance imaging device to the information processing device 1. Full sampling acquisition is a data acquisition method that collects data without thinning out k-space lines.

[0078] As shown in Figure 7, once fully sampled k-space data is acquired, the processing circuit 11, through the implementation of the image transformation function 112, performs image reconstruction on the fully sampled k-space data to generate a reconstructed image (step SB1). After step SB1 is performed, the processing circuit 11, through the implementation of the basis transformation function 113, performs principal component analysis with respect to the image spatial axis direction on the reconstructed image generated in step SB1 to calculate multiple PCA basis sets (step SB2). After step SB2 is performed, the processing circuit 11, through the implementation of the selection function 114, selects a null basis from among the multiple PCA basis sets calculated in step SB2 (step SB3). The details of the processing in steps SB2 and SB3 will now be explained.

[0079] Figure 8 schematically shows the processing in steps SB2 and SB3. As shown in Figure 8, in step SB1, a reconstructed image 31 based on fully sampled k-space data is generated. As shown in Figure 8, the processing circuit 11 generates multiple local images 33 from the reconstructed image 31, each corresponding to a plurality of local regions 32 set at different locations on the reconstructed image 31. Specifically, the processing circuit 11 sets local regions 32 of a predetermined block size at the corners of the reconstructed image 31 and replicates the pixel value distribution within the set local regions 32 as a local image 33. The processing circuit 11 slides the local regions 32 across the entire area on the reconstructed image 31. During this time, the processing circuit 11 replicates the pixel value distribution within the local regions 32 as a local image 33 at predetermined distance intervals. This generates multiple local images 33 corresponding to a plurality of local regions 32. The plurality of local images 33 are arranged along the image space axis in a three-dimensional space (hereinafter referred to as xy-s space) composed of a two-dimensional image axis and a one-dimensional image space axis. The image spatial axes are coordinate axes that define the position of the local region 32 or local image 33 in the reconstructed image 31. The sequence of multiple local images 33 has a strong correlation with respect to the image spatial axis direction.

[0080] The block size can be any size as long as it is smaller than the matrix size of the reconstructed image 31. For example, if the matrix size of the reconstructed image 31 is 512×512, the matrix size can be set to any size such as 5×5, 7×7, or 11×11. The number of local regions 32 (number of image spatial positions) is also not particularly limited. In Figure 8, for illustrative purposes, the block size is 5×5, meaning the number of pixels in the local image 33 is "25", and the number of image spatial positions is "50". Local regions 32 may or may not overlap with other local regions 32.

[0081] As shown in Figure 8, when multiple local images 33 are generated, the processing circuit 11 performs principal component analysis on the multiple local images 33 arranged in xy-s space. Specifically, first, for all image space positions k of the local image 33 (where k is the index of the image space position; 1 ≤ k ≤ 50), the pixel values ​​of 5 × 5 pixels are arranged to generate a single 25-dimensional vector. For example, for each image space position k, 50 25-dimensional vectors Vk(t) are obtained that show the change in pixel value along the pixel number. The number of elements in vector Vk(t) matches the block size, which is equal to "25" in the case of Figure 8. In this case, vector Vk(t) represents a point in 25-dimensional space.

[0082] As shown in Figure 8, once a vector Vk(t) representing the change in pixel value along the pixel number is obtained for all 50 spatial positions in the image, the processing circuit 11 performs principal component analysis on the 50 vectors Vk(t) and calculates the same number of PCA basis bases Uj as the block size (where j is the index of the PCA basis; 1 ≤ j ≤ 25). As shown in Figure 8, when the block size is "25", 25 PCA basis bases are calculated. Each PCA basis base Uj has the same number of elements uj,k as the block size. For convenience, the PCA basis indices are assumed to be assigned in order from the largest eigenvalue or variance value.

[0083] As shown in Figure 8, once the PCA basis Uj is calculated, the processing circuit 11 selects the PCA basis Uj with a smaller contribution rate than the criterion as the null basis Uj. For example, if the criterion is the top 3 variance values ​​of the eigenvalues, PCA basis U1-U3 are set as the primary basis and rejected, and PCA basis U4-U50 are set as the null basis and selected. As mentioned above, the criterion is not limited to the top 3. The processing circuit 11 assigns a flag to the primary basis to indicate that it is a primary basis and stores it in the memory device 15, and assigns a flag to the null basis to indicate that it is a null basis and stores it in the memory device 15. Although the processing circuit 11 is said to acquire a vector Vk(t) for each local image (image spatial position), it may also acquire a vector Vk(t) for each combined local image obtained by combining two or more adjacent local images.

[0084] When step SB3 is performed, the processing circuit 11 generates an output image based on the fully sampled k-space data and the null basis selected in step SB3 by realizing the basis utilization function 115 (step SB4). Specifically, in step SB4, the processing circuit 11 generates an output image by applying an image transformation to the fully sampled k-space data, using the null space spanned by the null basis selected in step SB3 as a constraint condition. Specifically, the processing circuit 11 minimizes the error function EI(x) exemplified in equation (1) or (2) under the constraint condition that the constraint term is zero or equal to or less than a predetermined value. More specifically, the processing circuit 11 sequentially changes the pixel value of each pixel in the updated image while applying the constraint term λ||Rx||2 2 A new image is determined that satisfies both the constraint and the minimization condition of the error function EI(x). The constraint and minimization conditions can be set in the same way as described above. The initial image of the new image may be a transformed image generated by applying an image transformation to the raw data, or it may be an image with an arbitrary pixel value distribution. In the case of Example 2, which uses k-space data obtained by full sampling acquisition, for example, the inverse discrete Fourier transform (DFT) may be used as the function F in equation (1) or (2).

[0085] When step SB4 is performed, the processing circuit 11, through the implementation of the image processing function 116, performs noise reduction processing on the output image generated in step SB4, thereby generating a noise-reduced image (step SB5). The noise reduction processing in step SB5 can be performed in the same way as the noise reduction processing in step SA6.

[0086] When step SB5 is performed, the processing circuit 11 determines whether or not to terminate the null space constraint reconstruction (step SB6). In step SB6, the processing circuit 11 determines whether or not the termination condition is met. The termination condition may be, for example, that the number of iterations has reached a predetermined number, or that the image quality evaluation index of the SNR of the noise reduction image is below a threshold, or any other arbitrary condition may be set. If the termination condition is not met, the processing circuit 11 determines not to terminate the null space constraint reconstruction, and if the termination condition is met, it determines to terminate the null space constraint reconstruction.

[0087] If it is determined that null space constraint reconstruction should not be terminated (SB6: NO), the processing circuit 11 performs principal component analysis with respect to the image spatial axis direction on the noise reduction image generated in step SB5 to calculate multiple PCA basis vectors (step SB2), selects a null basis vector from the multiple PCA basis vectors calculated in step SB2 (step SB3), generates an output image based on the fully sampled k-space data and the null basis vector selected in step SB3 (step SB4), and performs noise reduction processing on the output image generated in step SB4 (step SB5). Then the processing circuit 11 again determines whether or not to terminate null space constraint reconstruction (step SB6). In this way, the processing circuit 11 repeats steps SB3-SB6 until it is determined in step SB6 that the termination condition is met.

[0088] If it is determined that null space constraint reconstruction is complete (SB6:YES), the processing circuit 11 implements the output control function 118 to display the noise-reduced image generated in step SB5 on the display device 13, save it to the storage device 15, or transmit it to another computer via the communication interface 12.

[0089] With the above steps, the process including null space constraint reconstruction according to Example 2 is completed.

[0090] It should be noted that the process including null space constraint reconstruction according to Example 2 described above is just one example, and various modifications are possible. For example, in the above description, the image data subject to the basis transformation with respect to the image spatial axes was assumed to be a transformed image based on fully sampled k-space data, which is expected to be of high quality in order to ensure the accuracy of the basis transformation with respect to the image spatial axes. However, this embodiment is not limited to this, and the basis transformation may be applied to a transformed image based on sparsely sampled k-space data, or principal component analysis may be applied to past images. For example, if null space constraint reconstruction is applied to the k-space data after contrast in step SB4, principal component analysis may be applied to the transformed image based on the k-space data before contrast in step S2.

[0091] As described above, the processing circuit 11 according to Embodiment 2 calculates multiple basis vectors by applying a basis transformation with respect to the image spatial axis direction to a single image, selects a null basis with a relatively low contribution from among the calculated multiple basis vectors, and applies a null space constraint reconstruction to the k-space data with the null space spanned by the null basis as a constraint condition to generate an output image.

[0092] According to the above configuration, since the null space spanned by the null basis derived from the basis transformation in the image spatial axis direction is used as a constraint, during the reconstruction process, information with weak image spatial correlation contained in a single image is actively attenuated, and relatively, information with strong image spatial correlation is passively emphasized. Thus, according to Example 2, reconstruction using information in the image spatial axis direction becomes possible. Furthermore, since the projection amount of a single image into the null space is strongly estimated to be zero or noise, the null space-constrained reconstruction according to Example 2 requires less computational load compared to low-rank approximation model reconstruction using the main basis.

[0093] (Example 3) The processing circuit 11 according to Embodiment 3 applies a basis transformation to the image data in terms of the time axis direction and the image spatial axis direction in the basis transformation function 113.

[0094] Figure 9 schematically shows an example of processing including null space constraint reconstruction by the information processing device 1 according to Example 3. As shown in Figure 9, first, the processing circuit 11 acquires time-series sparse sampling + ACS line k space data by realizing the acquisition function 111. The time-series sparse sampling + ACS line k space data of Example 3 is the same as the time-series sparse sampling + ACS line k space data of Example 1.

[0095] As shown in Figure 9, the processing circuit 11 applies both a basis transformation in the time axis direction and a basis transformation in the image spatial axis direction to the time-series sparse sampling + ACS line k-space data, and selects a null basis derived from the basis transformation in the time axis direction (hereinafter referred to as the time null basis) and a null basis derived from the basis transformation in the image spatial axis direction (hereinafter referred to as the image spatial null basis).

[0096] Specifically, similar to Example 1, the processing circuit 11 extracts the central line data of the time series from the sparsely sampled time series + ACS line k spatial data (step SC1), performs image reconstruction on the central line data of the time series to generate a central image of the time series (step SC2), performs principal component analysis on the central image of the time series with respect to the time axis direction to calculate multiple PCA basis sets (step SC3), and selects a time null basis set from among the multiple PCA basis sets (step SC4).

[0097] In parallel with steps SC1-SC4, the processing circuit 11 generates an image (hereinafter referred to as the view-sharing image) by performing view-sharing reconstruction based on the time-series sparse sampling + ACS line k-space data through the implementation of the image conversion function 112 (step SC5). Specifically, in step SC5, the processing circuit 11 applies view-sharing to the sparse sampling + ACS line k-space data of any one frame from the time-series sparse sampling + ACS line k-space data, and interpolates the k-space data of the missing k-space lines in the sparse sampling + ACS line k-space data of that frame. Then, the processing circuit 11 applies image conversion to the k-space data after view-sharing to generate a converted image (hereinafter referred to as the view-sharing image).

[0098] When step SC5 is performed, the processing circuit 11 performs principal component analysis on the view-sharing image with respect to the image spatial axis direction to calculate multiple PCA basis sets (step SC6), similar to steps SB2-SB3 in Example 2, and selects an image spatial null basis set from among the multiple PCA basis sets (step SC7). The principal component analysis with respect to the image spatial axis direction can be performed in the same manner as in step SB2 in Example 2.

[0099] When steps SC4 and SC7 are performed, the processing circuit 11, by realizing the basis utilization function 115, generates a time-series output image based on the time-series sparse sampling + ACS line k-space data, the time null basis selected in step SC4, and the image space null basis selected in step SC7 (step SC8). Specifically, the processing circuit 11 minimizes the error function EI(x), exemplified in equation (4) below, under the constraint condition that the constraint term is zero or equal to or less than a predetermined value.

[0100]

number

[0101] As shown in equation (4), the constraint term for Example 3 is, for example, the constraint term λ1||R1x||2 with respect to the time null basis. 2 and constraint term λ²||R²x||2 with respect to the null basis in image space 2 It is defined by the sum of the following. Note that R1 is the matrix representation of the time null basis, λ1 is an eigenvalue of the time null basis R1, R2 is the matrix representation of the image space null basis, and λ2 is an eigenvalue of the image space null basis R2. The processing circuit 11 sequentially changes the pixel value of each pixel of the updated image x, while applying the constraint term λ1||R1x||2 2 and λ2||R2x||2 2 The updated image x is determined that satisfies both the constraint condition and the condition for minimizing the error function EI(x). Note that the processing circuit 11 is λ1||R1x||2 2 Constraints relating to λ²||R²x||2 2 The constraints related to λ1||R1x||2 may be checked alternately, or 2 and λ2||R2x||2 2 The constraints relating to may be repeatedly determined. Furthermore, in equation (4), as shown in equation (2), the diagonal weight matrix W1 may be used instead of λ1, and the diagonal weight matrix W2 may be used instead of λ2.

[0102] When step SC8 is performed, the processing circuit 11 implements the image processing function 116 to perform noise reduction processing on the time-series output image generated in step SC8 (step SC9). The noise reduction processing in step SC9 can be performed in the same way as the noise reduction processing in step SA6.

[0103] When step SC9 is performed, the processing circuit 11 determines whether or not to terminate the null space constraint reconstruction (step SC10). In step SC10, the processing circuit 11 determines whether or not the termination condition is met. The termination condition may be, for example, that the number of iterations has reached a predetermined number, or that the image quality evaluation index of the SNR of the noise reduction image is below a threshold, or any other arbitrary condition may be set. If the termination condition is not met, the processing circuit 11 determines not to terminate the null space constraint reconstruction, and if the termination condition is met, it determines to terminate the null space constraint reconstruction.

[0104] If it is determined that null space constraint reconstruction should not be terminated (SC10: NO), the processing circuit 11 performs data conversion on the time-series noise reduction image generated in step SC9 by implementing the data conversion function 117 to generate time-series computational k-space data (step SC11). When step SC11 is performed, the processing circuit 11 generates a new time-series output image based on the time-series computational k-space data generated in step SC11, the time null basis selected in step SC4, and the image space null basis selected in step SC7 (step SC8), and performs noise reduction processing on the new time-series output image generated in step SC8 (step SC9). Then the processing circuit 11 again determines whether or not to terminate null space constraint reconstruction (step SC10). In this way, the processing circuit 11 repeats steps SC8-SC11 until it is determined in step SC10 that the termination condition is met.

[0105] If it is determined that null space constraint reconstruction is complete (SC10:YES), the processing circuit 11 implements the output control function 118 to display the time-series noise-reduced image generated in step SC9 on the display device 13, save it to the storage device 15, or transmit it to another computer via the communication interface 12.

[0106] With the above steps, the process including null space constraint reconstruction according to Example 3 is completed.

[0107] It should be noted that the process including null space constraint reconstruction described in Example 3 above is just one example, and various modifications are possible. For example, in the above description, principal component analysis in the time axis direction was performed on the central image, but principal component analysis in the time axis direction may be performed on any transformed image, such as a transformed image based on sparse sampling + ACS line k-space data. Also, although principal component analysis in the image spatial axis direction was performed on the view-sharing image, it is not limited to this, and principal component analysis in the image spatial axis direction may be performed on any transformed image, similar to Example 2.

[0108] As described above, the processing circuit 11 according to Embodiment 3 calculates multiple bases by applying a basis transformation in the time axis direction to a time-series image based on time-series k-space data, and selects a time-null basis with a relatively low contribution from among the calculated multiple bases. The processing circuit 11 also calculates multiple bases by applying a basis transformation in the image-space axis direction to a single image, and selects an image-space null basis with a relatively low contribution from among the calculated multiple bases. Then, the processing circuit 11 applies a null-space constraint reconstruction to the time-series k-space data, using the null space spanned by the time-null basis and the null space spanned by the image-space null basis as constraint conditions, to generate a time-series output image.

[0109] According to the above configuration, reconstruction becomes possible using both information related to the time axis and information related to the image spatial axis. Therefore, an improvement in image quality is expected compared to Examples 1 and 2.

[0110] (Example 4) Example 4 is a modification of Example 3. The processing circuit 11 in Example 4 obtains an image space null basis using the k-space data (hereinafter referred to as T1-weighted k-space data) collected by T1-weighted data acquisition in null space constraint reconstruction of time-series k-space data (hereinafter referred to as T2-weighted k-space data) collected by T2-weighted data acquisition.

[0111] Figure 10 schematically shows an example of processing including null space constraint reconstruction by the information processing device 1 according to Embodiment 4. As shown in Figure 10, first, the processing circuit 11 acquires time-series T2-weighted k-space data and T1-weighted k-space data by realizing the acquisition function 111. The T2-weighted k-space data is acquired by performing T2-weighted data acquisition on a subject using a magnetic resonance imaging device. The T1-weighted k-space data is acquired by performing T1-weighted data acquisition on the same subject using a magnetic resonance imaging device. The order in which T2-weighted data acquisition and T1-weighted data acquisition are performed is not particularly limited; T2-weighted data acquisition may be performed before or after T1-weighted data acquisition.

[0112] In Example 4, since T2-weighted data acquisition is used for principal component analysis in the time axis direction, T2-weighted data acquisition requires video imaging and is preferably performed at a higher acceleration rate compared to T1-weighted data acquisition. In Example 4, since T1-weighted data acquisition is used for principal component analysis in the image spatial axis direction, T1-weighted data acquisition requires high sampling and is preferably performed at a lower acceleration rate compared to T2-weighted data acquisition. For example, T2-weighted data acquisition may be performed with sparse sampling at an acceleration rate of "6x", and T1-weighted data acquisition may be performed with sparse sampling at an acceleration rate of "2x". Of course, these acceleration rates are examples and are not limited to the above values. Also, T1-weighted data acquisition may be performed with an acceleration rate of "1x", i.e., full sampling. The T1-weighted k-space data may be the T1-weighted k-space data of one frame from a time-series T1-weighted data, or it may be the k-space data of one frame based on time-series T1-weighted data such as the average data of the time-series T1-weighted data.

[0113] As shown in Figure 10, the processing circuit 11, similar to Example 1, extracts time-series T2-weighted central line data from time-series T2-weighted k-space data (step SD1), performs image reconstruction on the time-series T2-weighted central line data to generate a time-series T2-weighted central image (step SD2), performs principal component analysis on the time-axis direction of the time-series T2-weighted central image to calculate multiple PCA basis sets (step SD3), and selects a time-null basis set from among the multiple PCA basis sets (step SD4).

[0114] In parallel with steps SD1-SD4, the processing circuit 11 generates a T1-weighted image by performing image reconstruction on the T1-weighted k-space data through the implementation of the image conversion function 112 (step SD5). In step SD5, the processing circuit 11 may generate a T2-weighted image by performing PI (parallel imaging) reconstruction on the T1-weighted k-space data to reduce aliasing artifacts caused by sparse sampling acquisition. Instead of PI reconstruction, the processing circuit 11 may zero-fill missing k-space lines in the T1-weighted k-space data and reconstruct the image based on the zero-filled k-space data to generate a T1-weighted image. Alternatively, if time-series T1-weighted k-space data has been collected, the processing circuit 11 may generate a T1-weighted image by performing view-sharing reconstruction on the time-series T1-weighted k-space data.

[0115] When step SD5 is performed, the processing circuit 11 performs principal component analysis on the T1-weighted image with respect to the image spatial axis direction to calculate multiple PCA basis sets (step SD6), similar to steps SB2-SB3 in Example 2, and selects an image spatial null basis set from among the multiple PCA basis sets (step SD7). The principal component analysis with respect to the image spatial axis direction can be performed in the same manner as in step SB2 in Example 2.

[0116] Once steps SD4 and SD7 are performed, the processing circuit 11, by realizing the basis utilization function 115, generates a time-series T2-enhanced output image based on the time-series T2-enhanced k-space data, the time null basis selected in step SD4, and the image-space null basis selected in step SD7 (step SD8). Null space constraint reconstruction in step SD8 can be performed in the same manner as the null space reconstruction in Example 3.

[0117] When step SD8 is performed, the processing circuit 11, through the implementation of the image processing function 116, performs noise reduction processing on the time-series output image generated in step SD8, thereby generating a time-series T2-weighted noise-reduced image (step SD9). The noise reduction processing in step SD9 can be performed in the same way as the noise reduction processing in step SA6.

[0118] When step SD9 is performed, the processing circuit 11 determines whether or not to terminate the null space constraint reconstruction (step SD10). In step SD10, the processing circuit 11 determines whether or not the termination condition is met. The termination condition may be, for example, that the number of iterations has reached a predetermined number, or that the image quality evaluation index of the SNR of the T2-weighted noise reduction image is below a threshold, or any other arbitrary condition may be set. If the termination condition is not met, the processing circuit 11 determines not to terminate the null space constraint reconstruction, and if the termination condition is met, it determines to terminate the null space constraint reconstruction.

[0119] If it is determined that null space constraint reconstruction should not be terminated (SD10: NO), the processing circuit 11 performs data conversion on the time-series T2-weighted noise-reduced image generated in step SD9 by implementing the data conversion function 117 to generate time-series T2-weighted calculation k-space data (step SD11). When step SD11 is performed, the processing circuit 11 generates a new time-series T2-weighted output image based on the time-series T2-weighted calculation k-space data generated in step SD11, the time null basis selected in step SD4, and the image space null basis selected in step SD7 (step SD8), and performs noise reduction processing on the new time-series T2-weighted output image generated in step SD8 (step SD9). Then the processing circuit 11 again determines whether or not to terminate null space constraint reconstruction (step SD10). In this way, the processing circuit 11 repeats steps SD8-SD11 until it is determined in step SD10 that the termination condition is met.

[0120] If it is determined that null space constraint reconstruction is complete (SD10:YES), the processing circuit 11 implements the output control function 118 to display the time-series T2-weighted noise-reduced image generated in step SD9 on the display device 13, save it to the storage device 15, or transmit it to another computer via the communication interface 12.

[0121] With the above steps, the process including null space constraint reconstruction according to Example 4 is completed.

[0122] The process including null space constraint reconstruction described in Example 4 above is merely an example, and various modifications are possible. For example, in the above example, T1-weighted k-space data and T2-weighted k-space data were used as the data acquisition method, but this example is not limited to this, and k-space data obtained by any data acquisition method may be used. For example, k-space data obtained by any data acquisition method, such as proton density-weighted k-space data obtained by proton density-weighted data acquisition or diffusion-weighted k-space data obtained by diffusion-weighted image acquisition, may be replaced with T1-weighted k-space data and / or T2-weighted k-space data.

[0123] As described above, the processing circuit 11 according to Example 4 calculates multiple bases by applying a basis transformation in the time axis direction to a time-series image based on time-series k-space data obtained by the first data acquisition method, and selects a time-null basis with a relatively low contribution from among the calculated multiple bases. The processing circuit 11 also calculates multiple bases by applying a basis transformation in the image-space axis direction to a single image obtained by the second data acquisition method, and selects an image-space null basis with a relatively low contribution from among the calculated multiple bases. Then, the processing circuit 11 applies a null-space constraint reconstruction to the time-series k-space data obtained by the first data acquisition method, using the null space spanned by the time-null basis and the null space spanned by the image-space null basis as constraint conditions, to generate a time-series output image.

[0124] With the above configuration, a time null basis can be obtained using images based on a data acquisition method that easily provides temporal correlation, and an image space null basis can be obtained using images based on a data acquisition method that easily provides image space correlation. Therefore, it is expected that the image quality of the output image will improve.

[0125] (Example 5) The processing circuit 11 in Example 5 applies a basis transformation to the image data with respect to the scan parameter value axis direction in the basis transformation function 113. The basis transformation with respect to the scan parameter value axis direction is a basis transformation applied to image data arranged along the scan parameter value axis. The scan parameter value axis refers to the coordinate axis related to the scan parameter value. As a scan parameter, for example, the b value, which is the applied intensity of MPG (motion probing gradient) in diffusion-weighted image acquisition, is used.

[0126] Figure 11 is a schematic diagram showing the basis transformation with respect to the parameter axis direction according to Embodiment 5. As shown in Figure 11, first, the processing circuit 11 generates multiple diffusion-weighted images 41 based on multiple k-space data for multiple b values. Multiple k-space data for multiple b values ​​are collected by a magnetic resonance imaging apparatus by performing diffusion-weighted image acquisition while sequentially changing the b value, which is a scan parameter. The multiple diffusion-weighted images 41 are arranged along the b value axis in a three-dimensional space (hereinafter referred to as xy-b space) composed of a two-dimensional image axis and a one-dimensional b value axis. The b value axis is a coordinate axis that defines the b value of the diffusion-weighted image acquisition from which the k-space data corresponding to the diffusion-weighted image 41 was collected. The multiple diffusion-weighted images 41 have a strong correlation with respect to the b value axis direction.

[0127] The processing circuit 11 performs principal component analysis on multiple diffusion-weighted images 41 arranged in xy-b space. Specifically, it stores the changes in pixel values ​​along the b-value axis for all pixels in the diffusion-weighted images 41. For example, a vector showing the change in pixel value along the b-value axis is obtained for each pixel. The number of elements in this vector matches the number of diffusion-weighted images. The processing circuit 11 then performs principal component analysis on multiple pixel value change vectors for multiple pixels to calculate PCA bases equal to the number of b-values ​​(number of diffusion-weighted images 41), and selects a null base (hereinafter referred to as the b-value null base) from among the PCA bases. Subsequently, the processing circuit 11 can generate an output image by performing image transformations on multiple k-space data related to multiple b-values, using the null space spanned by the b-value null base as a constraint.

[0128] Furthermore, the scan parameters in Example 5 are not limited to the b value alone; any scan parameters such as echo time (TE) or repetition time (TR) may be used.

[0129] As described above, the processing circuit 11 according to Example 5 calculates multiple bases by applying a basis transformation with respect to the scan parameter axis direction to multiple images based on multiple k-space data related to multiple scan parameter values, selects a null base with a relatively low contribution from among the calculated multiple bases, and applies a null space constraint reconstruction to the multiple k-space data with the null space spanned by the null base as a constraint condition to generate multiple output images.

[0130] According to the above configuration, since the null space spanned by the null basis derived from the basis transformation in the scan parameter axis direction is used as a constraint, during the reconstruction process, weakly correlated information regarding scan parameter values ​​included in the scan parameter sequence image is actively attenuated, and relatively, strongly correlated information regarding scan parameter values ​​is negatively emphasized. Thus, according to Example 5, reconstruction using information in the scan parameter value axis direction becomes possible.

[0131] (Example 6) The processing circuit 11 in Example 6 applies a basis transformation to the image data with respect to the contrast axis direction in the basis transformation function 113. The basis transformation with respect to the contrast axis direction is a basis transformation applied to image data arranged along the contrast axis. The contrast axis refers to the coordinate axis related to the contrast type. Examples of contrast types include T1-weighting, T2-weighting, and proton density-weighting.

[0132] Figure 12 schematically shows the basis transformation with respect to the contrast axis direction according to Example 6. As shown in Figure 12, multiple MR images with different contrast types are acquired. For example, a T1-weighted image 51 corresponding to T1 weighting, a T2-weighted image 52 corresponding to T2 weighting, and a proton density-weighted image 53 corresponding to proton density weighting are acquired. The T1-weighted image 51, T2-weighted image 52, and proton density-weighted image 53 are acquired by a magnetic resonance imaging apparatus by sequentially changing TR and TE while acquiring data. The T1-weighted image 51, T2-weighted image 52, and proton density-weighted image 53 are arranged along the contrast axis in a three-dimensional space (hereinafter referred to as xy-c space) composed of a two-dimensional image axis and a one-dimensional contrast axis. The T1-weighted image 51, T2-weighted image 52, and proton density-weighted image 53 have a strong correlation with respect to the contrast axis.

[0133] The processing circuit 11 performs principal component analysis on the T1-weighted image 51, T2-weighted image 52, and proton density-weighted image 53 arranged in the xy-c space. Specifically, it stores the changes in pixel values ​​along the contrast axis for all pixels in the T1-weighted image 51, T2-weighted image 52, and proton density-weighted image 53. For example, a vector showing the change in pixel value along the contrast axis is obtained for each pixel. The number of elements in this vector matches the number of images arranged in the xy-c space. The processing circuit 11 then performs principal component analysis on multiple pixel value change vectors for multiple pixels to calculate PCA basis vectors equal to the number of contrast species (number of images arranged in the xy-c space), and selects a null basis (hereinafter referred to as the contrast-null basis) from among the PCA basis vectors. Subsequently, the processing circuit 11 can generate an output image by performing image transformations on multiple k-space data for multiple contrast species, using the null space spanned by the contrast-null basis as a constraint condition.

[0134] Furthermore, the contrast type used in Example 6 is not limited to T1-weighted, T2-weighted, or proton density-weighted images; any contrast such as T1, T2, or center frequency (f0) may be used. Also, while the above example uses a single image for each contrast type, two or more images may be used for each contrast type. For example, two or more images of the same contrast type may have different scan parameter values ​​such as b-value, TE, or TR.

[0135] As described above, the processing circuit 11 according to Example 6 calculates multiple bases by applying a basis transformation in the contrast axis direction to multiple images based on multiple k-space data for multiple contrast types, selects a null base with a relatively low contribution from the calculated multiple bases, and applies a null space constraint reconstruction to the multiple k-space data with the null space spanned by the contrast-null base as a constraint condition to generate multiple output images.

[0136] According to the above configuration, since the null space spanned by the contrast null basis derived from the basis transformation along the contrast axis is used as a constraint, during the reconstruction process, weakly correlated information regarding contrast contained in the contrast sequence image is actively attenuated, and relatively, strongly correlated information regarding contrast is negatively emphasized. Thus, according to Example 6, reconstruction using information along the contrast axis becomes possible.

[0137] (Example 7) In some of the embodiments described above, null space constraint reconstruction using the null basis was applied to the k-space data corresponding to the image used to calculate the null basis. However, this embodiment is not limited to this, and null space constraint reconstruction using the null basis may be applied to k-space data other than the k-space data corresponding to the image used to calculate the null basis. For example, if data collection is performed N times (where N is any integer) for the same subject, a null basis may be calculated from a transformed image based on the k-space data obtained from the p-th (1≦p≦N) data collection, and null space constraint reconstruction using the null basis may be applied to the k-space data obtained from the q-th (1≦q≦N, q≠p) data collection. The temporal relationship between p and q is not particularly limited; p may be earlier or later than q. The p-th and q-th data collections do not need to be performed in the same examination; for example, the p-th data collection may have been performed in a past examination several days, weeks, months, or years prior to the examination of the q-th data collection.

[0138] Performing the p-th and q-th data acquisitions using the same data acquisition method can guarantee the accuracy of null space constraint reconstruction for the q-th data acquisition compared to performing them using different data acquisition methods. However, the p-th and q-th data acquisitions may be performed using different data acquisition methods. That is, the null basis obtained for images based on k-space data using the first data acquisition method may be reused for null space constraint reconstruction on k-space data using the second data acquisition method. Also, performing the p-th and q-th data acquisitions on the same subject can guarantee the accuracy of null space constraint reconstruction for the q-th data acquisition compared to performing them on different subjects. However, the p-th and q-th data acquisitions may be performed on different subjects. That is, the null basis obtained for the first subject may be reused for the other second subject.

[0139] (Example 8) The null space constraint reconstruction described in at least one of the above embodiments can be performed by a neural network such as a CNN. In this case, for example, the error function of equation (1), (2), (3), or (4) can be incorporated into a neural network such as a CNN using the MoDL (model-based reconstruction using deep learned priors) method. In this case, the null space constraint reconstruction and subsequent post-processing such as noise reduction can also be incorporated into the neural network.

[0140] As an example of comparison with the spatial constraint reconstruction according to this embodiment, a self-calibration type low-rank model reconstruction using the main space spanned by the main basis vectors can be cited. This self-calibration type low-rank model reconstruction requires time to construct the low-rank model. This is particularly noticeable when constructing a low-rank model using singular value decomposition. Furthermore, self-calibration type low-rank model reconstruction is considered difficult to incorporate into neural networks such as CNNs. This is because it is necessary to project image data onto the low-rank model once and evaluate the error between the projection amount and the previous projection amount. While pre-trained low-rank models have fewer processing time drawbacks, the dimensionality reduction effect is reduced because the basis vectors are not based on the image data being reconstructed. Thus, self-calibration type low-rank model reconstruction is difficult to use in practice.

[0141] On the other hand, the null space constraint reconstruction according to this embodiment is more practically usable than the self-calibration type low-rank model reconstruction described in the comparative example. For example, as shown in equation (1), in null space constraint reconstruction, there is no need to evaluate the error of the projection amount into null space, making it easy to incorporate into a neural network. Also, since the projection amount onto the null basis, which is not the main basis of the image data, is zero or noise, the computational load on projection amounts and other factors is reduced.

[0142] (Example 9) In some of the embodiments described above, the error function EI(x) or error function ED(x) includes a data consistency term, as shown in equations (1)-(4). However, this embodiment is not limited to this. Even if the data consistency term is unknown, for example, the following equation (5) may be considered as a constraint equation (constraint term) and the image x may be updated. Note that equation (5) corresponds to equation (1) and corresponds to an example of applying the null space R in the image space.

[0143]

number

[0144] The null space constraint reconstruction method described above assumes the existence of a data consistency term. However, regardless of the presence or absence of a data consistency term, methods such as projection onto convex sets (POCS) can be used to satisfy each constraint equation. For example, if you want to apply the constraint equations to a feature sequence rather than a time-series image, a data consistency term cannot be obtained, but POCS can still be used in such cases. It is also possible to apply POCS to time-series images.

[0145] For example, when using time-series k-space data as raw data, the processing circuit 11 according to Embodiment 9 first generates a time-series image by applying image transformation to the time-series k-space data through the implementation of the image transformation function 112. Next, the processing circuit 11 solves the minimization problem of the constraint equation (5) above, which is defined by the norm of the projection amount of the time-series image onto null space by the convex projection method, through the implementation of the base utilization function 115, and generates an output image.

[0146] When using POCS, the projection Rx of the feature sequence x onto the null space R is realized by POCS. The projection of the feature sequence onto the null space belongs to a convex set. The constraint equation in equation (5) involves the composition of I constraint equations as a projection operation onto the null space using POCS. The number I corresponds to the number of time-series images to be projected, in other words, the number of iterations of the sequential operation in null-space constraint reconstruction using equation (1), etc. The i-th (1≦i≦I) constraint equation represents the projection of the i-th (1≦i≦I) feature sequence onto the null space.

[0147] When POCS(i) represents the projection amount into null space R required to satisfy the i-th constraint equation without error, the projection amount into null space R is expressed as max(0, POCS(i) - POCS_threshold) (where the separately defined constant POCS_threshold satisfies POCS_threshold ≥ 0) or POCS(i) * POCS_weight (where the separately defined constant POCS_weight satisfies 0 ≤ POCS_weight < 1). POCS(i) - POCS_threshold represents uniformly reducing the projection amount into null space R by POCS_threshold, and if it is negative, the i-th projection operation is stopped by a max operation between the negative value and 0. POCS(i) * POCS_weight is equivalent to multiplying the projection amount into null space R by POCS_weight and then multiplying that projection amount by a constant. In this way, the influence of the constraint equations can be controlled by adjusting POCS_threshold and POCS_weight.

[0148] Similar to equation (2), even when the error function does not include a data consistency term, it is possible to replace the eigenvalues ​​λ with the weight matrix W, as shown in equation (6).

[0149]

number

[0150] For example, each element of the weight matrix W may be set to a value obtained by multiplying the eigenvalue λ by POCS_threshold, or by multiplying the reciprocal of the square root of the eigenvalue λ by POCS_weight.

[0151] These constraints can be implemented, for example, by a neural network that incorporates the projection operation of POCS as a linear operation that does not have coefficients to be learned. For example, it can be implemented by a neural network in which neural network layers that perform arbitrary processing (hereinafter referred to as NN layers) and neural network layers that correspond to POCS operations (hereinafter referred to as POCS layers) are alternately inserted.

[0152] Figure 13 shows an example configuration of the integrated neural network 60 according to Embodiment 9. As shown in Figure 13, the integrated neural network 60 is a deep neural network having an image transformation layer 61, an NN layer 62, a POCS layer 63, an NN layer 64, and a POCS layer 65. The image transformation layer 61 is a neural network that processes time-series k-space data into a feature sequence. The NN layers 62 and 64 are neural networks that perform arbitrary processing on the feature sequence. The NN layers 62 and 64 may calculate adjustment values ​​such as POCS_threshold and POCS_weight to be used in the subsequent POCS layers 63 and 65. The POCS layers 63 and 65 are neural networks that perform null-space constraint reconstruction by POCS calculations according to constraint equations such as equation (5) or (6), and generate a time-series output image from the feature sequence.

[0153] The integrated neural network 60 can be generated by pre-training or end-to-end training. Pre-training can be performed, for example, by following these steps: First, the processing circuit 11 individually trains each of the NN layer 62, POCS layer 63, NN layer 64, and POCS layer 65. Next, the processing circuit 11 connects the NN layer 62, POCS layer 63, NN layer 64, and POCS layer 65 as shown in Figure 13 and trains them as a single deep neural network. This makes it possible to generate the integrated neural network 60. In the case of end-to-end training, the processing circuit 11 should train a single deep neural network to take time-series k-space data as input and output a time-series output image.

[0154] By directly connecting a unit layer having NN layer 62 and POCS layer 63 with a unit layer having NN layer 64 and POCS layer 65, it becomes possible to improve the image quality of time-series images. The number of unit layers connected in series is not limited to two, but may be three or more. Furthermore, the integrated neural network 60 may include any layers other than the image conversion layer, NN layer, and POCS layer.

[0155] (Example 10) The signal data in Example 10 is assumed to be MRS (MR Spectroscopy) spectral data collected by a magnetic resonance imaging (MRI) system. The MRS spectral data is waveform data of the MRS spectrum, which represents the frequency series of signal intensity values. More specifically, the MRS spectrum represents the signal intensity value for each chemical shift frequency. The signal intensity value depends on the content of substances such as molecules and metabolites that have the corresponding chemical shift frequency. Furthermore, the data processing performed by the base utilization function 115 in Example 10 is assumed to be noise reduction.

[0156] First, through the implementation of the acquisition function 111, the processing circuit 11 acquires MRS spectral data collected by the magnetic resonance imaging device. The MRS spectral data includes multiple MRS spectra for multiple locations. Next, through the implementation of the basis transformation function 113, the processing circuit 11 applies a basis transformation to the MRS spectral data in the frequency axis direction. That is, the processing circuit 11 calculates multiple bases by applying a basis transformation to the multiple MRS spectra included in the MRS spectral data. Through the implementation of the selection function 114, the processing circuit 11 selects a specific base (null base) from among the multiple bases. Then, through the implementation of the basis utilization function 115, the processing circuit 11 applies noise reduction processing to the multiple MRS spectra using the null space spanned by the null base as a constraint condition to generate multiple noise-reduced MRS spectra.

[0157] As an example, suppose MRS spectra are collected at 25 locations in a two-dimensional xy space, and each MRS spectrum contains signal intensity values ​​for 10 molecules. In this case, the 25 MRS spectra are arranged along the frequency axis in a three-dimensional space (hereinafter referred to as xy-f space) composed of a two-dimensional spatial axis and a one-dimensional frequency axis. Principal component analysis is performed on the 25 MRS spectra arranged in xy-f space to calculate 25 PCA basis sets. Then, approximately 5 major basis sets are rejected from the 25 PCA basis sets, and the remaining PCA basis sets are selected as null basis sets.

[0158] Noise reduction processing may be performed, for example, using equations (1)-(4), or by minimizing the error function EI(x) shown in equation (5) or (6) above using POCS. In Example 10, x represents MRS spectral data to which arbitrary noise reduction processing has been applied. R represents the selected null basis. Specifically, in the noise reduction processing, the processing circuit 11 first performs arbitrary noise reduction processing on the MRS spectral data using a smoothing filter or the like. Then, the processing circuit 11 applies the noise-reduced MRS spectral data to the null space R to calculate the projection amount Rx, and calculates the error function EI(x) shown in equation (5) or (6) based on the projection amount Rx. The processing circuit 11 repeats the noise reduction processing and the calculation of the error function to determine the MRS spectral data x that minimizes the error function EI(x). This completes the noise reduction processing.

[0159] Figure 14 shows an example configuration of the integrated neural network 70 according to Example 10. As shown in Figure 14, the integrated neural network 70 is a deep neural network having an NN layer 71, a POCS layer 72, an NN layer 73, and a POCS layer 74. NN layers 71 and 73 perform noise reduction processing on the MRS spectral data. NN layers 62 and 64 may calculate adjustment values ​​such as POCS_threshold and POCS_weight to be used in the subsequent POCS layers 63 and 65. POCS layers 72 and 74 perform POCS calculations on the MRS spectral data from NN layers 71 and 73 according to constraint equations such as equation (5) or (6). The POCS layer 74 outputs the final noise-reduced MRS spectral data. Note that the number of NN layers 71, 73 and POCS layers 72, 74 is not limited to two; any number of one or more is acceptable. Alternatively, an NN layer may be connected to the POCS layer 74, and the output of the NN layer may be used as the final noise-reduced MRS spectral data.

[0160] (Example 11) The data processing performed by the base utilization function 115 in Example 11 is, for example, a segmentation process (or sub-region classification process) that quantifies and outputs whether each pixel (or sub-region) of the image data belongs to a specific region. A segmentation process that uses null space as a constraint condition will be called null space-constrained segmentation.

[0161] Figure 15 is a diagram showing a series of steps related to null space constraint segmentation by the information processing device according to Example 11. As shown in Figure 15, the processing circuit 11 acquires image data of a moving image and calculates multiple PCA bases by applying a time-axis basis transformation to the moving image through the implementation of the basis transformation function 113 (step SE1). Next, the processing circuit 11 selects a null base from among the multiple PCA bases through the implementation of the selection function 114 (step SE2).

[0162] As shown in Figure 15, the processing circuit 11, by realizing the basis utilization function 115, applies the image data of the moving image to the NN layer 151 to calculate an intermediate output value, and applies the intermediate output value to the POCS layer 152 to calculate segmentation information. The NN layer 151 calculates segmentation information from the moving image. The POCS layer 152 performs a POCS operation on the segmentation information from the NN layer 151, based on the null basis obtained in step SE2 and the weights of each null basis, according to constraint equations such as equation (5) or (6). The POCS layer 152 outputs the final segmentation information. The learning parameters of the NN layer 151 are trained to take a moving image as input and output segmentation information. The NN layer 151 and POCS layer 152 may be repeated multiple times.

[0163] The POCS layer 152 may also apply the POCS calculation to the intermediate output value of the segmentation process. Alternatively, the processing circuit 11 may calculate a null basis from the ECG signal acquired in parallel with the image and perform data processing on the image using the null basis based on the ECG signal. Alternatively, a null basis may be calculated from the water signal acquired by the MRS and perform data processing on the metabolite using the said null basis.

[0164] (Example 12) In some of the embodiments described above, the constraint formula was applied to data (e.g., data for one channel) that had the same shape as the input signal (e.g., data for one channel).

[0165] Figure 16 shows an example of the network configuration of the deep neural network 160 according to Example 12. As shown in Figure 16, the deep neural network 160 includes NN layer 161, POCS layer 162, NN layer 163, POCS layer 164, and NN layer 165. NN layer 161, POCS layer 162, NN layer 163, POCS layer 164, and NN layer 165 are connected in series. POCS layer 162 and POCS layer 164 apply constraints to each of the intermediate data (e.g., 64-channel data) with different shapes generated by NN layer 161 and NN layer 163, respectively.

[0166] Figure 17 shows an example of the network configuration of another deep neural network 170 according to Embodiment 12. As shown in Figure 17, the deep neural network 170 includes NN layer 171, POCS layer 172, NN layer 173, POCS layer 174, and NN layer 175. NN layer 171 and NN layer 173 have paths that go through POCS layer 172 and paths that connect directly. NN layer 173 and NN layer 175 have paths that go through POCS layer 174 and paths that connect directly. POCS layer 172 and POCS layer 174 may each apply constraints to a portion of the intermediate data (e.g., 32 channels of data out of 64) of the intermediate data (e.g., 64 channels of data) that has a different shape generated by NN layer 171 and NN layer 173, respectively.

[0167] (others) The information processing device 1 according to this embodiment can also be implemented by appropriately combining the above-described embodiments 1 to 12. For example, constraint terms based on at least two or more null bases of the null bases according to embodiments 1 to 12 may be incorporated into the error function.

[0168] According to at least one embodiment described above, a basis with a relatively low contribution rate can be utilized.

[0169] In the above description, the term "processor" refers to circuits such as CPUs, GPUs, or Application Specific Integrated Circuits (ASICs), programmable logic devices (e.g., Simple Programmable Logic Devices (SPLDs), Complex Programmable Logic Devices (CPLDs), and Field Programmable Gate Arrays (FPGAs)). The processor achieves its function by reading and executing a program stored in a memory circuit. Alternatively, instead of storing the program in a memory circuit, the processor may be configured to directly incorporate the program into its circuitry. In this case, the processor achieves its function by reading and executing the program incorporated into the circuitry. Furthermore, instead of executing a program, the processor may achieve the function corresponding to the program through a combination of logic circuits. In this embodiment, each processor is not limited to being configured as a single circuit; multiple independent circuits may be combined to form a single processor and achieve its function. Moreover, the multiple components shown in Figure 1 may be integrated into a single processor to achieve its function.

[0170] While several embodiments have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be implemented in a variety of other forms, and various omissions, substitutions, modifications, and combinations of embodiments are possible without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of Symbols]

[0171] 1. Information Processing Device 11 Processing Circuit 12 Communication Interfaces 13 Display equipment 14 Input Interfaces 15 Storage device 111 Acquisition function 112 Image Conversion Function 113. Basis transformation function 114 Selection Function 115 Basic usage functions 116 Image Processing Functions 117 Data Conversion Function 118 Output control function

Claims

1. An image conversion unit that performs image conversion on raw data to generate a converted image which is image data, A basis transformation unit that performs basis transformation on signal data to calculate a plurality of basis sets, wherein the basis transformation unit calculates the plurality of basis sets using the transformed image as the signal data, A selection unit that selects one or more specific bases from among the plurality of bases whose contribution rate is lower than a standard, A basis utilization unit that performs data processing using the aforementioned specific basis, wherein the data processing includes generating an output image from the raw data such that constraint conditions regarding the projection amount for the aforementioned specific basis are satisfied, An image processing unit that applies noise reduction processing to the output image to generate a noise-reduced image, A data conversion unit that performs data conversion on the noise-reduced image to generate computational raw data, It is equipped with, The basis utilization unit generates other output images by applying image transformation to the computational raw data based on the noise reduction image, using the specific space spanned by the specific basis as a constraint condition. Information processing device.

2. The information processing apparatus according to claim 1, wherein the basis utilization unit performs an update operation multiple times to relatively reduce an error function that includes a data consistency term based on the difference between the calculated raw data and the raw data, and a projection amount of the updated image onto each of the specific basis sets, and generates the output image from the raw data.

3. The information processing apparatus according to claim 2, wherein the base utilization unit minimizes the error function under the constraint that the projection amount is zero or less than or equal to a threshold.

4. The information processing apparatus according to claim 1, wherein the basis transformation unit performs the basis transformation with respect to the image spatial axis direction of the image data.

5. The information processing apparatus according to claim 1, wherein the basis conversion unit performs the basis conversion with respect to the scan parameter value axis direction of the image data.

6. The information processing apparatus according to claim 1, wherein the basis conversion unit performs the basis conversion with respect to the contrast axis direction of the image data.

7. The information processing apparatus according to claim 1, wherein the base utilization unit generates the output image from the raw data under the constraint condition that the amount of projection of the updated image onto a specific space spanned by the specific base is zero or noise.

8. A basis transformation unit that calculates a plurality of basis sets by applying a basis transformation to signal data, comprising a basis transformation unit that applies the basis transformation with respect to the image spatial axis direction of the image data which is the signal data, A selection unit that selects one or more specific bases from among the plurality of bases whose contribution rate is lower than a standard, A base utilization unit that performs data processing using the aforementioned specific base, It is equipped with, The aforementioned image data consists of multiple local images, each corresponding to a plurality of local regions set at different positions in a reconstructed image based on a single k-space data set. The basis transformation unit applies the basis transformation to the plurality of local images to calculate the plurality of basis vectors, The selection unit selects a specific basis from among the plurality of basis sets, The base utilization unit generates an output image by applying an image transformation to the single k-space data, using the specific space spanned by the specific base as a constraint condition. Information processing device.

9. The information processing apparatus according to claim 8, wherein the calculation of the plurality of bases by the base conversion unit, the selection of the specific base by the selection unit, and the generation of the output image by the base utilization unit are repeated until predetermined conditions are met.

10. A basis transformation unit that calculates a plurality of basis vectors by applying basis transformations to signal data, the basis transformation unit that applies the basis transformations with respect to the time axis direction and the image spatial axis direction of the image data which is the signal data, A selection unit that selects one or more specific bases from among the plurality of bases whose contribution rate is lower than a standard, A base utilization unit that performs data processing using the aforementioned specific base, It is equipped with, The aforementioned image data includes a first image data and a second image data, The first image data is a plurality of first reconstructed images based on a plurality of central lines extracted from a plurality of sparse k-space data for a plurality of time points, The second image data is a second reconstructed image based on k-space data generated by applying view sharing to the plurality of sparse k-space data, The basis transformation unit calculates a plurality of first basis sets by applying the basis transformation to the plurality of first reconstructed images, and calculates a plurality of second basis sets by applying the basis transformation to the second reconstructed image. The selection unit selects one or more specific first bases from the plurality of first bases, and selects one or more specific second bases from the plurality of second bases. The base utilization unit applies image transformation to the plurality of sparse k-space data, using the constraints of the specific first space spanned by the specific first base and the specific second space spanned by the specific second base, to generate a plurality of output images relating to the plurality of time points. Information processing device.

11. An image conversion unit that performs image conversion on first raw data to generate image data which is signal data, A basis transformation unit calculates multiple basis sets by applying basis transformation to the aforementioned signal data, A selection unit that selects one or more specific bases from among the plurality of bases whose contribution rate is lower than a standard, A base utilization unit that performs data processing using the aforementioned specific base, wherein the data processing involves applying an image transformation to a second raw data different from the first raw data, using a specific space spanned by the aforementioned specific base as a constraint condition, to generate an output image. An information processing device equipped with the following.

12. A basis transformation unit that performs basis transformation on signal data to calculate multiple basis sets, A selection unit that selects one or more specific bases from among the plurality of bases whose contribution rate is lower than a standard, A base unit that performs data processing using the aforementioned specific base, comprising a base unit that converts the signal data into a classification value using the aforementioned specific base, An information processing device equipped with the following.

13. A basis transformation unit that performs basis transformation on signal data to calculate multiple basis sets, A selection unit that selects one or more specific bases from among the plurality of bases whose contribution rate is lower than a standard, A basis utilization unit that performs data processing using the aforementioned specific basis, comprising a basis utilization unit that performs feature transformation using convex projection processing for each of the aforementioned specific basis, An information processing device equipped with the following.

14. A basis transformation unit that calculates multiple basis vectors by applying a basis transformation to signal data, comprising a basis transformation unit that applies the basis transformation in the frequency axis direction to the signal data, which is MRS spectral data, A selection unit that selects one or more specific bases from among the plurality of bases whose contribution rate is lower than a standard, A base utilization unit that performs data processing using the aforementioned specific base, It is equipped with, The aforementioned MRS spectral data includes multiple MRS spectra for multiple locations, The basis transformation unit applies the basis transformation to the plurality of MRS spectra to calculate the plurality of basis sets, The selection unit selects a specific basis from among the plurality of basis sets, The base utilization unit generates a plurality of noise-reduced MRS spectra by applying noise reduction processing to the plurality of MRS spectra using the specific space spanned by the specific base as a constraint condition. Information processing device.

15. The information processing apparatus according to any one of claims 1, 8, 10, 11, 12, 13, and 14, wherein the basis transformation is principal component analysis, singular value decomposition, independent component analysis, Fourier transform, or discrete cosine transform.

16. An image conversion step of applying image conversion to raw data to generate a converted image which is image data, A basis transformation step that calculates a plurality of basis sets by applying a basis transformation to signal data, the basis transformation step that calculates the plurality of basis sets using the transformed image as the signal data, A selection step of selecting one or more specific bases from among the aforementioned plurality of bases whose contribution rate is lower than the standard, A basis utilization step that performs data processing using the aforementioned specific basis, the data processing step comprising generating an output image from the raw data such that constraint conditions regarding the projection amount for the aforementioned specific basis are satisfied, An image processing step to generate a noise-reduced image by applying noise reduction processing to the output image, The system comprises a data conversion step, which involves applying data conversion to the noise-reduced image to generate computational raw data, The basis utilization step involves applying an image transformation to the computational raw data based on the noise reduction image, using the specific space spanned by the specific basis as a constraint, to generate another output image. Information processing methods implemented by computers.

17. A basis transformation step for calculating multiple basis sets by applying a basis transformation to signal data, comprising a basis transformation step for applying the basis transformation with respect to the image spatial axis direction of the image data which is the signal data, A selection step of selecting one or more specific bases from among the aforementioned plurality of bases whose contribution rate is lower than the standard, The system comprises a base utilization step which performs data processing using the aforementioned specific base, The aforementioned image data consists of multiple local images, each corresponding to a plurality of local regions set at different positions in a reconstructed image based on a single k-space data set. The basis transformation step involves applying the basis transformation to the plurality of local images to calculate the plurality of basis vectors, The selection step involves selecting a specific basis from among the plurality of basis sets, The basis utilization step generates an output image by applying an image transformation to the single k-space data, using the specific space spanned by the specific basis as a constraint condition. Information processing methods implemented by computers.

18. A basis transformation unit that calculates multiple basis sets by applying basis transformations to signal data, comprising a basis transformation step of applying the basis transformations with respect to the time axis direction and the image spatial axis direction of the image data which is the signal data, A selection step of selecting one or more specific bases from among the aforementioned plurality of bases whose contribution rate is lower than the standard, The system comprises a base utilization step which performs data processing using the aforementioned specific base, The aforementioned image data includes a first image data and a second image data, The first image data is a plurality of first reconstructed images based on a plurality of central lines extracted from a plurality of sparse k-space data for a plurality of time points, The second image data is a second reconstructed image based on k-space data generated by applying view sharing to the plurality of sparse k-space data, The basis transformation step involves applying the basis transformation to the plurality of first reconstructed images to calculate a plurality of first basis sets, and applying the basis transformation to the second reconstructed image to calculate a plurality of second basis sets. The selection step involves selecting one or more specific first bases from the plurality of first bases, and selecting one or more specific second bases from the plurality of second bases. The basis utilization step involves applying an image transformation to the plurality of sparse k-space data, using the constraints of a specific first space spanned by the specific first basis and a specific second space spanned by the specific second basis, to generate a plurality of output images relating to the plurality of time points. Information processing methods implemented by computers.

19. An image conversion step of applying image conversion to first raw data to generate image data which is signal data, A basis transformation step is performed on the signal data to calculate multiple basis sets. A selection step of selecting one or more specific bases from among the aforementioned plurality of bases whose contribution rate is lower than the standard, A basis utilization step that performs data processing using the aforementioned specific basis, wherein the data processing involves applying an image transformation to a second raw data different from the first raw data, using a specific space spanned by the aforementioned specific basis as a constraint condition, to generate an output image. A computer-implemented information processing method that includes the following features.

20. A basis transformation process that calculates multiple basis sets by applying a basis transformation to signal data, A selection step of selecting one or more specific bases from among the aforementioned plurality of bases whose contribution rate is lower than the standard, A base utilization step that performs data processing using the aforementioned specific base, comprising a base utilization step that converts the signal data into a classification value using the aforementioned specific base, A computer-implemented information processing method that includes the following features.

21. A basis transformation process that calculates multiple basis sets by applying a basis transformation to signal data, A selection step of selecting one or more specific bases from among the aforementioned plurality of bases whose contribution rate is lower than the standard, A basis utilization step that performs data processing using the aforementioned specific basis, comprising a basis utilization step that performs feature transformation using convex projection processing for each of the aforementioned specific basis, A computer-implemented information processing method that includes the following features.

22. A basis transformation step for calculating multiple basis sets by applying a basis transformation to signal data, comprising a basis transformation step for applying the basis transformation in the frequency axis direction to the signal data, which is MRS spectral data, A selection step of selecting one or more specific bases from among the aforementioned plurality of bases whose contribution rate is lower than the standard, The system comprises a base utilization step which performs data processing using the aforementioned specific base, The aforementioned MRS spectral data includes multiple MRS spectra for multiple locations, The basis transformation step involves applying the basis transformation to the plurality of MRS spectra to calculate the plurality of basis sets, The selection step involves selecting a specific basis from among the plurality of basis sets, The basis utilization step involves applying noise reduction processing to the multiple MRS spectra using a specific space spanned by the specific basis as a constraint condition to generate a plurality of noise-reduced MRS spectra. Information processing methods implemented by computers.

Images