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

By combining successive reconstruction processing and noise removal, the problems of increased noise and reduced reconstruction accuracy in MRI imaging are solved. High-precision reconstruction and noise removal are achieved under various interleaving patterns, simplifying the construction of the learning model and improving the processing efficiency and quality of MRI imaging.

CN116098604BActive Publication Date: 2026-11-17FUJIFILM CORP
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202211029081.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2021-11-08
Filing Date
2022-08-25
Publication Date
2026-11-17
Estimated Expiration
2042-08-25

AI Technical Summary

Technical Problem

In existing MRI imaging techniques using parallel imaging, noise increases and reconstruction accuracy decreases as the inter-segment division ratio increases. This is especially true in inter-segment division patterns not included in the learning data, where reconstruction accuracy further declines, artifacts and noise remain, and the construction becomes more complex and computation time increases.

Method used

Successive reconstruction processing is introduced into the MRI device. The noise removal unit performs noise removal based on the statistical properties of the image and noise. Combined with the data consistency preservation unit, the measurement data of each receiving coil is repeatedly processed to maintain the consistency of the measurement data. The initial value is updated successively through initial value setting, synthesis, noise removal and data merging to achieve high-precision reconstruction.

Benefits of technology

It achieves high-precision reconstruction and noise removal under various interleaving patterns, reduces the dependence on complex learning models, expands the universality of the processed objects, and prevents over-smoothing noise removal, maintaining data consistency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116098604B_ABST
    Figure CN116098604B_ABST
Patent Text Reader

Abstract

This invention provides a magnetic resonance imaging apparatus, an image processing apparatus, and an image processing method. It provides a method for high-precision reconstruction and noise removal of various interval patterns including equal-interval intervals. The image processing unit, which processes measurement data acquired by the MRI apparatus, performs image reconstruction using measurement data of each channel measured with a given interval pattern and the sensitivity distribution of each receiving coil. During this process, noise removal of the reconstructed image and calculations to maintain consistency between the measurement data of each channel generated from the denoised image and the original measurement data are performed sequentially. Therefore, high-precision image restoration and noise removal can be performed without relying on the interval pattern.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to magnetic resonance imaging (MRI) devices, and more particularly to techniques for reconstructing and noise removal of measurement data obtained from intermittent measurements. Background Technology

[0002] As a technique for shortening imaging time in MRI examinations, k-space undersampling reconstruction is a method that intermittently acquires signals in the spatial frequency region (k-space) and reconstructs the unmeasured portion through signal processing. Parallel imaging (PI) is widely used in this method, where multiple receiving coils (channels) simultaneously receive signals, utilizing differences in sensitivity distribution for reconstruction.

[0003] As data recovery methods for the PI method, besides the SENSE method which recovers equally spaced intervals, there are also known methods such as the POCSENSE method and the SPIRiT method (e.g., Non-Patent Literature 1, Non-Patent Literature 2) that address arbitrary interval patterns through repeated reconstruction. In the POCSENSE and SPIRiT methods, the inverse problem representing the relationship between the measurement data, the sensitivity map, and the reconstructed image is solved through repeated calculations, thereby recovering the image without interval division. However, in the PI method, increasing the interval division ratio generally leads to a problem where, although the imaging time is reduced, the noise increases.

[0004] On the other hand, in order to reduce the increase in noise caused by interleaving, various methods using convolutional neural networks (CNNs) have been proposed. For example, in Patent Document 1, a reconstruction method based on a learned deep NN is proposed, which combines a CNN for noise removal and a layer (DC layer) for processing to maintain the consistency between the denoised data and the acquired data.

[0005] Existing technical documents

[0006] Patent documents

[0007] Patent Document 1: US Patent No. 10712416

[0008] Non-patent literature

[0009] Non-patent document 1: "POCSENSE: POCS-based reconstruction for sensitivityencoded magnetic resonance imaging", Magnetic Resonance in Medicine, 2004December, 52(6)PP.1397-1406

[0010] Non-Patent Literature 2: "SPIRiT: Iterative Self-consistent Parallel Imaging Reconstruction from Arbitrary k-Space", Magnetic Resonance in Medicine, 2010 August, 64(2) pp. 451-471

[0011] Generally, machine learning is used to reconstruct the network by combining images with and without interpolation, thus learning the overall network containing repeated processing.

[0012] However, the interval patterns can vary, including those with equal intervals and those that are random. In interval patterns not included in the training data, the reconstruction accuracy decreases, and artifacts and noise remain. While increasing the training data is considered, it's difficult to handle countless patterns, leading to a more complex CNN construction and increased computation time. Summary of the Invention

[0013] The present invention is made in view of the above-mentioned matters, and its object is to provide a method for high-precision reconstruction and noise removal of various interval patterns, including equal interval intervals.

[0014] To address the aforementioned issues, this invention incorporates noise removal processing into the repetitive calculations used to maintain consistency with the measurement data. The noise removal processing is based on the statistical properties of the image and the noise.

[0015] That is, the first aspect of the present invention is an MRI apparatus, for example comprising: a measurement unit having a plurality of receiving coils for collecting MRI signals of a subject at each receiving coil; a control unit for controlling the collection pattern of the MRI signals; and an image processing unit for creating an image of the subject using measurement data composed of the MRI signals collected at each receiving coil. The image processing unit includes: a successive reconstruction unit for reconstructing the image by repeatedly processing the measurement data of each receiving coil.

[0016] The successive reconstruction unit includes: an initial value setting unit that sets an initial value for repeated operations based on measurement data of each receiving coil; a synthesis unit that synthesizes images of each receiving coil; a noise removal unit that removes noise from the image based on the statistical properties of the image before or after synthesis and the noise contained in the image; and a data consistency maintenance unit that uses the denoised image to create denoised k-space data for each receiving coil, merges the denoised k-space data with the measurement data to create estimated measurement data, and the successive reconstruction unit updates the initial value based on the estimated measurement data to perform successive processing.

[0017] Typically, the measurement data is undersampled measurement data with a given interval pattern. The successive reconstruction unit receives the sensitivity maps of each receiving coil along with the measurement data, and the synthesis unit uses the measurement data and sensitivity maps to reconstruct the image.

[0018] Furthermore, a second aspect of the present invention is an image processing apparatus for processing MRI images, comprising: a successive reconstruction unit that reconstructs an image by repeatedly performing calculations using measurement data acquired from each receiving coil of the MRI apparatus. The successive reconstruction unit has the same structure and function as the successive reconstruction unit of the MRI apparatus described above.

[0019] Furthermore, a third aspect of the present invention is an image processing method that processes measurement data collected by multiple receiving coils of an MRI image to create a denoised reconstructed image. This image processing method includes a successive reconstruction process to maintain consistency between the denoised k-space data and the measurement data. The successive reconstruction process includes the following steps: setting initial values ​​for processing based on the measurement data of each receiving coil; synthesizing images of each receiving coil; removing noise from the image based on the statistical properties of the images before or after synthesis and the noise contained in the image data; and using the denoised images to create denoised k-space data for each receiving coil, merging the denoised k-space data and the measurement data to create estimated measurement data. The successive reconstruction process updates the initial values ​​based on the estimated measurement data, thereby repeating the processing.

[0020] Invention Effects

[0021] According to the present invention, by performing denoising processing on image data obtained in a consistent repetitive operation, the denoising processing is not affected by the curling that depends on the interval division pattern. Therefore, noise removal can be performed with good accuracy regardless of the interval division pattern. Furthermore, since curling is removed by successive processing including a synthesis process using coil sensitivity distribution, it is also possible to handle equally spaced interval division.

[0022] Furthermore, by combining denoising processing with data consistency preservation processing, over-smoothing due to denoising processing can be prevented, and noise can be removed with good accuracy. As a result, various interval division patterns, including equal-interval interval division, can be reconstructed with high accuracy and noise removal can be achieved.

[0023] Furthermore, since deep learning models such as CNNs can be limited to processing a portion of repetitive computations, the generalizability of the processing objects can be expanded without requiring a huge learning model. Attached Figure Description

[0024] Figure 1 This is an overall structural diagram illustrating one embodiment of the MRI device using the present invention.

[0025] Figure 2 This is a block diagram of the computer according to the first embodiment.

[0026] Figure 3 This is a diagram illustrating the processing flow of the first embodiment.

[0027] Figure 4 This is a diagram illustrating the processing of the first embodiment.

[0028] Figure 5 The diagrams are examples of division patterns of measurement data obtained by the measurement unit of an MRI device. (a) shows an equal-interval division pattern, (b) shows a non-equal-interval division pattern, and (c) and (d) show division patterns obtained from measurements where some data is missing or removed.

[0029] Figure 6 This is a diagram illustrating an example of the processing in the synthesis section.

[0030] Figure 7 This is a diagram illustrating an example of noise reduction using CNN.

[0031] Figure 8 This is a diagram illustrating an example of the structure of a CNN.

[0032] Figure 9 This is a diagram illustrating a modified example of the processing of the first embodiment.

[0033] Figure 10 The diagram shows other variations of the processing of the first embodiment.

[0034] Figure 11 This is a block diagram of the computer according to the second embodiment.

[0035] Figure 12 This is a diagram illustrating the processing flow of the second embodiment.

[0036] Figure 13 This is a diagram illustrating an example of noise estimation processing.

[0037] Figure 14 This is a flowchart illustrating a modified example of the processing in the second embodiment.

[0038] Explanation of reference numerals in the attached figures

[0039] 10: MRI device; 100: Measurement unit; 200: Image processing unit (computer); 210: Measurement control unit; 230: Image processing unit; 231: Data receiving unit; 233: Successive reconstruction unit; 234: Noise estimation unit; 235: Initial value setting unit; 236: Synthesis unit; 237: Noise removal unit; 238: Data consistency maintenance unit. Detailed Implementation

[0040] The embodiments of the MRI apparatus and image processing apparatus of the present invention will be described below with reference to the accompanying drawings. An embodiment using the MRI apparatus of the present invention will be initially described.

[0041] The MRI device 10 of this embodiment is as follows: Figure 1 As shown, it is generally divided into: a measurement unit 100, which measures the nuclear magnetic resonance signal generated from the subject 101; and a computer (processing unit) 200, which controls the measurement unit 100 and uses the nuclear magnetic resonance signal measured by the measurement unit 100 to perform image reconstruction, correction, and other operations.

[0042] The measuring unit 100 includes: a static magnetic field coil 102 that generates a static magnetic field in the space where the subject 101 is placed; a transmitting unit (105, 107) that transmits high-frequency magnetic field pulses to the subject 101 disposed in the static magnetic field; a receiving unit (106, 108) that receives the nuclear magnetic resonance signal generated by the subject; and a tilting magnetic field coil 103 that provides a magnetic field gradient to the static magnetic field generated by the static magnetic field coil 102 in order to impart position information to the nuclear magnetic resonance signal.

[0043] The static magnetic field coil 102 consists of a static magnetic field coil of constant or superconductivity, a static magnetic field generating magnet, etc. Depending on the direction of the generated static magnetic field, it can be a vertical magnetic field or a horizontal magnetic field, etc. The shape of the coil and the overall appearance of the device vary depending on the method. This embodiment is applicable to MRI devices of any type.

[0044] The transmitting unit includes: a high-frequency transmitting coil 105 (hereinafter referred to as the transmitting coil) that transmits a high-frequency magnetic field to the measurement area of ​​the subject 101; and a transmitter 107 that includes a high-frequency oscillator, an amplifier, etc. The receiving unit includes: a high-frequency receiving coil 106 (hereinafter referred to as the receiving coil) that receives the nuclear magnetic resonance signal generated from the subject 101; and a receiver 108 that includes an orthogonal detector circuit, an A / D converter, etc. In this embodiment, the receiving coil is composed of multiple channels (small receiving coils), each connected to an orthogonal detector circuit and an A / D converter constituting the receiver 108. The nuclear magnetic resonance signal received by the receiver 108 from each channel (i.e., each receiving coil) is transferred to the computer 200 as a complex digital signal.

[0045] The tilting magnetic field coil 103 has three sets of tilting magnetic field coils that apply tilting magnetic fields in the x, y, and z directions respectively, and are connected to the tilting magnetic field power supply 112. Furthermore, the MRI device may include a shimming coil 104 for adjusting the static magnetic field distribution and a shimming power supply 113 for driving it.

[0046] Furthermore, the measurement unit 100 includes a sequence control device 114 for controlling the operation of the measurement unit 100. The sequence control device 114 controls the operation of the tilted magnetic field power supply unit 112, the transmitter 107, and the receiver 108, and controls the timing of the application of the tilted magnetic field and the high-frequency magnetic field, as well as the reception of the nuclear magnetic resonance signal. The nuclear magnetic resonance signal can be encoded by the tilted magnetic field, and the position and pattern in the k-space where the nuclear magnetic resonance signal is configured as a digital signal can be determined by the number of encoding steps and the amount of each encoding.

[0047] The timing diagram of the control performed by the sequence control device is called a pulse sequence, which is preset according to the measurement and stored in the storage device of the computer 200.

[0048] Computer 200 controls the overall operation of MRI device 10 and performs various calculations on the received MRI signals. Therefore, computer 200... Figure 1 As shown, it includes functional units such as a measurement control unit 210, an image processing unit 230, and a display control unit 250. In this embodiment, the image processing unit (computer) 230 performs tasks such as image synthesis using the sensitivity distribution of the receiving coil, noise removal processing, and data consistency maintenance processing between the measurement data and the restored image.

[0049] Computer 200 is an information processing device equipped with CPU, memory, storage device, etc., and is connected to display 201, external storage device 203, input device 205, etc.

[0050] Display 201 is an interface for displaying the results obtained through calculation and processing to the operator. Input device 205 is an interface for the operator to input the conditions, parameters, etc., required for the measurement and calculation processes performed in this embodiment. The user can input measurement parameters such as the multiplier (also called the division rate) in the PI method via input device 205. External storage device 203, together with the internal storage device of computer 200, stores the data used in the various calculation and processing performed by computer 200, the data obtained through calculation and processing, the input conditions, parameters, etc.

[0051] The functions of each part of the computer 200 can be implemented as software installed on the computer 200, and the CPU loads the program (software) stored in the storage device into the memory for execution. Various data used in the processing of each function, and various data generated during processing, are stored in the storage device or external storage device 203. Furthermore, some of the functions implemented by the computer 200 can be implemented using hardware such as ASIC (Application Specific Integrated Circuit) or FPGA (Field-Programmable Gate Array).

[0052] Some of the functions implemented by the computer 200, such as the functions of the image processing unit 230, can also be implemented by a separate computer or other image processing device capable of transmitting and receiving data with the MRI device 10, and such image processing device is also included in the present invention.

[0053] The image processing unit 230 of this embodiment performs successive reconstruction of the measurement data of each receiving coil obtained by the measurement unit 100 of the MRI apparatus 10 with a given division pattern and division rate. This successive reconstruction includes image synthesis using the sensitivity distribution of the receiving coil, noise removal, and data consistency preservation processing. Furthermore, in these processes, transformations are performed such as inverse Fourier transform from k-space data to real space data, Fourier transform from real space data to k-space data, and transformation to sparse space.

[0054] The operation of the MRI device 10 with the above structure is the same as that of a known MRI device. The measurement control unit 210 receives the imaging conditions and pulse sequence selected or pre-set by inspection protocol corresponding to the imaging method and the imaging object, and sends the imaging conditions and pulse sequence to the sequence control device 114. The sequence control device 114 controls the transmitter 107, the tilting magnetic field power supply unit 112, the receiver 108, etc., and the measurement unit 100 collects the measurement data composed of nuclear magnetic resonance signals.

[0055] At this point, with the interval pattern and interval rate set as imaging conditions, the number and amount of codes assigned to the nuclear magnetic resonance signal are controlled to become the set interval pattern and interval rate.

[0056] Measurement data is collected using multiple small receiving coils, and the measurement data for each channel corresponding to each small receiving coil is configured in a k-space of a given matrix size for image reconstruction. That is, image reconstruction using the measurement data is performed in the image processing unit 230 of the computer 200. At this time, the image processing unit 230 repeatedly performs various operations such as transforming the measurement data into real-space data (image data), synthesizing the image for each channel, removing noise contained in the image, and maintaining consistency. Each process included in this successive reconstruction is implemented using a given algorithm.

[0057] The structure of the image processing unit 230 and the specific implementation of the processing are described below.

[0058] <First Implementation Method>

[0059] This embodiment uses measurement data of each channel measured with a given division pattern and the sensitivity distribution of each receiving coil to perform image reconstruction. At this time, denoising of the reconstructed image and operations to maintain the consistency between the measurement data of each channel generated from the denoised image and the original measurement data are performed sequentially. Therefore, high-precision image restoration and denoising can be performed without relying on the division pattern. Furthermore, no learning model corresponding to various division patterns is required, reducing the burden of preparing training data.

[0060] refer to Figure 2 The structure of the image processing unit 230 of this embodiment will be explained below. As shown in the figure, the image processing unit 230 includes a data receiving unit 231 and a successive reconstruction unit 233. The successive reconstruction unit 233 includes an initial value setting unit 235, a synthesis unit 236, a noise removal unit 237, and a data consistency retention unit 238.

[0061] The data receiving unit 231 receives measurement data (k-space data) from each channel. The initial value setting unit 235 sets the initial values ​​for the repeated operations performed by the successive reconstruction unit 233 based on the measurement data received by the data receiving unit 231. The initial values ​​for the repeated operations are the data that is processed in any step of the repeated operations, and there are several ways depending on the step. The following describes the case where the image data input to the compositing unit 236 is used as the initial values.

[0062] The compositing unit 236 combines the image data of each channel, which are set as the initial values ​​by the initial value setting unit 235, into one image data.

[0063] The noise removal unit 237 performs noise removal (denoising) on ​​the synthesized image data. There are various methods for denoising, but in this embodiment, the denoising process is performed based on the statistical properties of the image and the statistical properties of the noise contained in the image. Details of the method will be described later.

[0064] The data consistency maintenance unit 238 transforms the denoised image into k-space data for each channel, merges the k-space data with the original measurement data, performs an inverse transformation to obtain real spatial data, and updates the initial values. This maintains consistency with the measurement data.

[0065] The following is for reference Figure 3 as well as Figure 4 The details of the processing by the image processing unit 230 in this embodiment will now be explained. Figure 3 This is a diagram illustrating the processing flow of the image processing unit 230. Figure 4 This is a diagram illustrating the processing.

[0066] [Data Acceptance]

[0067] In the measurement unit 100, if measurement data is collected by taking a picture with a given interval pattern and interval rate, the data receiving unit 231 receives the measurement data of each channel.

[0068] The division ratio (speed factor) is the reciprocal of the ratio of the number of measured data points to the number of data points in each grid cell of the k-space (if the ratio is 1 / 2, then the speed factor is 2).

[0069] Except for the pattern Figure 5 Patterns such as those shown include: an equal-interval division pattern (a) in which k-space data is divided at equal intervals; a non-equal-interval division pattern (b) in which data is randomly sampled with consistent or inconsistent probability densities; a pattern (c) in which a portion of the data is removed due to a measurement interruption; and a pattern (d) in which rows are removed based on a given condition using an equal-interval or non-equal-interval pattern. In pattern (d), for example, there are effects such as body motion in the given condition. In this case, there is a method to obtain the navigation echo corresponding to the measurement point or measurement row and remove the measurement point or measurement row with an error greater than a certain value between it and the navigation echo used as a reference. Furthermore, various body motion detection methods can be used.

[0070] also, Figure 5 Although not shown in the diagram, in order to obtain the sensitivity maps of each coil from the measurement data, there are patterns that perform a full scan only on the central region and equal intervals only on its periphery. Furthermore, there are non-Cartesian scans that scan the k-space radially or spirally.

[0071] When the division rate and division pattern are preset, the image processing unit 230 uses the preset information to perform processing. Alternatively, the data receiving unit 231 can receive information representing measurement points and non-measurement points, such as the mask representing the measurement points, together with the measurement data.

[0072] [Step S301]

[0073] The initial value setting unit 235 sets the initial value I0 of the repeated operation of the successive reconstruction unit 233 based on the input measurement data. In this embodiment, since the repeated operation is performed based on real space data, the initial value setting unit 235 transforms the measurement data of each receiving coil (hereinafter also referred to as the measurement data of each channel) into real space data and sets it as the initial value. The measurement data is data obtained by dividing a portion of the k-space, so the unmeasured data is estimated and transformed into real space data.

[0074] For example, zero-filling is performed on unmeasured points in k-space, followed by an inverse Fourier transform to generate true spatial data. Alternatively, after generating an image using known parallel imaging operations such as the SENSE method (called parallel reconstruction), this image data can be multiplied by the sensitivity maps of each receiving coil to generate image data for each channel, which is then used as the initial values ​​for each channel's image. In cases of non-equal interval division, parallel reconstruction is performed only on the equally spaced portions.

[0075] Furthermore, according to the inventors' research according to the present invention, the following results can be obtained: when using an image created by zero-filling as the initial value, the number of repeated operations needs to be increased in order to remove the wrinkles; on the other hand, when using a parallel reconstructed image as the initial value, the repeated operations converge quickly because the wrinkles are removed.

[0076] The initial value setting unit 235 can further perform Gaussian filtering, bilateral filtering, and other filtering processes on the image described above. At this time, the filtering intensity can be changed using the G-factor diagram, which characterizes the noise magnification of parallel imaging. This denoising process performed by the initial value setting unit 235 is different from the denoising process performed by the noise removal unit 237, which will be described later. By performing parallel reconstruction and filtering during initial value setting, wrinkles and noise are reduced beforehand before repeated processing begins, thus reducing the number of repetitions.

[0077] [Steps S302~S305]

[0078] The successive reconstruction unit 233 uses the initial value I0 set by the initial value setting unit 235 as the initial value of the image data of each channel to perform the following repeated calculations.

[0079] [Step S302] Synthesis

[0080] First, the synthesis unit 236 uses the sensitivity map of the receiving coil to synthesize the image data of the receiving coil. The synthesis method is as follows: Figure 6 As shown, a so-called MAC (Multi Array Coil) synthesis is performed, in which the image data In of each channel is multiplied by a coefficient corresponding to the sensitivity map, and the results are added between channels to create one image data M. n In addition, I n and M n The subscript n is the reference numeral for the nth repeated operation (the same applies below), the initial I n Use the initial value I0. In addition to MAC synthesis, image data synthesis can also be performed by convolving the kernel corresponding to the sensitivity map with the k-space data of each channel and then adding them together in the k-space.

[0081] [Step S303] Noise Removal

[0082] Subsequently, the noise processing unit 237 processes the synthesized image data M n Denoising is performed in real-world space. Denoising is based on statistical properties. These statistical properties of images and noise are typically reflected in the distribution of pixel values ​​and the distribution of noise. Here, a given algorithm is used for denoising, which utilizes the difference between the distribution of a vector (composed of all or part of the pixel values ​​of the reconstructed image, or a vector obtained by transforming it) and the noise distribution. Alternatively, since statistical properties are inherent in past data, the parameters of a given algorithm that separates noise from the reconstructed image can be adjusted based on past data.

[0083] Specifically, examples include: a scheme that combines wavelet transform and soft thresholding (method 1); sparse modeling and sparse coating, which use lexicon representations of local structures for reconstruction (method 2); and deep learning models such as CNN (Convolutional Neural Network) that are trained to remove Gaussian noise (method 3).

[0084] In method 1, the reconstructed image is sparse, meaning there are many zero elements in the wavelet transform. In contrast, the noise is not sparse, but its value is smaller compared to the elements of the reconstructed image. This difference in statistical properties between the two is utilized. Therefore, as shown in equation (1), the image data (containing noise) M is first subjected to a wavelet transform (Ψ) to separate it into multiple elements. Each element is then subjected to soft thresholding (S). λ Remove elements with small values. Perform an inverse wavelet transform (Ψ) on them.-1 ), to generate image data M'.

[0085]

Mathematical Formula 1

[0086] M′=Ψ -1 (S λ (Ψ(M))) (1)

[0087] In addition, in equation (1), M represents the image data using a vector obtained by arranging all pixels, and S λ This is a function that applies a soft thresholding threshold λ to each element of the vector. The soft threshold λ can be arbitrarily adjusted according to the denoising intensity; it can be a default value or set by the user, or it can be adjusted to correspond to the amount of noise estimated using the methods described in the implementation methods below. Compared to using a deep learning model, these methods 1 and 2 can remove noise with less computational load.

[0088] One example of method 3 is the use of a CNN. This CNN learns by combining images with low noise (the correct image) and high noise from the same location in the past with images with high noise levels, in order to remove noise. More specifically, examples of high noise images include images captured at the same location as the correct image with a higher noise reduction ratio, and images created by artificially adding complex Gaussian noise to the correct image (complex images or absolute value images). Figure 7 An example of noise processing for an absolute image using a CNN is shown. In this example, an absolute image and a phase image are calculated from image M (S701, S702). The pixel values ​​of the absolute image are divided by the maximum value, and this is used as the input to the CNN (S703, S704). The CNN output, i.e., the denoised image, is multiplied by the maximum value (S705), and then multiplied by the phase image calculated in S702 (S706) to obtain the denoised image (complex image).

[0089] exist Figure 7 The example shown demonstrates denoising using a CNN only for absolute value images, but denoising can also be performed on complex images. When processing complex images, a CNN can be constructed using two channels (real and imaginary parts) for both input and output, or a known network for processing complex data can be used. By processing as a complex image, denoising including phase can be performed, improving the accuracy of noise removal.

[0090] There are no particular restrictions on the structure or number of layers of a CNN. For example, a general CNN structure that connects convolutional layers and active layers sequentially can be used. However, suitable alternatives, such as... Figure 8As shown, by setting it to a residual learning form that adds the inputs and outputs the results, the network learns the difference between itself and the correct image, thus improving the noise removal accuracy.

[0091] Furthermore, while increasing the number of CNN layers or using complex networks such as U-Net that connect multiple resolution layers can lead to higher accuracy, in this embodiment, since the error in noise removal is also reduced in the data consistency preservation unit 238 described later, the number of CNN layers can be reduced; for example, a CNN with 3 to 5 layers can be used. Since the noise processing in this embodiment is repeated as part of the successive reconstruction process, the reduced number of layers not only reduces noise processing but also lightens the overall processing of successive reconstruction.

[0092] Method 3 (using CNN) can mainly remove Gaussian noise, but it can also learn further to treat artifacts such as creases, ringing artifacts, and motion artifacts in the image as noise and remove them as well. In this way, artifacts can be reduced along with noise.

[0093] [Step S304]FT

[0094] Next, the data consistency maintenance unit 238 performs a process to restore the denoised image data M' into k-space data for each channel. Specifically, after multiplying the image data M' by the sensitivity diagram of each receiving coil (S3041), a Fourier transform is performed to obtain the k-space data K' for each channel (S3042).

[0095] The data consistency maintenance unit 238 merges the two data sets (S3043) to maintain the consistency between the k-space data K' and the measurement data K0 received by the data receiving unit 231. For example, as... Figure 4 As shown, the data of the unmeasured points (the estimated data) in the k-space data K' are merged with the data of the actual measured points in the measurement data K0 (the data with zero-filled unmeasured points to set the matrix size of k-space). That is, the measurement points in the k-space data K' of each channel are replaced with measurement data at the same location (Method 1).

[0096] Alternatively, the data from the measurement points within the k-space data of each channel can be weighted and summed with the data from the same location within the measurement data K0 (Method 2). That is, the data can be merged according to the following formula (2). In Method 1, λ is used in formula (2). DC =0, Method 2 is λ DC >0.

[0097]

Mathematical Formula 2

[0098]

[0099] In equation (2), U represents the operation of multiplying the measured points in k-space by 1 and the unmeasured points by 0. E is the identity matrix, and (EU) becomes the matrix representing the operation of multiplying the measured points in k-space by 0 and the unmeasured points by 1. In fact, the matrix is ​​not calculated, but the result of equation (2) is obtained at high speed by calculating each element in k-space.

[0100] Although Method 1 is fast, the data with noise removed is only the data that has been supplemented for the unmeasured parts. In contrast, Method 2 has the advantage of removing noise from the measurement data itself.

[0101] The data consistency maintenance unit 238 performs an inverse Fourier transform on the k-space data Kn of each merged channel to produce real spatial data.

[0102] [Steps S305, S306]

[0103] The successive reconstruction unit 233 updates the initial values ​​using the real spatial data of each channel created by the data consistency preservation unit 238 through merging (S306), and repeats the above steps S302 to S304. At this time, the difference between the initial values ​​before and after the update is taken, and if the difference is less than the preset error, the repeated operation ends (S305). With the end of the operation, the reconstructed image Mn synthesized using the real spatial data In before the update becomes the output of the successive reconstruction unit 233.

[0104] As explained above, the MRI apparatus (image processing unit) of this embodiment performs repeated operations of denoising processing and data consistency preservation processing, enabling successive reconstruction that maintains consistency with the measurement data regardless of the division pattern during measurement. Furthermore, it prevents excessive denoising. Moreover, since the wrinkles in equal-interval division are removed through successive reconstruction processing that includes synthesis processing using coil sensitivity distribution and data consistency preservation processing, it can also handle equal-interval division.

[0105] <Modification 1 of the first embodiment>

[0106] In the first embodiment, the image data of each channel used by the compositing unit 236 in MAC compositing were set as initial values, but the initial values ​​could also be... Figure 4 Either the real space data or the k-space data used in each step of the repeated operation shown.

[0107] exist Figure 9 The steps for setting initial values ​​are shown. Figure 9 and Figure 4 The corresponding processing will replace Figure 4The steps for setting the initial value in (a) initial value setting are shown as (a1), (a2), and (a3). The initial value setting unit 232 sets the initial value used in any of these steps.

[0108] In case (a1), the image reconstructed using the measurement data received by the data receiving unit 231 is used as the initial value, and the denoising process S303 begins. The reconstruction based on the measurement data can be MAC synthesis or reconstruction using parallel imaging operations. In cases (a2) and (a3), the k-space data is set as the initial value, either before or after merging. In case (a2), after denoising the reconstructed image Mn, or without denoising, the k-space data of each channel (estimated k-space data for the unmeasured points) is transformed and set as the initial value K'0 of the k-space data K' before merging. In case (a3), after merging the k-space data K'0 of each channel in (a2) with the measurement data K0, it is set as the initial value K'0 of the estimated k-space data K'.

[0109] Regardless of the case, only the starting point of the repeated calculation differs; the processing is the same, and the same effect as the first embodiment can be obtained. In the first embodiment or (a1), since the real space data is set as the initial value, it is easy to implement: the configuration accepts data after various known corrections such as distortion correction in the real space; and adds processing to set the initial value after filtering processing to perform noise removal. In (a2) or (a3), since the k-space data measured from the device is set as the initial value unchanged, the system can be simplified.

[0110] <Modification 2 of the first embodiment>

[0111] In the first embodiment, the images of each channel are synthesized and the synthesized image is denoised. However, in this embodiment, the image of each channel is denoised.

[0112] exist Figure 10 The processing of this embodiment is illustrated. In Figure 10 In, with Figure 4 The same processing is illustrated with the same reference numerals, and its detailed description is omitted. In this embodiment, the case where image data of each channel based on measurement data is set as the initial value is also shown, but the same changes as in the above-described variation 1 can be made.

[0113] In this embodiment, the compositing unit 236 creates a reconstructed image (an image that unfolds the folds) for each channel (reconstruction is performed for each channel). In contrast, the compositing unit 236 in the first embodiment generates an image based on data from each channel, while the process in this embodiment uses information from the sensitivity map to create a reconstructed image for each channel. As a method for creating a reconstructed image for each channel, techniques such as convolving the kernel calculated from the sensitivity map with k-space data, as in the SPIRiT method, are used.

[0114] The noise removal unit 237 performs noise removal processing on the reconstructed image of each channel created by the synthesis unit 236 (S303). The noise removal processing is the same as in the first embodiment, based on the statistical properties of the image and noise, and uses a given algorithm (including CNN). The denoised image data is restored to k-space data, obtaining k-space data K' for each channel (S3042). The subsequent merging (S3043), inverse Fourier transform (S3044), and initial value update (S305) in the data consistency preservation unit 238 are the same as in the first embodiment.

[0115] The successive reconstruction unit 233 takes the difference between the initial values ​​before and after the update, and ends the repeated calculation if the difference is less than the preset error. Then, the compositing unit 236 performs MAC compositing on the reconstructed images Mn of each channel that were just updated, and ends the processing.

[0116] This variation improves image quality by using reconstruction methods such as SPIRIT, which are less prone to artifacts like wrinkles.

[0117] <Second Implementation Method>

[0118] This embodiment adds a function to estimate the noise level of the image to the structure of the first embodiment. The feature is that it can adjust various coefficients and weights (collectively referred to as parameters) used in repeated calculations based on the estimated noise level.

[0119] exist Figure 11 The structure of the image processing unit 230 in this embodiment is shown. Figure 11 In, it has the same Figure 2 Elements with the same function are shown with the same reference numerals in the accompanying drawings, and repeated descriptions are omitted. For example... Figure 11 As shown, the image processing unit 230 of this embodiment includes a noise estimation unit 234.

[0120] The processing flow in this embodiment is as follows: Figure 12 As shown, the estimation of noise level and the adjustment based on noise level (S310) are added. Otherwise, it is the same as the first embodiment and its variations. Therefore, the following description focuses on the different processing.

[0121] The noise estimation unit 234 uses the image reconstructed through parallel reconstruction to estimate the noise level. The image reconstructed through this parallel reconstruction can be used as the reconstructed image generated by the synthesis unit 236 in the first repeated operation, or as the reconstructed image used in the creation of the channel image data can be used if the image data of each channel is set to the initial value by the initial value setting unit 235 after parallel reconstruction.

[0122] One method for estimating noise levels from an image is to calculate the standard deviation and mode of the noise based on the distribution (histogram) of the image's signal intensity, using this as the noise level σ. Noise is like... Figure 13 As shown in the histogram, since the signal intensity appears at a higher frequency than the image signal in areas of low signal strength, this distribution can be used to strictly distinguish it from the signal, thereby calculating the amount of noise.

[0123] Alternatively, instead of using the signal intensity distribution, information excluding the subject portion can be extracted from the reconstructed image as described above to calculate the noise level. For example, the background portion of the image can be extracted, and the standard deviation and mode of the signal intensity in the background portion can be used as the noise level. Furthermore, since a spatial differential image of the reconstructed image can be obtained, after removing the signal components originating from the subject that change more gradually compared to the noise, the noise level can be calculated based on this signal value.

[0124] When calculating the noise level from the reconstructed image, corrections can be made to the image using the G-factor and R-factor (inter-division ratio) from the parallel reconstruction. Since the G-factor is a graphical representation of the spatial distribution of noise amplification during parallel reconstruction, the noise level, independent of the G-factor, can be calculated by dividing each pixel value of the image by the corresponding G-factor value. Although the inter-division ratio is a single value, since noise generally increases proportionally to the square root of the inter-division ratio, the noise level, independent of it, can be calculated by dividing by the square root of the inter-division ratio. Because these corrections reduce dependence on the reconstruction process, parameters can be adjusted to correspond to the noise level of the measurement data itself, resulting in stable denoising accuracy.

[0125] In addition to the methods mentioned above, various other methods can be used to estimate the amount of noise, based on reconstructed images, measurement data, or images calculated from them.

[0126] The noise level calculated by the noise level estimation unit 234 is used in the adjustment of functions, algorithms, CNN coefficients, thresholds, and CNN switching in each step of the repeated calculation in the successive reconstruction unit 233 (mainly the noise removal unit 237).

[0127] For example, when wavelet transform and soft thresholding are performed in the noise removal unit 237, the threshold λ for soft thresholding is adjusted to λ = W·σ, etc. W is a constant, which can be set by default or adjusted by the user. Furthermore, λ can also be determined by a polynomial of the noise amount σ, etc.

[0128] In the case where the noise removal unit 236 is a CNN, for example, multiple CNNs can be prepared, such as a CNN that has learned from teaching data with high noise levels and a CNN that has learned from teaching data with low noise levels, and the CNN can be switched according to the noise level σ. Specifically, the total number of CNNs and other network structures, as well as the weights of each neuron, are switched.

[0129] Furthermore, in the first embodiment, two methods were described as processing for the data consistency maintenance unit 238, but it is also possible to make the common formula (2) of these methods “λ” DC The value of σ varies depending on the noise level, which can be used to switch methods or adjust the weighting of data during merging. For example, adjustments are made to increase the weight of the denoised k-space data when there is a lot of noise, so that the merging can proceed.

[0130] The successive reconstruction unit 233 performs repeated calculations under the condition that the noise level estimated by the noise level estimation unit 234 has been adjusted. The processing of the successive reconstruction unit 233 is the same as that in the first embodiment and its variations, and the description is omitted.

[0131] According to this embodiment, the intensity of denoising can be adjusted according to the amount of noise contained in the image, thereby improving the accuracy of denoising. Furthermore, by repeatedly performing denoising processing together with data consistency maintenance processing, excessive denoising can be prevented.

[0132] <Modifications of the Second Embodiment>

[0133] exist Figure 12 In the second embodiment shown, noise estimation processing S310 is performed before repeated calculations, but it can also be done as follows: Figure 14 As shown, the noise level estimation and adjustment are set up as a process within the repeated calculation. In this case, Figure 11 In the structural diagram, the noise estimation unit 234 is changed to be included in the successive reconstruction unit 233.

[0134] In this modified example, after the initial value is changed following the processing performed by the data consistency preservation unit 238, the updated image data is again subjected to noise estimation and parameter adjustments, and subsequent processing (denoising, merging, etc.) is performed. Furthermore, in Figure 14In this process, step S310 can be set as a pre-process of the synthesis process S302, but it can also be set as a post-process of the synthesis process S302.

[0135] According to this variation, the amount of computation for repeated operations increases, but noise reduction can be performed more appropriately.

[0136] The above describes embodiments of the MRI apparatus of the present invention and the processing method performed in its image processing unit. The processing methods that can be used in each step described in these embodiments can be appropriately combined, thereby achieving high efficiency in repetitive operations.

[0137] Furthermore, the image processing apparatus of the present invention possesses the functions of the image processing unit 230 of the aforementioned MRI apparatus (e.g., Figure 2 The elements shown can be built on general-purpose computers and workstations. The processing content is the same, so repeated descriptions are omitted. Such an image processing device can possess general image processing functions in addition to the functions described above, and may also have... Figure 1 The auxiliary devices (input / output devices, storage devices) of the computer 200 shown are self-evident.

Claims

1. A magnetic resonance imaging device, characterized in that, have: The measuring unit has multiple receiving coils, and each receiving coil collects the nuclear magnetic resonance signal of the subject. The control unit controls the collection pattern of the nuclear magnetic resonance signal; The image processing unit uses measurement data consisting of nuclear magnetic resonance signals collected by each of the receiving coils to perform parallel reconstruction processing to create an image of the subject as a parallel reconstructed image; and The noise estimation unit estimates the noise level of the parallel reconstructed image. The image processing unit has: The successive reconstruction unit reconstructs the image by repeatedly processing the measurement data from each of the receiving coils. The successive reconstruction unit has the following features: The initial value setting unit sets the initial value for repeated calculations based on the measurement data of each of the receiving coils. A synthesis unit that synthesizes the images of each of the receiving coils; The noise removal unit removes noise from the image based on the statistical properties of the image or the synthesized image of each receiving coil and the noise contained in the image, and parameters adjusted according to the noise amount estimated by the noise amount estimation unit. and The data consistency preservation unit uses the denoised image to create denoised k-space data for each of the receiving coils, and combines the denoised k-space data with the measurement data to create estimated measurement data. The successive reconstruction unit updates the initial value based on the estimated measurement data, thereby performing successive processing.

2. The magnetic resonance imaging device according to claim 1, characterized in that, The measurement data is undersampled measurement data with a given interval pattern. The successive reconstruction unit receives the sensitivity maps of each receiving coil along with the measurement data, and the synthesis unit uses the measurement data and the sensitivity maps to reconstruct an image.

3. The magnetic resonance imaging device according to claim 1, characterized in that, The initial value setting unit sets the estimated image of each receiving coil, obtained by transforming the measurement data of each receiving coil into real space data through inverse Fourier transform, as the initial value of the image of each receiving coil used by the synthesis unit.

4. The magnetic resonance imaging device according to claim 2, characterized in that, The initial value setting unit sets the estimated image of each receiving coil, which is constructed using the image reconstructed from the measurement data of each receiving coil and the sensitivity map of the receiving coil, as the initial value of the image of each receiving coil used by the synthesis unit.

5. The magnetic resonance imaging device according to claim 1, characterized in that, The initial value setting unit sets the image reconstructed based on the measurement data of each receiving coil as the initial value of the synthesized image used by the noise removal unit.

6. The magnetic resonance imaging device according to claim 1, characterized in that, The initial value setting unit sets the denoised k-space data of each receiving coil, or the estimated k-space data obtained by combining the denoised k-space data and the measurement data, as the initial value.

7. The magnetic resonance imaging device according to claim 1, characterized in that, The noise removal unit performs noise removal by utilizing a vector composed of at least a portion of the pixel values ​​of the image being processed, or the difference between the distribution of such vector and the noise distribution.

8. The magnetic resonance imaging apparatus according to claim 7, characterized in that, The noise removal unit performs a sparse transformation on the image being processed to a sparse space, and then performs soft thresholding on the data after the sparse transformation to remove noise.

9. The magnetic resonance imaging device according to claim 1, characterized in that, The noise removal unit includes a neural network that has been trained using past data as training data, and performs noise removal based on the past data.

10. The magnetic resonance imaging apparatus according to claim 9, characterized in that, The neural network learns to remove Gaussian noise.

11. The magnetic resonance imaging apparatus according to claim 9, characterized in that, The neural network is further trained to remove at least one of the following: wrinkle artifacts, ringing artifacts, and motion artifacts.

12. The magnetic resonance imaging apparatus according to claim 1, characterized in that, The data consistency preservation unit replaces the data of the measured position (i.e., the measurement point) in the denoised k-space data with the data of the same position in the measurement data to create the estimated measurement data.

13. The magnetic resonance imaging apparatus according to claim 1, characterized in that, The data consistency preservation unit performs a weighted summation of the data of the measured point in the denoised k-space data and the data of the measured point at the same position in the measurement data to generate the estimated measurement data.

14. The magnetic resonance imaging apparatus according to claim 1, characterized in that, The noise estimation unit uses the signal intensity distribution of the image before the noise reduction process or the image after removing the subject from the image to estimate the noise level.

15. The magnetic resonance imaging apparatus according to claim 1, characterized in that, The noise estimation unit adjusts, based on the estimated noise level, the threshold used by the noise removal unit for noise removal, the coefficients of the formula for calculating the estimated measurement data calculated by the data consistency preservation unit, and, in the case where the noise removal unit includes a neural network, at least one of the following:

16. The magnetic resonance imaging apparatus according to claim 1, characterized in that, The noise estimation unit is included in the successive reconstruction unit, and the noise estimation is repeated each time the initial value is updated.

17. An image processing apparatus for processing measurement data collected by multiple receiving coils of a magnetic resonance imaging apparatus to reconstruct a denoised image. The image processing apparatus is characterized by having: A noise estimation unit that estimates the noise level of the parallel reconstructed image; and The successive reconstruction unit reconstructs the image by repeatedly processing the measurement data from each of the receiving coils. The successive reconstruction unit has the following features: The initial value setting unit sets the initial value for repeated calculations based on the measurement data of each of the receiving coils. A synthesis unit that synthesizes the images of each of the receiving coils; The noise removal unit removes noise from the image based on the statistical properties of the image or the synthesized image of each receiving coil and the noise contained in the image, and parameters adjusted according to the noise amount estimated by the noise amount estimation unit. and The data consistency preservation unit uses the denoised image to create denoised k-space data for each of the receiving coils, and combines the denoised k-space data with the measurement data to create estimated measurement data. The successive reconstruction unit updates the initial value based on the estimated measurement data, thereby performing successive processing.

18. An image processing method for processing measurement data collected by multiple receiving coils of a magnetic resonance imaging device to create a denoised reconstructed image. The image processing method is characterized in that... It includes: noise estimation processing to estimate the noise level of the parallel reconstructed image; and successive processing to maintain the consistency between the denoised k-space data and the measurement data. The successive processing includes the following steps: The initial values ​​for processing are set based on the measurement data of each of the receiving coils; The images of each of the receiving coils are combined; The noise contained in the image is removed based on the statistical properties of the image or the synthesized image of each of the receiving coils and the noise contained in the image, and the parameters adjusted according to the noise amount estimated by the noise amount estimation process. and The denoised image is used to create denoised k-space data for each receiving coil. This denoised k-space data is then combined with the measurement data to create estimated measurement data. The successive processing updates the initial value based on the estimated measurement data, thereby repeating the processing.

Citation Information

Patent Citations

  • Methods and systems for magnetic resonance image reconstruction using an extended sensitivity model and a deep neural network

    US10712416B1

  • Highly-scalable image reconstruction using deep convolutional neural networks with bandpass filtering

    US20190257905A1