Noise removal device, noise removal method, machine learning device, machine learning method, and program

Abstract

To provide a technique for generating an MRI image from which noise is removed by considering a feature of an MRI image of a living body.SOLUTION: A noise removal device (1) comprises: at least one processor (12) which executes noise removal processing of removing noise of an MRI image; and at least one memory (13) which stores a model utilized for the noise removal processing. The model is the model using the MRI image as an input image and the MRI image from which the noise is removed as an output image. The device is configured such that a pixel value of each pixel of the output image is included in a range equal to or greater than the minimum value and equal to or less than the maximum value of the pixel value of the pixel in the vicinity of the pixel of the input image corresponding to the pixel.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a noise removal device, a noise removal method, a machine learning device, a machine learning method, and a program. [Background technology]

[0002] In today's medical field, MRI images are widely used as a diagnostic method. When acquiring MRI images, especially when a sufficient signal-to-noise ratio (SNR) cannot be ensured with a single scan, a method is known to improve the SNR by taking multiple scans and averaging them.

[0003] For example, when acquiring a diffusion-weighted MRI image, it is common to perform about three imaging sessions and average the results. However, this increases the imaging time by the number of imaging sessions. Non-Patent Document 1, for example, discloses a technique for shortening imaging time, which uses a mathematical technique based on deep learning to remove noise from low-SNR images, such as images that are not averaged, and convert them into high-SNR images. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Sagawa H. et al., Deep Learning-based Noise Reduction for Fast Volume Diffusion Tensor Imaging: Assessing the Noise Reduction Effect and Reliability of Diffusion Metrics, Magnetic Resonance in Medical Sciences; published online (2020). Summary of the Invention [Problem to be solved by the invention]

[0005] However, conventional technologies using AI (artificial intelligence) such as deep learning may result in pixel values ​​of certain pixels being abnormally high or low for MRI images of living organisms, which may accidentally produce erroneous information (such as abnormal findings that do not actually exist). Conversely, there is also a possibility that information across multiple images that should be reflected in secondary indicators may be lost.

[0006] An object of one aspect of the present invention is to realize a technology for generating an MRI image from which noise has been removed, taking into consideration the characteristics of the MRI image of a living body. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems, a noise removal device according to one aspect of the present invention includes at least one processor that executes a noise removal process to remove noise from an MRI image, and at least one memory that stores a model used in the noise removal process, wherein the model is a model in which the MRI image is used as an input image and the noise-removed MRI image is used as an output image, and is configured so that the pixel value of each pixel of the output image is within a range between the minimum value and the maximum value of the pixel values ​​of pixels in the vicinity of the pixel of the input image that corresponds to the pixel.

[0008] In order to solve the above-mentioned problems, a noise removal device according to one embodiment of the present invention comprises at least one processor that executes a noise removal process for removing noise from an MRI image, and at least one memory that stores a model used in the noise removal process, wherein the model uses the MRI image as an input image and the MRI image from which noise has been removed as an output image, and includes a network that acts globally on the input image and n filters (n is a natural number equal to or greater than 2) that act locally on the input image, and the noise removal process includes a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0009] Furthermore, a noise removal method according to one aspect of the present invention includes a noise removal process in which at least one processor removes noise from an MRI image using a model stored in at least one memory, the model being a model in which the MRI image is an input image and the noise-removed MRI image is an output image, and the pixel value of each pixel of the output image is configured to be within a range from the minimum value to the maximum value of pixel values ​​of pixels in the vicinity of the pixel of the input image corresponding to the pixel.

[0010] Furthermore, a noise removal method according to one aspect of the present invention includes a noise removal process in which at least one processor removes noise from an MRI image using a model stored in at least one memory, the model being a model in which the MRI image is an input image and the noise-removed MRI image is an output image, the model including a network that acts globally on the input image and n filters (n is a natural number equal to or greater than 2) that act locally on the input image, and the noise removal process including a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0011] Furthermore, a machine learning device according to one embodiment of the present invention includes at least one processor that executes a model construction process to construct a model using supervised machine learning, in which an MRI image is used as an input image and an MRI image from which noise has been removed is used as an output image, and at least one memory that stores the model, wherein the model is a model in which the MRI image is used as an input image and the MRI image from which noise has been removed is used as an output image, and is configured such that the pixel value of each pixel of the output image is within a range from the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel of the input image that corresponds to the pixel.

[0012] Furthermore, a machine learning device according to one embodiment of the present invention includes at least one processor that executes a model construction process to construct a model using supervised machine learning, in which an MRI image is used as an input image and an MRI image from which noise has been removed is used as an output image, and at least one memory that stores the model, wherein the model is a model in which the MRI image is used as an input image and the MRI image from which noise has been removed is used as an output image, and includes a network that acts globally on the input image and n filters (n is a natural number greater than or equal to 2) that act locally on the input image, and the noise removal process includes a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0013] Furthermore, a machine learning method according to one aspect of the present invention includes a model construction process in which at least one processor constructs a model using a model stored in at least one memory by supervised learning, the model having an MRI image as an input image and an MRI image with noise removed as an output image, the model being a model having the MRI image as an input image and the MRI image with noise removed as an output image, and configured such that the pixel value of each pixel of the output image is within a range between the minimum value and the maximum value of the pixel values ​​of pixels in the vicinity of the pixel of the input image corresponding to the pixel.

[0014] Furthermore, a machine learning method according to one embodiment of the present invention includes a model construction process in which at least one processor constructs a model using a model stored in at least one memory by supervised learning, the model having an MRI image as an input image and an MRI image from which noise has been removed as an output image, the model having the MRI image as an input image and the MRI image from which noise has been removed as an output image, the model including a network that acts globally on the input image and n filters (n is a natural number greater than or equal to 2) that act locally on the input image, and the noise removal process includes a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0015] The noise removal device and machine learning device according to each aspect of the present invention may be realized by a computer. In this case, the control program for each device, which causes the computer to operate as each part (software element) of each device, and the computer-readable recording medium on which the control program is recorded, also fall within the scope of the present invention. [Effects of the Invention]

[0016] According to one aspect of the present invention, it is possible to realize a technique for generating an MRI image from which noise has been removed, taking into account the characteristics of the MRI image of a living body. [Brief explanation of the drawings]

[0017] [Figure 1] 1 is a block diagram showing the configuration of a noise removal device according to a first embodiment of the present invention. [Figure 2] 2 is a schematic diagram showing the flow of noise removal processing performed by the noise removal device according to the first embodiment. FIG. [Figure 3] FIG. 3 is a diagram illustrating an example of a filter used in the model according to the first embodiment. [Figure 4]1 is a flowchart showing the flow of noise removal processing S1 according to the first embodiment. [Figure 5] FIG. 10 is a block diagram showing the configuration of a machine learning device according to a fourth embodiment of the present invention. [Figure 6] 10 is a flowchart showing the flow of a model construction process S2 according to the fourth embodiment. [Figure 7] FIG. 10 is a schematic diagram showing a diffusion tensor learned by a machine learning device according to a fifth embodiment of the present invention. [Figure 8] 13 is a flowchart showing the flow of a model construction process S3 according to the fifth embodiment. [Figure 9] FIG. 20 is a schematic diagram showing the flow of a model construction process according to the sixth embodiment. [Figure 10] 13 is a flowchart showing the flow of a model construction process S4 according to the sixth embodiment. [Figure 11] 1 shows an input image, an output image, and a teacher image of a head MRI image according to Example 1 of the present invention. [Figure 12] 10 is an FA parameter map that is a secondary index of a head diffusion-weighted image according to Example 2 of the present invention. [Figure 13] FIG. 10 is a diagram showing changes in parameters that are secondary indices of a diffusion-weighted image before and after noise removal according to Example 3 of the present invention. [Figure 14] FIG. 4 is a schematic diagram showing the flow of noise removal processing performed by a model M of the noise removal device according to Example 4-1. [Figure 15] FIG. 10 is a schematic diagram showing the flow of noise removal processing performed by a model M of the noise removal device according to Example 4-2. [Figure 16] FIG. 10 is a schematic diagram showing the flow of noise removal processing performed by a model M of the noise removal device according to Example 4-3. [Figure 17] 10 is a graph showing pixel differences between a target image and output images Y1, Y2, and Y3. [Figure 18] 10 is a graph showing the evaluation results using SSIM, which is an index showing structural similarity. [Figure 19]10 is a graph showing the difference between the apparent diffusion coefficient ADC and a target image. DETAILED DESCRIPTION OF THE INVENTION

[0018] [Embodiment 1] An embodiment of the present invention will be described in detail below with reference to the drawings. Embodiment 1 is a basic embodiment for embodiments 2 and 3 described below. FIG. 1 is a block diagram showing the configuration of a noise removal device 1 according to embodiment 1. The noise removal device 1 removes noise from captured MRI (Magnetic Resonance Imaging) images to generate MRI images with a high SNR (Signal to Noise Ratio). Note that in this embodiment, "removing noise" does not necessarily mean completely removing noise, but also includes reducing noise.

[0019] (Configuration of noise removal device) As shown in FIG. 1, the noise removal device 1 includes an input / output interface 11, at least one processor 12, and a memory 13. The processor 12 acquires captured MRI images via the input / output interface 11. The processor 12 also executes a noise removal process to remove noise from the MRI images. The memory 13 stores a model M1 used in the noise removal process. The model M1 is a trained model that has been trained through supervised learning. A method for training the model M1 will be described later.

[0020] 1, the processor 12 and the memory 13 (model M1) are depicted as being incorporated into a single noise removal device 1, but they do not necessarily have to be incorporated into a single device. For example, they may be located in different devices or locations and connected to each other so that information can be communicated. Furthermore, some or all of them may be located on the cloud.

[0021] In this embodiment, the MRI images from which noise is removed are digital MRI images that are primarily captured in medical settings. That is, an MRI image is composed of a large number of pixels arranged in a matrix, and each pixel has a pixel value. For monochrome MRI images, the pixel value is, for example, a luminance value with 256 gradations (any integer between 0 and 255) indicating the degree of brightness. For color MRI images, the pixel value is, for example, the luminance value of each of R, G, and B. In this embodiment, noise in MRI images refers to, for example, noise known as artifacts.

[0022] Fig. 2 is a schematic diagram showing the flow of noise removal processing performed by the noise removal device 1. As shown in Fig. 2, model M (also referred to as "model M1" in this embodiment) includes a network and a filter, and executes the processing shown in processing paths (I), (II), (III), and (IV). Specifically, model M1 includes a network N that acts globally on input image X, and n filters F (n is a natural number equal to or greater than 2) that act locally on input image X.

[0023] More specifically, in processing path (I), processor 12 applies n filters F1, F2, . . . , F n (called filters F) to generate n intermediate images X1, X2, ... X from the input image X. n (second process) The n filters F have different filter coefficients F(A) as will be described later.

[0024] On the other hand, the processing path (II) in FIG. 2 is a path processed by a network (for example, a CNN network) N that operates globally on the input image X. In the processing path (II), the processor 12 uses this network N to derive n weights, or weighting coefficients V(n), from the input image X (first processing). The weighting coefficients V(n) are calculated based on the weights of n intermediate images X1, X2, ...X. n The processor 12 uses n weights (weighting coefficients V(n)) to calculate the weighted average of n intermediate images X1, X2, ... Xn An intermediate output image having a pixel value y is generated by weighting the n filters (third process). The coefficients defining the network and the coefficients defining each of the n filters are determined by machine learning. In other words, the model M1 is constructed by machine learning, which will be described later.

[0025] Furthermore, the processor 12 performs a limiting process on the intermediate output image using the max pooling value M shown in processing path (III) in FIG. 2 and the minimum pooling value m shown in processing path (IV). Specifically, for each region including a pixel of interest, a pixel value y' is generated, which is the minimum value between the maximum value M obtained by the max pooling process and the pixel value y of the intermediate output image. Next, for each region including a pixel of interest, a pixel value y'' is generated, which is the maximum value between the minimum value m obtained by the minimum pooling process and the pixel value y'. In this way, an output image Y having the pixel value y'' is generated. In other words, the processor 12 generates the final output image Y by performing a process of limiting the minimum and maximum values ​​on the output pixel value y of the intermediate output image.

[0026] As described above, model M1 is a model that uses an MRI image acquired by processor 12 as an input image and an MRI image obtained by removing noise from the MRI image as an output image. Model M1 is configured so that the pixel value of each pixel in the output image is within a range between the minimum value and the maximum value of the pixel values ​​of pixels in the vicinity of the corresponding pixel in the input image.

[0027] 3 is a diagram showing an example of a filter F used in the processing path (I) of the model M1 according to the first embodiment. Specifically, it is a diagram showing an example of pixel values ​​of pixels in a region (adjacent region) DIN near a pixel of interest in an MRI image input to the noise removal device 1, and the pixel value of the pixel of interest in a corresponding region DOUT in an intermediate image processed using the filter F. The pixel values ​​of the pixels in the input image (hereinafter also referred to as "input pixel values") and the output pixel values ​​are, for example, luminance values ​​of a monochrome image. In the example shown in FIG. 3, the "neighborhood of a pixel" is a region of ±1 pixel in the X and Y directions of the pixel (pixel of interest), that is, a 3 × 3 pixel region centered on the pixel of interest.

[0028] As shown in Figure 3, the coordinates of a pixel arranged in the left-right direction (X direction) and the up-down direction (Y direction) of the figure are represented as p(x, y). x is the left-right coordinate of the pixel, and y is the up-down coordinate. When the input pixel value of the pixel at coordinate (x, y) is p(x, y), the output pixel value of that coordinate is p'(x, y), and the filter coefficient of filter F is F(A) = (a1, a2, ... a9), p'(x, y) is calculated using the following formula. p'(x, y)=F(A)·p(x, y)=a1×p(x-1, y-1)+a2×p(x, y-1)+…a8×p(x, y+1)+a9×p(x+1, y+1) The output pixel value p'(x, y) is obtained by weighted addition of neighboring pixel values. The sum of the filter coefficients F(A) does not necessarily have to be 1. The filter coefficients F(A) of the n filters F are determined by machine learning.

[0029] For example, in the example shown in Figure 3, it is assumed that the filter coefficients F(A) are a1 = a2 = a3 = a4 = a6 = a7 = a8 = a9 = 0.1, a5 = 0.2, and the output value of the pixel of interest is p'(x, y) = 21.2. Note that the output values ​​of other pixels are not shown because the input values ​​of the neighboring 3x3 pixels are indefinite, but they are calculated in the same way using this filter F.

[0030] 3, the neighboring region is a region of ±1 pixel in the X and Y directions of the pixel of interest, but is not limited to this. For example, the neighboring region is a region from (x-n1) to (x+n2) in the X direction, and a region from (y-n3) to (y+n4) in the Y direction. Here, n1, n2, n3, and n4 are integers equal to or greater than 0. However, since setting a region that is too wide restricts the conditions too much, it is preferable that n1, n2, n3, and n4 are each approximately 1 to 3.

[0031] The reason for performing such processing in processing path (I) is to determine the specific output value through a relatively simple calculation process that uses only local (neighborhood) values ​​as input, so as to be unaffected by errors in the global range of the image, which are difficult to predict.

[0032] On the other hand, the processing path (II) is a set of relatively simple intermediate images X1, X2...X generated in the processing path (I). n The role of the filter (I) is to appropriately weight and average the weighted coefficients V(n) to derive a weighting coefficient V(n) that can remove noise over a global range of the image. The images generated using the filter F in the processing path (I) are images that focus only on local areas and attempt to perform noise removal processing, so these alone do not necessarily achieve appropriate noise removal over a global range. Therefore, in order to achieve appropriate noise removal both locally and globally, it is preferable to further provide processing such as the processing path (II) and optimize it in parallel with the processing path (I), so that the intermediate images X generated in the processing path (I) can be optimally combined based on global information.

[0033] Furthermore, processing paths (III) and (IV) have the role of forcibly restricting pixel values ​​that are considered to be abnormal within the range of pixel values ​​of neighboring pixels when the pixel values ​​of the image generated by the above processing paths (I) and (II) are included.

[0034] For example, if no conditions are applied to the total value of the filter coefficients F(A), the total of the learned filter coefficients F(A) may exceed 1, depending on the results of learning to remove noise from the entire image. This may cause the output pixel value of the pixel of interest to exceed the maximum value of the input pixel values ​​of its neighboring pixels. Even in such cases, the output pixel value of the generated image can be forcibly limited to within the range of the input pixel values ​​of the neighboring pixels in the final processing steps (III) and (IV), thereby preventing the generation of pixel values ​​that are considered abnormal.

[0035] Note that the total value of the filter coefficients F(A) determined by learning may not exceed 1. In that case, the output pixel value of the image processed by filter F falls within the range of the maximum and minimum input pixel values ​​of the neighboring pixels, and the output pixel value of the image processed by processing path (II) also falls within the range of the maximum and minimum input pixel values ​​of the neighboring pixels. Furthermore, even if the total value of the filter coefficients F(A) slightly exceeds 1, it is expected that the resulting output pixel value will fall within a range that does not deviate significantly from the range of the maximum and minimum input pixel values ​​of the neighboring pixels. In such a case, the effect of the present invention can be achieved without executing processing paths (III) and (IV).

[0036] Furthermore, for example, if a model is trained under the condition that the sum of the filter coefficients F(A) of filter F in processing path (I) is 1 or less, the output pixel value of the image processed by filter F will be equal to or less than the weighted average of the input pixel values ​​of the neighboring pixels. In other words, the output pixel value of the image processed by filter F will fall within the range of the maximum and minimum input pixel values ​​of the neighboring pixels, and the output pixel value of the image processed by processing path (II) will also fall within the range of the maximum and minimum input pixel values ​​of the neighboring pixels. Therefore, if a model is trained under such conditions, there is no need to execute processing paths (III) and (IV).

[0037] As described above, the model M1 is configured so that the pixel value of a pixel in the output image is included in the range of the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel in the input image.

[0038] Since the structure of a living body has continuity with its surroundings, it is unlikely that a single pixel in the area surrounding the pixel of interest will have an exceptionally high or low pixel value. Therefore, by taking into account the characteristics of MRI images of a living body and setting the above-mentioned restrictions on the range of possible output pixel values, it is possible to prevent pixel values ​​that are clearly judged to be abnormal from being output.

[0039] As described above, the noise removal device 1 according to the first embodiment can realize a technique for generating an MRI image from which noise has been removed, taking into consideration the characteristics of the MRI image of a living body.

[0040] Furthermore, while conventional techniques require multiple MRI images to obtain an image with reduced noise, the use of the noise removal device 1 according to this embodiment makes it possible to obtain an MRI image from a single MRI image with noise removed to the same degree as conventional techniques. This reduces the imaging time and the image synthesis time. Furthermore, it is possible to eliminate the risk of imaging conditions changing between multiple imaging sessions, and also reduces the burden on the patient.

[0041] (Noise removal method) Next, a noise removal method executed by the noise removal device 1 will be described with reference to the drawings. The noise removal method includes a noise removal process in which at least one processor 12 removes noise from an MRI image using a model M1 stored in at least one memory. FIG. 4 is a flowchart showing the flow of the noise removal process S1 for removing noise from an MRI image. As shown in FIG. 4, in step S11, the processor 12 acquires a captured MRI image via the input / output interface 11. Next, in step S12, the processor 12 generates an MRI image in which noise from the MRI image has been removed using the model M1.

[0042] Model M1 is a model that uses an MRI image as an input image and an MRI image with noise removed as an output image. As described above, model M1 is configured so that the pixel value of each pixel in the output image is within the range from the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the corresponding pixel in the input image.

[0043] According to the noise removal method of this embodiment, it is possible to obtain the same effects as those obtained by the noise removal device 1 described above.

[0044] [Embodiment 2] Next, a description will be given of a noise removal device 1 according to a second embodiment of the present invention. Note that the configuration of the noise removal device 1 according to the second embodiment is the same as the configuration of the noise removal device 1 according to the first embodiment, and therefore a description thereof will be omitted here.

[0045] (Functions of the noise removal device of the second embodiment) The noise removal device 1 according to the second embodiment differs from the noise removal device 1 according to the first embodiment in that a model M2 is stored in the memory 13. In addition to the learning performed when constructing the model M1, the model M2 is constructed (trained) so as to reduce the error (also simply referred to as "difference") between the secondary index derived from the output image and the secondary index derived from the corresponding MRI image of the training data from which noise has been removed. In the second embodiment, the secondary index is a diffusion tensor. Note that in the present embodiment, the input image is a diffusion-weighted image.

[0046] The noise removal device 1 according to the second embodiment can achieve the same effects as those achieved by the noise removal device 1 according to the first embodiment. Furthermore, according to this embodiment, a model M2 is used that is constructed to reduce the error between the secondary index (e.g., diffusion tensor) of the output image and the secondary index (e.g., diffusion tensor) of the teacher image. Therefore, it is possible to realize a noise removal device that generates an output image with higher accuracy. Note that accuracy in this embodiment is a measure of how close the generated image is to the teacher data.

[0047] (Noise removal method) Next, a noise removal method executed by the noise removal device 1 according to the second embodiment will be described. The noise removal method according to the second embodiment is basically the same as the noise removal process S1 described in the first embodiment. However, in the noise removal method according to the second embodiment, the input image is a diffusion-weighted image, and the model M2 is constructed so as to reduce the error between the secondary indices derived from the output image and the secondary indices derived from the corresponding MRI images of the training data from which noise has been removed. As described above, the secondary indices are diffusion tensors.

[0048] The noise removal method according to the second embodiment can achieve the same effects as those achieved by the noise removal method according to the first embodiment. Furthermore, this embodiment uses a model M2 constructed to reduce the error between the secondary index (e.g., diffusion tensor) of the output image and the secondary index (e.g., diffusion tensor) of the teacher image. Therefore, it is possible to realize a noise removal method that generates a more accurate output image.

[0049] [Embodiment 3] Next, a description will be given of a noise removal device 1 according to a third embodiment of the present invention. Note that the configuration of the noise removal device 1 according to the third embodiment is the same as the configuration of the noise removal device 1 according to the first embodiment, and therefore a description thereof will be omitted here.

[0050] (Functions of the noise removal device of the third embodiment) The noise removal device 1 according to the third embodiment differs from the noise removal device 1 according to the first embodiment in that a model M3 is stored in the memory 13. The model M3 is a model trained using, as training data, an MRI image captured and an MRI image obtained by removing noise using multiple MRI images captured under predetermined conditions, including the captured MRI image. The model M3 is constructed (trained) so as to reduce errors between output images of multiple MRI images captured under predetermined conditions. Note that the model M3 may be constructed by simultaneously performing the training performed when the model M1 or model M2 was constructed.

[0051] The noise removal device 1 according to the third embodiment can achieve the same effects as those achieved by the noise removal device 1 according to the first embodiment. Furthermore, according to this embodiment, a model M3 constructed to reduce the error between multiple output images is used. Furthermore, a model M3 constructed to reduce the error between each output image and a teacher image may also be used. Therefore, a noise removal device that generates output images with higher accuracy can be realized.

[0052] (Noise removal method) Next, a noise removal method executed by the noise removal device 1 according to embodiment 3 will be described. The noise removal method according to embodiment 3 is basically the same as the noise removal process S1 described in embodiment 1. However, in the noise removal method according to embodiment 3, the model M3 is a model trained using, as training data, an MRI image that has been captured and an MRI image obtained by removing noise from multiple MRI images captured under predetermined conditions, including the captured MRI image, and is constructed so as to reduce errors between output images of multiple MRI images captured under predetermined conditions.

[0053] The noise removal method according to the third embodiment can achieve the same effects as those achieved by the noise removal method according to the first embodiment. Furthermore, according to this embodiment, a model M3 is used that is constructed to reduce the error between output images of multiple MRI images captured under predetermined conditions. The model M3 may further be constructed to reduce the difference between each output image and the teacher image. Therefore, a noise removal method that generates output images with higher accuracy can be realized.

[0054] [Embodiment 4] (machine learning device) Next, a machine learning device according to a fourth embodiment of the present invention will be described in detail with reference to the drawings. This embodiment is the basis for the machine learning devices described in the fifth and sixth embodiments below.

[0055] (Configuration of machine learning device) FIG. 5 is a block diagram showing the configuration of a machine learning device 2 according to this embodiment. As shown in FIG. 5, the machine learning device 2 includes an input / output interface 21, at least one processor 22, and at least one memory 23. The memory 23 stores a model that uses an MRI image as an input image and an MRI image from which noise has been removed as an output image. The model is a model that executes noise removal processing using an MRI image as an input image and an MRI image from which noise has been removed as an output image. The model is configured so that the pixel value of each pixel in the output image is within a range from the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel in the input image corresponding to the pixel. The machine learning device 2 according to this embodiment is a device for constructing the model M1 described in the first embodiment. In the following description, the model of this embodiment, including the pre-learning model, will be referred to as the "model M1" for convenience.

[0056] The processor 22 acquires, as training data, the captured MRI image and an MRI image obtained by removing noise from multiple MRI images including the captured MRI image, via the input / output interface 21. The processor 22 executes a model construction process to construct a trained model M1 through supervised machine learning using the training data. That is, the processor 22 modifies coefficients that define the model M1 so as to reduce the error between the output image and the training image from which noise has been removed.

[0057] One piece of training data is an MRI image obtained by capturing an image and an MRI image (trainer image) obtained by removing noise from multiple MRI images including the captured MRI image. A data set that collects multiple pieces of training data is called a training data set. The processor 22 trains the model M1 using this training data set to construct a trained model M1. As described in the first embodiment, when a captured MRI image is input, the trained model M1 outputs an MRI image obtained by removing noise from the input MRI image (performs a noise removal process). Note that training the model M1 means updating the coefficients and parameters (hereinafter referred to as "coefficients") included in the model M1 until predetermined conditions are satisfied.

[0058] In medical practice, the common method for acquiring MRI images is to capture multiple MRI images under common conditions and then add these multiple MRI images to obtain a single MRI image with noise removed. The "common conditions" refer to predetermined imaging device conditions, and all multiple MRI images are taken under the same imaging device conditions. However, because the patient's position and physical condition vary at each imaging time, the multiple MRI images taken are not necessarily identical.

[0059] For example, the processor 22 may acquire, as training data, a set of one of the multiple MRI images acquired by such a method and an MRI image obtained by removing noise using multiple MRI images including the acquired MRI image. In this case, since the training data can use an MRI image that has already been acquired, there is an advantage that it is not necessary to newly create the training data.

[0060] The specific configuration of such a model M1 is not limited. As an example, the model M1 includes a network that acts globally on the input image and n filters (n is a natural number of 2 or more) that act locally on the input image. The network and filters are the network N and filter F described with reference to FIG. 2 in the first embodiment.

[0061] Furthermore, the noise removal process performed by processor 22 includes, as described in embodiment 1, a first process of deriving n weights from the input image using a network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating an output image by taking a weighted average of the n intermediate images using the n weights.

[0062] 2, the first process is a process in which the processor 22 derives n weights, or weighting coefficients V(n), from the input image X using a network N. The second process is a process in which the processor 22 derives n intermediate images X1, X2, ... X from the input image X using n filters F that act locally on the input image X.n The n filters F are obtained by changing the numerical values ​​of the filter F(A). In the third process, the processor 22 generates n intermediate images X1, X2, ... X using n weights (weighting coefficients V(n)). n The output image Y may be generated by further limiting the minimum and maximum values ​​of the pixel values ​​y of the intermediate output image.

[0063] Processor 22 updates the coefficients defining model M1 so as to reduce the error in pixel values ​​between output image Y and the teacher image from which noise has been removed. Then, an output image is generated using the input image of the next teacher data, and the coefficients defining model M1 are updated so as to reduce the error between the output image and the teacher image from which noise has been removed. Machine learning device 2 repeats this process to train model M1.

[0064] According to the machine learning device 2 having the above configuration, the pixel value of each pixel in the output image is restricted to be within a range between the minimum and maximum values ​​of the pixel values ​​of pixels in the vicinity of the corresponding pixel in the input image. This makes it possible to prevent pixel values ​​that are clearly determined to be abnormal from being output. Therefore, it is possible to construct a model that generates MRI images from which noise has been removed, taking into account the characteristics of MRI images of living organisms.

[0065] (machine learning methods) Next, a machine learning method according to the fourth embodiment will be described with reference to the drawings. The machine learning method includes a model construction process for constructing a model M1 through supervised learning. The model construction process is performed using the above-described machine learning device 2. The model is a model in which at least one processor uses an MRI image as an input image, utilizing a model stored in at least one memory, and outputs an MRI image from which noise has been removed.

[0066] FIG. 6 is a flowchart showing the flow of the construction process S2 of the model M1. As shown in FIG. 6, in step S21, the processor 22 acquires, as training data, an MRI image captured and a training image obtained by removing noise using multiple MRI images including the captured MRI image. Next, in step S22, the processor 22 updates the model M1 that removes noise through supervised learning. That is, the training data is input, an output image is obtained, and coefficients defining the model M1 are updated so that the error between the output image and the training image is reduced. Next, in step S23, the processor 22 determines whether a termination condition is satisfied. The termination condition may be, for example, that the error is within a predetermined range or that a predetermined number of learning rounds have been completed. If it is determined in step S23 that the termination condition is satisfied (step S23: YES), the construction process S2 is terminated. If it is determined in step S23 that the termination condition is not satisfied (step S23: NO), the process returns to step S21.

[0067] Model M1 is a model that uses an MRI image as an input image and an MRI image with noise removed as an output image. Model M1 is configured so that the pixel value of each pixel in the output image falls within a range between the minimum and maximum pixel values ​​of pixels in the vicinity of the corresponding pixel in the input image. This configuration is the same as that described in embodiment 1, so a description thereof will be omitted here.

[0068] As described above, according to the machine learning method of the fourth embodiment, by imposing the above-mentioned restrictions on the range of possible output pixel values, it is possible to construct a model that generates MRI images from which noise has been removed, taking into account the characteristics of MRI images of living organisms.

[0069] [Embodiment 5] Next, a machine learning device 2 according to a fifth embodiment of the present invention will be described below with reference to the drawings. Note that the configuration of the machine learning device 2 according to the fifth embodiment is the same as the configuration of the machine learning device 2 described in the fourth embodiment, and therefore a description thereof will be omitted.

[0070] (Functions of the machine learning device of embodiment 5) The machine learning device 2 according to this embodiment differs from the machine learning device 2 according to the fourth embodiment in that, in addition to the learning described in the fourth embodiment, the processor 22 trains a model so as to reduce the error between a secondary index derived from an output image and a secondary index derived from an MRI image from which noise has been removed from the corresponding training data. The machine learning device 2 according to the fifth embodiment is a device for constructing the model M2 described in the second embodiment. Hereinafter, the models according to this embodiment, including untrained models, will be referred to as the "model M2" for convenience.

[0071] When an MRI image is a diffusion-weighted image, the diffusion tensor can be derived as a secondary index from the diffusion-weighted image. A diffusion-weighted image is an image created using the direction and speed of water diffusion in the imaging subject as parameters. The diffusion tensor is a numerical representation of the flow of water, and is one of the important biological information contained in the original image.

[0072] In this embodiment, the training data in machine learning is a pair of a first MRI image and a second MRI image from which noise has been removed. For example, the first MRI image is one of multiple images captured under common conditions. The second MRI image is an image from which noise has been removed, obtained from the multiple images captured under common conditions. In this embodiment, the model construction process includes a derivation process that derives secondary indices of the MRI image output from model M2 when the first MRI image is input, and secondary indices of the second MRI image, and an update process that updates coefficients that define the model so that the difference between the two secondary indices is reduced.

[0073] Consider a case where the output image Y described in the fourth embodiment is a diffusion-weighted image. FIG. 7 is a schematic diagram showing a diffusion tensor learned by a machine learning device according to the fifth embodiment. As shown in FIG. 7, a diffusion tensor Ty is derived from the output image Y. Meanwhile, a diffusion tensor Tz is derived from training data Z (which is also a diffusion-weighted image). The diffusion tensor can be derived for each pixel.

[0074] As shown in Figure 7, the diffusion tensor is a three-dimensional diffusion vector denoted by λ and λ. 2κ , λ 3κ These may be expressed as ellipsoids Ty and Tz with three axes. The diffusion tensor is an ellipsoid in the case of anisotropic diffusion, and a sphere in the case of isotropic diffusion. In addition to comparing the pixel values ​​of the output image Y and the training data Z to reduce the error, it is also possible to generate an output image that is closer to the training data by comparing the vector values ​​of the ellipsoids Ty and Tz to reduce the error.

[0075] (Model construction method) Next, the construction process S3 of the model M2 according to this embodiment will be described with reference to the drawings. The construction process S3 may be performed in addition to the model construction process S2 described in embodiment 4. Fig. 8 is a flowchart showing the flow of the model construction process S3 according to embodiment 5.

[0076] As shown in FIG. 8, in step S31, the processor 22 derives secondary indices of the output image and the training data. That is, the processor 22 derives secondary indices of the MRI image output from the model M2 when a first MRI image is input, and secondary indices of the second MRI image (derivation process). Next, in step S32, the processor 22 trains the model so as to reduce the error between the secondary indices of the output image and the secondary indices of the training data. That is, the processor 22 updates the coefficients defining the model M2 so as to reduce the error between the two secondary indices (update process). As an example, the MRI image is a diffusion-weighted image, and the secondary indices are diffusion tensors.

[0077] The machine learning device 2 according to the above-described embodiment 5 can achieve the same effects as the machine learning device 2 according to embodiment 4. Furthermore, according to this embodiment, it is possible to construct a model M2 that reduces the error between the secondary index (e.g., diffusion tensor) of the output image Y and the secondary index (e.g., diffusion tensor) of the teacher image Z. Therefore, it is possible to realize a machine learning device 2 that constructs a model M2 that generates a more accurate output image.

[0078] [Embodiment 6] Next, a machine learning device 2 according to a sixth embodiment of the present invention will be described below with reference to the drawings. Note that the configuration of the machine learning device 2 according to the sixth embodiment is similar to the configuration of the machine learning device 2 described in the fourth embodiment, and therefore a description thereof will be omitted.

[0079] (Functions of the machine learning device of the sixth embodiment) The machine learning device 2 of this embodiment differs from the machine learning device 2 of embodiment 4 in that the multiple MRI images serving as training data are MRI images captured under common conditions, and the processor 22 constructs a model so as to reduce the error between output images of the multiple MRI images captured under the common conditions. The machine learning device 2 of embodiment 6 is a device for constructing the model M3 described in embodiment 3. Hereinafter, the models of this embodiment, including untrained models, will be referred to as "model M3" for convenience.

[0080] FIG. 9 is a schematic diagram illustrating the flow of a model construction process performed by a machine learning device 2 according to a sixth embodiment. As described above, in the case of conventional technology, n MRI images X1, X2, ..., XN are captured under common conditions when capturing MRI images. The training data for machine learning in this embodiment are the multiple MRI images X1, X2, ..., XN captured under these common conditions. Furthermore, the model construction process includes updating coefficients defining the model M3 so as to minimize differences between the MRI images Y1, Y2, ..., YN output from the model M3 when each of the multiple MRI images is input. That is, the processor 22 generates n output images Y1, Y2, ..., YN using the model M3 from each of the n MRI images X1, X2, ..., XN. The processor 22 then compares the resulting output images Y1, Y2, ..., YN and constructs the model M3 to minimize errors.

[0081] Here, comparing output images means comparing any (or all) combinations, such as comparing Y1 with Y2, comparing Y1 with Y3, etc., as shown in Fig. 9. Note that the results of comparing each of the output images Y1, Y2, ..., YN with the teacher image Z may also be added.

[0082] The n MRI images X1, X2, ..., XN captured under common conditions are images with equivalent information. Therefore, the output images Y1, Y2, ..., YN should also be images with equivalent information. Therefore, by constructing the model M3 so that the error between the n output images Y1, Y2, ..., YN is reduced, the model's capabilities can be further improved. Furthermore, as described above, the model M3 may be trained to increase the degree of correspondence between each of the output images Y1, Y2, ..., YN and the training image Z.

[0083] The machine learning device 2 according to the sixth embodiment described above can achieve the same effects as the machine learning device 2 according to the fourth embodiment. Furthermore, according to this embodiment, it is possible to construct a model M3 that improves the degree of match between a plurality of output images Y1, Y2, ..., YN. Furthermore, it is possible to construct a model M3 that reduces the error between each of the output images Y1, Y2, ..., YN and the teacher image Z. Therefore, it is possible to realize a machine learning device that constructs a model M3 that generates output images with higher accuracy.

[0084] (Model construction method) Next, a method for generating a model M3 according to the sixth embodiment will be described with reference to the drawings. FIG. 10 is a flowchart showing the flow of the construction process S4 of the model M3 according to the sixth embodiment. As shown in FIG. 10, in step S41, the processor 22 acquires a plurality of MRI images captured under common conditions. These MRI images are MRI images captured under common conditions that served as the basis for the MRI images of the training data from which noise has been removed. Next, in step S42, the processor 22 generates an output image from the plurality of MRI images. Next, in step S43, the processor 22 trains the model M3 so as to reduce the error between the output images of the plurality of MRI images captured under common conditions.

[0085] According to the model construction method of the sixth embodiment, it is possible to obtain the same effects as the model construction method of the fourth embodiment. Furthermore, according to this embodiment, it is possible to construct a model M3 that reduces the error between output images of multiple MRI images taken under common conditions. Therefore, it is possible to construct a model M3 that generates output images with higher accuracy.

[0086] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Example]

[0087] An embodiment of the present invention will be described below. Fig. 11 shows an input image 121, an output image 122, and a teacher image 123 of a head MRI image according to the first embodiment. The input image 121 is an MRI image that has been captured. The output image 122 is an MRI image that has been generated by inputting the input image 121 to the noise removal device according to the first embodiment. The teacher image 123 is an MRI image obtained by combining a plurality of MRI images captured under conditions common to those of the input image 121 and removing noise.

[0088] The SNR of the output image 122 is improved to the extent that it is comparable to that of the teacher image 123. In other words, it was found that by using the noise removal device according to this embodiment, an MRI image from which noise has been removed to the same extent as with the prior art can be obtained from a single MRI image. [Example]

[0089] Another embodiment of the present invention will be described below. Fig. 12 shows an FA parameter map, which is a secondary index of a head diffusion-weighted image according to a second embodiment of the present invention. From the left, Fig. 12 shows an input image 131, an output image 132, and a teacher image 133, along with enlarged views of each. The input image 131 is an FA parameter map of an acquired diffusion-weighted image. The output image 132 is an FA parameter map of a diffusion-weighted image generated by inputting the input image 131 to the noise reduction device according to the second embodiment. The teacher image 133 is an FA parameter map of a diffusion-weighted image obtained by combining multiple diffusion-weighted images acquired under conditions common to the input image 131 and removing noise.

[0090] The SNR of the output image 132 is reduced to a degree comparable to that of the teacher image 133. In other words, it was found that by using the noise removal device according to this embodiment, an FA parameter map of a diffusion-weighted image in which noise has been removed to the same extent as that obtained by the conventional technology can be obtained from a single diffusion-weighted image. [Example]

[0091] Other embodiments of the present invention will be described below. Fig. 13 is a diagram showing changes in parameters, which are secondary indices of diffusion-weighted images according to a third embodiment of the present invention, before and after noise removal. In the diagram, AD (Axial Diffusivity) indicates the major axis diameter of the diffusion tensor ellipsoid, RD (Radial Diffusivity) indicates the minor axis diameter of the ellipsoid, and MD (Mean Diffusivity) indicates the average of AD and RD. FA (Fractional Anisotropy) indicates the kurtosis of the ellipsoid. AK (Axial Kurtosis), RK (Radial Kurtosis), and MK (Mean Kurtosis) are indices separate from the diffusion tensor, and correspond to the major axis diameter of the diffusion tensor, the minor axis diameter of the diffusion tensor, and the average of the diameter lengths in all directions, respectively, when water molecule diffusion is approximated by a fourth-order tensor.

[0092] The vertical axis of each figure represents the root mean square error (RMSE), with a smaller value indicating closer proximity to the training data. These parameter values ​​are the average of the diffusion-weighted images of 10 people, and the boxes represent error bars. 1 on the horizontal axis represents the value obtained from the input image, and 2 on the horizontal axis represents the value obtained from the output image. As can be seen from Figure 13, the parameter values ​​obtained from the output image are smaller than the parameter values ​​obtained from the input image, and values ​​closer to the training data are obtained. In other words, it was found that the output image processed by the noise removal device according to the embodiment also yields values ​​closer to the training data than the input image for the secondary indices of the diffusion-weighted images. [Example]

[0093] (Example 4-1) Another embodiment of the present invention will be described below. Fig. 14 is a schematic diagram showing the flow of noise removal processing performed by model M of a noise removal device according to embodiment 4-1 of the present invention. In this embodiment 4-1, the flow of noise removal processing is similar to the flow of noise removal processing shown in Fig. 2, but differs in the following respects.

[0094] In the processing path (I), the model M uses 24 filters F1, F2, ..., F3 to generate a new pixel value by weighting and combining pixel values ​​of 3 × 3 (9) regions of the MRI image. 24 (referred to as filters F). That is, n, which indicates the number of filters F, is 24, and the 24 filters F each have a different filter coefficient F(A).

[0095] In this example, two MRI images X1 and X2 taken consecutively are input to the model M. The model M generates 24 intermediate images X1 from the input image X1 using 24 filters F. 11 ~X 1n (n=24) were generated, and these were multiplied by the 24 weights (total of 1) obtained in the processing path (II) to obtain the weighted average intermediate image X1. Also, 24 intermediate images X2 were generated from the input image X2. 21 ~X 2n(n=24) were generated, and these were each multiplied by the 24 weights (totaling 1) obtained in processing path (II) to obtain a weighted average intermediate image X2. Then, intermediate images X1 and X2 were subjected to max pooling processing in processing path (III) and minimum pooling processing in processing path (IV), respectively, and the average (Av) of intermediate images X1 and X2 obtained after these processes was taken to obtain output image Y1. Model M was trained so that the output image Y1 obtained in this way would approach the target image. The target image was an image obtained by adding together five MRI images taken consecutively.

[0096] In the processing of processing path (II), input image X1 and input image X2 were used as inputs, and processing was performed using an extended 2D convolutional layer (5 × 5, dilate = 2), a 2D convolutional layer (5 × 5, stride = 2), and a deconvolutional layer (2 × 2, stride = 2) as processing layers, and weights were output using a 2D convolutional layer (5 × 5) and a softmax layer as output layers.

[0097] In conventional technology, when capturing MRI images, five consecutively captured MRI images are added together to generate a single MRI image with reduced noise. However, in this embodiment, model M is trained to generate image Y1 with reduced noise from two consecutively captured MRI images X1 and X2. The training data in this case is an image with reduced noise created by adding five consecutively captured MRI images, which is the target image described above. Specifically, a target image serving as training data is generated by adding five consecutively captured MRI images with reduced noise, and model M is trained to generate an image similar to the training image using any two of the five MRI images. Note that, to evaluate the degree of approximation to the target image, the mean square error (MSE) between corresponding pixels is calculated, and model M is trained to minimize the sum of the MSEs.

[0098] (Example 4-2) 15 is a schematic diagram showing the flow of noise removal processing performed by model M of a noise removal device according to Example 4-2. The noise removal processing shown in Example 4-2 includes a first process of deriving n weights using two sets of three input images captured at different imaging positions using a network, a second process of generating n intermediate images from the input images using n filters, and a third process of generating an output image by taking a weighted average of the n intermediate images using the n weights. Here, the set of three input images captured at different imaging positions may be captured consecutively.

[0099] More specifically, differences between the noise removal process shown in Example 4-2 and the noise removal process shown in Fig. 14 will be described. In Example 4-2, two sets of three MRI images captured consecutively at different imaging cross-sectional positions were used as input MRI images. That is, a set of three MRI images represented by slice(k-1), slicek, and slice(k+1) with different imaging cross-sectional positions was used as input image X1, and a set of three MRI images captured consecutively at the same time with different imaging cross-sectional positions represented by slice(k-1), slicek, and slice(k+1) was used as input image X2. The processing of input image X1 and input image X2 is the same as the processing described in Example 4-1 above.

[0100] The model M was trained so that the output image Y2 obtained in this way would approach the target image. Specifically, the training of the model M (model construction process) includes a derivation process that derives the difference between the MRI image output from the model M when two sets of three input images captured at different imaging positions are input and the training data (target image), and an update process that updates the coefficients that define the model M so that the difference becomes smaller.

[0101] The target image is an image obtained by adding together five consecutively captured MRI images, as in Example 4-1. The slices of the MRI images indicated by slice(k-1), slicek, and slice(k+1) are spaced approximately 7 mm apart. In this example, learning is performed using not only the information on the slice itself but also the information on the slices above and below it, which is believed to improve accuracy.

[0102] (Example 4-3) 16 is a schematic diagram showing the flow of noise removal processing performed by model M of the noise removal device according to Example 4-3. The noise removal processing shown in Example 4-3 is basically the same as the noise removal processing described in Example 4-2, but differs from the noise removal processing shown in FIG. 15 in the following points.

[0103] In Example 4-3, as in Example 4-2, two sets of three MRI images were used as input images, each set consisting of six MRI images taken consecutively at different cross-sectional positions. However, in Example 4-3, two sets of six MRI images were processed together in the processing path (I). In other words, 24 intermediate images X1 to X2 were generated from the input images X1 and X2. n (n=24) were generated. These were each multiplied by the 24 weights (totaling 1) obtained in processing path (II) to obtain a weighted average of intermediate image X. In processing path (II), input images X1 and X2 were processed individually up to the processing layer, and then processed together in the output layer to obtain 24 weights. Furthermore, max pooling processing in processing path (III) and minimum pooling processing in processing path (IV) were performed on intermediate image X to obtain output image Y3. A model M was trained so that the output image Y3 obtained in this manner would approximate the target image. The training method (model construction process) was the same as that described in Example 4-2. The target image was an image obtained by adding together five MRI images taken consecutively, as in Example 4-1.

[0104] FIGS. 17 to 19 show the results of evaluating how close the images output by the trained model M obtained in Examples 4-1 to 4-3 above are to the target image.

[0105] Figure 17 is a graph showing the pixel differences (normalized pixel values) between the target image and the output images Y1, Y2, and Y3. The smaller the difference, the better the evaluation. X1+X2 on the left side of the graph is the image obtained by averaging the input image X1 and the input image X2. The magnitude of the difference was significantly large for X1+X2, while Y1, Y2, and Y3 were all similar and there was no significant difference (Wilcoxon sign-rank test, P<0.05 considered significant, same below). The b factor of the image was 800 (sec / mm 2 ) was decided.

[0106] Figure 18 is a graph showing the evaluation results using SSIM (Structural Similarity), an index showing structural similarity. The larger the SSIM, the better the evaluation. SSIM was significantly higher for Y3 than for X1+X2, and significantly lower for Y1 and Y2 than for X1+X2. The b factor of the image was 800 (sec / mm 2 ) was decided.

[0107] Figure 19 shows the difference (mm ) between the apparent diffusion coefficient ADC (Apparent Diffusion Coefficient) and the target image. 2 / s). Since it is a difference, the smaller the better. ADC is related to the diffusion coefficient of water molecules and differs between diseased tissue and healthy tissue. By comparing ADC, it is possible to determine not only the appearance of the image but also whether diffusion information is preserved. The results show that Y3 is significantly smaller than the other data. Also, Y1 and Y2 are significantly higher than X1 + X2. Note that when the b factor is 0 (sec / mm 2 ) and 800 (sec / mm 2 ADC was calculated using two images of the same subject.

[0108] From the above results, it was found that Y3 produced an image relatively close to the target image. Furthermore, by using processing route (II), it was found that it is possible to produce an image equivalent to that obtained by adding multiple images, while reducing the number of model layers directly involved in image output.

[0109] 〔summary〕 A noise removal device according to aspect 1 of the present invention comprises at least one processor that executes a noise removal process to remove noise from an MRI image, and at least one memory that stores a model to be used in the noise removal process, wherein the model is a model in which the MRI image is used as an input image and the noise-removed MRI image is used as an output image, and the pixel value of each pixel of the output image is configured to be within a range between the minimum value and the maximum value of the pixel values ​​of pixels in the vicinity of the pixel of the input image that corresponds to the pixel.

[0110] According to the above configuration, it is possible to realize a technique for generating an MRI image from which noise has been removed, taking into consideration the characteristics of the MRI image of a living body.

[0111] In a noise removal device according to aspect 2 of the present invention, the model may include a network that acts globally on the input image and n filters (n is a natural number greater than or equal to 2) that act locally on the input image, and the noise removal process may include a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0112] According to the above configuration, the same effects as those of the noise removal device according to the first aspect can be obtained.

[0113] In the noise removal device according to the third aspect of the present invention, the coefficients that define the network and the coefficients that define each of the n filters may be determined by machine learning.

[0114] According to the above configuration, the same effects as those of the noise removal device according to the first aspect can be obtained.

[0115] A noise removal device according to a fourth aspect of the present invention comprises at least one processor that executes a noise removal process for removing noise from an MRI image, and at least one memory that stores a model used in the noise removal process, wherein the model has the MRI image as an input image and an MRI image from which noise has been removed as an output image, and includes a network that acts globally on the input image and n filters (n is a natural number greater than or equal to 2) that act locally on the input image, and the noise removal process includes a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0116] According to the above configuration, the same effects as those of the noise removal device according to the first aspect can be obtained.

[0117] A noise removal method according to aspect 5 of the present invention includes a noise removal process in which at least one processor removes noise from an MRI image using a model stored in at least one memory, the model being a model in which the MRI image is an input image and the noise-removed MRI image is an output image, and the pixel value of each pixel of the output image is configured to be within a range from the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel of the input image corresponding to the pixel.

[0118] According to the above configuration, it is possible to realize a noise removal method that generates an MRI image from which noise has been removed, taking into consideration the characteristics of an MRI image of a living body.

[0119] A noise removal method according to a sixth aspect of the present invention includes a noise removal process in which at least one processor removes noise from an MRI image using a model stored in at least one memory, the model being a model in which the MRI image is an input image and the noise-removed MRI image is an output image, the model including a network that acts globally on the input image and n filters (n is a natural number greater than or equal to 2) that act locally on the input image, the noise removal process including a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0120] According to the above configuration, the same effects as those of the noise removal method according to the fifth aspect can be obtained.

[0121] A program according to aspect 7 of the present invention is a program for causing a computer to function as the noise removal device according to any one of aspects 1 to 4, and is a program for causing the computer to execute the noise removal process.

[0122] A recording medium according to an eighth aspect of the present invention is a computer-readable non-transitory recording medium having the program according to the seventh aspect recorded thereon.

[0123] A machine learning device according to aspect 9 of the present invention comprises at least one processor that executes a model construction process using supervised machine learning to construct a model in which an MRI image is an input image and an MRI image with noise removed is an output image, and at least one memory that stores the model, wherein the model is a model in which the MRI image is an input image and the MRI image with noise removed is an output image, and is configured such that the pixel value of each pixel of the output image is within a range between the minimum and maximum pixel values ​​of pixels in the vicinity of the pixel of the input image corresponding to the pixel.

[0124] According to the above configuration, it is possible to construct a model that generates MRI images from which noise has been removed, taking into consideration the characteristics of MRI images of a living body.

[0125] In a machine learning device according to aspect 10 of the present invention, the training data in the machine learning may be a pair of a first MRI image and a second MRI image from which noise has been removed, and the model construction process may include a derivation process for deriving secondary indices of the MRI image output from the model when the first MRI image is input, and secondary indices of the second MRI image, and an update process for updating coefficients that define the model so that the difference between the two secondary indices is reduced.

[0126] According to the above configuration, a machine learning device can be realized that learns to reduce the error in the secondary indicators between the output image and the teacher image, thereby constructing a model that generates more accurate output images.

[0127] In the machine learning device according to aspect 11 of the present invention, the MRI image may be a diffusion-weighted image, and the secondary index may be a diffusion tensor.

[0128] According to the above configuration, it is possible to realize a machine learning device suitable for removing noise from diffusion-weighted images.

[0129] In a machine learning device according to aspect 12 of the present invention, the training data in the machine learning may be a plurality of MRI images taken under common conditions, and the model construction process may include a process of updating coefficients that define the model so that the differences between MRI images output from the model when each of the plurality of MRI images is input are reduced.

[0130] According to the above configuration, since learning is performed to reduce the error between multiple output images, it is possible to realize a machine learning device that constructs a model that generates output images with higher accuracy.

[0131] In a machine learning device according to aspect 13 of the present invention, the model may include a network that acts globally on the input image and n filters (n is a natural number greater than or equal to 2) that act locally on the input image, and the noise removal process may include a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0132] According to the above configuration, the same effects as those of the machine learning device according to aspect 9 can be obtained.

[0133] A machine learning device according to aspect 14 of the present invention comprises at least one processor that executes a model construction process using supervised machine learning to construct a model having an MRI image as an input image and an MRI image that has been subjected to noise reduction as an output image, and at least one memory that stores the model, wherein the model includes a network that acts globally on the input image and n filters (n is a natural number greater than or equal to 2) that act locally on the input image, and the noise reduction process includes a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0134] According to the above configuration, it is possible to construct a model that generates MRI images from which noise has been removed, taking into account the characteristics of MRI images of a living body.

[0135] A machine learning method according to aspect 15 of the present invention includes a model construction process in which at least one processor constructs a model using a model stored in at least one memory through supervised learning, in which an MRI image is used as an input image and a noise-removed MRI image is used as an output image, wherein the model is configured so that the pixel value of each pixel in the output image is within a range greater than or equal to the minimum value and less than or equal to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel in the input image corresponding to the pixel.

[0136] According to the above configuration, it is possible to construct a model that generates MRI images from which noise has been removed, taking into account the characteristics of MRI images of a living body.

[0137] A machine learning method according to aspect 16 of the present invention includes a model construction process in which at least one processor constructs a model using a model stored in at least one memory through supervised learning, the model having an MRI image as an input image and an MRI image that has been subjected to noise reduction processing as an output image, the model having the MRI image as an input image and an MRI image that has been subjected to noise reduction processing as an output image, the model including a network that acts globally on the input image and n filters (n is a natural number greater than or equal to 2) that act locally on the input image, the noise reduction processing including a first process of deriving n weights from the input image using the network, a second process of generating n intermediate images from the input image using each of the n filters, and a third process of generating the output image by taking a weighted average of the n intermediate images using the n weights.

[0138] According to the above configuration, the same effects as those of the machine learning method according to aspect 15 can be obtained.

[0139] A program according to aspect 17 of the present invention is a program for causing a computer to function as a machine learning device according to any one of aspects 9 to 14, and is a program for causing a computer to execute the model building process.

[0140] A recording medium according to an eighteenth aspect of the present invention is a computer-readable non-transitory recording medium having the program according to the seventeenth aspect recorded thereon. [Explanation of symbols]

[0141] 1. Noise reduction device 11 Input / Output Interface 12 processors 13. Memory 2. Machine learning equipment 21 Input / Output Interface 22 processors 23 Memory

Claims

1. At least one processor that performs a noise reduction process to reduce noise in an MRI image, and at least one memory that stores a model used in the noise reduction process, The model uses the MRI image as an input image and an MRI image from which noise has been removed as an output image, and performs a limiting process using max pooling values ​​and minimum pooling values ​​on an intermediate output image so that the pixel value of each pixel in the output image is within a range from the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel in the input image corresponding to the pixel. A noise removal device characterized by:

2. the model includes a network that operates globally on the input image and n filters (n is a natural number equal to or greater than 2) that operate locally on the input image; The noise removal process includes: a first process of deriving n weights from the input image using the network; a second process of generating n intermediate images from the input image using each of the n filters; a third process of generating an intermediate output image by weighted averaging the n intermediate images using the n weights; and a limiting process for generating the output image by performing a limiting process using a max pooling value and a minimum pooling value on the intermediate output image.

2. The noise removal device according to claim 1.

3. The coefficients defining the network and the coefficients defining each of the n filters are determined by machine learning.

3. The noise removal device according to claim 2.

4. a denoising process in which at least one processor denoises the MRI image utilizing the model stored in the at least one memory; The model uses the MRI image as an input image and an MRI image from which noise has been removed as an output image, and performs a limiting process using max pooling values ​​and minimum pooling values ​​on an intermediate output image so that the pixel value of each pixel in the output image is within a range from the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel in the input image corresponding to the pixel. A noise removal method comprising:

5. a denoising process in which at least one processor denoises the MRI image utilizing the model stored in the at least one memory; The model uses the MRI image as an input image and an MRI image from which noise has been removed as an output image, and includes a network that acts globally on the input image and n filters (n is a natural number of 2 or greater) that act locally on the input image; The noise removal process includes: a first process of deriving n weights from the input image using the network; a second process of generating n intermediate images from the input image using each of the n filters; a third process of generating an intermediate output image by weighted averaging the n intermediate images using the n weights; and a limiting process for generating the output image by performing a limiting process using a max pooling value and a minimum pooling value on the intermediate output image so that the pixel value of each pixel of the output image falls within a range between the minimum value and the maximum value of pixel values ​​of pixels in the vicinity of a pixel of the input image corresponding to the pixel. A noise removal method comprising:

6. 4. A program for causing a computer to function as the noise removal device according to claim 1, the program causing the computer to execute the noise removal process.

7. A computer-readable non-transitory recording medium on which the program according to claim 6 is recorded.

8. The system includes at least one processor that executes a model construction process to construct a model using a supervised machine learning method, the model having an MRI image as an input image and an MRI image from which noise has been removed as an output image, and at least one memory that stores the model; The model uses the MRI image as an input image and an MRI image from which noise has been removed as an output image, and performs a limiting process using max pooling values ​​and minimum pooling values ​​on an intermediate output image so that the pixel value of each pixel in the output image is within a range from the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel in the input image corresponding to the pixel. A machine learning device characterized by:

9. the training data in the machine learning is a set of a first MRI image and a second MRI image from which noise has been removed; The model construction process includes a derivation process for deriving secondary indices of an MRI image output from the model when a first MRI image is input, and secondary indices of the second MRI image; and an update process for updating coefficients defining the model so that the difference between the two secondary indicators becomes smaller. The machine learning device according to claim 8 .

10. the MRI image is a diffusion-weighted image; The secondary index is the diffusion tensor. The machine learning device according to claim 9 .

11. The training data in the machine learning is a plurality of MRI images taken under common conditions, the model construction process includes a process of updating coefficients defining the model so that differences between MRI images output from the model when each of the plurality of MRI images is input are reduced. The machine learning device according to any one of claims 8 to 10.

12. the model includes a network that operates globally on the input image and n filters (n is a natural number equal to or greater than 2) that operate locally on the input image; The noise removal process includes: a first process of deriving n weights from the input image using the network; a second process of generating n intermediate images from the input image using each of the n filters; a third process of generating an intermediate output image by weighted averaging the n intermediate images using the n weights; and a limiting process for generating the output image by performing a limiting process using a max pooling value and a minimum pooling value on the intermediate output image. The machine learning device according to any one of claims 8 to 11.

13. The system includes at least one processor that executes a model construction process to construct a model using a supervised machine learning method, the model having an MRI image as an input image and an MRI image that has been subjected to noise removal as an output image, and at least one memory that stores the model; the model includes a network that operates globally on the input image and n filters (n is a natural number equal to or greater than 2) that operate locally on the input image; The noise removal process includes: a first process of deriving n weights from the input image using the network; a second process of generating n intermediate images from the input image using each of the n filters; a third process of generating an intermediate output image by weighted averaging the n intermediate images using the n weights; and a limiting process for generating the output image by performing a limiting process using a max pooling value and a minimum pooling value on the intermediate output image so that the pixel value of each pixel of the output image falls within a range between the minimum value and the maximum value of pixel values ​​of pixels in the vicinity of a pixel of the input image corresponding to the pixel. A machine learning device characterized by:

14. a model construction process in which at least one processor constructs a model by supervised learning using the model stored in at least one memory, the model having an MRI image as an input image and an MRI image from which noise has been removed as an output image; The model uses the MRI image as an input image and an MRI image from which noise has been removed as an output image, and performs a limiting process using max pooling values ​​and minimum pooling values ​​on an intermediate output image so that the pixel value of each pixel in the output image is within a range from the minimum value to the maximum value of the pixel values ​​of pixels in the vicinity of the pixel in the input image corresponding to the pixel. A machine learning method characterized by:

15. a model construction process in which at least one processor constructs a model by supervised learning using the model stored in at least one memory, the model having an MRI image as an input image and an MRI image that has been subjected to noise removal as an output image; The model uses the MRI image as an input image and an MRI image that has been subjected to noise removal processing as an output image, and includes a network that acts globally on the input image and n filters (n is a natural number of 2 or greater) that act locally on the input image; The noise removal process includes: a first process of deriving n weights from the input image using the network; a second process of generating n intermediate images from the input image using each of the n filters; a third process of generating an intermediate output image by weighted averaging the n intermediate images using the n weights; and a limiting process for generating the output image by performing a limiting process using a max pooling value and a minimum pooling value on the intermediate output image so that the pixel value of each pixel of the output image falls within a range between the minimum value and the maximum value of pixel values ​​of pixels in the vicinity of a pixel of the input image corresponding to the pixel. A machine learning method characterized by:

16. 14. A program for causing a computer to function as the machine learning device according to claim 8, the program causing the computer to execute the model building process.

17. A computer-readable non-transitory recording medium on which the program according to claim 16 is recorded.

18. the model includes a network that operates globally on the input image and n filters (n is a natural number equal to or greater than 2) that operate locally on the input image; The noise removal process includes: a first process of deriving n weights using two sets of three input images captured at different imaging positions using the network; a second process of generating n intermediate images from the input image using each of the n filters; a third process of generating an intermediate output image by weighted averaging the n intermediate images using the n weights; The noise removal device according to claim 1 , further comprising: a limiting process for generating the output image by performing a limiting process using a max pooling value and a minimum pooling value on the intermediate output image.

19. The training data in the machine learning is an image obtained by adding together five consecutively captured MRI images, The model construction process includes a derivation process for deriving a difference between an MRI image output from the model and the training data when two sets of three input images captured at different imaging positions are input; an update process for updating coefficients defining the model so that the difference becomes smaller. The machine learning device according to claim 8 .

Images