Medical image processing device

Abstract

The present invention addresses the problem of providing a medical image processing device for automating the measurement of an FA value.　The present invention provides, as a solution, a medical image processing device comprising: a diffusion tensor image MRI data acquisition unit that acquires diffusion tensor image MRI data resulting from imaging a lumbar section of a patient who has lower back and leg pain; a coordinate information estimation unit that estimates coordinate information for a nerve root in a designated axial cross section; and a measurement unit that measures the FA value of the nerve root for which the coordinate information was estimated.

Description

Medical image processing equipment

[0001] The present invention relates to a medical image processing device, a method for acquiring information on spinal nerves, a method for generating diffusion tensor tractography showing the course of nerve fibers of spinal nerves, and a method and program for generating a 3D FA map.

[0002] Low back pain is a common national complaint in Japan, ranking first among both men and women. The 2019 Low Back Pain Diagnostic Guidelines state that magnetic resonance imaging (MRI) is essential for diagnosing low back pain. Lumbar nerve disorders can cause lower back and lower limb pain, but asymptomatic disc degeneration and herniation are often observed. In conventional MRI diagnoses, nerve root compression on the image often does not necessarily indicate the cause of pain. This makes functional assessment, such as visualizing damaged nerves and quantifying pain, difficult.

[0003] Diffusion tensor imaging (DTI), which uses MRI to emphasize the directionality of water molecule diffusion within biological tissue, can track the diffusion anisotropy of nerve fibers and visualize them noninvasively. In particular, it has been reported that DTI can clearly visualize the three-dimensional course of the lumbar plexus. Furthermore, it has been reported that tractography is disrupted in lumbar foraminal stenosis, which is considered difficult to diagnose, and that the fractional anisotropy (FA) value, an index indicating the strength of diffusion anisotropy, decreases and correlates with clinical symptoms (Non-Patent Documents 1-3).

[0004] Eguchi Y, Ohtori S, Orita S, et al. , Preliminary Results. American Journal of Neuroradiology 2011;32:1824-1829. Eguchi Y, Ohtori S, Yamashita M, et al. , European Spine Journal 2010;19:1874-1882. Eguchi Y, Ohtori S, Yamashita M, et al. , Neuroradiology 2011;53:633-641.

[0005] However, measuring FA values ​​and creating tractography from DTI data are manual operations that require time and experience, and it is difficult to say that they are yet widely used in clinical practice.In order to solve the above problems, the present invention aims to provide a medical image processing device for automating FA value measurement, a method for acquiring spinal nerve information, a method for generating diffusion tensor tractography that shows the course of nerve fibers in spinal nerves, and a method and program for generating a 3D FA map.

[0006] The present invention includes the following aspects [1] to

[14] and [6-1]. [1] A medical image processing device comprising: a diffusion tensor imaging MRI data acquisition unit that acquires diffusion tensor imaging MRI data of the lumbar spine of a patient with lower back and lower limb pain; a coordinate information estimation unit that estimates coordinate information of a nerve root in a specified axial cross section; and a measurement unit that measures an FA value of the nerve root whose coordinate information has been estimated, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information; the coordinate information estimation unit estimates the coordinate information of the nerve root in the specified axial cross section using a trained model that has been trained using as training data images in which two nerve root locations and a spinal canal location included in the same axial cross section image of a diffusion-weighted MRI image of the lumbar spine have been annotated; and the measurement unit measures the FA value of the nerve root whose coordinate information has been estimated based on the scalar information and the coordinate information, based on the coordinate information of the nerve root estimated by the coordinate information estimation unit. [2] The medical image processing device according to [1], wherein the same axial cross-sectional image of the MRI diffusion-weighted image of the lumbar spine is a cross-sectional image of at least one of the third lumbar vertebra (L3), the fourth lumbar vertebra (L4), and the fifth lumbar vertebra (L5). [3] The medical image processing device according to [2], wherein the same axial cross-sectional image of the MRI diffusion-weighted image of the lumbar spine is a cross-sectional image of the third lumbar vertebra (L3), the fourth lumbar vertebra (L4), and the fifth lumbar vertebra (L5). [4] The medical image processing device according to any one of [1] to [3], wherein the same axial cross-sectional image of the MRI diffusion-weighted image of the lumbar spine is a plurality of cross-sectional images centered on the center of the same nerve root. [5] The medical image processing device according to [4], wherein the same axial cross-sectional image of the MRI diffusion-weighted image of the lumbar spine is a proximal, central, and distal cross-sectional image of the same nerve root. [6] The medical image processing device according to any one of [1] to [5], wherein the annotation of the nerve root portion and the annotation of the spinal canal portion are different in the training data. [6-1] The medical image processing device according to any one of [1] to [5], wherein the coordinate information estimation unit estimates the coordinates of the nerve root after removing only the spinal canal portion by image processing.[7] The medical image processing device according to any one of [1] to [6] and [6-1], wherein the specified axial sections are a plurality of consecutive axial sections in the lumbar spine of the patient, and further comprising a tractography generation unit that associates coordinate information of nerve roots in each axial section estimated by the coordinate information estimation unit with scalar information, vector information and coordinate information of FA map information created from the diffusion tensor imaging MRI data, and generates a diffusion tensor tractography showing the course of nerve fibers. [8] The medical image processing device according to any one of [1] to [6] and [6-1], wherein the specified axial sections are a plurality of consecutive axial sections in the lumbar spine of the patient, and further comprising a 3D FA map generation unit that associates coordinate information of nerve roots in each axial section estimated by the coordinate information estimation unit with FA values ​​measured by the measurement unit, and reconstructs the nerve roots in three-dimensional space to generate a 3D FA map. [9] A method for acquiring information on spinal nerves, the method comprising the steps of acquiring diffusion tensor imaging MRI data of the lumbar spine of a patient with lower back and lower limb pain, estimating coordinate information of a nerve root in a specified axial cross section, and measuring an FA value of the nerve root whose coordinate information has been estimated, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information, the step of estimating the coordinate information estimates the coordinate information of the nerve root in the specified axial cross section using a trained model trained using as training data an image in which two nerve root locations and a spinal canal location included in the same axial cross section image of a diffusion-weighted MRI image of the lumbar spine have been annotated, and the step of measuring the FA value of the nerve root measures the FA value of the nerve root whose coordinate information has been estimated based on the scalar information and the coordinate information, based on the coordinate information of the nerve root estimated in the step of estimating the coordinate information.

[10] The method according to [9], wherein the information on the spinal nerve obtained by the method according to [9] is an index for estimating the degree of spinal nerve damage from the measured FA value.

[11] A method for generating a diffusion tensor tractography showing the course of nerve fibers of spinal nerves, the method comprising the steps of: acquiring diffusion tensor imaging MRI data of the lumbar spine of a patient with lower back and lower limb pain; estimating coordinate information of nerve roots in each of a plurality of consecutive axial cross sections in the lumbar spine of a specified patient; associating the information; and superimposing the axial cross sections to form a three-dimensional image to generate a diffusion tensor tractography showing the course of nerve fibers, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information; and estimating the coordinate information estimates the coordinate information of the nerve roots in the specified axial cross section using a trained model trained using as training data images in which annotated images of two nerve root sites and a spinal canal site included in the same axial cross section image of the lumbar spine diffusion-weighted MRI image are used; The method, wherein the step of associating information associates the coordinate information of the nerve root in each axial cross section estimated by the step of estimating the coordinate information of the nerve root with the scalar information, vector information, and coordinate information of the FA value map created from the diffusion tensor image MRI data.

[12] A method for generating a 3D FA map, comprising the steps of: acquiring diffusion tensor imaging MRI data of the lumbar region of a patient with lower back and lower limb pain; estimating coordinate information of a nerve root in each of a plurality of consecutive axial sections in the lumbar region of a specified patient; measuring an FA value of the nerve root whose coordinate information has been estimated; and reconstructing the nerve root in a three-dimensional space using the coordinate information of the nerve root whose coordinate information has been estimated and the measured FA value corresponding to the coordinate information, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information; and estimating the coordinate information estimates the coordinate information of the nerve root in the specified axial section using a trained model trained using as training data an image in which two nerve root regions and a spinal canal region included in the same axial section image of the lumbar diffusion-weighted MRI image have been annotated,

[13] The method for generating a 3D FA map according to

[12] , further comprising the steps of: comparing the measured FA values ​​of two nerve roots whose coordinate information is estimated based on the scalar information and the coordinate information, based on the coordinate information of the nerve root estimated in the coordinate information estimating step; and displaying a nerve root region where there is a predetermined difference between the measured FA values ​​of the two nerve roots whose coordinate information is estimated included in the same axial cross-sectional image, and the other nerve root regions, in a manner that allows the nerve root to be distinguished from other nerve root regions.

[14] A program for causing a computer to execute each step of the method for acquiring spinal nerve information according to [9] or

[10] , the method for generating a diffusion tensor tractography showing the course of nerve fibers of a spinal nerve according to

[11] , or the method for generating a 3D FA map according to

[12] or

[13] .

[0007] The present invention provides a medical image processing device for automating the measurement of FA values, a method for acquiring information on spinal nerves, a method for generating a diffusion tensor tractography showing the course of nerve fibers in spinal nerves, and a method and program for generating a 3D FA map. In a preferred embodiment of the present invention, the creation of tractography can also be automated.

[0008] Figure 1 shows parameters representing anisotropic diffusion at coordinates (X, Y, Z). This shows that the direction of water molecule movement is restricted, resulting in anisotropic diffusion. Figure 2(A) shows the annotation target area. Figure 2(B) shows an example of annotation performed on two nerve root regions and a spinal canal region. Figure 3 shows the structure of a U-Net. Figure 4 shows the original image (left) and the predicted image (right) actually obtained in the evaluation of the trained model. Figure 5 shows the diffusion-weighted image and annotation image used as training data in the trained model [1-1]. Figure 6 shows the results and extracted image from the trained model [1-1]. Figure 7 is a schematic diagram of a method for automatically measuring FA values. Figure 8 shows the results of a comparison between automatic (AI) and manual FA value measurement. Figure 9 shows an example of a match between disruption points in tractography and a 3D FA map.

[0009] [Medical Image Processing Apparatus] The medical image processing apparatus of the present invention is a medical image processing apparatus comprising: a diffusion tensor imaging MRI data acquisition unit that acquires diffusion tensor imaging MRI data obtained by photographing the lumbar spine of a patient with lower back and lower limb pain; a coordinate information estimation unit that estimates coordinate information of a nerve root in a specified axial cross section; and a measurement unit that measures an FA value of the nerve root whose coordinate information has been estimated, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information; the coordinate information estimation unit estimates the coordinate information of the nerve root in the specified axial cross section using a trained model that has been trained using as training data an image in which two nerve root sites and a spinal canal site included in the same axial cross section image of a diffusion-weighted MRI image of the lumbar spine have been annotated; and the measurement unit measures the FA value of the nerve root whose coordinate information has been estimated based on the scalar information and the coordinate information, based on the coordinate information of the nerve root estimated by the coordinate information estimation unit.

[0010] <Patient> The medical image processing device of the present invention is intended for patients suffering from lower back and lower limb pain. Lower back and lower limb pain refers to either or both lower back pain and lower limb pain. Examples of patients suffering from lower back and lower limb pain include patients who experience subjective symptoms of lower back and lower limb pain. Lower back and lower limb pain is primarily caused by lumbar nerve compression, lumbar foraminal stenosis, or the like in one or more of the third lumbar vertebra (L3), fourth lumbar vertebra (L4), and fifth lumbar vertebra (L5) in the lumbar spine. One object of the present invention is to provide a medical image processing device for visualizing and quantitatively evaluating lower back and lower limb pain caused by nerve damage due to lumbar nerve compression, lumbar foraminal stenosis, or the like in the lumbar spine. It is expected that visualization and quantitative evaluation will contribute to auxiliary diagnosis.

[0011] <Anisotropic Diffusion> Anisotropic diffusion will be described in detail below. In nerve fibers, the axon cell membrane and myelin sheath prevent water diffusion in a direction perpendicular to the nerve fiber bundle, so isotropic diffusion of water molecules is lost in the nerve fiber, resulting in anisotropic diffusion. The FA (Fractional Anisotropy) value is an index indicating the strength of anisotropic diffusion. The FA value ranges from 0 to 1, with values ​​closer to 1 indicating stronger anisotropic diffusion, and 0 indicating completely isotropic diffusion. It is known that nerve damage in nerve fibers reduces the diffusion anisotropy along the nerve fiber and decreases the FA value. Therefore, it is known that the FA value can quantitatively evaluate nerve damage. Therefore, the FA value is preferably used as an index of the degree of spinal nerve damage.

[0012] The parameters representing anisotropic diffusion in the coordinate system (X, Y, Z) are shown in Figure 1. Each axis of the coordinate system (X, Y, Z) corresponds to the coordinate system of the MRI data. Figure 1 shows that the direction of water molecule movement is restricted, resulting in anisotropic diffusion. As shown in the figure, the direction showing the strongest diffusion is defined as X', and the two directions perpendicular to the X' direction are defined as Y' and Z', and the diffusion coefficient in each direction is defined as λ １ , λ ２ , λ ３ (λ １ ≧λ ２ ≧λ ３ ) λ １ , λ ２ , λ ３ are called eigenvalues. Furthermore, each eigenvalue (λ １ , λ ２ , λ ３ ) The vectors in each direction corresponding to e １ , e ２ , e ３ These are called eigenvectors. In this specification, the eigenvalues ​​(λ １ , λ ２ , λ ３ ) and eigenvectors (e １ , e ２ , e ３The combination of these vectors is called vector information. Once the vector information is obtained, the degree of anisotropic diffusion within a voxel can be quantitatively calculated.

[0013] Eigenvalue (λ １ , λ ２ , λ ３ ) and eigenvectors (e １ , e ２ , e ３ ) can quantitatively express anisotropic diffusion at specific coordinates (X, Y, Z). The diffusion tensor D at the coordinates (X, Y, Z) is given by a 3x3 determinant shown in Eq. 1. Here, for example, D ＸＸ represents the diffusion coefficient when a motion problem gradient (MPG) is applied in the X-axis direction. Similarly, the diffusion coefficients are calculated for each of the six directions XX, XY, XZ, YY, YZ, and ZZ (D ＸＸ , D ＸＹ , D ＸＺ , D ＹＹ , D ＹＺ , D ＺＺ ), when the diffusion tensor D shown in Eq. 1 is diagonalized, each eigenvalue (λ, λ) is obtained as shown in Eq. 2. ２ , λ ３ ) can be obtained.

[0014]

[0015] Each eigenvalue (λ １ , λ ２ , λ ３ ), the apparent diffusion coefficient (ADC) can be calculated by equation (Eq. 3), and the FA value (fractional anisotropy value) can be calculated by equation (Eq. 4).

[0016]

[0017] <Diffusion Tensor Imaging MRI Data Acquisition Unit> The medical image processing device of the present invention includes a diffusion tensor imaging MRI data acquisition unit. In this specification, "MRI" refers to magnetic resonance imaging. MRI is measured in voxel units, and the aforementioned coordinates usually refer to voxel coordinates. The diffusion tensor imaging MRI data acquisition unit acquires diffusion tensor imaging MRI data acquired by an MRI device. The diffusion tensor imaging MRI data may be data acquired immediately before or data acquired in the past. In the case of data acquired in the past, data stored in a computer-readable medium, for example, can be used. The diffusion tensor imaging MRI data can be acquired, for example, using a known MRI device. The diffusion tensor imaging MRI data is acquired from the lumbar spine of a patient with lower back and lower limb pain. The lumbar region of a patient refers to the area including the spinal canal and its surroundings, more preferably the area including the spinal canal of the third lumbar vertebra (L3), the fourth lumbar vertebra (L4), and the fifth lumbar vertebra (L5) and their surroundings.

[0018] (Diffusion Tensor Imaging MRI Data) Diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information. Diffusion-weighted image information A diffusion-weighted image (DWI) is an image that emphasizes the movement (diffusion) of water molecules, and is created, for example, by applying a motion problem gradient (MPG) from a fixed direction to the diffusion motion of protons.

[0019] FA Map Information The FA map information includes scalar information, vector information, and coordinate information. That is, the FA map information includes, for each coordinate (X, Y, Z), scalar information indicating the magnitude of the FA value and the aforementioned eigenvalue (λ １ , λ ２ , λ ３ ) and eigenvectors (e １ , e ２ , e ３ ) combinations are associated.

[0020] <Coordinate Information Estimation Unit> The medical image processing device of the present invention includes a coordinate information estimation unit that estimates coordinate information of a nerve root in an axial plane specified in the diffusion-weighted image. An axial plane is a plane that is perpendicular to the body axis, which is a line connecting the head and tail of a measurement subject. The axial plane is specified, for example, by a measurement target designation unit that is further included in the medical image processing device of the present invention. The axial plane can be designated by an operator of the medical image processing device via the measurement target designation unit, for example. The medical image processing device may also be configured to automatically designate the axial plane.

[0021] The coordinate information estimation unit estimates the coordinate information of the nerve root in the specified axial section using a trained model trained using training data of images annotated with nerve root regions (usually two regions) and spinal canal regions (usually one region) included in the same axial section image of a lumbar vertebrae diffusion-weighted MRI image. The trained model has an algorithm in which an artificial neural network having a multi-layered neuron structure performs calculations, extracts an extraction pattern from the training data, and applies the pattern to new data to output an estimation result.

[0022] In one preferred embodiment, the coordinate information estimation unit estimates the coordinates of the nerve roots after removing only the spinal canal region through image processing. Specifically, since the spinal canal region represents the largest area in the image, removing the largest area allows removing only the spinal canal region. This is particularly suitable when the annotations for the nerve root region and the spinal canal region are identical in the training data described below.

[0023] (Teacher Data) The teacher data and its preferred embodiments are described in detail below. The trained model used by the coordinate information estimation unit is trained using as teacher data images annotated with two nerve root regions and one spinal canal region, which are typically included in the same axial cross-sectional image of a diffusion-weighted MRI image of the lumbar spine. Using an image annotated with two nerve root regions and one spinal canal region improves the accuracy of the estimated nerve root coordinate information. The annotations for the nerve root regions and the spinal canal region may be the same or different. In one preferred embodiment, the annotations for the nerve root regions and the spinal canal region are different. In another preferred embodiment, the annotations for the nerve root regions and the spinal canal region are the same. The MRI diffusion-weighted images of the lumbar spine used as teacher data can be obtained from either healthy individuals or patients. From the viewpoint of data standardization, the data is preferably obtained from healthy individuals. The axial cross-sectional image is at least one axial cross-sectional image of the third lumbar vertebra (L3), the fourth lumbar vertebra (L4), and the fifth lumbar vertebra (L5), and more preferably three types of axial cross-sectional images of the third lumbar vertebra (L3), the fourth lumbar vertebra (L4), and the fifth lumbar vertebra (L5). Furthermore, the MRI diffusion-weighted images of the lumbar region used as training data are preferably multiple cross-sectional images centered on the center of the same nerve root. Among these, cross-sectional images of the proximal, central, and distal portions of the same nerve root are more preferred. In healthy nerves, the FA value tends to increase in the proximal, central, and distal portions of the nerve root. Therefore, it is desirable to compare the FA values ​​of the healthy side and the compressed side in the same slice at multiple locations centered on the center of the same nerve root and compare the relative changes.

[0024] <Measurement Unit> The medical image processing device of the present invention includes a measurement unit that measures the FA value of the nerve root whose coordinate information is estimated based on the scalar information and the coordinate information of the nerve root estimated by the coordinate information estimation unit. Specifically, the measurement unit identifies the scalar information associated with the coordinate information in the FA map information corresponding to the coordinate information of the nerve root whose coordinate information is estimated as the FA value of the nerve root whose coordinate information is estimated. The coordinate information of the estimated nerve root usually includes multiple voxels. In such a case, the average value of the FA values ​​of each voxel is identified as the FA value of the nerve root whose coordinate information is estimated.

[0025] <Diffusion Tensor Tractography> Diffusion tensor tractography (DTT; hereinafter simply referred to as tractography) is an image that visualizes anisotropic nerve fibers by tracking them three-dimensionally based on the diffusion tensor. Tractography can visualize the disruption of nerve trajectories. Tractography involves setting a starting point (seed) for nerve tracking in an arbitrary region set as a ROI (Region of Interest), and calculating the eigenvalue (λ) of the starting point. １ , λ ２ , λ ３ ) and eigenvectors (e １ , e ２ , e ３ ) is calculated. The vector e is the direction in which the diffusion is maximum. １ The next eigenvalue of the coordinate encountered along １ , λ ２ , λ ３ ) and eigenvectors (e １ , e ２ , e ３ ) is calculated. Then, the vector e １ In this way, the course of nerve fibers can be tracked, and this method is also called nerve tracking.

[0026] <Tractography Generation Unit> The medical image processing device of the present invention may further include a tractography generation unit. The tractography generation unit associates the coordinate information of the nerve root in each of a plurality of consecutive axial cross sections estimated by the coordinate information estimation unit with the scalar information, vector information, and coordinate information, and generates a diffusion tensor tractography showing the course of nerve fibers. Specifically, for example, the diffusion tensor tractography is generated as follows: First, in one axial cross section selected from a plurality of consecutive axial cross sections, an ROI is set as a starting point for nerve tracking based on the coordinate information of the nerve root whose coordinate information has been estimated. Then, an eigenvalue (λ) within the ROI is calculated from the vector information associated with the coordinate information of the nerve root. １ , λ ２ , λ ３ ) and eigenvectors (e １ , e ２ , e ３ ) is calculated. This determines the direction of the diffusion tensor within the ROI. Next, the direction e where the diffusion is maximum is calculated. １ Start tracking along the direction e １ Similarly, the diffusion tensor (i.e., each eigenvalue and each eigenvector) is calculated for the voxel corresponding to the coordinate information of the nerve root encountered in the above. １ , λ ２ , λ ３ ) and eigenvalues ​​(λ １ , λ ２ , λ ３ ) is included. In a plurality of continuous axial sections, the vector e, which is the direction in which the diffusion is maximum, is included. １ The same process is repeated along the line, and nerve tracking is performed one after another to obtain a three-dimensionally constructed tractography. On the other hand, when a voxel with no continuity is encountered, nerve tracking is terminated. In another embodiment, the eigenvector (e １ , e ２ , e ３ ), eigenvalues ​​(λ １ , λ ２ , λ ３ ) and eigenvalues ​​(λ１ , λ ２ , λ ３ ) to generate fiber tracking data, and then exclude the fiber tracking data of nerves that pass through coordinates other than the coordinate information of the nerve root in each axial cross section for which coordinate information has been estimated, thereby generating diffusion tensor tractography.

[0027] <3D FA Map Generator> The medical image processing device of the present invention may further include a 3D FA map generator. When the 3D FA map generator is included, the specified axial planes are a plurality of consecutive axial planes in the patient's lumbar spine, and the 3D FA map generator associates the coordinate information of the nerve root in each of the consecutive axial planes estimated by the coordinate information estimation unit with the FA value measured by the measurement unit, reconstructs the nerve root in three-dimensional space, and generates a 3D FA map. The 3D FA map generator preferably compares the measured FA values ​​of two nerve roots whose coordinate information is estimated in the same axial cross-sectional image, and displays nerve roots whose measured FA values ​​are different by a predetermined amount from other nerve roots whose coordinate information is estimated in the same axial cross-sectional image. The predetermined difference is, for example, a difference of 0.1 or more in the FA value. Here, displaying nerve roots for which there is a predetermined difference in the measured FA values ​​of two nerve roots for which coordinate information has been estimated in a manner that allows them to be distinguished from other nerve roots for which the difference is less than the predetermined value means (1) displaying only nerve roots for which the difference is greater than the predetermined value in a manner that allows them to be distinguished, (2) displaying only nerve roots for which the difference is less than the predetermined value in a manner that allows them to be distinguished from each other, so that they do not exist, and (3) displaying nerve roots for which the difference is greater than the predetermined value in a manner that allows them to be distinguished from each other, for example, by displaying them in different colors or by displaying them in different shades of color, so that they are visually distinguished from each other.

[0028] The medical image processing apparatus of the present invention is realized by a computer equipped with, for example, a CPU, RAM, ROM, a hard disk, a display, a keyboard, a mouse, a communication interface, and the like.

[0029] [Method for Acquiring Spinal Nerve Information] The present invention also provides a method for acquiring spinal nerve information, the method comprising the steps of acquiring diffusion tensor imaging MRI data of the lumbar spine of a patient suffering from lower back and lower limb pain, estimating coordinate information of a nerve root in a specified axial plane, and measuring an FA value of the nerve root whose coordinate information has been estimated, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information, the step of estimating coordinate information estimates the coordinate information of the nerve root in the specified axial plane using a trained model trained using as training data images in which two nerve root regions and a spinal canal region included in the same axial plane image of the lumbar diffusion-weighted MRI image have been annotated, and the step of measuring the FA value of the nerve root measures the FA value of the nerve root whose coordinate information has been estimated based on the scalar information and the coordinate information, based on the coordinate information of the nerve root estimated in the step of estimating coordinate information.

[0030] The information on the spinal nerves obtained by the method of the present invention is suitably used as an index for estimating the degree of spinal nerve damage from the measured FA value.

[0031] The method for acquiring information on spinal nerves of the present invention can be carried out using, for example, the above-described medical image processing device.

[0032] [Method for generating a diffusion tensor tractography showing the course of nerve fibers in a spinal nerve] The present invention also provides a method for generating a diffusion tensor tractography showing the course of nerve fibers in a spinal nerve. The method of the present invention for generating a diffusion tensor tractography showing the course of nerve fibers of spinal nerves comprises the steps of: acquiring diffusion tensor imaging MRI data of the lumbar region of a patient with lower back and lower limb pain; estimating coordinate information of nerve roots in each of a plurality of consecutive axial cross sections in the lumbar region of a specified patient; associating the information; and superimposing the axial cross sections to form a three-dimensional image to generate a diffusion tensor tractography showing the course of nerve fibers, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information; and the step of estimating the coordinate information estimates the coordinate information of the nerve roots in the specified axial cross section using a trained model trained using as training data an image in which two nerve root locations and a spinal canal location included in the same axial cross section image of the lumbar diffusion-weighted MRI image are annotated, The step of associating the information associates the coordinate information of the nerve root in each axial cross section estimated by the step of estimating the coordinate information of the nerve root with the scalar information, vector information, and coordinate information of the FA value map created from the diffusion tensor image MRI data.

[0033] The method of the present invention for generating a diffusion tensor tractography showing the course of nerve fibers of spinal nerves can be implemented using, for example, the above-described medical image processing device.

[0034] [Method for Generating a 3D FA Map] The present invention also provides a method for generating a 3D FA map, which depicts an object having FA value information of a nerve root along the course of the nerve fiber.

[0035] A method for generating a 3D FA map of the present invention includes the steps of acquiring diffusion tensor imaging MRI data of the lumbar region of a patient with lower back and lower limb pain, estimating coordinate information of a nerve root in each of a plurality of consecutive axial sections in the lumbar region of a specified patient, measuring FA values ​​of the nerve roots whose coordinate information has been estimated, and reconstructing the nerve root in a three-dimensional space using the coordinate information of the nerve roots whose coordinate information has been estimated and the measured FA values ​​corresponding to the coordinate information, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information, and the step of estimating the coordinate information estimates the coordinate information of the nerve root in the specified axial section using a trained model trained using as training data an image in which two nerve root regions and a spinal canal region included in the same axial section image of the lumbar diffusion-weighted MRI image have been annotated, The step of measuring the FA value of the nerve root measures the FA value of the nerve root whose coordinate information has been estimated based on the scalar information and the coordinate information, based on the coordinate information of the nerve root estimated by the step of estimating the coordinate information.

[0036] Reconstruction of the nerve root in three-dimensional space is performed using the coordinate information of the estimated nerve root and the measured FA value corresponding to the coordinate information. Reconstruction of the nerve root in three-dimensional space is performed by creating an object of the nerve root whose coordinate information is estimated and arranging the created object in the sequential order of the axial cross sections. At this time, the coordinate information of the nerve root in each axial cross section is associated with the measured FA value corresponding to the coordinate information, and the created object has FA value information of the corresponding nerve root. Object creation is performed by acquiring the contours of the left and right nerve roots for each axial cross section image from the coordinate information of the estimated nerve root, and based on the acquired contour information. The contour may be the contour of the nerve root ROI, or may be a rectangle that circumscribes the ROI, i.e., surrounds the nerve root. For example, in the case of a rectangle, the coordinates of the upper left corner of the rectangle and the vertical and horizontal lengths are acquired as contour information. The acquired information may be reduced before being used to create the object. This makes it possible to display a 3D FA map even if the structure becomes too large during nerve root reconstruction.

[0037] The method for generating a 3D FA map may further include the steps of: comparing measured FA values ​​of two nerve roots whose coordinate information included in the same axial cross-sectional image is estimated; and distinguishably displaying nerve root regions where there is a predetermined difference between the measured FA values ​​of two nerve roots whose coordinate information included in the same axial cross-sectional image is estimated from other nerve root regions.

[0038] The predetermined difference is, for example, a difference of 0.1 or more in the FA value. If there is a predetermined difference in the FA value, the nerve root is displayed so as to be distinguishable from other nerve roots whose FA values ​​are less than the predetermined difference. Here, "displaying a nerve root whose coordinate information is estimated and whose measured FA values ​​have a predetermined difference so as to be distinguishable from other nerve roots whose difference is less than the predetermined difference" means (1) displaying only nerve roots whose difference is less than the predetermined difference so as to be distinguishable, (2) displaying only nerve roots whose difference is less than the predetermined difference so as to be distinguishable from other nerve roots whose difference is less than the predetermined difference, and (3) visually distinguishing between nerve roots whose difference is less than the predetermined difference and those whose difference is less than the predetermined difference, for example, by displaying them in different colors or by displaying them in different shades of color.

[0039] Specifically, the generation of the 3D FA map can be performed using, for example, known software. For example, it can be performed as follows using Three.js, software that draws 3D objects on a web browser using html and JavaScript files. First, a 40x40x40 zero-filled array is created using JavaScript. A CSV file containing nerve root position coordinate information and FA value information is read, and if there is a predetermined difference in the FA value, it is determined to display it in a distinguishable manner, such as by displaying it in a different color. The THREE.Scene function is used to create a 3D scene, which is a space for displaying objects. A camera and light are created to set the viewpoint for viewing the 3D space and the position and brightness of the light that illuminates the scene. The object is placed in the 3D space based on the array of nerve root coordinate information. At this time, the getPosition function is used to calculate the x, y, and z coordinates of the box (object). These are then stored in document.body. It is added to the HTML document with the appendChild(renderer.domElement) property, and the Render function continuously redraws the entire scene to the browser.

[0040] The 3D FA map generated by the method of the present invention is preferably used as a means for visualizing the extent of spinal nerve damage. The method for generating a 3D FA map of the present invention can be performed using, for example, the above-mentioned medical image processing device.

[0041] [Program] The program of the present invention is a program for causing a computer to execute the steps of the method for acquiring spinal nerve information, the steps of the method for generating a diffusion tensor tractography showing the course of nerve fibers of the spinal nerve, or the steps of the method for generating a 3D FA map. Examples of the computer include a desktop or portable personal computer (PC) or a workstation. The computer may be constructed by combining multiple computers of any type connected via a communication network such as the Internet or an intranet. The program is provided by being recorded on a recording medium such as a CD-ROM, DVD-ROM, or semiconductor memory. Alternatively, the program may be provided as a data signal via the communication network.

[0042] [1] Construction of a prediction model A model for predicting the location of the L3-L5 nerve root was constructed using an axial diffusion-weighted image of the area including the L3-L5 nerve as input.

[0043] (1) Software and hardware used: MRIconGL: www.nitrc.org / projects / mricrogl; free software Diffusion Toolkit: trackvis.org / download / ; free software med2image: github.com / FNNDSC / med2image; open source Python package milesial / Pytorch-UNet: github.com / milesial / Pytorch-UNet

[0044] (2) Creation of training data. DTI data from 90 patients who underwent MRI scans at the Department of Radiology, Chiba University Hospital, was used. The MRI data, written to a DVD in the hospital, was loaded onto a PC in DICOM format and converted to a Nifti format file using dcm2niix, included in MRIconGL. The Nifti format file containing the DTI data obtained by the conversion and the Gradient Table file were then input into the free software Diffusion Toolkit for data reconstruction. The resulting Nifti format file containing the diffusion-weighted data was used to generate 256 x 256 pixel JPEG images of all axial slices from the diffusion-weighted data using med2image. All generated images were blacked out except for the 116 x 106 pixel region shown in Figure 2(A). The resulting axial diffusion-weighted images were visually inspected one by one, and for images in which the presence of a nerve root was confirmed, two nerve root regions and the spinal canal region were annotated (Figure 2(B)). Because the target nerve root region is small, the purpose of this was to reduce unnecessary regions and increase feature counts, thereby improving learning accuracy. This procedure was performed on the DTI data of 90 patients, resulting in a total of 839 diffusion-weighted images and annotation images of two nerve root regions and the spinal canal region. The 839 images were obtained from approximately 90 patient data at three locations (L3, L4, and L5), each at three locations: proximal, central, and distal.

[0045] (3) Construction of trained model U-Net was used as the segmentation model. U-Net is a network for estimating image segmentation, and uses Fully Convolution Network, Deconvolution, and Skip-Connection. The structure of U-Net is shown in Figure 3. The structure of U-Net is an hourglass-shaped Encoder-Decoder network that forms a feature pyramid.

[0046] The segmentation model used was milesial / Pytorch-UNet. Using Python 3.6.0 and Pytorch 1.12.1, Pytorch-UNet is a custom model trained from scratch using 5,000 images and U-Net, which achieved a Dice coefficient of 0.98 on over 100,000 test images.

[0047] The development and experiment environments are as follows: OS: Windows 11 GPU: NVIDIA GeForce RTX 3070Ti CUDA: 11.3 CuDNN: 8.5.0 Learning rate: 0.00001 Number of Epochs: 100 Batch size: 1 Optimizer: RMSProp Loss function: CrossEntropyLoss

[0048] (4) Evaluation of the trained model The F-measure (also known as the Dice coefficient) was used as the evaluation index for the model. Table 1 shows the confusion matrix.

[0049]

[0050] A confusion matrix is ​​a table used to show the prediction accuracy in binary classification. The recall obtained from this matrix is ​​defined by equation (1), precision by equation (2), and the harmonic mean of recall and precision (F-measure) by equation (3).

[0051]

[0052] In a confusion matrix, recall represents the proportion of predicted data that is true compared to data that is actually true, and precision represents the proportion of data that is actually true compared to data predicted to be true. In other words, the higher the recall in data prediction, the fewer correct answers are missed, and the higher the precision, the fewer incorrect answers there are. The higher the F-value, which is the harmonic mean of these, the higher the similarity between the predicted data and the original data. The F-value was calculated using Pytorch.

[0053] Table 2 shows the most accurate results among the F-values ​​and Train Loss values ​​obtained by learning a total of 839 data sets of pairs of diffusion-weighted images and annotation images five times.

[0054]

[0055] The F-score and Train Loss values ​​shown here are the averages of the values ​​calculated for each epoch. The original image and the predicted image obtained are shown in Figure 4. The left side of Figure 4 is the original image, and the right side is the predicted image obtained.

[0056] Although the predicted image shows a larger area, it can be said that a total of three areas have been predicted: two nerve root areas and one spinal canal area.

[0057] (5) Preliminary Study of Annotation Conditions: The annotation conditions for the training data were preliminarily studied. (i) In the training data, we compared annotations for two nerve root locations and one spinal canal location (a total of three locations) with annotations for only two nerve root locations. The former resulted in improved prediction accuracy for the nerve root location. This is thought to be because the model captured the pattern of the positional relationship between the spinal canal and nerve root locations and used it as additional context information for training. (ii) In the training data, we compared annotation target regions of 256 x 256 pixels with those of 116 x 106 pixels. As a result, we were able to narrow the field of view (FOV) to the area around the spinal cord, making the nerve root location more clearly defined.

[0058] [1-1] Improvement of trained model In "(2) Creation of training data" in [1] above, annotation was performed with separate labels for two nerve root regions and the spinal canal region (dural canal). Similar to [1] above, this was performed on the DTI data of 90 individuals, resulting in a total of 839 diffusion-weighted images and corresponding annotation images for the two nerve root regions and the spinal canal region. Figure 5 shows the diffusion-weighted images and annotation images.

[0059] Similar to "(3) Construction of a trained model" in [1] above, the model was trained by inputting diffusion-weighted images as input data and annotation images as training data into the network, and the model was made to learn how to label the input data. The input diffusion-weighted images and annotation image pairs for 90 cases were divided into Training: Validation: Test in a ratio of 50: 25: 15 (cases).

[0060] The learning rate at which the accuracy was best = 10 －４ Resnet34 was used as the feature extraction model. The segmentation task was evaluated using the Dice coefficient as an index, similar to "(4) Evaluation of the trained model" in [1] above. The Dice coefficient was 0.780±0.105 on average for the test data. Figure 6 shows the results and extracted images.

[0061] The development environment and parameters are as follows: OS: Windows 10 GPU: Tesla T4 Python: 3.8.16 Python: 1.12.1 Sementation models Python: 0.2.1 Learning rate: 10 －２ , 10 －３ , 10 －４ Number of training times: 40 Batch size: 6 Classification classes: 3 (background, spinal canal, nerve root)

[0062] [2] Automatic measurement of FA value The FA value was automatically measured using the trained model constructed in [1] above, and compared with the FA value obtained by manual measurement.

[0063] (1) Software used and development environment The software used is the same as in [1] above. The development environment is as follows: OS: MacOS Big Sur Processor: 3.1 GHz Intel Core i5 Memory: 16 GB 1867 MHz DDR3

[0064] (2) Data Used Data from one patient was extracted from the DTI data of a total of 90 patients who underwent MRI scans at the Radiology Department of Chiba University Hospital, which was used to build the above prediction model.

[0065] (3) Generation of Diffusion-Weighted Images MRI data was loaded into a PC in DICOM format and converted to a Nifti format file using dcm2niix included in MRIconGL. Next, the Nifti format file containing the DTI data obtained by conversion and the Gradient Table file were input into the Diffusion Toolkit to reconstruct the data. The parameters of the Diffusion Toolkit were as follows: Imaging model: DTI Maximum b value: 1000 Gradient table: gradient_table Orientation patch: Invert Y The Nifti format file containing the diffusion weighted data obtained at that time was used to generate all slice images of the axial section from the diffusion weighted data as JPEG images of 256 x 256 pixels using med2image.

[0066] (4-1) Manual measurement of FA values ​​Each axial diffusion-weighted image was visually checked, and for images in which the presence of a nerve root was recognized, two nerve root locations were identified and designated as regions of interest (ROIs). All coordinate information and FA values ​​of the ROIs were obtained, and the average FA value was calculated.

[0067] (4-2) Automatic measurement of FA value Using the trained model constructed in [1] above, the generated diffusion-weighted images of the axial section were read one by one, and two nerve root regions and one spinal canal region were extracted and generated as JPEG files. The generated predicted image was binarized, and only the spinal canal region was removed by image processing to remove the largest area in the image. Using all images after image processing, all coordinate information and FA values ​​of the predicted portion were obtained, and the average FA value was calculated. Figure 7 shows a schematic diagram of the method for automatically measuring FA values.

[0068] (5) Results Figure 8 shows the comparison results between automatic measurement of FA values ​​(AI measurement) and manual measurement of FA values. It was revealed that the FA values ​​obtained by automatic measurement (AI measurement) and manual measurement were equivalent.

[0069] [3] Creation of lumbar nerve tractography (1) Software used and development environment In addition to the software used in [1] above, the following software was used: Trackvis: trackvis.org / download / ; free software The development environment is as follows: OS: MacOS Big Sur Processor: 3.1GHz Intel Core i5 Memory: 16GB 1867MHz DDR3

[0070] (2) Data Used Data from one patient was extracted from the DTI data of a total of 90 patients who underwent MRI scans at the Radiology Department of Chiba University Hospital, which was used to build the above prediction model.

[0071] (3) Generation of Diffusion-Weighted Images MRI data was loaded into a PC in DICOM format and converted to a Nifti format file using dcm2niix included in MRIconGL. Next, the Nifti format file containing the DTI data obtained by conversion and the Gradient Table file were input into the Diffusion Toolkit to reconstruct the data. The parameters of the Diffusion Toolkit were as follows: Imaging model: DTI Maximum b value: 1000 Gradient table: gradient_table Orientation patch: Invert Y The Nifti format file containing the diffusion weighted data obtained at that time was used to generate all slice images of the axial section from the diffusion weighted data as JPEG images of 256 x 256 pixels using med2image.

[0072] (4) Automatic Creation of Lumbar Nerve Tractography Using the trained model constructed in [1] above, the generated axial diffusion-weighted images were read one by one, and two nerve root regions and one spinal canal region were extracted and generated as JPEG files. The generated predicted image was binarized, and only the spinal canal region was removed using image processing to remove the largest area in the image. Based on the coordinate information of the nerve root region remaining in the image, an ROI file for use in Trackvis was created in Nifti format. Finally, a lumbar nerve tractography was created by reading the ROI file created using the Trackvis command. The above process was implemented using a Python script.

[0073] (5) Results The average time required for three automated lumbar nerve tractography runs, and the average time required for three manual runs as a control, are shown in Table 3. The difference in average time required for manual and automated lumbar nerve tractography runs was 160 seconds, meaning that automation allows for creation in significantly less time than manual methods.

[0074]

[0075] [4] Creation of 3D FA Map (1) Data Used The DTI data of 20 patients who underwent MRI scans at the Radiology Department of Chiba University Hospital and for whom tractography was available were used. (2) Creation of 3D FA Map The trained model constructed in [1] above was used to extract coordinate information for the nerve roots. FA values ​​at the corresponding positions were also calculated. Specifically, the following function in the Python library nibabel was used to obtain a numpy array containing the respective FA values ​​and ROI information. Function: def get_nii_fdata(nii_path): nii = nib.load(nii_path) return nii.get_fdata()

[0076] Slice images with pixel values ​​only in the ROI portion of each axial image were created and saved. The two largest contours were extracted from the generated image, and the first was stored as the left side and the second as the right side. This image processing was performed using the OpenCV cv2.findContours function. Furthermore, a circumscribing rectangle was drawn around the obtained contour, and the coordinates of the top left corner, vertical, and horizontal lengths of the rectangle enclosing each nerve root were recorded. The cv2.boundingRect() function was used to draw the rectangle. To compare the FA values ​​between the left and right sides, the averages of the FA values ​​corresponding to the obtained ROI positions were calculated and compared. If the FA value on the left side was 0.1 or more lower than the FA value on the right side, left_flag = 1; if the FA value on the right side was 0.1 or more lower than the FA value on the left side, right_flag = 1. If there was no difference, both flags were set to 0. The coordinate information was saved in a csv file in the form of [left_or_right, x, y, z, fa, flag].

[0077] Next, reconstruction was performed using Three.js from the coordinate and FA value information obtained from the csv file. Three.js is a Javascript library that generates objects from coordinate information. First, a 40 x 40 x 40 zero-filled array was created, and 1 was stored only in the locations listed in the csv file. In addition, 2 was stored in the locations where flag = 1. Next, an object was generated by specifying that locations where the element was 1 were gray and locations where the element was 2 were red.

[0078] FIG. 9 shows an example of the agreement of the disruption points between the tractography and the 3D FA map.

Claims

1. A medical image processing device comprising: a diffusion tensor imaging MRI data acquisition unit that acquires diffusion tensor imaging MRI data of the lumbar spine of a patient suffering from lower back and lower limb pain; a coordinate information estimation unit that estimates coordinate information of a nerve root in a specified axial cross section; and a measurement unit that measures an FA value of the nerve root whose coordinate information has been estimated, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information, the coordinate information estimation unit estimates the coordinate information of the nerve root in the specified axial cross section using a trained model trained using as teacher data an image in which two nerve root locations and a spinal canal location included in the same axial cross section image of a diffusion-weighted MRI image of the lumbar spine are annotated, and the measurement unit measures the FA value of the nerve root whose coordinate information has been estimated based on the scalar information and the coordinate information, based on the coordinate information of the nerve root estimated by the coordinate information estimation unit.

2. The medical image processing device according to claim 1, wherein the same axial cross-sectional image of the MRI diffusion-weighted image of the lumbar region is at least one cross-sectional image of the third lumbar vertebra (L3), the fourth lumbar vertebra (L4) and the fifth lumbar vertebra (L5).

3. The medical image processing device according to claim 2, wherein the same axial cross-sectional images of the MRI diffusion-weighted image of the lumbar region are cross-sectional images of the third lumbar vertebra (L3), the fourth lumbar vertebra (L4) and the fifth lumbar vertebra (L5).

4. The medical image processing device according to claim 1, wherein the same axial cross-sectional images of the MRI diffusion-weighted image of the lumbar spine are multiple cross-sectional images centered on the center of the same nerve root.

5. The medical image processing device according to claim 4, wherein the same axial cross-sectional images of the MRI diffusion-weighted image of the lumbar spine are proximal, central, and distal cross-sectional images of the same nerve root.

6. The medical image processing device according to claim 1, wherein in the training data, the annotation of the nerve root region and the annotation of the spinal canal region are different.

7. The medical image processing device of claim 1, further comprising: a tractography generation unit that associates the coordinate information of the nerve root in each axial section estimated by the coordinate information estimation unit with scalar information, vector information and coordinate information of FA map information created from the diffusion tensor image MRI data, and generates a diffusion tensor tractography showing the course of nerve fibers, wherein the specified axial sections are multiple consecutive axial sections in the lumbar spine of the patient.

8. The medical image processing device according to claim 1, wherein the specified axial sections are a plurality of consecutive axial sections in the lumbar spine of the patient, and further comprising a 3D FA map generating unit that associates coordinate information of the nerve root in each axial section estimated by the coordinate information estimation unit with the FA value measured by the measurement unit, reconstructs the nerve root in three-dimensional space, and generates a 3D FA map.

9. A method for acquiring information on spinal nerves, comprising the steps of acquiring diffusion tensor imaging MRI data of the lumbar spine of a patient having lower back and lower limb pain, estimating coordinate information of a nerve root in a specified axial cross section, and measuring an FA value of the nerve root whose coordinate information has been estimated, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information, the estimating coordinate information estimates the coordinate information of the nerve root in the specified axial cross section using a trained model trained using as training data an image in which two nerve root sites and a spinal canal site included in the same axial cross section image of a diffusion-weighted MRI image of the lumbar spine are annotated, and the measuring step measures the FA value of the nerve root based on the scalar information and the coordinate information, based on the coordinate information of the nerve root estimated by the estimating coordinate information step.

10. The method according to claim 9, wherein the information on the spinal nerve obtained by the method according to claim 9 is an index for estimating the degree of spinal nerve damage from the measured FA value.

11. A method for generating a diffusion tensor tractography showing the course of nerve fibers of a spinal nerve, comprising the steps of: acquiring diffusion tensor imaging MRI data of the lumbar region of a patient with lower back and lower limb pain; estimating coordinate information of nerve roots in each of a plurality of consecutive axial sections in the lumbar region of a specified patient; associating information; and superimposing each axial section to construct a three-dimensional structure to generate a diffusion tensor tractography showing the course of nerve fibers, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information; and the step of estimating the coordinate information estimates the coordinate information of the nerve root in the specified axial section using a trained model trained using as teacher data an image in which annotated images of two nerve root sites and a spinal canal site included in the same axial section image of a diffusion-weighted MRI image of the lumbar region are used; The method, wherein the step of associating information associates the coordinate information of the nerve root in each axial section estimated by the step of estimating the coordinate information of the nerve root with scalar information, vector information and coordinate information of an FA value map created from the diffusion tensor image MRI data.

12. A method for generating a 3D FA map, comprising the steps of acquiring diffusion tensor imaging MRI data of the lumbar region of a patient having low back and lower limb pain, estimating coordinate information of a nerve root in each of a plurality of consecutive axial sections in the lumbar region of a specified patient, measuring an FA value of the nerve root whose coordinate information has been estimated, and reconstructing the nerve root in a three-dimensional space using the coordinate information of the nerve root whose coordinate information has been estimated and the measured FA value corresponding to the coordinate information, wherein the diffusion tensor imaging MRI data includes diffusion-weighted image information and FA map information including scalar information, vector information, and coordinate information, and the step of estimating the coordinate information estimates the coordinate information of the nerve root in the specified axial section using a trained model trained using as teacher data an image in which annotations have been made on two nerve root sites and a spinal canal site included in the same axial section image of a diffusion-weighted MRI image of the lumbar region, The method includes measuring the FA value of the nerve root based on the coordinate information of the nerve root estimated by the coordinate information estimating step, and measuring the FA value of the nerve root whose coordinate information is estimated based on the scalar information and the coordinate information.

13. The method for generating a 3D FA map according to claim 12, further comprising: a step of comparing measured FA values ​​of two nerve roots whose coordinate information included in the same axial cross-sectional image is estimated; and a step of distinguishably displaying nerve root regions where there is a predetermined difference between the measured FA values ​​of two nerve roots whose coordinate information included in the same axial cross-sectional image is estimated and other nerve root regions.

14. A program for causing a computer to execute each step of the method for acquiring information on spinal nerves according to claim 9, the method for generating a diffusion tensor tractography showing the course of nerve fibers of spinal nerves according to claim 11, or the method for generating a 3D FA map according to claim 12.

Images