JP2005230456A - Method and apparatus for supporting diagnosis of cerebropathy - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To objectively and automatically determine an input image specification or the like, and to automate diagnosis by image treatment. <P>SOLUTION: A positioning is performed to correct a spatial deviation to an inputted MRI brain image (Step 1), a gray matter tissue is extracted from the image (Step 2), the first image smoothing is performed (Step 3), an anatomical normalization is performed to the image (Step 4), the second smoothing is performed (Step 5), a correction of the concentration value is performed (Step 6), and the brain image after the correction and an MRI brain image of a normal person are statistically compared (Step 7) to provide a diagnostic result. Then, a check of the inputted images such as resolution and the like (Step 0), a check of the gray matter tissue extract result (Step 21), and a check (Step 41) of the anatomical normalized result are automatically performed about the above brain image to automate the diagnostic support. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a brain disease diagnosis support method and apparatus, and more particularly to a brain suitable for application to diagnosis support by disease by inputting a brain image by MRI (Magnetic Resonance Imaging) or the like and processing the image. The present invention relates to a disease diagnosis support method and apparatus.

  With the arrival of an aging society, the number of patients with dementia is increasing year by year. There are various types of dementia diseases, and in diagnosis, it is necessary to distinguish between them and perform appropriate treatment according to the disease.

  On the other hand, in order to meet such demands, in recent years, information on the state of the brain can be acquired by nuclear medicine examinations such as SPECT (Single Photon Emission Computed Tomography) and PET (Positron Emission Tomography), CT (Computererized Tomography) and MRI. (For example, see Patent Document 1).

  As a result, it has been clarified that the phenomenon that blood flow in a specific part of the brain decreases or the tissue atrophy varies depending on the disease, and a quantitative evaluation method for these is required.

  For example, a decrease in blood flow in a local region of the brain can be tested by comparing with SPECT or PET images.

  As for tissue atrophy, the volume of a specific part can be obtained from an MRI image, and the relative size can be compared to determine the presence or absence of an abnormality.

  For example, for MRI images, based on the VBM (Voxel-Based Morphometry) method, patient images and healthy subject images are standardized by various image processing and then statistically compared to determine local atrophy sites of brain tissue. There is a way to extract. When this method is used, a doctor can make a diagnosis from the distribution of atrophy sites and the degree of atrophy.

  The processing result for diagnosis becomes important information for life, and accordingly, the corresponding reliability is required. In particular, high-accuracy techniques such as the VBM method for MRI images require complex image processing steps, and the system uses specifications such as the resolution of the input image, the dynamic range of the density value, and the image direction. Therefore, it is necessary to thoroughly examine whether or not a good processing result is obtained in each step of the entire processing flow.

JP 2003-107161 A

  However, the confirmation of the specifications of the input image and the determination of the processing result of the image are generally performed by visual observation in the past, and with such a determination method, a subjective factor is included in the determination result, or There is a possibility that an operational error such as missing a processing error may occur. In addition, there is a case where a large amount of patient data is desired to be collectively processed for the construction of a database or the like. However, in such a case, there is a problem that a lot of manpower is required to examine the processing result.

  The present invention has been made in order to solve the above-described conventional problems, and at the same time, at least the input image specification that has been made by visual observation can be objectively confirmed, thereby improving the reliability of the entire image processing. Therefore, an object of the present invention is to provide a diagnosis support method and apparatus for brain diseases that can realize automation of diagnosis by image processing.

  The present invention relates to a brain disease diagnosis support method for supporting diagnosis by inputting a brain image of a subject, processing the image, and presenting a diagnosis result. This problem is solved by automatically checking at least one of resolution, dynamic range, and image direction.

  The present invention is also a brain disease diagnosis support apparatus that supports diagnosis by inputting a brain image of a subject, processing the image, and outputting a diagnosis result, the brain of the subject being input By providing an automatic evaluation means for automatically checking at least one of resolution, dynamic range, and image direction for an image, the above-mentioned problems are similarly solved.

  In the present invention, the image direction is checked in the following assumption, (Assumption 1) in the order of the symmetry of the axis, in the order of lateral direction → sagittal direction → axis (vertical) direction. (Assumption 2) Symmetry is in front of the head (Assumption 3) The tissue exists up to the lower end of the head image, but there is no tissue at the upper end.

  Further, when the brain image of the subject is input as a slice image, the continuity of all slice images may be automatically checked.

  The image processing may include brain tissue extraction processing and anatomical standardization processing, and at least one processing result may be automatically checked.

  At that time, the automatic check may be performed by comparing the brain image as the processing result with a corresponding reference image created in advance.

  In the present invention, the brain image may be an MRI brain image. In this case, after inputting the MRI brain image of the subject, the gray matter tissue is extracted from the MRI brain image to create a gray matter brain image, and the gray matter brain image is subjected to anatomical standardization In addition, statistical comparison with a brain image corresponding to a healthy person prepared in advance may be performed, and after inputting an MRI brain image of a subject, anatomical standardization is performed on the MRI brain image. Then, after the gray matter tissue is extracted from the standardized MRI brain image to create a gray matter brain image, statistical comparison may be performed with a brain image corresponding to a healthy person prepared in advance.

  The present invention also provides a computer-readable program for executing the above-described brain disease diagnosis support method on a computer.

  The present invention also provides a computer-readable program for realizing the above-mentioned brain disease diagnosis support apparatus by a computer.

  The present invention further provides a recording medium storing any one of the above computer-readable programs.

  According to the present invention, it has become possible to objectively confirm at least the input image specification that has been done visually, and thus the reliability of the entire image processing is improved, and thus automating diagnosis support by image processing is realized. Will be able to.

  In addition, the reliability of the entire processing flow can be improved by making it possible to determine whether good processing results have been obtained in each processing step based on quantitative and objective evaluation as necessary. In addition, it is possible to provide diagnosis support for each disease using MRI images or the like for doctors who do not have specialized knowledge of image processing or statistics.

  Further, by automating image processing for diagnosis support, it becomes possible to rapidly process brain image data of a large number of subjects such as patients.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram showing a diagnosis support system (device) for brain disease, which is an embodiment according to the present invention.

  The diagnosis support system of this embodiment includes a user interface 10, an image / statistic processing unit 20, and a database unit 30. The user interface 10 has an image input function 12 for inputting an MRI image as an input image, and a result display function 14 for displaying a result processed by the processing unit 20, and the processing unit 20 is connected to the user interface 10 and the like. It has an input / output evaluation function (automatic evaluation means) 22 for evaluating an input / output image, an image processing function 24 for processing an input MRI image, and a statistical processing function 26 for performing various statistical calculations. . The database unit 30 stores a standard brain image template 32, a gray matter brain image template 34, a healthy person image database 36, and the like that are used for processing to be described later by the processing unit 20.

  In this embodiment, FIG. 2A shows a basic processing flow from outputting a diagnosis result based on the MRI brain image of the subject to supporting the diagnosis.

  As will be described in detail later, first, an MRI brain image of a subject that has been subjected to predetermined pre-processing (in the figure, “brain” is omitted) is input, and spatial deviation is corrected with respect to the brain image. Alignment is performed (step 1). Next, a tissue necessary for diagnosis of dementia disease is extracted from the brain image after alignment (step 2), and a first image smoothing process is performed on the extracted brain image (step 3).

  Next, anatomical standardization is performed on the first smoothed brain image (step 4), and second image smoothing is performed on the standardized brain image (step 5). Next, the density value is corrected for the second smoothed image (step 6), the corrected brain image is statistically compared with the MRI brain image of the healthy person (step 7), and the comparison result is diagnosed. The result is output and used for diagnosis support.

  In this embodiment, the MRI brain image is statistically processed, and the validity of the input / output data is evaluated in the process of steps 1 to 7 described above in which statistical comparison with the image of a healthy person is performed to extract an abnormal region of the brain. For this purpose, an input image check (step 0), a tissue extraction result check (step 21), and an anatomical standardization result check (step 41) are further performed.

  As specific evaluation items, (A) input image check includes (a) resolution, (b) dynamic range, (c) automatic confirmation of image direction, and (d) continuity check of all slice images. In addition, each of (B) brain tissue extraction result check and (C) anatomical standardization result check performed to evaluate each processing result has a comparison with a reference image.

  In the present embodiment, each of the basic processes in steps 1 to 7 and the check processes in steps 0, 21, and 42 can be performed by a program in the processing unit 20 formed of a computer.

  Hereinafter, the basic processing flow will be described in detail. However, for convenience, as shown in FIG. 2B, a specific example is a case where the tissue extracted in Step 2 is a gray matter tissue necessary for diagnosis of Alzheimer-type dementia disease or the like.

  First, when inputting an MRI brain image, preprocessing of an MRI brain image acquired in advance from a subject is performed. Specifically, as shown in FIG. 3, an image of a slice image obtained by cutting out the entire brain and a part of the brain, the slice is captured in a predetermined thickness so as to include the entire brain of the subject, for example, 100 to 200 A T1-weighted MRI image is input. In addition, the resampling of the slice image is performed so that the length of each side of the voxel in each slice image becomes equal in advance. Here, a voxel is a coordinate unit of an image having “thickness” and corresponds to a pixel in a two-dimensional image.

  After inputting the pre-processed MRI brain image, it is checked whether or not the imaging direction and resolution of the slice image meet the conditions set in the system in advance. In the imaging direction, there are sagittal: sagittal section (vertical cut from the side) and coronal: coronal section (vertical cut from the front) in addition to the axial (transverse): transverse section of FIG.

  As described above, when it is confirmed that the MRI brain image is input in conformity with the set condition, the alignment process of step 1 is performed.

  In order to improve the accuracy when comparing the input brain image with a standard brain image template in the processing described later, a transformation as shown in FIG. 4 is performed by linear transformation (affine transformation). This corresponds to correction of the spatial position and angle.

  More specifically, the four types shown in the x, y, and z directions shown in each of the x, y, and z directions that minimize the sum of squares of errors between the input brain image and the standard brain image template 32 read from the database unit 30 are shown in total. Twelve conversion parameters are obtained. Next, the input brain image is affine-transformed using the obtained parameters, and the input brain image is spatially aligned with a standard brain image whose position, size, etc. are preset. be able to.

  After the above alignment is completed, the gray matter extraction process in step 2 is performed.

  The input T1-weighted MRI brain image includes three types of tissues: gray gray matter corresponding to nerve cells, white matter corresponding to brighter nerve fibers, and cerebrospinal fluid close to black. Therefore, in the diagnosis of dementia, attention is paid to the gray matter tissue and processing for extracting the tissue is performed.

  In this extraction processing, in order to separate these tissues, clustering into these three clusters is performed by image processing as shown in FIG.

  For this clustering process, the following two types of models are assumed.

  One is a model of density values.

  This is a model that the distribution of voxel density values differs for each tissue. When the tissues are arranged in order of increasing concentration value (close to white), the order is white matter, gray matter, and cerebrospinal fluid. In this case, it is assumed that the density value histogram after the separation is a normal distribution.

  The second is a model of the existence probability of three organizations with respect to a spatial position.

  In the human brain, the tissue distribution with respect to the spatial position is roughly similar, although there are individual differences. Therefore, when many brain images of many people are collected and examined, it can be seen which tissue has a high probability that the voxel corresponding to the spatial coordinates is. For example, the existence probability at a certain coordinate (x, y) = (5, 10) when the image size in units of voxels of a slice image is X = 256, Y = 256 is (white matter, gray matter, Cerebrospinal fluid) = (20%, 70%, 10%). That is, it can be said that the model expresses the difference in spatial distribution due to individual differences by probability.

  Here, it is assumed that each voxel belongs to one of the organizations and the existence probability of each organization is known in advance according to the spatial position.

  An optimal tissue distribution is estimated so that both of the above two assumptions hold. Specifically, each voxel value is distributed to each organization so as to maximize the following equation.

ここで、ｒ_ijk：クラスタｋのボクセル（i，j）がｆ_ijである尤度関数
ｓ_ijk：ボクセル（i，j）がクラスタｋに属する事前確率
ｆ_ij：ボクセル（i，j）の濃度値
ｂ_ijk：ボクセル（i，j）組織ｋに属する確率
ｈ_k：クラスタｋに属するボクセル数
ｃ_k：各クラスタのボクセルの分散
ｖ_k：各クラスタのボクセルの平均値

Where r _ijk : the likelihood function with voxel (i, j) of cluster k being f _ij
s _ijk : Prior probability that voxel (i, j) belongs to cluster k
f _ij : concentration value of voxel (i, j)
b _ijk : Probability of belonging to voxel (i, j) organization k
h _k : number of voxels belonging to cluster k
c _k : Voxel distribution of each cluster
v _k : average value of voxels in each cluster

  In addition to the above two models, the details of the tissue extraction method, in which a non-uniform noise model peculiar to MRI is also introduced, is described in Ashburner J, Friston KJ: Voxel-Based Morphometry ... The Methods. Neuroimage 11 (6Ptl): pp. 805-821, 2000.

  As described above, the gray matter tissue is extracted three-dimensionally from the brain images of many healthy subjects by using the existence probability calculated for each voxel for each of the gray matter, white matter, and cerebrospinal fluid tissues as a template. The obtained brain image (hereinafter also referred to as gray matter brain image) is obtained.

  As described above, the image smoothing process in step 3 is performed on the brain image from which the gray matter tissue has been extracted.

  Here, for the purpose of improving the S / N ratio of the image and making the smoothness equal to the template image used for the next anatomical standardization, the image is smoothed by a three-dimensional Gaussian kernel. The FWHM (half width) of the filter used for smoothing is about 8 mm.

  As a specific process, a three-dimensional brain image and a three-dimensional convolution of a three-dimensional Gaussian function are performed. This is possible by sequentially performing a one-dimensional convolution in the x, y, and z directions.

  Hereinafter, a method of performing convolution of a one-dimensional Gaussian kernel will be described with reference to a conceptual diagram of smoothing shown in FIG.

  When the one-dimensional discrete Gaussian function is g, it is expressed by the following equation (2). In this equation, j is a position corresponding to a voxel when the center of the Gaussian kernel is set to j = 0. If the one-dimensional image signal (voxel value) is f, the convolution h between f and the one-dimensional Gaussian kernel g is expressed by a product-sum operation as shown in Equation (2). Here, d is the length of the Gaussian kernel that performs the convolution operation, and is determined by the implementation. For example, in the case of an implementation in which d is about 6 times FWHM, when FWHM = 8 mm and voxel size is 2 × 2 × 2 mm, d = 8 × 6/2 = 24 [voxel].

  The above is one-dimensional convolution, but as shown in the conceptual diagram of FIG. 7, the same processing is performed sequentially for each of the x-axis direction, the y-axis direction, and the z-axis direction for a three-dimensional brain image. By executing the above, three-dimensional convolution can be realized.

  After adjusting the smoothness of the gray matter brain image as described above, a process called anatomical standardization in Step 4 is performed. In order to absorb the anatomical differences in the brain images that exist between individuals, this is a global correction for the whole brain size and a local correction for a partial size. .

  Specifically, as conceptually shown in FIG. 8, the square sum of errors from the standard gray matter brain image template 34 read from the database unit 30 using linear transformation and nonlinear transformation. Image processing is performed so that is minimized. The gray matter brain image template 34 used here is an average image obtained from brain images obtained by extracting gray matter tissue from many healthy subjects. In this anatomical standardization process, first, global correction of the position, size, and angle by linear transformation is performed, and then the shape of local irregularities and the like is corrected by nonlinear transformation.

  The linear transformation performed here is an affine transformation similar to the alignment in step 1 described above. In addition, as shown in FIG. 9, the non-linear transformation estimates a deformation field composed of low-frequency components of DCT (discrete cosine transform) in each of the x direction and the y direction. The image is converted.

  The second image smoothing process of step 5 is performed on the gray matter brain image (hereinafter also referred to as a standardized brain image) subjected to anatomical standardization as described above.

  This is a process for improving the S / N ratio of the above-mentioned standardized brain image and making the smoothness of the image equal to the image group of the healthy person used as a standard when performing comparison later. The FWHM is set to about 12 to 15 mm.

  Specifically, it can be realized by performing the same process as the image smoothing process in step 3 except that the value of FWHM is different. By performing the second image smoothing in this way, individual differences that do not completely match in the anatomical standardization processing can be reduced.

  The density value correction in step 6 is performed on the standardized brain image subjected to the second image smoothing as described above. Here, the correction of the voxel density value corresponding to the pixel value in units of voxels is performed.

  This is a process to match the distribution of voxel values in the image group of healthy subjects used as a standard when performing comparison later, and corrects the voxel values of the entire brain. Specifically, as shown in FIG. 10, the density value correction is performed for all voxels by the following conversion formula, as shown in FIG.

x ′ = (MEANnormal / MEANsubject) · x (4)
Where x: density value before correction x ′: density value after correction MEANnormal: average of all voxel density values of healthy subject image group MEANsubject: average of all voxel density values of the image to be processed

  In this way, after correcting the voxel density value of the input image (standardized brain image) to that of the healthy subject image group, a process of removing artifacts is performed. As shown in the image of FIG. 11A, this artifact is an error region generated at the voxel position where the brain tissue does not originally exist by the image smoothing (2) having a large half-value width performed in the step 5. is there.

  Specifically, as shown in FIG. 4B, the standardized brain image template of the gray matter tissue used in the anatomical standardization in the step 4 is used as a mask, and the mask is used as a standardized brain image. These artifacts are removed by multiplying each voxel value of.

  After correcting the voxel density value as described above, the statistical processing in step 7 is performed. Here, the MRI brain image of the subject who has undergone overall standardization through each of the above steps 1 to 6, and the healthy person previously collected and stored as the healthy person image database 36 in the database unit 30 A comparison test with the MRI brain image group is performed. It is desirable that the group of healthy subject images to be used is composed of a subject close to the age of the subject.

  Specifically, as shown in the image of FIG. 12, a comparison test of 1: N (N is the total number of healthy subject images) is performed for such healthy subject image groups and voxel units, and there is a statistically significant difference. Detect voxels that are seen (presumed to be abnormal).

  First, for all voxels, Z scores expressed by the following equations are calculated.

  Thus, the Z score is a value obtained by scaling the difference between the voxel value of the subject image and the average voxel value of the corresponding voxel of the healthy subject image group by the standard deviation, and this is a relative value of the gray matter volume. This indicates the degree of the decline.

Next, an appropriate critical value Z ′ is determined, and the Z score is Z ′ <Z (6)
The voxels that satisfy the above are obtained, and the voxels exhibiting a statistically significant difference are obtained. As the critical value, Z ′ = 2 that can be estimated to be abnormal with a probability of about 95% or more is used. Thereby, the presence of an abnormal part can be recognized and extracted.

  In addition, the healthy person image database 36 used at step 7 is the alignment of the steps 1 to 6 with respect to each of the healthy person image groups separately collected in advance → gray matter tissue extraction → image smoothing (1) → Anatomic standardization-> image smoothing (2)-> density correction is performed in order and created and stored in the same manner.

  Also, in this diagnosis support system, these collected healthy person images are classified by age, for example, every 5 years or every 10 years, and the average value and standard deviation calculated for each group are stored in a storage device. By doing so, the test based on the Z score can be performed.

  In that case, if the subject's age is 76 years, for example, if the subject's age is 76 years old, it is 74 to 78 years old (width) May be collected and compared.

  When using the Z score in this way, it is only necessary to have the above average value and standard deviation data for each voxel, so there is no need to store the image data itself after data creation. There are also advantages.

  In the present embodiment, in order to automate the diagnosis support based on the basic processing flow in steps 1 to 7 described in detail above, the input image check in step 0, the tissue extraction result check in step 21 and the anatomy in step 41 are performed. In addition, an automatic check of scientific standardization results is performed. The details of each check will be described below.

(A) Input image check:
Evaluate the validity of the input image specifications. It is necessary to evaluate the relevance for the following reasons. One is that the MRI brain image has a difference in image quality due to a difference in magnetic field intensity, but the resolution and dynamic range need to satisfy the standard. These specifications greatly affect the reliability of the final processing result.

  Another reason is that there can be various forms of brain image data holding methods. As a method for holding MRI images as three-dimensional information, it is common to hold images on the x and y planes as many as the number of slices (this corresponds to the z direction). , Sagittal, coronal images can be taken. Also, it is possible to invert the left and right and up and down for each, and to change the direction in the z direction. Furthermore, image format conversion may be performed for image processing and storage, but the image slice is not necessarily composed of only one direction of image slice. For example, the image may be a 101-slice image in which one sagittal image near the center is added to 100 consecutive slices of a transverse image in order to easily identify an individual.

  For this reason, the user has to check visually whether the image to be input satisfies the specs (preset) assumed by the system, and thus the user has a heavy workload. .

  Therefore, in the present embodiment, the function of comprehensively evaluating automatically (checking) the items of resolution, dynamic range, automatic confirmation of the image direction, and checking the continuity of all slice images (by the difference between adjacent frames) is described above. This is realized by an input / output evaluation function (software) 22 of the processing unit 20. Each of these items will be described below.

(A) Image resolution:
As described above, the head MRI image is three-dimensional information, and the minimum unit is a three-dimensional unit called a voxel obtained by adding a thickness to a pixel in the two-dimensional image. The number of voxels X, Y, and Z on the x, y, and z axes is taken as the resolution.

  The resolution of the input image is acquired from the header portion (recording area) of the image file. This will be described for the case of using the DICOM format and ANALYZE format that are often used in medical images.

  In the case of the DICOM format, as shown in the image of FIG. 13, data for one slice image is stored in one file, and the three-dimensional data is composed of files of the number of slices. As shown in the figure, the DICOM format is a set of records having a tag and a value, and the tag acquires resolutions X and Y from the value of the record corresponding to the resolution. Since the number of slices corresponds to the number of files, the resolution Z is obtained from the number of files.

In the case of the ANALYZE format, as shown in FIG. 14, there is one header file and one data file. The structure of the header file byte sequence is header key, image dimension, data It consists of three parts of history, and each length and element are fixed. Among these, image resolution X, Y, Z is image Since it is stored in the array of dimension part dim [1] -dim [3], this is acquired.

(B) Dynamic range dB:
The dynamic range is generally a ratio between the minimum value and the maximum value of a signal expressed in dB, and is expressed by the following equation.

dB = 20log (A / B)
Where A is the maximum voxel value
B is the minimum voxel value

  When obtaining the maximum value and the minimum value of the voxel values in the image, a histogram is used to remove abnormal density values as noise components, as shown in FIG. Remove outliers above. Specifically, in the case of the Z score, a value greatly deviating from the average value is removed as noise by a method similar to the method determined to be abnormal by Expression (6).

  Each of the above values (a) and (b) is compared with a value assumed by a preset system, and a warning is issued if the value is different from the condition assumed by the system.

(C) Image orientation evaluation:
The input image is composed of a set of slice images. As described above, the slice image usually has three directions: transverse (cross section), sagittal (sagittal section), and coronal (coronal section). Further, upside down and left side upside down can be considered for each cross section.

  This inter-image direction evaluation (check) is performed by identifying the cross-sectional axis direction and its upside-down inversion for the input image, and confirming that the direction of the input image matches that assumed by the system. This is to prevent the erroneous input as exemplified.

  Here, the following assumptions are made for the head MRI image in order to automatically identify the image direction.

  (Assumption 1) The symmetry of the axis is, in descending order, side direction → sagittal direction → axis (vertical) direction.

  (Assumption 2) The eyeball is symmetric and is in front of the head.

  (Assumption 3) The tissue exists up to the lower end of the head image, but there is no tissue at the upper end.

  The basis of Assumption 1 is as follows. As shown in FIG. 17 showing the symmetry image in the head three-dimensional image, the human head is generally bilaterally symmetric, and therefore the degree of symmetry is large in the lateral direction. The other sagittal and axial directions are basically asymmetric, but as can be seen from the transverse image in the figure, there is some symmetry in the sagittal direction from the top of the head to the cerebrum. Therefore, when the degree of symmetry is represented by the thickness of the arrow, the side direction, the sagittal direction, and the axial direction are arranged in descending order.

  Assumption 2 is a characteristic regarding the position of the eyeball, and FIG. 18 shows images in the respective cross-sectional directions.

  Assumption 3 is a feature on the obtained image because there is a neck at the bottom of the head, and there is tissue to the edge of the image, but there is no continuous tissue outside the scalp above. .

  Based on the above assumptions, automatic image direction checking is realized according to the flowchart shown in FIG.

  First, in order to evaluate the objectivity with respect to each axis of the inputted three-dimensional image, the asymmetry is calculated as follows (step 11).

  (1) The total number of effective voxels K is calculated. Here, the number of effective voxels is the number of voxels corresponding to a portion where a human tissue exists on a three-dimensional image. K is calculated as the number of voxels in the entire image minus the number of voxels outside the head.

  Specifically, for the input image that has been binarized, the pixel value 0 corresponding to the voxel outside the head is counted from the left and right sides of each scanning line of each slice image. FIG. 20 schematically shows this concept by paying attention to a scanning line having a certain cross section. The length in which the voxel value is continuous at 0 is obtained from both directions, and this is subtracted from the number of voxels in the entire scanning line. Thus, the number of effective voxels on the scanning line is obtained. This is performed for the entire image, and the total number K is obtained by integrating the number of effective voxels of each scanning line.

  When calculating the total effective voxel K, for example, a median filter using a 3 × 3 × 3 mask is applied to remove isolated point noise in advance so as not to be affected by noise. It is also effective.

  (2) As schematically shown in FIG. 21, a plane that is perpendicular to the x, y, and z axes, and one of which is closest to K / 2 when the effective voxels are divided by the plane: Ax (x = mx), Ay (y = my), and Az (z = mz) are obtained as temporary symmetry planes.

  The concept of how to obtain the provisional symmetry plane is illustrated in FIG. 22 with respect to the x axis. The number of effective voxels divided by the temporary symmetry plane (x = x ′) is a (x ′) and b (x ′), respectively, and the absolute value | a (x) −b (x) | Is set as the position mx of the provisional symmetry plane.

  (3) For each of the provisional symmetry planes Ax, Ay, Az, asymmetry degrees ASYNx, ASYNy, ASYNz representing the degree of deviation from symmetry are obtained.

  These are defined as the sum of squares of residuals (differences) of voxel values at positions that are plane-symmetrical with respect to the provisional symmetry planes Ax, Ay, and Az, as in the following equations.

  Note that, as shown in FIG. 23, the coordinate symmetric about the plane x = mx of the coordinate (x, y, z) is (2mx−x, y, z), and the same applies to the y axis and the z axis. is there.

  Next, the axis is determined using the asymmetry obtained now (step 12). From Assumption 1, the respective axes are assigned to the lateral direction, the sagittal direction, and the axial (vertical) direction in ascending order of asymmetry ASYNx, ASYNy, ASYNz of each axis, that is, in descending order of symmetry.

  Next, the position of the eyeball is extracted (step 13), and the top and bottom of the image are determined based on the information and the like (step 14). The eye position is extracted by the following procedure.

(1) Filter processing An eyeball extraction filter is applied to a three-dimensional head image to search for eyeball candidates. The filter for eyeball extraction is a spherical filter, and the value is close to 1 in the shell part, and the value is 0 in the center part and the outer part of the shell. When the radius from the center is r, it is expressed by the following equation. It is.

  Here, R1, R2, R3, and Ra are constants and are determined empirically. A two-dimensional representation of this is shown in FIG. The reason why such a filter is used is that, in the MRI image, the central part of the eyeball is black (value is close to 0), and the outer side is white even when compared with the surrounding area (value is large). It is.

  For the convolution calculation performed on the entire image by this filter, the coordinates of N points are selected in descending order of the calculation value, and these are set as candidate points at the center of the eyeball. Here, N is a constant determined empirically.

  Note that the values of R1, R2, R3, and Ra may be determined adaptively from the size of the head in a given input image.

(2) Extraction of Eyeball Corresponding Points It is assumed that the left and right eyeballs are symmetric with respect to the symmetry plane in FIG. In the above-described calculation of symmetry, among Ax, Ay, and Az, the most symmetric surface is A, and the combination of this surface and the most symmetric eyeball candidate point is the true eyeball.

  Specifically, considering that a pair is formed for the candidate N points, N (N−1) / 2 pairs are obtained, but symmetry with respect to the plane A is calculated for all these pairs. Here, the following three elements are considered for symmetry.

  i) Angle: The angle θ formed by the vector connecting the pair of candidate points and the normal vector of the surface A is close to zero.

ii) Distance: The distances w ₁ and w ₂ from the surface A to the two candidate points are equal.

iii) Size: The calculation values (values obtained by the eyeball extraction filter) n ₁ and n _{2 of the} two candidate points are large.

  The symmetry of the pair is defined by the following equation considering the above three elements.

SYNeye = n ₁ n ₂ cos θ / (| w ₁ −w ₂ | +1)

  A pair of candidate points having the largest SYNeye is regarded as a true eyeball.

(3) Determination of front-rear direction In the above-described axis determination process, the sagittal axis is determined, but the front-rear and vertical directions are not yet determined. Therefore, the front-rear direction is determined with the surface closest to the line segment connecting the left and right eyeballs obtained in (2) as the coronal front surface.

(4) Determination of vertical direction Further, the vertical direction is determined. As shown in FIG. 25, when the number of effective voxels in the sagittal cross section is plotted with respect to the axial direction, it can be seen that the number of effective voxels decreases sharply at the upper part compared to the lower part and finally becomes zero. Therefore, since the axis of the input image has already been determined, the number of effective voxels for each slice is plotted in the vertical direction, and the vertical direction is determined under the following conditions.

i) If the end is 0, that is the top
ii) If neither end is 0, the lower sum of the number of valid voxels for the previous three slices is

  By the above, it is possible to estimate the three axes, the sagittal direction, and the direction of the axis (vertical) direction for the input image.

  Next, it is determined whether or not the direction estimated with the above and the axis assumed by the system (set as a condition) match the direction, and the validity of the direction of the input image is evaluated. (Step 15), if they do not match, call attention (issue a warning).

(D) Checking the continuity of all slice images When managing brain images for one person as a set of slice images, as shown in FIG. 26, usually a fixed direction (in this case, transverse) as in image sequence 1 Often represented as a sequence of consecutive slice images. However, as shown in the image sequence 2, after an image in a certain direction continues, an image in another direction (in this case, a sagittal image) may be attached to show the entire image.

  Accordingly, when the image sequence 1 is assumed, if such an image sequence 2 is input, a large error occurs in the subsequent result. Therefore, the degree of continuity between slices is defined by the inter-slice difference as in the following equation to prevent discontinuous images from entering.

但し、ｉはスライス番号であり、ｆ(x,y,z)は座標(x,y,z)における画像のボクセル値を表わす。

Here, i is a slice number, and f (x, y, z) represents a voxel value of an image at coordinates (x, y, z).

  Such a degree of continuity Di is calculated over all slices, and if any of them exceeds a preset threshold value, attention is urged that there are discontinuous slices before and after.

(B) Checking tissue extraction results:
In diagnosis support based on an MRI brain image, in order to extract an abnormal part, as shown in the example in FIG. 5, there are many cases where processing is divided into tissue components constituting the brain. For this, as described in the case where the gray matter tissue is extracted in step 2, an algorithm based on the prior probability of distribution for each tissue component is used, and each extracted tissue component is obtained as an output result of the separation process. Is obtained. Here, as to the validity of the obtained image, as shown in the image of FIG. 27, the correlation coefficient with the reference image is calculated as an evaluation function. In step 21, the correlation between the gray matter brain image extracted in step 2 and the gray matter reference image is calculated. Specifically, the following equation is calculated.

ここで、ｆ(x,y,z)は出力された抽出画像、ｇ(x,y,z)はリファレンス画像の各ボクセル値を、Ｘ、Ｙ、Ｚは画像のｘ軸、ｙ軸、ｚ軸方向の大きさ（位置）を表わし、

はそれぞれ出力画像、リファレンス画像の平均値を、σ_f，σ_gはそれぞれ出力画像、リファレンス画像の標準偏差を表わす。なお、このリファレンス画像には、例えば、多くの被検者の対応する脳画像から得られた平均画像等を用いる。

Here, f (x, y, z) is the output extracted image, g (x, y, z) is the voxel value of the reference image, X, Y, Z are the x-axis, y-axis, z of the image Represents the size (position) in the axial direction,

Are the average values of the output image and the reference image, respectively, and σ _f and σ _g are the standard deviations of the output image and the reference image, respectively. For example, an average image obtained from brain images corresponding to many subjects is used as the reference image.

If the obtained correlation coefficient ρ is equal to or less than the threshold ρ _s , it is considered that there is a problem in the tissue extraction process, and attention is urged. This threshold value ρ _s is a constant value determined empirically.

(C) Check anatomical standardization results:
When analyzing an MRI brain image, as described in step 4, the anatomical standardization with other average brain images is performed so that the spatial position is matched. There are many. For this purpose, as described above, linear transformation, nonlinear transformation, an algorithm combining them, or the like is used, and a corresponding standardized image is obtained as an output result. Here, as to the validity of the obtained image, as shown in FIG. 28, the correlation coefficient with the reference image is calculated as an evaluation function, as in the above-described check of the tissue extraction result. In step 41, the correlation between the gray matter brain image standardized in step 4 and the reference image is calculated. Specifically, the following formula using the same symbol is calculated for convenience.

ここで、ｆ(x,y,z)は出力された標準化画像、ｇ(x,y,z)はリファレンス画像の各ボクセル値を、Ｘ、Ｙ、Ｚは画像のｘ軸、ｙ軸、ｚ軸方向の大きさを表わし、

はそれぞれ出力画像、リファレンス画像の平均値を、σ_f、σ_gはそれぞれ出力画像、リファレンス画像の標準偏差を表わす。リファレンス画像には、例えば、多くの被検者の対応する脳画像から得られた平均画像等を用いる。

Here, f (x, y, z) is the output standardized image, g (x, y, z) is the voxel value of the reference image, X, Y, Z are the x-axis, y-axis, z of the image Represents the size in the axial direction,

Are the average values of the output image and the reference image, respectively, and σ _f and σ _g are the standard deviations of the output image and the reference image, respectively. As the reference image, for example, an average image obtained from brain images corresponding to many subjects is used.

If the calculated correlation coefficient ρ is equal to or less than the threshold value ρ _N , it is considered that there is a problem in the tissue extraction process, and attention is urged. This ρ _N is a constant value determined empirically.

  As described in detail above, according to the present embodiment, the MRI brain image of the subject is subjected to various processes such as extraction of gray matter tissue, anatomical standardization, image smoothing, and the like, and then is normalized. When comparing the MRI brain image of a group of people with a Z score and outputting diagnosis results for diagnosis support, the input image specifications, gray matter tissue extraction results, and anatomical standardization results can be automatically checked. As a result, a series of processes up to diagnosis support can be automated.

  Although the present invention has been specifically described above, the present invention is not limited to that shown in the above embodiment, and various modifications can be made without departing from the scope of the invention.

  For example, the first image smoothing in step 3 in the basic processing flow has the meaning of reducing the noise of the input image and making the smoothness equal to the gray matter template image used in anatomical standardization. If they are close to each other, this smoothing process can be omitted. In this case, there is an advantage that there is no loss of information due to smoothing.

  The order of the gray matter tissue extraction and anatomical standardization processing may be reversed. In this case, the standard brain image template used in anatomical standardization is not related to gray matter, but is related to the whole brain before extraction. The standard whole brain image template is created from an average image of many healthy person images or an average image of many healthy person images and many patient images.

  In the basic processing flow, when there is a large difference in size between the input brain image and the gray matter template image, for example, gray matter extraction may not be performed correctly. However, when the anatomical standardization is performed first as described above, there is an advantage that the spatial correspondence with the gray matter template is improved.

  In the above embodiment, the test method using the Z score as the evaluation value of the statistical comparison is shown. However, the present invention is not limited to this, and the above-described other general test method such as the two-sample t test is used. May be.

  As described above, according to the present invention, it has become possible to objectively confirm at least the input image specifications, which has been done by visual observation in the past, thereby improving the reliability of the entire image processing, and thus diagnosis support by image processing. Can be automated.

  In addition, if necessary, it is possible to clarify the checkpoint by the user by making it possible to quantitatively and objectively evaluate whether or not a good processing result has been obtained in each processing step. . Therefore, it becomes possible to improve the reliability of the entire processing flow, and it becomes possible to provide diagnosis support for each disease by MRI images or the like for doctors who do not have specialized knowledge of image processing and statistics. .

  Further, by automating the processing flow even in a complicated algorithm, it is possible to eliminate a human error during the intermediate processing and obtain a highly reliable processing result.

  In addition, when an MRI brain image is used as an input image, it is not necessary to administer a contrast medium containing a radioisotope, such as SPECT, for MRI imaging, and therefore, there is no exposure to the physical burden. There is also an advantage of less. This also makes it easy to collect patient data and healthy person data for comparison. In addition, an image obtained by MRI has an advantage that resolution is higher than that of a SPECT image. Furthermore, it can be mechanically processed in all the processing steps, and quantitative and objective results can be obtained.

The block diagram which shows the outline | summary of the diagnostic assistance system of one Embodiment which concerns on this invention The flowchart which shows the processing flow of the diagnostic assistance by this embodiment Flow chart showing the processing flow of diagnosis support when extracting gray matter tissue Conceptual diagram showing brain slice images and voxel characteristics Conceptual diagram schematically showing the characteristics of affine transformation used for image alignment Conceptual diagram schematically showing features of processing to extract gray matter from input brain images Conceptual diagram schematically showing the characteristics of one-dimensional image smoothing processing Conceptual diagram schematically showing the characteristics of three-dimensional smoothing processing Conceptual diagram schematically showing features of anatomical standardization Conceptual diagram showing the characteristics of nonlinear transformation Conceptual diagram showing the characteristics of voxel density correction processing Conceptual diagram showing artifacts caused by image smoothing and how to remove them Conceptual diagram showing the characteristics of the comparison test for each corresponding voxel Conceptual diagram showing the characteristics of the data structure of DICOM format Conceptual diagram showing the characteristics of the data structure of the ANALYZE format Diagram for explaining the concept of noise removal Conceptual diagram showing an example of an incorrect input image Conceptual diagram showing the characteristics of symmetry in the head 3D image Conceptual diagram showing the characteristics of the position of the eyeball in brain image sections in different directions The flowchart which shows the process sequence which identifies the image direction of an input image Conceptual diagram explaining how to count the number of effective voxels Conceptual diagram showing the features of the provisional symmetry plane for each axis on the brain image Conceptual diagram showing a search example for determining a symmetry plane on a brain image Diagram for explaining the relationship of plane-symmetric coordinates Conceptual diagram showing the characteristics of the filter that extracts the eyeball position from the brain image in two dimensions Conceptual diagram showing an example of the change in the axial direction of the number of effective voxels Conceptual diagram showing the characteristics of the continuity check method for all slice images Conceptual diagram showing the characteristics of the evaluation method of the tissue extraction process result Conceptual diagram showing the characteristics of the anatomical standardization processing result evaluation method

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... User interface 20 ... Processing part 22 ... Input / output evaluation function 30 ... Database part 32 ... Standard brain image template 34 ... Gray matter brain image template 36 ... Healthy person image database

Claims (19)

A brain disease diagnosis support method for supporting diagnosis by inputting a brain image of a subject, performing image processing and presenting a diagnosis result,
A diagnosis support method for a brain disease, comprising: automatically checking at least one of resolution, dynamic range, and image direction for an input brain image of a subject. Checking the image direction is as follows. (Assumption 1) The symmetry of the axis is in the descending order: side direction → sagittal direction → axis (vertical) direction.
(Assumption 2) The eyeball is symmetric and is in front of the head.
(Assumption 3) There is a tissue up to the lower end of the head image, but there is no tissue at the upper end.
The diagnosis support method for brain disease according to claim 1, wherein the diagnosis support method is performed under the following conditions.   The brain disease diagnosis support method according to claim 1 or 2, wherein when the brain image of the subject is input as a slice image, the continuity of all slice images is automatically checked.   4. The brain disease diagnosis support method according to claim 1, wherein the image processing includes brain tissue extraction processing and anatomical standardization processing, and at least one processing result is automatically checked.   5. The brain disease diagnosis support method according to claim 4, wherein the automatic check is performed by comparing the brain image as the processing result with a corresponding reference image created in advance.   6. The brain disease diagnosis support method according to claim 1, wherein the brain image is an MRI brain image. After inputting the MRI brain image of the subject,
Extracting gray matter tissue from the MRI brain image to create a gray matter brain image;
7. The brain disease diagnosis support method according to claim 6, wherein the gray matter brain image is subjected to anatomical standardization and then statistically compared with a brain image corresponding to a healthy person prepared in advance. . After inputting the MRI brain image of the subject,
Anatomic standardization is applied to the MRI brain image,
The gray matter tissue is extracted from the standardized MRI brain image to create a gray matter brain image, and then statistically compared with a brain image corresponding to a healthy person prepared in advance. The method for assisting diagnosis of brain disease according to claim 1. A brain disease diagnosis support apparatus that supports diagnosis by inputting a brain image of a subject, processing the image, and outputting a diagnosis result,
A brain disease diagnosis support apparatus, comprising: an automatic evaluation unit that automatically checks at least one of resolution, dynamic range, and image direction for an input brain image of a subject. The automatic evaluation means checks the image direction as follows: Assumption (Assumption 1) The order of symmetry of the axis is as follows: side direction → sagittal direction → axis (vertical) direction.
(Assumption 2) The eyeball is symmetric and is in front of the head.
(Assumption 3) There is a tissue up to the lower end of the head image, but there is no tissue at the upper end.
The brain disease diagnosis support apparatus according to claim 9, wherein the brain disease diagnosis support apparatus has a function of   The automatic evaluation unit has a function of automatically checking the continuity of all slice images when the brain image of the subject is input as a slice image. Diagnosis support device for brain diseases.   The image processing includes brain tissue extraction processing and anatomical standardization processing, and the automatic evaluation means has a function of automatically checking at least one processing result. Or a diagnosis support apparatus for brain disease according to 11;   13. The automatic evaluation unit has a function of comparing the brain image as the processing result with a corresponding reference image created in advance and performing the automatic check. Diagnosis support device for brain diseases.   14. The brain disease diagnosis support apparatus according to claim 9, wherein the brain image is an MRI brain image. After inputting the MRI brain image of the subject,
Extracting gray matter tissue from the MRI brain image to create a gray matter brain image;
15. The brain disease diagnosis support apparatus according to claim 14, wherein the gray matter brain image is subjected to anatomical standardization and then statistically compared with a brain image corresponding to a healthy person prepared in advance. . After inputting the MRI brain image of the subject,
Anatomic standardization is applied to the MRI brain image,
15. A gray matter tissue is extracted from the standardized MRI brain image to create a gray matter brain image, and then statistically compared with a brain image corresponding to a healthy person prepared in advance. The diagnosis support apparatus for brain diseases according to 1.   A computer-readable program for executing the brain disease diagnosis support method according to claim 1 on a computer.   A computer-readable program for realizing the brain disease diagnosis support apparatus according to claim 9 by a computer.   A recording medium in which the program according to claim 17 or 18 is stored.

Images