JP2018038515A - Diagnosis support device, diagnosis support program, and diagnosis support method for alzheimer's disease - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a diagnosis support device for supporting accurate diagnosis of AD of a subject whose facility or data base is different from that of a control group that serves as a criterion for prediction.SOLUTION: A diagnosis support device 1 includes: subject strength acquisition means 11 for acquiring functional connection strength of a default mode network in the inner temporal lobe region of a subject; standardization means 12 for calculating a z value indicating a position in the control strength distribution of the subject strength for each voxel value of the inner temporal lobe region based on the average and standard deviation of functional connection strength of a default mode network in the inner temporal lobe region of a control group of non-Alzheimer's disease; and index calculation means 13 for calculating an index for predicting whether or not the subject has Alzheimer's disease based on the z value.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a diagnosis support apparatus, a diagnosis support program, and a diagnosis support method that support diagnosis of Alzheimer's disease.

  According to estimates by the Ministry of Health, Labor and Welfare, there are 4.6 million dementia patients in Japan, and 4 million reserves, and countermeasures for dementia are an urgent issue in Japan, which is facing an aging society. Early detection is important because dementia is caused by irreversible impairment of brain function. About half of dementia is Alzheimer's Disease (hereinafter referred to as “AD”). Conventional MRI diagnosis in AD has been performed by evaluating the degree of hippocampal atrophy using T1-weighted images. For example, Patent Document 1 below discloses an image diagnosis support apparatus that can perform analysis for AD diagnosis even when an input image includes an abnormal site such as a cerebral infarction or a brain tumor. Yes.

  However, even if AD is discovered at the stage of hippocampal atrophy, the cognitive function has already been irreversibly lowered, so a technique for catching signs of AD at an earlier stage is required. As one candidate, the resting fMRI technique has recently attracted attention. It has been reported that the functional connectivity of the hippocampus as assessed by resting fMRI images has already been reduced in mild cognitive impairment, the pre-stage of AD, and is considered useful for early detection of AD. (Non-Patent Documents 1 and 2). Along with this need, various methods have been developed as an analysis method for comparing image data such as fMRI between a patient and a control group. However, these methods have not been unified, and a standard method has not yet been established (Non-patent Document 1). As an example of brain image analysis such as resting fMRI, machine learning represented by multi-voxel pattern analysis such as support vector machine can be used to compare resting fMRI images with subjects and control groups. In order to compare AD with mild cognitive impairment, a study has been conducted to distinguish it from a healthy state. It has been reported that when the prediction model derived by these machine learning is verified by leave-one-out cross-validation (LOOCV), a correct answer rate of 60% to over 90% is obtained. (For example, non-patent documents 3 to 5).

JP 2013-165765 JP

Onoda Keiichi, “Clinical application of resting fMRI: functional association and cognitive function”, Journal of Senile Dementia, 19 (8), 2014, p.109-111 Greicius MD, other names, “Default-mode network activity distinguishes Alzheimer's disease from healthy aging: evidence from functional MRI”, Proc. Natl. Acad. Sci. USA. 101 (13), March 30, 2004, p.4637 -42 Khazaee A, 2 others, “Identifying patients with Alzheimer's disease using resting-state fMRI and graph theory”, Clinical Neurophysiology, 126 (11), November 2015, p.2132- 2141 Dyrba M and three others, “Multimodal analysis of functional and structural disconnection in Alzheimer ’s disease using multiple kernel SVM”, Hum. Brain Mapp, 36 (6), June 2015, p.2118-2131 Li Y, 4 others, “Exploring the functional brain network of Alzheimer ’s disease: based on the computational experiment”, PLoS One, 8, 2013, e73186

  In a prediction model constructed by using a support vector machine as an ADNI database, which is an image database of North American AD, as a learning data, a correct answer rate of 90% or more is obtained by LOOCV (Non-patent Document 3). On the other hand, the present inventors tried to discriminate between an AD patient and a control group by using a prediction model constructed by a support vector machine using an ADNI database for an AD patient who actually visited Shimane University Hospital. However, the verification result at LOOCV was a chance level with a correct answer rate of 50%. The conventional multi-voxel pattern analysis disclosed in Non-Patent Documents 3 to 5 targets one facility or one database, and does not consider whether AD can be predicted for subjects with different facilities or databases. Regarding the AD patients who visited Shimane University Medical Hospital, the reason for the low correct answer rate in the prediction model obtained by the conventional multi-voxel pattern analysis is that the fMRI image at rest is due to differences in MRI scanner manufacturers and the minute measurement variables. It is conceivable that even an MRI scanner of the same model is affected by variations in characteristics among individuals. Therefore, it is considered that the conventional technique cannot predict AD of subjects whose facilities or databases are different from those of the control group.

  Considering the above factors, reference data is required for each facility or database, but since machine learning requires a large amount of data, it is necessary to acquire control data suitable for machine learning for each facility or database. Is also difficult.

  Therefore, an object of the present invention is to provide a diagnosis support apparatus, a diagnosis support program, and a diagnosis support method that support accurate diagnosis of AD of a subject whose facility or database is different from a control group serving as a reference for prediction.

In order to solve the above problems, the present invention has the following aspects.
[Claim 1]
A test strength acquisition means for acquiring a functional coupling strength (hereinafter referred to as “test strength”) of a default mode network of the medial temporal lobe region of the subject;
Based on the mean and standard deviation of the functional binding strength (hereinafter referred to as “control strength”) of the default mode network of the medial temporal lobe region of the control group that is not Alzheimer's disease, the test strength in the distribution of the control strength Standardization means for calculating a z-value indicating a position for each voxel of the medial temporal lobe region;
Index calculating means for calculating an index for predicting whether or not the subject has Alzheimer's disease based on the z value;
A diagnostic support apparatus for Alzheimer's disease.
[Section 2]
The index calculating means includes
The voxels are arranged in ascending order of the z value, and the upper average of the z values from the highest voxel with the smallest z value to the upper x (0 <x ≦ 100)% voxels is calculated as the index. Item 2. The diagnosis support apparatus according to Item 1.
[Section 3]
Item 3. The diagnosis support device according to Item 2, wherein x is a predetermined value within a range of 10 to 25.
[Claim 4]
The test strength acquisition means acquires the test strength from a test image of fMRI at rest of the subject's brain,
The standardization means obtains the mean and standard deviation from a control image of resting brain fMRI of the control group,
The diagnosis support apparatus according to any one of Items 1 to 3, wherein the test image and the control image are preprocessed so that the shape, size, and number of voxels of the medial temporal lobe region are the same.
[Section 5]
Item 5. The diagnosis support apparatus according to any one of Items 1 to 4, wherein the standardization unit calculates the z value z _i in the i-th voxel of the medial temporal lobe region by the following equation.
z _i = (y _i −m _i ) / σ _i
y _i : Functional coupling strength of the default mode network of the subject's medial temporal lobe region in the i th voxel m _i : Function of the default mode network in the i th voxel of the medial temporal lobe region of the control group Mean σ _i : standard deviation of functional bond strength of the default mode network in the i-th voxel in the medial temporal lobe region of the control group [Claim 6]
The diagnosis support apparatus according to any one of Items 1 to 5, further comprising a prediction unit that predicts whether or not the subject has Alzheimer's disease based on the index.
[Claim 7]
Computer
A test strength acquisition means for acquiring a functional coupling strength (hereinafter referred to as “test strength”) of a default mode network of the medial temporal lobe region of the subject;
Based on the mean and standard deviation of the functional binding strength (hereinafter referred to as “control strength”) of the default mode network of the medial temporal lobe region of the control group that is not Alzheimer's disease, the test strength in the distribution of the control strength A standardization means for calculating a z-value indicating a position for each voxel of the medial temporal lobe region; and
Index calculating means for calculating an index for predicting whether or not the subject has Alzheimer's disease based on the z value;
A diagnostic support program for Alzheimer's disease.
[Section 8]
A test strength acquisition step of acquiring a functional coupling strength (hereinafter referred to as “test strength”) of a default mode network of the subject's medial temporal lobe region;
Based on the mean and standard deviation of the functional binding strength (hereinafter referred to as “control strength”) of the default mode network of the medial temporal lobe region of the control group that is not Alzheimer's disease, the test strength in the distribution of the control strength A standardization step of calculating a z-value indicating a position for each voxel of the medial temporal lobe region;
An index calculating step for calculating an index for predicting whether or not the subject has Alzheimer's disease based on the z value;
A diagnostic support method for Alzheimer's disease, comprising:

  According to the present invention, it is possible to assist in accurately diagnosing AD of a subject whose facility or database is different from a control group serving as a reference for prediction.

1 is a block diagram showing a hardware configuration of an Alzheimer's disease diagnosis support system according to an embodiment of the present invention. It is a functional block diagram of the diagnosis assistance apparatus which concerns on embodiment of this invention. It is a flowchart which shows the procedure of the diagnosis assistance method which concerns on embodiment of this invention. It is an image of a default mode network including the medial temporal region of an Alzheimer patient (AD) and a control (HC) (cited from Non-Patent Document 2). It is a graph which shows an example of the frequency distribution of the functional coupling | bonding after normalization in the voxel of a certain test subject's medial temporal area. It is a graph which shows an example of the high-order average with respect to the order | rank of a voxel. (A) is a graph showing an average of voxel frequencies with respect to z values of ADNI-AD and an average of voxel frequencies with respect to z values of ADNI-HC, and (b) is a graph of voxel frequencies with respect to z values of Shimane-AD. It is a graph which shows the average and the average of the voxel frequency with respect to z value of Shimane-HC. (A) is a box-and-whisker plot showing the distribution of the upper average of ADNI-AD and the distribution of the upper average of ADNI-HC, and (b) is the distribution of the upper average of Shimane-AD and Shimane-HC. It is a box-and-whisker diagram which shows distribution of an upper average. (A) is the ROC curve of AD prediction for ADNI-AD and ADNI-HC in the embodiment of the present invention, and (b) is the AD prediction for Shimane-AD and Shimane-HC in the embodiment of the present invention. It is a ROC curve. (A) is a box-and-whisker diagram showing the distribution of the above-mentioned index values of Shimane-AD and Shimane-HC in the comparative example, and (b) is an AD prediction for Shimane-AD and Shimane-HC in the comparative example. It is a ROC curve.

  Embodiments of the present invention will be described below with reference to the accompanying drawings. In addition, this invention is not limited to the following embodiment.

〔Definition of terms〕
First, terms used in the present specification, claims, and drawings will be described.
In the present invention, “Alzheimer's disease” or “AD” is McKhann GM, Knopman DS, Chertkow H, et al .: The diagnosis of dementia due to Alzheimer's disease: Recommendations from the National Institute on Aging-Alzheimer's Association workgroups on diagnostic guidelines for Alzheimer's disease. Alzheimers Dement 2011; 7: 263-269. Dementia caused by AD, mild cognitive impairment (MCI due to AD), and preclinical stages of AD AD). Preferred are dementia due to AD and mild cognitive impairment in the background of AD.

  The “subject” of the present invention may be a person who is suspected of AD or mild cognitive impairment, or may be a person who is not suspected. As another aspect, the present invention relates to MMSE (Mini-Mental State Examination) and DASC-21 (Dementia Assessment Sheet for Community-based Integrated Care System-21 items). Sheet), MoCA (Montreal Cognitive Assessment), Mini-Cog, HDS-R (Hasegawa's Dementia Scale-Revised: revised Hasegawa dementia scale), etc., in subjects with mild cognitive impairment or suspected dementia It may be a target. For example, MMSE is 23 points or less, DASC-21 is 31 points or more, Mini-Cog is 2 points or less, or HDS-R 20 points or less, and dementia is suspected. In addition, for example, MMSE is 27 points or less (except 23 points or less), or MoCA is 25 points or less, and mild cognitive impairment is suspected.

  The `` control group '' of the present invention is a group of persons without functional and organic abnormalities in the brain, and is clinically functional in the brain even if it has healthy individuals and other wounds, Including at least two persons who are judged not to have an abnormality. Preferably 20 or more, more preferably 30 or more. Moreover, it is preferable that the same person is not included in the control group.

  In the present invention, the “inner temporal lobe region” mainly refers to the left and right regions including the hippocampus, the hippocampus, and the amygdala, preferably the left and right regions composed of the hippocampus, the hippocampus, and the amygdala. .

〔overall structure〕
FIG. 1 is a block diagram showing a hardware configuration of a diagnosis support system 100 for Alzheimer's disease (hereinafter referred to as “AD”) according to the present embodiment. The diagnosis support system 100 mainly includes an AD diagnosis support apparatus (hereinafter referred to as “diagnosis support apparatus”) 1, an fMRI apparatus 2, and a display apparatus 3. The diagnosis support apparatus 1 includes the fMRI apparatus 2 and a display. It is connected to the device 3.

  The diagnosis support apparatus 1 is an apparatus that supports diagnosis of whether or not a subject has Alzheimer's disease, and is configured by, for example, a general-purpose personal computer. The diagnosis support apparatus 1 includes a CPU 101, a memory 102, and a hard disk drive (HDD) 103 as hardware configurations. The CPU 101 executes various arithmetic processes by reading the program stored in the HDD 103 into the memory 102 and executing it. The memory 102 temporarily stores programs and data read by the CPU 101 from the HDD 103. The HDD 103 is a nonvolatile large-capacity auxiliary storage device. As an auxiliary storage device, an SSD (Solid State Drive) or a flash memory may be used instead of the HDD 103.

  The fMRI apparatus 2 is an apparatus that obtains an fMRI image in which an active site of the brain is visualized using functional nuclear magnetic resonance imaging (functional MRI). As the fMRI apparatus 2, a known fMRI apparatus equipped with hardware and software necessary for fMRI measurement in a normal MRI apparatus can be used. In the present embodiment, the fMRI apparatus 2 generates an fMRI image (hereinafter referred to as “test image”) when the subject's brain is at rest, and transmits the test image to the diagnosis support apparatus 1.

  The display device 3 is composed of a liquid crystal display, for example, and displays the calculation processing result by the diagnosis support device 1. The diagnosis support apparatus 1 may be connected to an input device such as a keyboard (not shown) or a printing device such as a printer.

[Configuration of diagnosis support device]
FIG. 2 is a functional block diagram of the diagnosis support apparatus 1. As shown in FIG. 2, the diagnosis support apparatus 1 includes a test intensity acquisition unit 11, a standardization unit 12, an index calculation unit 13, a prediction unit 14, and an output unit 15. Each of these means is realized by the CPU 101 reading the diagnostic support program 16 stored in the HDD 103 into the memory 102 and executing it. That is, the diagnosis support program 16 is a program that causes the diagnosis support apparatus 1, which is a computer, to function as each of the above means, particularly the test intensity acquisition means 11, the standardization means 12, and the index calculation means 13. The diagnosis support program 16 may be supplied to the diagnosis support apparatus 1 via a computer-readable recording medium (CD-ROM or the like) that records the program, or the diagnosis support apparatus 1 via a communication network such as the Internet. May be supplied. Note that at least a part of the test intensity acquisition unit 11, the standardization unit 12, the index calculation unit 13, the prediction unit 14, and the output unit 15 may be realized by hardware such as an integrated circuit.

  The test strength acquisition means 11 has a function of acquiring the functional coupling strength (hereinafter referred to as “test strength”) of the default mode network of the subject's medial temporal lobe region. In order to realize this function, in this embodiment, the test intensity acquisition unit 11 includes a reception unit 11a, a preprocessing unit 11b, and an independent component analysis unit 11c.

  The standardization means 12 is based on the average and standard deviation of the functional binding strength (hereinafter referred to as “control strength”) of the default mode network in the medial temporal lobe region of the non-AD control group, and the control strength of the control strength It has a function of calculating a z value indicating a position in the intensity distribution for each voxel in the medial temporal lobe region. In order to realize this function, in the present embodiment, the standardization means 12 includes a preprocessing means 12a, an independent component analysis means 12b, an average / standard deviation calculation means 12c, and a z value calculation means 12d.

  The index calculation unit 13 has a function of calculating an index for predicting whether or not the subject is AD based on the z value calculated by the standardization unit 12. The prediction unit 14 has a function of predicting whether or not the subject is AD based on the index. The output unit 15 has a function of outputting the prediction result 19 by the prediction unit 14 to the display device 3, and thereby the prediction result as to whether or not the subject is AD is displayed on the display device 3.

  Specific processing contents by the above functional blocks will be described in detail in the description of the diagnosis support method below.

  The HDD 103 stores a contrast image 17 and a threshold 18 in addition to the diagnosis support program 16. The control image 17 is a resting fMRI image of a control group that is not AD. In the control group, the subject and the facility may be the same or different. That is, the control image 17 may be a resting fMRI image of the control group generated by the fMRI apparatus 2 or a resting fMRI image generated by an fMRI apparatus different from the fMRI apparatus 2. The threshold value 18 is numerical data that is referred to for the prediction means 14 to perform a prediction process. Specific numerical values will be described later.

[Calculation processing of diagnosis support apparatus, diagnosis support method]
FIG. 3 is a flowchart showing the procedure of the AD diagnosis support method according to the present embodiment. Below, the form which implements the said diagnostic assistance method by the diagnostic assistance apparatus 1 shown in FIG. 2 is demonstrated.

  As shown in FIG. 3, the diagnosis support method according to the present embodiment includes a test intensity acquisition step S1, a standardization step S2, an index calculation step S3, a prediction step S4, and an output step S7. Hereinafter, each step will be described.

(Test strength acquisition)
In the test strength acquisition step S1, the test strength (DMN binding strength of the subject's medial temporal lobe region) is acquired. In the present embodiment, first, the reception unit 11 a of the test intensity acquisition unit 11 illustrated in FIG. 2 receives the test image 10 transmitted from the fMRI apparatus 2 and stores the test image 10 in the memory 102.

  Subsequently, the pre-processing unit 11 b performs pre-processing on the test image 10 so that the test image 10 is suitable for the arithmetic processing performed by the standardization unit 12 and the index calculation unit 13. Specifically, the shape, size and number of voxels of the medial temporal lobe region of the test image 10 are preprocessed so as to be the same as the control image 17 preprocessed by the preprocessing unit 12a of the standardization unit 12 described later. The In this embodiment, the shape of the brain of the test image 10 is converted into the shape of a standard brain by a known method. This pre-processing includes head fine movement correction, spatial standardization, spatial smoothing and the like. Although the shape, size, and number of voxels of the medial temporal lobe region to be analyzed are not particularly limited, in the present embodiment, they are 500 to 3000 voxels that continuously span the hippocampus, the hippocampal gyrus, and the amygdala. In this case, one voxel is about 2 mm × 2 m × 2 mm or 3 mm × 3 mm × 3 mm. The ratio of the hippocampus, parahippocampal gyrus, and amygdala on both the left and right sides included in the 1000 to 1500 voxels is 42% to 45% for the hippocampus and 45% to 48% for the hippocampus when the selected number of voxels is 100%. %, And the amygdala is about 7% to 13%. In addition, the preprocessing unit 11b assigns a number to each voxel in the medial temporal lobe region.

  Subsequently, the independent component analysis unit 11 c calculates the functional coupling strength (hereinafter referred to as “DMN coupling strength”) of the default mode network from the test image 10 by independent component analysis. In the present embodiment, the independent component analysis unit 11c calculates the DMN bond strength for each voxel in the subject's medial temporal lobe region. FIG. 4 is an image of DMN in the brain of Alzheimer's patient (AD) and non-AD control (HC), and the arrows in the figure indicate DMN binding strength in the medial temporal region. The independent component analysis means 11c may calculate the DMN bond strength of the entire brain including the medial temporal lobe region, or may perform mask processing on the test image 10 to obtain the DMN bond strength of only the medial temporal lobe region. May be calculated.

  As described above, in the test intensity acquisition step S1, the test intensity acquisition unit 11 acquires the test intensity. The test intensity is output to the z value calculation unit 12d of the standardization unit 12.

  As a modified example, in the test intensity acquisition step S1, the processing by the reception unit 11a, the preprocessing unit 11b, and the independent component analysis unit 11c is performed by a device other than the diagnosis support device 1, and the test is performed from the other device. The intensity data may be transmitted to the diagnosis support apparatus 1. In this case, the test intensity acquisition unit 11 performs a process of receiving the test intensity data from the other device and storing it in the memory 102 and / or the HDD 103. That is, in this specification, “acquisition” is a concept including both an aspect of generating data and the like and an aspect of receiving data and the like from another device. Alternatively, the processing by the reception unit 11a and the preprocessing unit 11b is performed by another device, the test image 10 that has been preprocessed is transmitted to the diagnosis support device 1, and the processing by the independent component analysis unit 11c is performed by the diagnosis support device 1. You may implement in.

(Standardization)
Subsequently, in the standardization step S2, a z value indicating the position of the test intensity in the distribution of the control intensity is calculated for each voxel in the medial temporal lobe region based on the average and standard deviation of the control intensity. In the present embodiment, first, the preprocessing unit 12 a of the standardization unit 12 shown in FIG. 2 reads a plurality of contrast images 17 stored in the HDD 103 into the memory 102 and performs preprocessing on each contrast image 17. Specifically, the shape, size, and number of voxels of the medial temporal lobe region of the control image 17 are preprocessed so as to be the same as the test image 10 preprocessed by the preprocessing unit 11b. In addition, the pre-processing means 12a assigns a number to each voxel of the control image 17 so that the coordinates of the same number of voxels in each control image 17 and the test image 10 are the same.

  Subsequently, the independent component analysis means 12b calculates the DMN bond strength of the medial temporal lobe region from the control image 17 by independent component analysis. In the present embodiment, the independent component analysis unit 12 b calculates the DMN bond strength for each voxel in the medial temporal lobe region of each control image 17. The independent component analyzing means 12b may calculate the DMN bond strength of the entire brain including the medial temporal lobe region, or the mask image is applied to the control image 17 so that the DMN bond strength of only the medial temporal lobe region is calculated. May be calculated. As described above, the independent component analyzing unit 12b calculates the DMN binding strength (hereinafter referred to as “control strength”) of the medial temporal lobe region of each control group.

Subsequently, the average / standard deviation calculating unit 12c calculates the average and standard deviation of each control intensity calculated by the independent component analyzing unit 12b for each voxel in the medial temporal lobe region. For example, if the number of subjects included in the control group (ie, the number of control images 17) is n, and the control intensity of the k-th control of the i-th (i = 1 to 1000) voxel is y _ik , The average _mi and standard deviation σ _i of the i-th voxel are calculated by the following formula.

Subsequently, z value calculation means 12d indicates the position of the test intensity input from the test intensity acquisition means 11 in the control intensity distribution based on the average and standard deviation calculated by the average / standard deviation calculation means 12c. A value is calculated for each voxel in the medial temporal lobe region. In the present embodiment, the z value z _i in the i-th voxel is calculated by the following equation.

z _i = (y _i −m _i ) / σ _i
y _i : Functional coupling strength of the default mode network of the subject's medial temporal lobe region in the i th voxel m _i : Function of the default mode network in the i th voxel of the medial temporal lobe region of the control group Mean σ _i : standard deviation of the functional bond strength of the default mode network in the i-th voxel in the medial temporal lobe region of the control group

  The z-value is a dimensionless number, the z-value of the voxel whose test intensity is equal to the average of the control intensity is 0, the z-value of the voxel whose test intensity is greater than the average of the control intensity is positive, and the test intensity is the average of the control intensity Smaller voxels have negative z values. In addition, if z value is a value which shows the position in distribution of the test intensity | strength of control intensity | strength, it will not specifically limit, For example, a deviation value may be sufficient.

  As described above, in the standardization step S2, the standardization means 12 calculates the z value indicating the position of the test intensity in the distribution of the control intensity based on the average and standard deviation of the control intensity for each voxel in the medial temporal lobe region. To calculate. The z value is output to the index calculation means 13 together with the voxel number.

  As a modification, in the standardization step S2, at least one of the processes by the preprocessing unit 12a, the independent component analysis unit 12b, and the average / standard deviation calculation unit 12c is performed by a device other than the diagnosis support device 1. Also good. For example, each process by the pre-processing unit 12a, the independent component analyzing unit 12b, and the average / standard deviation calculating unit 12c is performed by the other apparatus, and the control intensity average and standard deviation data calculated by the other apparatus are obtained. You may transmit to the diagnostic assistance apparatus 1. FIG. In this case, the standardization means 12 calculates the z value based on the average and standard deviation received from another device.

  Of the processes in the standardization step S2, the processes by the preprocessing unit 12a, the independent component analysis unit 12b, and the average / standard deviation calculation unit 12c may be performed before or simultaneously with the test intensity acquisition step S1.

  Further, the z value may be calculated for the entire brain of the test image 10. In this case, after the z value is calculated, a mask process is performed to extract the z value of only the medial temporal lobe region, and the extracted z value may be output to the index calculation means 13.

(Index calculation)
Subsequently, in an index calculation step S3, an index for predicting whether or not the subject is AD is calculated based on the z value calculated in the standardization step S2. In the present embodiment, first, the index calculation unit 13 forms a histogram of the z values input from the standardization unit 12, and calculates the frequency distribution of voxels with respect to the z values.

  FIG. 5 is a graph showing an example of a frequency distribution of voxels. The horizontal axis represents the z value, and the vertical axis represents the voxel frequency (number of voxels) corresponding to the z value.

  Subsequently, the index calculating unit 13 arranges the voxels in ascending order of the z value, and calculates the upper average of the z values from the highest voxel having the smallest z value to the upper x (0 <x ≦ 100)% voxels. . Although it will not specifically limit if x is in the range of more than 0 and 100 or less, it is preferable that x is a predetermined value in the range of 10-25, and is x = 10 in this embodiment.

  For example, in the voxel frequency distribution shown in FIG. 5, if the minimum value of the z value is −5 and the maximum value of the z value is +4, the index calculating unit 13 selects the voxel having the z value of −5. The voxel with the z value of +4 is set as the least significant voxel, and the most significant voxel to the least significant voxel are arranged in ascending order of the z value. Subsequently, the index calculation unit 13 obtains, for each voxel, an upper average of z values from the highest-order voxel to the voxel. That is, for the i-th (i = 1 to 1000) voxel, the upper average of z values from the highest-order voxel to the i-th voxel is obtained.

FIG. 6 is a graph showing an example of a higher average with respect to the rank of voxels. The horizontal axis represents the order of voxels in%, with the highest voxel corresponding to 0% and the lowest voxel corresponding to 100%. The vertical axis represents the upper average of z values in each voxel. For example, the value corresponding to the upper 10% voxel (z ₁₀ ) is the 100th voxel from the highest voxel (the upper 10% voxel). The average of z values up to is shown. Since the z value increases as the rank of the voxel decreases, the upper average corresponding to the lower voxel increases.

In the present embodiment, the index calculation means 13 calculates z ₁₀ , which is an upper average of z values from the highest voxel to the top 10% voxel, as an index for predicting whether or not the subject is AD. . The calculated index is output to the prediction unit 14. Note that the index calculation unit 13 may calculate the index without performing the above histogram formation.

(AD prediction)
Subsequently, in the prediction step S4, it is predicted whether or not the subject is AD based on the index calculated in the index calculation step S3. In the present embodiment, the prediction unit 14 compares the index (z ₁₀ ) calculated by the index calculation unit 13 with the threshold value 18 stored in the HDD 103.

  As a result, when the index is smaller than the threshold 18 (Yes in Step S4), the prediction unit 14 predicts that the subject is highly likely to be AD (Step S5). On the other hand, when the index is equal to or larger than the threshold 18 (No in step S4), the predicting unit 14 predicts that the possibility that the subject is AD is low (step S6).

  The threshold value 18 is set as follows. As shown in the examples below, the upper average of the z-values from the top voxel to the lower voxel is the difference between the AD patient and the control regardless of whether the subject and the control group facility or database are the same. It is known that there is a significant difference between them. In particular, the significant average between AD patients and controls is significant in the top average of z values from the top voxel to the top 10-25% voxels. For example, the top average of z values from the top voxel to the top 10% voxels is 75% of AD patients less than about -2, while 75% of controls are about -2 or more. Also, the upper mean z-value from the top voxel to the top 25% voxels is 75% for AD patients less than about -1.5, while 75% for controls is about -1.5 or higher. .

Therefore, in this embodiment, the threshold value 18 is set to −2, and the predicting unit 14 compares the index z ₁₀ input from the index calculating unit 13 with −2 to predict whether or not the subject is AD. . Thereby, it is possible to obtain a stable sensitivity exceeding 70% regardless of whether or not the subject's facility or database is the same as the control group.

(Output of prediction results)
Subsequently, in the output step S7, the prediction result predicted in the prediction step S4 is output. In this embodiment, the output means 15 outputs the prediction result 19 by the prediction means 14 to the display device 3 shown in FIG. Thereby, the prediction result of whether or not the subject is AD is displayed on the display device 3. Although the specific description of a prediction result is not specifically limited, For example, you may display that possibility that it is AD is high or possibility that it is AD is low. Or you may display by numerical values, such as a percentage which shows that a test subject is AD. In this case, the numerical value may be increased as the index calculated by the index calculation unit 13 is smaller. As another example, the prediction result may be displayed using symbols indicating evaluations such as A to D. A doctor or the like can use this prediction result as one of diagnostic materials for determining whether or not the subject has AD.

  The output mode of the prediction result is not particularly limited. For example, the prediction result 19 may be output to a printing apparatus such as a printer, and the prediction result may be printed on a paper medium.

[Additional Notes]
As described above, in the above-described embodiment, it is possible to predict the AD of a subject whose facility or database is different from the control group serving as a reference for prediction with high probability. Therefore, accurate diagnosis of the subject's AD can be supported.

  Also, in machine learning in conventional brain image research, a technique called multi-voxel pattern analysis has been mainly used. However, when this technique is used for AD prediction, an enormous number of rests are used. Sometimes learning data of fMRI image is required. On the other hand, in the above embodiment, the subject's AD can be predicted without performing complicated arithmetic processing such as machine learning, and a large number of control groups are not required.

  It should be noted that the present invention is not limited to the above-described embodiment, and various modifications are possible within the scope shown in the claims, and a mode obtained by appropriately combining technical means disclosed in the above-described embodiment is also described in this embodiment. It is included in the technical scope of the invention.

  For example, in the above embodiment, the index calculation means 13 calculates the “upper average” of z values from the highest voxel to the top x% voxels as an index for predicting whether or not the subject is AD. However, the present invention is not limited to this. When the number of voxels in the medial temporal lobe is the same in the test image and the resting fMRI image in the control group, the index calculation means 13 calculates the “total” of z values from the top voxel to the top x% voxels. , It may be calculated as the index.

  In the above embodiment, the index calculation unit 13 arranges the voxels in ascending order of the z value and calculates the voxel having the smallest z value as the most significant voxel in order to calculate the index. The voxels may be arranged in order of increasing z value, and the voxel having the largest z value may be set as the most significant voxel. In this case, the index calculation means 13 calculates the lower average of z values from the lowest voxel with the smallest z value to the lower x% voxels as the index.

  Further, the diagnosis support program 16 in the above embodiment only needs to perform arithmetic processing according to the procedure shown in the flowchart of FIG. 3, and individual processing such as preprocessing, independent component analysis, average / standard deviation calculation, z value calculation, and the like. The arithmetic processing may be performed using known software.

  In the present invention, the prediction unit 14 of the diagnosis support apparatus 1 is not an essential component, and the index calculated by the index calculation unit 13 may be output by the output unit 15. In this case, a doctor or the like may predict whether or not the subject is AD based on the output index.

  In the following, it was verified that the AD diagnosis support apparatus according to the present invention can predict AD of subjects having different facilities and databases from the control group with high probability. First, by using the fMRI apparatus shown in FIG. 1, resting fMRI images of the AD patient and the whole brain of the control who visited the Shimane University Hospital were acquired. In addition, the HDD 103 of the diagnosis support apparatus 1 shown in FIG. 1 and FIG. 2 stores, as a control image 17, a resting fMRI image of the control whole brain in the ADNI database in North America that can be used freely by researchers. In addition, resting fMRI images of the brains of AD patients in the ADNI database were also acquired. Table 1 shows the numbers of AD patients (AD) and controls (HC), age, sex ratio, and MMSE score in this example.

  Subsequently, AD patients (hereinafter referred to as “Shimane-AD”) and a control group (hereinafter referred to as “Shimane-HC”) who visited a hospital attached to the Shimane University School of Medicine, and AD patients in the ADNI database (hereinafter referred to as “ The whole brain fMRI image of “ADNI-AD” was input to the diagnosis support apparatus 1, and the DMN bond strength of Shimane-AD, Shimane-HC, and ADNI-AD was calculated by the test strength acquisition means 11. For the control group in the ADNI database (hereinafter referred to as “ADNI-HC”), the DMNI binding strength (control strength) of ADNI-HC was calculated from the control image 17. Further, the average and standard deviation of the control intensity were calculated by the average / standard deviation calculation means 12c.

  Subsequently, on the basis of the mean and standard deviation of the control intensity calculated by the average / standard deviation calculating means 12c, the z-value calculating means 12d determines each Shimane-AD DMN bond strength for each voxel in the medial temporal lobe region. The z value was calculated. Similarly, each DMN bond strength of Shimane-HC, each DMN bond strength of ADNI-AD, and each DMN bond strength of ADNI-HC are also z-valued for each voxel in the medial temporal lobe region by the z-value calculating means 12d. The value was calculated. Specifically, z values were determined for the medial temporal lobe region 1295 voxels. The distribution of the 1295 voxels is continuous, and the left hippocampus 273 voxels, the right hippocampus 288 voxels, the left hippocampal gyrus 286 voxels, the right hippocampal gyrus 316 voxels, the left amygdala 62 voxels, and the right amygdala 70 voxels (determining the z value) In addition, when the number of 1295 voxels was 100%, the hippocampus was about 43%, the hippocampal circulation was about 46%, and the amygdala was about 10%. One voxel was 3 mm × 3 mm × 3 mm.

  FIG. 7 shows a graph in which the voxel frequency is converted into a histogram for the calculated z value. FIG. 7A is a graph showing the average of the voxel frequency with respect to the ADNI-AD z value and the average of the voxel frequency with respect to the ADNI-HC z value. FIG. 7B is a graph showing an average of voxel frequencies with respect to the z value of Shimane-AD and an average of voxel frequencies with respect to the z value of Shimane-HC. In these graphs, both the ADNI database and the Shimane University School of Medicine show that the distribution of voxel frequency versus z-value is such that the average of AD patients is slightly shifted to the left relative to the average of controls.

  Subsequently, the index calculation means 13 calculates an upper average of z values from ADNI-AD, ADNI-HC, Shimane-AD, and Shimane-HC from the highest voxel to the upper x% voxels, x = 10, 25, Calculation was performed for each of 50 and 100. The calculation result is shown in FIG. FIG. 8 (a) is a box-and-whisker diagram showing the distribution of the upper average of ADNI-AD and the distribution of the upper average of ADNI-HC, and FIG. 8 (b) shows the distribution of the upper average of Shimane-AD, and It is a box-and-whisker diagram which shows distribution of the upper average of Shimane-HC.

  As shown in FIG. 8A, when x = 10 and 25, the 75th percentile value of the upper average of ADNI-AD and the 25th percentile value of the upper average of ADNI-HC are substantially equal, and x = 50 and 100 Even in this case, the upper average box of ADNI-AD hardly overlaps the upper average box of ADNI-HC. Similarly, as shown in FIG. 8B, when x = 10 and x = 25, the 75th percentile value of Shimane-AD and the 25th percentile value of Shimane-HC are substantially equal, and x = 50 and 100 Even in this case, the Shimane-AD upper average box overlaps only partially with the Shimane-HC upper average box.

  Here, since the reference group for calculating the upper average is a control group of the ADNI database, the upper average of z values from the highest voxel to the lower voxel is determined by the facility or database of the subject and the control group. It can be seen that there is a significant difference between AD patients and controls, whether or not they are the same. In particular, when x = 10 and 25, it can be seen that significant differences between AD patients and controls are significant.

  8A and 8B, in both the ADNI database and the Shimane University School of Medicine, the upper average of z values from the highest voxel to the top 10% voxel is approximately −2 for 75% of AD patients. While 75% of the controls were about -2 or higher, and the top average z-value from the top voxel to the top 25% voxels was 75% of AD patients was about -1.5 While decreasing, it was found that 75% of the controls were greater than about -1.5.

  Therefore, the threshold value was set to -2 when x = 10 and -1.5 when x = 25. And the index calculated about each subject of Shimane-AD and Shimane-HC was compared with the threshold value, it was predicted whether each subject was AD, and it was verified whether the prediction result was correct. Specifically, for each subject, the predicted result of the subject was compared with the AD / HC classification of the subject that was grasped in advance by LOOCV. As a result, the correct answer rate of the prediction result exceeded 70% in all cases of x = 10, x = 25, and x = 50. The correct answer rate is 66.0% when x = 100, 72.3% when x = 50, 74.5% when x = 25, and 76.6% when x = 10, and x is small. As a result, the correct answer rate also increased.

  Thus, it was found that in this example, it is possible to predict whether or not a subject is AD with a correct answer rate exceeding 70% regardless of whether or not the subject's facility or database is the same as that of the control group.

FIG. 9A is an ROC curve (Receiver Operating Characteristic) of AD prediction for ADNI-AD and ADNI-HC in the present example, and FIG. 9B is Shimane-AD and Shimane in the present example. -ROC curve of AD prediction for HC. Table 2 shows the AUC value corresponding to each ROC curve. In any case, high sensitivity and specificity could be obtained particularly when x = 10 and x = 25. The AD prediction for ADNI-AD and ADNI-HC, and the AD prediction for Shimane-AD and Shimane-HC both showed good AUC values in the range of Moderate accuracy.

  Further, as shown in FIG. 8B, even in the case of x = 50 and x = 100, the Shimane-AD upper average box partially overlaps with the Shimane-HC upper average box. Therefore, in any case, it was found that by setting the threshold value to, for example, the intermediate value of the overlapping portion, it is possible to predict whether or not the subject is AD with a higher correct answer rate than the conventional technique (50%).

  On the other hand, as a comparative example, AD prediction for Shimane-AD and Shimane-HC was performed using a conventional technique. Specifically, the resting fMRI image of the ADNI database was machine-learned by multi-voxel pattern analysis based on a linear support vector machine. And the index value for AD prediction was computed with respect to Shimane-AD and Shimane-HC using the prediction model built by the said machine learning.

  FIG. 10A is a box-and-whisker diagram showing the distribution of the index values of Shimane-AD and Shimane-HC in the comparative example, and FIG. 10B is Shimane-AD and Shimane-HC in the comparative example. It is a ROC curve of AD prediction for. As shown in FIG. 10A, most of the Shimane-AD box overlaps with the upper average box of Shimane-HC. Therefore, as shown in FIG. 10 (b), the sensitivity and specificity of AD prediction for Shimane-AD and Shimane-HC were low, and the correct answer rate of LOOCV was a chance level (about 50%). Moreover, AUC value with respect to Shimane-AD and Shimane-HC was also 54.6, and was the level of Low accuracy.

  As described above, when the subject's facility or database is different from the control group, the AD diagnosis support apparatus according to the present invention can predict AD with higher probability than the conventional technique. In addition, according to the present invention, it is possible to perform early diagnosis of AD by fMRI, particularly diagnosis of mild cognitive impairment by fMDI, even for a subject whose database or facility is different from the control group serving as a reference for prediction. Early detection of AD is expected. In addition, since the medial temporal region is easy to identify anatomically and is a relatively narrow region, changes in this region in AD are less likely to cause spatial individual differences compared to other regions. It is considered suitable.

DESCRIPTION OF SYMBOLS 1 Diagnosis support apparatus 2 fMRI apparatus 3 Display apparatus 10 Test image 11 Test intensity acquisition means 11a Reception means 11b Preprocessing means 11c Independent component analysis means 12 Standardization means 12a Preprocessing means 12b Independent component analysis means 12c Average / standard deviation calculation means 12d z value calculation means 13 index calculation means 14 prediction means 15 output means 16 diagnosis support program 17 reference image 18 threshold 19 prediction result 100 diagnosis support system

Claims (8)

A test strength acquisition means for acquiring a functional coupling strength (hereinafter referred to as “test strength”) of a default mode network of the medial temporal lobe region of the subject;
Based on the mean and standard deviation of the functional binding strength (hereinafter referred to as “control strength”) of the default mode network of the medial temporal lobe region of the control group that is not Alzheimer's disease, the test strength in the distribution of the control strength Standardization means for calculating a z-value indicating a position for each voxel of the medial temporal lobe region;
Index calculating means for calculating an index for predicting whether or not the subject has Alzheimer's disease based on the z value;
A diagnostic support apparatus for Alzheimer's disease. The index calculating means includes
The voxels are arranged in ascending order of the z value, and the upper average of the z values from the highest voxel with the smallest z value to the upper x (0 <x ≦ 100)% voxels is calculated as the index. The diagnosis support apparatus according to claim 1.   The diagnosis support apparatus according to claim 2, wherein x is a predetermined value within a range of 10 to 25. The test strength acquisition means acquires the test strength from a test image of fMRI at rest of the subject's brain,
The standardization means obtains the mean and standard deviation from a control image of resting brain fMRI of the control group,
The diagnosis support apparatus according to claim 1, wherein the test image and the control image are preprocessed so that the shape, size, and number of voxels of the medial temporal lobe region are the same. The diagnosis support apparatus according to claim 1, wherein the standardization unit calculates the z value z _i in the i-th voxel of the medial temporal lobe region by the following formula.
z _i = (y _i −m _i ) / σ _i
y _i : Functional coupling strength of the default mode network of the subject's medial temporal lobe region in the i th voxel m _i : Function of the default mode network in the i th voxel of the medial temporal lobe region of the control group Mean σ _i : standard deviation of the functional bond strength of the default mode network in the i-th voxel in the medial temporal lobe region of the control group   The diagnosis support apparatus according to claim 1, further comprising a prediction unit that predicts whether or not the subject has Alzheimer's disease based on the index. Computer
A test strength acquisition means for acquiring a functional coupling strength (hereinafter referred to as “test strength”) of a default mode network of the medial temporal lobe region of the subject;
Based on the mean and standard deviation of the functional binding strength (hereinafter referred to as “control strength”) of the default mode network of the medial temporal lobe region of the control group that is not Alzheimer's disease, the test strength in the distribution of the control strength A standardization means for calculating a z-value indicating a position for each voxel of the medial temporal lobe region; and
Index calculating means for calculating an index for predicting whether or not the subject has Alzheimer's disease based on the z value;
A diagnostic support program for Alzheimer's disease. A test strength acquisition step of acquiring a functional coupling strength (hereinafter referred to as “test strength”) of a default mode network of the subject's medial temporal lobe region;
Based on the mean and standard deviation of the functional binding strength (hereinafter referred to as “control strength”) of the default mode network of the medial temporal lobe region of the control group that is not Alzheimer's disease, the test strength in the distribution of the control strength A standardization step of calculating a z-value indicating a position for each voxel of the medial temporal lobe region;
An index calculating step for calculating an index for predicting whether or not the subject has Alzheimer's disease based on the z value;
A diagnostic support method for Alzheimer's disease, comprising: