KR20230120605A - Dementia related information generation method using volume predicted from 2-dimensional magnetic resonance imaging and analysis apparatus - Google Patents

Abstract

A dementia information calculation method by using a volume predicted from a 2-dimensional magnetic resonance imaging (MRI) may comprise: a step in which an analysis apparatus receives 2D MRI slices of a target; a step in which the analysis apparatus extracts interest areas by inputting the 2D MRI slices into a segmentation model; a step in which the analysis apparatus predicts the volume of at least one area of the interest areas by inputting pixel information of the interest areas into a first learning model which has learned in advance; and a step in which the analysis apparatus calculates dementia information of the target by inputting the volume of at least one area into a second learning model which has learned in advance. Therefore, useful information related to dementia diagnosis may be provided.

Description

Dementia information calculation method and analysis device using volume predicted from 2-dimensional MRI

The technique to be described below is a technique for predicting whether or not dementia exists using the volume of a region of interest extracted from a 2D MRI.

Dementia refers to a syndrome that causes the future of cognitive functions such as memory, language, and judgment. There are many types of dementia, but Alzheimer's disease is the most common form of dementia.

Brain imaging such as Magnetic Resonance Imaging (MRI) and positron emission tomography (PET) is used to diagnose dementia. Typically, cortical thickness is used as a factor related to dementia.

Korean Patent Publication No. 10-2020-0062589

Conventionally, brain imaging such as 3D (dimensional) MRI and PET-CT was used for early diagnosis of dementia. However, 3D MRI uses high-spec equipment and requires high time and cost. Furthermore, 2D MRI scanners used in general health examination centers or primary hospitals generate about 20 2D images, making accurate diagnosis difficult.

The technique described below aims to provide a technique for predicting dementia or calculating dementia information based on the volume of a region of interest predicted from a small number of 2D MRI images.

A method for calculating dementia information using a volume predicted from a 2D MRI includes the steps of receiving 2D MRI slices of a subject by an analysis device, inputting the 2D MRI slices into a segmentation model by the analysis device, and extracting regions of interest, the analysis predicting, by a device, the volume of at least one region among the regions of interest by inputting pixel information of the regions of interest into a first learning model pre-learned; and the analyzing device predicting the volume of the at least one region in advance and calculating dementia information of the subject by inputting the information into the learned second learning model.

The analysis device that calculates dementia information using the volume predicted from 2D MRI is an interface device that receives 2D MRI slices of the subject, a segmentation model that extracts the region of interest from the 2D MRI slice, and predicts the volume by receiving information about the region of interest. A storage device for storing a first learning model for calculating dementia information and a second learning model for calculating dementia information, extracting regions of interest by inputting the received 2D MRI slices to the segmentation model, and converting pixel information of the extracted regions of interest into the first learning model. 1 includes an arithmetic device for predicting the volume of at least one region among the regions of interest by inputting the input to a learning model, and calculating dementia information of the subject by inputting the volume of the at least one region to the second learning model .

The regions of interest include at least one of frontal gray matter, temporal gray matter, parietal gray matter, occipital gray matter, lateral ventricle, hippocampus, and extracerebral cerebrospinal fluid region.

The technology described below calculates dementia-related information using a learning model that analyzes a small number of 2D MRI images. The technology described below can provide useful information for diagnosing dementia based on brain MRI calculated by a low-end 2D MRI scanner.

1 is an example of an entire process of calculating dementia information of a subject using 2D brain MRI.
2 is an example of a process of learning a segmentation model for extracting a region of interest from a brain image.
3 is an example of a learning process of a learning model that predicts cortical thickness and brain region volume based on regions of interest.
4 is an example of a learning process of a learning model that calculates dementia information based on the volume of a brain region.
5 is an example of an analysis device that calculates dementia information of a subject using brain images.

Since the technology to be described below can have various changes and various embodiments, specific embodiments will be illustrated in the drawings and described in detail. However, this is not intended to limit the technology described below to specific embodiments, and it should be understood to include all modifications, equivalents, or substitutes included in the spirit and scope of the technology described below.

Terms such as first, second, A, B, etc. may be used to describe various elements, but the elements are not limited by the above terms, and are merely used to distinguish one element from another. used only as For example, without departing from the scope of the technology described below, a first element may be referred to as a second element, and similarly, the second element may be referred to as a first element. The terms and/or include any combination of a plurality of related recited items or any of a plurality of related recited items.

In terms used in this specification, singular expressions should be understood to include plural expressions unless clearly interpreted differently in context, and terms such as “comprising” refer to the described features, numbers, steps, operations, and components. , parts or combinations thereof, but it should be understood that it does not exclude the possibility of the presence or addition of one or more other features or numbers, step-action components, parts or combinations thereof.

Prior to a detailed description of the drawings, it is to be clarified that the classification of components in the present specification is merely a classification for each main function in charge of each component. That is, two or more components to be described below may be combined into one component, or one component may be divided into two or more for each more subdivided function. In addition, each component to be described below may additionally perform some or all of the functions of other components in addition to its main function, and some of the main functions of each component may be performed by other components. Of course, it may be dedicated and performed by .

In addition, in performing a method or method of operation, each process constituting the method may occur in a different order from the specified order unless a specific order is clearly described in context. That is, each process may occur in the same order as specified, may be performed substantially simultaneously, or may be performed in the reverse order.

The technology described below is a technology for calculating dementia information from brain MRI.

The following dementia includes Alzheimer's dementia.

Dementia information refers to information for diagnosing or predicting dementia. For example, the dementia information includes whether or not the subject has dementia, whether the dementia is a high-risk group (possibility of future onset), dementia-related indicators, and the like. Dementia-related indicators may include various scores, brain atrophy, and the like that have been studied as significant in diagnosing dementia or predicting dementia.

Hereinafter, a device that analyzes a brain MRI and predicts whether a subject has dementia is referred to as an analysis device. The analysis device may take the form of a computer device such as a PC, a smart device, a network server, and a data processing dedicated chipset.

The analysis device can predict dementia based on brain images using a plurality of learning models. The learning model is a machine learning model. Thus, the classification model can be any of various types of models. For example, the learning model is any one of a decision tree, a random forest (RF), a K-nearest neighbor (KNN), a Naive Bayes, a support vector machine (SVM), a deep neural network (DNN), and a regression model. It may be a model implemented in the manner of There are various types of DNN models. For example, DNN includes a segmentation model for extracting a region of interest from an image, a classification model for performing classification based on features of an image, and the like.

1 is an example of an entire process 100 of calculating dementia information of a subject using 2D brain MRI.

The analysis device may calculate dementia information for a subject based on a small number of 2D MRI slices. In this case, the number of 2D MRI slices may be the number calculated by a general 2D MRI scanner. The researcher built a learning model using 20 2D MRI slices. Therefore, it is assumed that the number of 2D MRI slices is 20 for convenience of description below.

The analysis device receives the subject's 2D brain MRI (110). For example, the analysis device may receive 20 2D MRI slices of the subject.

The analysis device extracts a region of interest (ROI) from the received 2D MRI slices (120). The analysis device may extract a plurality of set regions of interest using a previously learned segmentation model. Regions of interest may be regions for calculating the volume of a specific region. Furthermore, the regions of interest may include regions for calculating cortical thickness.

The regions of interest are frontal gray matter, temporal gray matter, parietal gray matter, occipital gray matter, lateral ventricle, hippocampus and extracerebral cerebrospinal fluid. It may include at least one of extracerebral cerebrospinal fluid.

The analysis device may calculate a certain brain image index using the extracted regions of interest (130). Here, the brain image index may include the number of extracted pixels of the region of interest. The analysis device may normalize the region of interest to a preset size. For example, the analyzer may calculate the size of the entire brain and normalize the size of the region of interest to a constant level. The total size of the brain may be calculated by summing up the sizes of all of the aforementioned regions of interest and white matter of the regions of interest. Meanwhile, since various methodologies for brain size determination have been studied, the analysis device may calculate the subject's brain size using one of various image processing techniques. The analysis device may calculate brain image indexes such as the number of pixels based on the normalized ROI(s).

The analysis device may estimate the volume of the brain region using a pre-learned learning model (140). The analysis device may estimate the volume in each region of interest using a plurality of learning models. For example, the analysis device may calculate the volume of each region by individually using a learning model for the lateral ventricle, a learning model for the hippocampus, and a learning model for the extracerebral cerebrospinal fluid region.

In addition, the analysis device may use a learning model for predicting cortical thickness. In this case, the analysis device may calculate the thickness of the cerebral cortex in the corresponding region by inputting the brain image index for the region of interest to the learning model. The analysis device may predict cortical thickness in each region of interest using a plurality of learning models. For example, the analysis device may calculate the cortical thickness of each region using individual learning models for the frontal, temporal, parietal, and occipital regions.

Meanwhile, the analysis device may calculate the volume of the brain region or the thickness of the cerebral cortex by additionally using the subject's clinical information (gender, age, etc.) in addition to the brain image. In this case, each learning model must be trained in advance to calculate a brain region volume or cerebral cortical thickness using clinical information in addition to brain MRI.

The analysis device may finally calculate dementia information for the subject based on the volume of the region(s) of interest (150).

The dementia information calculation process of FIG. 1 uses a plurality of learning models to calculate dementia information of a subject. The researcher built a number of learning models for the above-mentioned process, and verified the performance and dementia prediction performance of individual models. Hereinafter, the construction process of the learning models used by the researcher to calculate dementia information of the subject will be described.

The researcher collected data of the population who visited the affiliated medical institution (Samsung Seoul Hospital). The population was selected as subjects who had both MRI and PET scans taken within a certain period of time. Table 1 below shows information about the population. The researcher used data from a total of 1,120 people. The population consisted of 529 patients with Alzheimer's disease dementia (ADD) and 591 normal controls (NC). Table 1 shows clinical information on population composition and cortical thickness and brain region volume by region of interest. Clinical information includes gender, age and education level. The thickness of the cerebral cortex includes the thickness of the frontal, temporal, parietal and occipital regions. The brain region volume includes the volume of the lateral ventricle, hippocampus and extracerebral cerebrospinal fluid region.

total population ADD (Patients with Dementia) NC (Normal) average Standard Deviation(%) average Standard Deviation(%) average Standard Deviation(%) Subjects (persons) 1,120 (100.0%) 529 (47.2%) 591 (52.8%) Female (person) 659 (58.8%) 306 (57.8%) 353 (59.7%) Age (years) 69.0 10.9 69.9 10.2 68.2 11.5 Education (years) 11.7 4.8 11.3 4.9 12.1 4.7 Frontal cortical thickness (mm) 3.087 0.152 3.007 0.154 3.158 0.108 Parietal cortical thickness (mm) 2.984 0.192 2.877 0.200 3.080 0.120 Temporal cortical thickness (mm) 3.186 0.189 3.068 0.185 3.291 0.117 Occipital cortical thickness (mm) 2.897 0.204 2.793 0.186 2.991 0.171 Lateral ventricle volume (mm ³ ) 36123.93346 20197.7947 44637.0 20985.2 28507.7 16056.7 Hippocampus volume (mm ³ ) 2,627 632 2,177 523 3,031 411 Frontal extracerebral CSF volume (mm ³ ) 62450.29197 21740.67107 72456.38 21692.64 52262.27 16416 Temporal extracerebral CSF volume (mm ³ ) 37218.44005 12610.11551 42089 12551.5 32259.33 10580.77 Parietal extracerebral CSF volume (mm ³ ) 36737.31928 12420.93104 42753.81 12495.3 30611.44 8839.807 Occipital extracerebral CSF volume (mm ³ ) 15064.27969 4565.302592 16663.86 4881.548 13435.62 3545.16 Anterior lateral ventricle volume (mm ³ ) 17740.63456 9031.29889 9074.209 9074.209 14515.32 7762.385 Posterior lateral ventricle volume (mm ³ ) 20459.65506 12092.68325 12581.57 12581.57 15613.22 9357.293

The researcher built a segmentation model using the data of 980 people among the data of the population. The data of 980 persons are 3D MRI data generated using a 3.0 T MRI scanner (Philips 3.0T Achieva). The researcher performed structural image analysis using the 3D segmentation mask of the CIVET pipeline. In CIVET, cortical thickness is computed as the Euclidean distance between the inner and outer surfaces of the cerebral cortex. Regions of interest include (i) frontal, temporal, parietal and occipital gray matter, (ii) lateral ventricle (iii) hippocampus and (iv) extracerebral cerebrospinal fluid region. The researcher used the regions of interest extracted from the data of 980 people as the ground truth for learning the segmentation model.

2 is an example of a process 200 of learning a segmentation model for extracting a region of interest from a brain image. Hereinafter, the learning process of the model will be described as being performed by the learning device. The learning device refers to a computer device capable of processing image data and performing a learning process of a machine learning model.

The learning device learns a segmentation model using MRI images of multiple subjects. However, FIG. 2 illustrates a process of learning a segmentation model using MRI of one subject as an example.

The learning device receives a 3D MRI of a specific subject belonging to the aforementioned population (210).

The learning device prepares a learning data set (220). The training data set consists of an input slice (original slice) and an answer image for the corresponding slice for each slice. The learning device selects 20 slices from 480 slices of 3D MRI. The learning device may select slices matching 20 slices obtained from the 2D MRI scanner among 480 slices. In addition, the learning device receives the correct answer images for the selected 20 slices. The answer image is an image obtained by dividing regions of interest with respect to the input slice. As the correct answer image, an image displayed (annotated) by a medical imaging expert by dividing the region of interest may be used.

The learning device performs learning of the segmentation model using the learning data (230). The learning device may learn a segmentation model using input slices and answer images for 20 slices. The segmentation model receives an input slice and learns to extract (separate) regions of interest of an answer image.

The segmentation model may be a semantic segmentation model. For example, the segmentation model may be a fully convolutional network (FCN)-based model such as U-net. The researcher extracted a region of interest from a 2D MRI slice using a segmentation model of a U-net structure.

The segmentation model extracts multiple regions of interest. The researcher performed curriculum learning to accurately extract various areas of interest with different characteristics. Curriculum learning is a learning strategy that imitates the human learning process to learn data with easy difficulty first and gradually learns difficult data.

The researcher built a segmentation model and verified its performance according to the following five learning sequences.

turn order of learning One All ROIs 2 Hippocampus→Hippocampus and LV→All regions of interest 3 Hippocampus and lateral ventricles → all regions of interest 4 Lateral ventricle → hippocampus and lateral ventricle → all regions of interest 5 Lateral ventricles → all regions of interest

The performance of the segmentation model was measured by IoU (intersection over union) and DSC (dice similarity coefficient) calculated based on the correct answer and the region of interest extracted by the segmentation model. IoU and DSC may be expressed by Equations 1 and 2 below, respectively.

In the above equation, A _GT is the correct region of interest, A _pred is the region of interest (prediction region) identified by the segmentation model, and A _overlap is the overlapping region between A _GT and A _pred .

The segmentation model showed IoU and DSC as shown in Table 3 below for the six regions of interest. The segmentation model showed the highest performance in lateral ventricle extraction.

area of interest IoU DSC frontal area 0.71356 0.83168 temporal region 0.67835 0.80696 parietal region 0.64923 0.7857 occipital 0.57339 0.72749 lateral ventricle 0.87904 0.93524 hippocampus 0.6961 0.8188 average 0.69828 0.81765

Table 4 below shows the extraction performance of the segmentation model for the extracerebral cerebrospinal fluid region. The segmentation model showed the highest performance in the lateral ventricle area.

area of interest IoU DSC Frontal extracerebral CSF 0.468802 0.620074 Temporal extracerebral CSF 0.325105 0.47922 parietal extracerebral CSF 0.459181 0.604894 Occipital extracerebral CSF 0.293677 0.443401 anterior lateral ventricle extracerebral CSF 0.771025 0.861737 posterior lateral ventricle extracerebral CSF 0.731659 0.835934 average 0.508241 0.640877

Table 5 below is the result of analyzing the performance of the learning sequence in Table 2. The learning method in the order of learning the lateral ventricle first showed slightly higher performance than the learning method in the other order. Therefore, it can be seen that the technique of performing curriculum learning by dividing the region of interest is significant in the performance of the segmentation model.

turn order of learning IoU One All ROIs 0.653 2 Hippocampus→Hippocampus and LV→All regions of interest 0.565 3 Hippocampus and lateral ventricles → all regions of interest 0.665 4 Lateral ventricle → hippocampus and lateral ventricle → all regions of interest 0.683 5 Lateral ventricles → all regions of interest 0.689

3 is an example of a learning process 300 of a learning model that predicts cortical thickness and brain region volume based on regions of interest. In FIG. 3 , the learning model predicts the thickness of the cerebral cortex and the volume of the brain region based on the above-mentioned region of interest.

The learning of FIG. 3 is based on a state in which the aforementioned region of interest is extracted from MRI of the brain of a subject belonging to a population. That is, the learning device may use the region of interest calculated using the segmentation model of FIG. 2 .

The learning device extracts input data based on regions of interest extracted from brain MRI (310). The input data may include brain image indexes and clinical information extracted from regions of interest extracted by the segmentation model.

The brain imaging index includes size information for six regions of interest (frontal gray matter, temporal gray matter, parietal gray matter, occipital gray matter, lateral ventricle, hippocampus, and extracerebral cerebrospinal fluid region). The size of the six regions of interest may be determined by the number of pixels in the region of interest image. That is, the size of the region of interest may be the sum of all pixels belonging to the region. Meanwhile, the size of the region of interest may be normalized based on the size of the entire brain. Normalization is the process of correcting the size of the brain, which differs from person to person, to a certain size.

The summed number of pixels of a specific region of interest may be the sum of all pixels of the corresponding region of interest for each of the 20 slices. Referring to the frontal gray matter as a reference, the summed number of pixels of the frontal gray matter may be a value obtained by summing all pixels belonging to the frontal gray matter region in each of 20 slices. Assume that the indexes of 20 slices are 0 to 19. The summed number of pixels in the frontal gray matter is the number of pixels belonging to the frontal gray matter region of slice 0, the number of pixels belonging to the frontal gray matter region of slice 1, ..., the number of pixels belonging to the frontal gray matter region of slice 19. It may be the sum of all values. Furthermore, the number of pixels of the ROI may be an average value of the number of pixels of all slices.

Clinical information may include age and gender of the subject. Clinical information may also include APOE4 genotype.

The researcher used the combined number of pixels, brain size, and clinical information of the seven regions of interest as input data.

The learning device performs a process of learning a learning model using the extracted input data (320). The learning device repeats the learning process of the learning model using input data extracted from brain MRIs of a plurality of subjects.

The learning device inputs the extracted input data to the learning model and performs learning while comparing the value (predicted value) output by the learning model with the correct answer. The trained model is trained to predict cortical thickness and volume of specific brain regions. The correct answer is the information calculated by MRI and PET-CT of subjects belonging to the population (see Table 1).

A learning model may be built as an individual model according to the information to be predicted. A learning model is a machine learning model and can be implemented as one of many types of models. The researcher used a regression model. The researcher developed a model that predicts cortical thickness in the frontal region (anterior temporal model), a model that predicts cortical thickness in the temporal region (temporal region model), a model that predicts cortical thickness in the parietal region (parietal region model), and a model that predicts cortical thickness in the occipital region (parietal region model). A model predicting cortical thickness (occipital region model), a model predicting the volume of the lateral ventricle (lateral ventricle model), a model predicting the volume of the hippocampus (hippocampal model), and a model predicting the volume of the extracerebral cerebrospinal fluid region (extracerebral cerebrospinal fluid). model) was built individually.

For example, the frontal model may predict the thickness of the cerebral cortex in the frontal region by receiving the summed number of pixels of the frontal gray matter and clinical information. In addition, the frontal model can predict the thickness of the cerebral cortex in the frontal region by receiving the summed number of pixels of the frontal gray matter, brain size, and clinical information.

For example, the extracerebral cerebrospinal fluid model can estimate the volume of the extracerebral cerebrospinal fluid region by receiving the summed number of pixels and clinical information of the extracerebral cerebrospinal fluid region. In addition, the extracerebral cerebrospinal fluid model can estimate the volume of the extracerebral cerebrospinal fluid region by receiving inputs of the summed number of pixels of the extracerebral cerebrospinal fluid region, brain size, and clinical information.

The researcher verified the performance of the model by constructing a model using brain size as input data and a model not using it. Table 6 below is a model that evaluated the performance of the learning model built by the researcher. The performance indicator is the Pearson Correlation Coefficient. In Table 6, model 1 is a model that does not use brain size, and model 2 is a model that uses even brain size. It can be seen that the performance of model 1 or model 2 is slightly higher depending on the area of interest. Accordingly, Model 1 or Model 2 may be selectively used as a model for predicting cortical thickness and brain region volume according to a region of interest. However, there was no significant difference between model 1 and model 2 in terms of performance.

ROI model 1 model 2 frontal gray matter 0.774 0.779 temporal gray matter 0.786 0.788 parietal gray matter 0.801 0.799 occipital gray matter 0.649 0.638 lateral ventricle 0.880 0.869 hippocampus 0.663 0.667

The researcher verified the performance of the learning model to predict the cortical thickness of the region of interest or the volume of the region of interest. The researcher compared the predicted value based on 2D MRI and the correct answer value measured on 3D MRI using a learning model. The comparison results are shown in Table 7 below. The thickness and volume of the cerebral cortex predicted by the learning model showed a significant correlation with the actual correct answer value. In particular, volume showed a high correlation.

division ROI Correlation (Pearson r) r2 slope cortical thickness frontal gray matter 0.779 0.607 0.394 cortical thickness temporal gray matter 0.788 0.621 0.396 cortical thickness parietal gray matter 0.801 0.641 0.519 cortical thickness occipital gray matter 0.649 0.394 0.285 volume lateral ventricle 0.879 0.754 0.757 volume hippocampus 0.667 0.445 0.266 volume extracerebral cerebrospinal fluid 0.873 0.762 0.608

4 is an example of a learning process 400 of a learning model that calculates dementia information based on the volume of a brain region. A learning model that finally calculates dementia information may be implemented as any one of various machine learning model types. The researcher implemented a model that calculates dementia information with a deep learning model.

The learning of FIG. 4 is based on the premise of acquiring cortical thickness, specific brain region volume, and clinical information for subjects belonging to the population. For example, the learning device may use the volume of a specific brain region for each region of the subject calculated using the learning model of FIG. 3 .

The learning device acquires input data that is learning data (410).

The learning device extracts the volume of a specific brain region extracted from 2D brain MRI of subjects belonging to the population. The volume of the brain region may include the volume of the lateral ventricle, hippocampus, and extracerebral cerebrospinal fluid region. In addition, the learning device acquires clinical information of subjects. Clinical information may include age, gender and education level. The learning data may include the volume of a specific brain region calculated using the learning model of FIG. 3 . Alternatively, the learning data may include the thickness of the cerebral cortex and the volume of a specific brain region calculated from brain images of actual subjects. The researcher used the thickness of the cerebral cortex and the volume of a specific brain region calculated from the 3D MRI analysis results of actual subjects as learning data.

The learning device performs a process of learning a learning model using the extracted input data (420). The learning device repeats the learning process of the learning model using input data extracted from a plurality of subjects.

The learning device inputs the extracted input data to the learning model and performs learning while comparing the value (predicted value) output by the learning model with the correct answer. The learning model is trained to predict whether or not the subject has dementia. The learning model outputs a binary classification result of whether or not the subject has dementia. The correct answer is the information on whether or not the subjects belonging to the population have dementia (see Table 1).

Researchers have built multiple learning models using different input data sets. The researcher used the data of 924 people (80%) of the population as learning data, and the data of 196 people (20%) as verification data. The researcher classified model groups using significantly different ROIs, and individually built models using various input data in each group. A number of learning models built by the researcher are shown in Table 8 below. Table 8 shows the model groups using four different ROIs and the learning data used to build different models in each group. Table 8 below is a model that predicts whether a subject has dementia (dementia or normal). That is, the learning model in FIG. 4 corresponds to a model that binary classifies whether the subject is dementia or normal.

model type Training data (input data)
6ROIs
(F/T/P/O/L/H) learning model 1 Frontal cortical thickness, temporal cortical thickness, parietal cortical thickness, occipital cortical thickness, lateral ventricular volume and hippocampal volume learning model 2 Learning model 1 + clinical information (age, gender) learning model 3 Learning Model 2 + Clinical Information (APOE4)
7ROI
(F/T/P/O/L/H/E) learning model 4 Frontal cortical thickness, temporal cortical thickness, parietal cortical thickness, occipital cortical thickness, lateral ventricle volume, hippocampal volume, and extracerebral cerebrospinal fluid area volume learning model 5 Learning Model 4 + Clinical Information (Age, Gender) learning model 6 Learning Model 5 + Clinical Information (APOE4)
3ROIs
(L/H/E) learning model 7 Volume of the lateral ventricle, volume of the hippocampus and volume of the extracerebral cerebrospinal fluid area learning model 8 Learning Model 7 + Clinical Information (Age, Gender) learning model 9 Learning Model 8 + Clinical Information (APOE4)
4ROIs
(F/T/P/O) Learning Model 10 Frontal cortical thickness, temporal cortical thickness, parietal cortical thickness, and occipital cortical thickness Learning model 11 Learning model 10 + clinical information (age, gender) Learning model 12 Learning Model 11 + Clinical Information (APOE4)

In Table 8, ROI is abbreviated. F is the frontal region, T is the temporal region, P is the parietal region, O is the occipital region, L is the lateral ventricle, H is the hippocampus, and E is the extracerebral cerebrospinal fluid area. (Extracerebral cerebrospinal fluid)).

Meanwhile, unlike Table 8, the learning model may be a model that takes the thickness and/or volume of some (at least one) of the above-mentioned regions of interest as input data. That is, as the learning model, various types of models may be constructed according to which region of interest is to be used.

The researcher verified the built learning model. Meanwhile, in the verification process, the researcher used the cortical thickness and brain region volume predicted by the learning model of FIG. 3 as input data for the learning model of FIG. 4 . Researchers performed 10-fold cross-validation. The verification results are shown in Table 9 below. Verification compared the result predicted by the learning model with the correct answer. As performance indicators, AUC (area under the receiver operating characteristic curve) and AUPRC (area under the precision-recall curve) were used.

The researcher compared the performance of the conventional 3D MRI-based classification model and the aforementioned 2D MRI-based learning model. Conventional 3D MRI-based classification models used previously studied models (Rebsamen, M., et al., Direct cortical thickness estimation using deep learning-based anatomy segmentation and cortex parcellation. Human Brain Mapping, 2020. 41(17) : p. 4804-4814.).

Looking at the results of Table 9, the performance difference between the 3D MRI-based model and the 2D MRI-based model is not large. Therefore, it can be said that the 2D MRI-based model developed by the researcher is also sufficiently significant in predicting dementia. In addition, models using only cortical thickness of regions of interest extracted from 2D MRI (learning models 10 to 12), models using only volumes of regions of interest (learning models (learning models 7 to 9) and cortical thickness and volume of regions of interest) All of the models used (learning models 1 to 6) showed a significant level of performance, and other models showed slightly higher performance than the model using only cortical thickness (learning models 10 to 12).

model type Conventional 3D MRI-based learning model 2D MRI-based learning model AUC AUPRC AUC AUPRC learning model 1 0.938 0.933 0.913 0.883 learning model 2 0.936 0.935 0.922 0.900 learning model 3 0.935 0.945 0.916 0.878 learning model 4 0.928 0.921 0.915 0.889 learning model 5 0.941 0.943 0.918 0.902 learning model 6 0.947 0.949 0.927 0.904 learning model 7 0.909 0.904 0.841 0.754 learning model 8 0.929 0.928 0.901 0.863 learning model 9 0.935 0.935 0.906 0.868 Learning Model 10 0.845 0.872 0.869 0.821 Learning model 11 0.863 0.885 0.877 0.828 Learning model 12 0.892 0.902 0.876 0.825

In addition, the researcher also constructed a model to calculate indicators related to dementia in previous studies. The researcher used the data of 924 people (80%) of the population as learning data, and the data of 196 people (20%) as verification data. The researcher constructed a model that calculates a constant index by receiving the thickness and/or volume of the cerebral cortex for the region of interest calculated using the learning model of FIG. 3 described above. In FIG. 4, the learning model corresponds to a model that calculates a constant index (score).

In a previous study, the W-score corresponds to an index for discriminating brain atrophy information (Renaud La Joie et al., Region-specific hierarchy between atrophy, hypometabolism, and β-amyloid (Aβ) load in Alzheimer's disease dementia, J Neurosci . 2012 Nov 14;32(46):16265-73. Reference). A subject with a W-score of 1.65 or higher can be judged to have brain atrophy. The researcher built a learning model that calculates the W-score. Table 10 below shows the correlation between the W-score of the learning model constructed by the researcher and the W-score calculated using 3D MRI of the same subject. Table 10 compares the results of calculating scores for each area of interest. In Table 10, model 1 uses age and gender as score adjustment variables, and model 2 uses age, gender, and education level as score adjustment variables.

whole area frontal area temporal region parietal region occipital hippocampus lateral ventricle extracerebral cerebrospinal fluid area Model1 0.826 0.759 0.768 0.819 0.639 0.721 0.918 0.774 Model2 0.834 0.762 0.783 0.821 0.651 0.706 0.916 0.764

In previous studies, the AD-score corresponds to an indicator for determining whether or not there is dementia ( Jin San Lee et al., Machine Learning-based Individual Assessment of Cortical Atrophy Pattern in Alzheimer's Disease Spectrum: Development of the Classifier and Longitudinal Evaluation, J Neurosci . SCIENTIFIC REPORTS, Vol.8(1): 4161, 2018. Reference).

The researcher built a learning model that calculates the AD-score using the above-mentioned population data. That is, this corresponds to the case in which the learning model in FIG. 4 calculates the AD-score. The researcher verified the built model using 196 test data. As a result of verification, the learning model that calculates the AD-score showed high performance with AUC = 0.92 and Accuracy = 0.854.

Furthermore, the researcher calculated the AD-score using the learning model that calculates the above-mentioned W-score. The researcher used the learning model to derive AD-scores by using W-scores for the seven areas of interest as input values. The researcher analyzed the correlation between the AD-score calculated using the learning model and the AD-score according to the previous research method. Table 11 below shows the performance of the learning model that calculates the W-score. Model 1 uses age and gender as score adjustment variables, and Model 2 uses age, gender, and education level as score adjustment variables. Looking at Table 11, it can be seen that even when using a model that calculates the W-score, it has a significant correlation with the correct value.

Model Classification PCA AD score association r r2 slope model 1 2 components 0.897 0.805 1.009 3 components 0.895 0.794 1.011 model 2 2 components 0.908 0.825 0.825 3 components 0.911 0.830 0.824

5 is an example of an analysis device 500 that calculates dementia information of a subject using a brain image. The analysis device 500 may be physically implemented in various forms. For example, the analysis device 500 may have a form of a computer device such as a PC, a smart device, a network server, and a chipset dedicated to data processing. Meanwhile, the analysis device 500 may be connected to brain imaging equipment or may be an integral device.

The analysis device 500 may include a storage device 510, a memory 520, an arithmetic device 530, an interface device 540, a communication device 550, and an output device 560.

The storage device 510 may store the 2D MRI of the subject generated by the medical imaging equipment.

The storage device 510 may store clinical information of the subject. Clinical information may include at least one of age, gender, level of education, and APOE4 genotype.

The storage device 510 may store a segmentation model for extracting an ROI from a brain image (2D MRI slice). A segmentation model is a pretrained model.

The storage device 510 may store a learning model that predicts the thickness of the cerebral cortex and/or the volume of a specific brain region based on information of the region of interest and clinical information (age, gender). A learning model predicting cortical thickness and/or volume of a specific brain region is called a first learning model. As described above, the first learning model uses the summed number of pixels of the region of interest (at least one of frontal gray matter, temporal gray matter, parietal gray matter, occipital gray matter, lateral ventricle, hippocampus, and extracerebral cerebrospinal fluid region) and clinical information as input data.

The storage device 510 may store a learning model for calculating dementia information based on the thickness of the cerebral cortex and/or the volume of a specific brain region. Finally, a learning model predicting dementia is called a second learning model. As described above, the second learning model may be implemented in various models according to the type of input data. The second learning model is (i) a model that uses only cortical thickness and volume of specific brain regions as input data, (ii) cortical thickness, volume of specific brain regions, and clinical information (age, gender, and APEO4) as input data. (iii) a model using only cortical thickness in specific regions of interest as input data, (iv) a model using cortical thickness and clinical information (age, gender, and APEO4) in specific regions of interest as input data, (v) It may be any one of a model using only the volume of specific regions of interest as input data, and (vi) a model using the volume and clinical information (age, gender, and APEO4) of specific regions of interest as input data. The second learning model may be any one of the models described in Table 7.

Dementia information may be any one of information such as a prediction result of dementia or not of a subject, a prediction result of dementia high-risk group, dementia-related indicators (brain atrophy information, W-score, AD-score, etc.).

The memory 520 may store data and information generated in the course of the analysis device 500 calculating dementia information from brain images.

The interface device 540 is a device that receives certain commands and data from the outside. The interface device 540 may receive the subject's 2D MRI and clinical information from a physically connected input device or external storage device. The number of input 2D MRI slices is less than 20. The interface device 540 may transmit dementia information predicted based on 2D MRI to an external object.

The communication device 550 refers to a component that receives and transmits certain information through a wired or wireless network. The communication device 550 may receive 2D MRI and clinical information of a subject from an external object. The number of 2D MRI slices is less than 20. The communication device 550 may transmit dementia information predicted based on 2D MRI to an external object such as a user terminal.

The interface device 540 may be a device that internally receives data received from the communication device 550 .

The output device 560 is a device that outputs certain information. The output device 560 may output an interface necessary for a data processing process, a brain image, a region of interest divided from a brain image, an index related to dementia calculated based on a region of interest, dementia information, and the like.

The arithmetic device 530 may pre-process the subject's brain image (2D MRI) at regular intervals. The arithmetic device 530 may generate a mask for segmenting the entire brain region through a data pre-processing process. The computing device 530 may segment the entire brain region from the brain image using a mask for the entire brain region.

The arithmetic unit 530 may normalize the size and resolution of the subject's 2D MRI to a constant.

The computing device 530 may extract a region of interest by inputting the 2D MRI slice of the subject to the learned segmentation model. Regions of interest include frontal gray matter, temporal gray matter, parietal gray matter, occipital gray matter, lateral ventricles, hippocampus and extracerebral cerebrospinal fluid regions.

The computing device 530 may extract size information of the regions of interest. The calculator 530 may calculate the summed number of pixels for each of the regions of interest. The calculation device 530 may calculate the summed number of pixels by summing the number of pixels identified in the entire 2D slice for each region of interest (eg, the frontal part).

The computing device 530 may calculate the size of the entire brain based on the size of the region of interest and other regions in the brain MRI. The brain size may be determined by summing the sizes of the seven regions of interest and the white matter of the regions of interest.

The computing device 530 may predict the thickness and/or volume of the cerebral cortex of the region of interest by inputting the summed number of pixels and clinical information (age and gender) of the region of interest to the first learning model. The computing device 530 may predict the thickness and/or volume of the cerebral cortex of the region of interest by inputting the summed number of pixels and clinical information (age and gender) for each region of interest into an individual learning model.

In addition, the calculation device 530 may predict the thickness and/or volume of the cerebral cortex of the region of interest by inputting the summed number of pixels, brain size, and clinical information (age and gender) of the region of interest to the first learning model. The computing device 530 may predict the thickness or volume of the cerebral cortex of the region of interest by inputting the summed number of pixels, brain size, and clinical information (age and gender) for each region of interest into an individual learning model.

The arithmetic device 530 may input the thickness and/or volume of the cerebral cortex predicted by the first learning model to the second learning model to calculate dementia information of the subject.

For example, the arithmetic unit 530 sets the thickness of the cerebral cortex of the frontal region, the thickness of the cerebral cortex of the temporal region, the thickness of the cerebral cortex of the parietal region, the thickness of the cerebral cortex of the occipital region, the volume of the lateral ventricle, the volume of the hippocampus, and the volume of the extracerebral cerebrospinal fluid region to second. It is possible to calculate the subject's dementia information by inputting it to the learning model.

Alternatively, the arithmetic device 530 may calculate dementia information of the subject by inputting the volume of the lateral ventricle, the volume of the hippocampus, and the volume of the extracerebral cerebrospinal fluid region to the second learning model.

At this time, the computing device 530 may calculate dementia information of the subject by further inputting clinical information (at least one of age, gender, and APOE4 genotype) in addition to the thickness or volume of the cerebral cortex to the second learning model.

The arithmetic device 530 may be a device such as a processor, an AP, or a chip in which a program is embedded that processes data and performs certain arithmetic operations.

In addition, the 2D MRI-based dementia information calculation method and the 2D MRI-based dementia prediction method as described above may be implemented as a program (or application) including an executable algorithm that can be executed on a computer. The program may be stored and provided in a temporary or non-transitory computer readable medium.

A non-transitory readable medium is not a medium that stores data for a short moment, such as a register, cache, or memory, but a medium that stores data semi-permanently and can be read by a device. Specifically, the various applications or programs described above are CD, DVD, hard disk, Blu-ray disk, USB, memory card, ROM (read-only memory), PROM (programmable read only memory), EPROM (Erasable PROM, EPROM) Alternatively, it may be stored and provided in a non-transitory readable medium such as EEPROM (Electrically EPROM) or flash memory.

Temporary readable media include static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and enhanced SDRAM (Enhanced SDRAM). SDRAM, ESDRAM), Synchronous DRAM (Synclink DRAM, SLDRAM) and Direct Rambus RAM (DRRAM).

This embodiment and the drawings accompanying this specification clearly represent only a part of the technical idea included in the foregoing technology, and those skilled in the art can easily understand it within the scope of the technical idea included in the specification and drawings of the above technology. It will be obvious that all variations and specific examples that can be inferred are included in the scope of the above-described technology.

Claims (10)

receiving 2D (dimension) Magnetic Resonance Imaging (MRI) slices of the subject by an analysis device;
extracting regions of interest by inputting the 2D MRI slices to a segmentation model by the analysis device;
predicting a volume of at least one of the regions of interest by the analysis device by inputting pixel information of the regions of interest to a first learning model trained in advance; and
The analysis device inputs the volume of the at least one region into a pre-learned second learning model to calculate dementia information of the subject,
The regions of interest include at least one of frontal gray matter, temporal gray matter, parietal gray matter, occipital gray matter, lateral ventricle, hippocampus, and extracerebral cerebrospinal fluid region. According to claim 1,
The first learning models predict the volume of the at least one region by receiving the sum number of pixels of the regions of interest, the age of the subject, and the gender of the subject Dementia information calculation method using a volume predicted by 2D MRI. According to claim 1,
The first learning model is a method for calculating dementia information using a volume predicted from a two-dimensional MRI including a plurality of learning models prepared in advance for each region of interest. According to claim 1,
The second learning model calculates dementia information of the subject by receiving the volume of the lateral ventricle, the volume of the hippocampus, and the volume of the extracerebral cerebrospinal fluid region, and calculating dementia information using a volume predicted by a two-dimensional MRI. According to claim 1,
The second learning model calculates dementia information using a volume predicted from a 2-dimensional MRI that calculates dementia information of the subject by receiving additional information of at least one of the subject's age, the subject's gender, and the subject's APOE4 genotype. method. An interface device for receiving 2D (dimension) Magnetic Resonance Imaging (MRI) slices of a subject;
a storage device for storing a segmentation model that extracts a region of interest from a 2D MRI slice, a first learning model that predicts a volume by receiving region of interest information, and a second learning model that calculates dementia information; and
The received 2D MRI slices are input to the segmentation model to extract regions of interest, and pixel information of the extracted regions of interest is input to the first learning model to predict a volume of at least one region among the regions of interest; An arithmetic device for calculating dementia information of the subject by inputting the volume of the at least one region to the second learning model,
The regions of interest include at least one of frontal gray matter, temporal gray matter, parietal gray matter, occipital gray matter, lateral ventricle, hippocampus, and extracerebral cerebrospinal fluid region, using a volume predicted by a two-dimensional MRI analysis device for calculating dementia information. According to claim 6
The first learning models calculate dementia information using the predicted volume from 2D MRI predicting the volume of the at least one region by receiving inputs of the sum of pixels of the regions of interest, the age of the target and the gender of the target. analysis device. According to claim 6
The first learning model is an analysis device for calculating dementia information using a volume predicted from a two-dimensional MRI including a plurality of learning models prepared in advance for each region of interest. According to claim 6
The second learning model receives the volume of the lateral ventricle, the volume of the hippocampus, and the volume of the extracerebral cerebrospinal fluid region and calculates the dementia information using the volume predicted by the two-dimensional MRI that calculates the dementia information of the subject Analysis device. According to claim 6
The second learning model further receives information on at least one of the subject's age, the subject's gender, and the subject's APOE4 genotype, and uses the volume predicted by the 2D MRI to calculate the subject's dementia information. An analysis device that calculates

Images