KR102532851B1 - Pet image analyzing apparatus based on time series for determining brain disease, and method thereof - Google Patents

Abstract

The present invention relates to a time-series-based PET image analysis device and method for determining brain disease, wherein the tomography is performed for a predetermined time according to a PET technique for taking tomography of the brain of a subject to whom an amyloid agent is administered, so that continuous tomography is performed. An image acquisition unit that acquires a PET image composed of a plurality of tomography images captured by , and time information obtained from each of the plurality of tomography images and a score value corresponding to a pixel combination for each region of the tomography image An input layer unit including a time-series vector generation module for generating a predetermined image vector based on the time-series vector, and a plurality of input layers for receiving at least one image vector sequentially one by one based on the time-series information and classifying and storing the image vectors according to input time points; , A predetermined state value is generated based on the input value of the input layer corresponding to each input time point, and the generated state value corresponds to the input value of the input layer corresponding to the current time point and a time point immediately before the current time point. A hidden layer unit including a plurality of hidden layers generating state values for the current time point based on state values, and a state value of the hidden layer corresponding to the input layer having the last input time point, It is characterized in that it includes an output layer unit generating a predetermined output value for an immediately subsequent time point, and a probability calculation unit calculating a probability value corresponding to at least one class among a plurality of preset classes based on the output value.
Accordingly, there is an effect of deriving prediction results with high accuracy and sensitivity comparable to dementia determination using the entire PET image using only PET images captured for a short period of time by analyzing time-series data over time.

Description

Time-series-based PET image analysis device and method for determining brain disease

The present invention applies a neural network module that predicts near-future information based on past information using time-series data over time to calculate predictive data on the degree of response to an amyloid agent, time-series-based PET image analysis for brain disease determination device and its method.

Among degenerative brain diseases, especially degenerative brain diseases that show a positive reaction to amyloid, symptoms appear only after the condition deteriorates, and the symptoms are also difficult to accurately determine, so it is difficult to detect them early.

Amyloid-positive degenerative brain disease is a representative brain disease caused by damage to brain cells. It mainly occurs in old age. Unlike other cells, brain cells do not regenerate, so prevention or early detection and treatment is the safest and most efficient way. It is known as a way to treat.

Recently, with the development of imaging medical examination methods such as tomography devices such as ultrasound, CT, and MRI, the importance of devices for diagnosing diseases of the human body through medical image analysis and processing technology has been expanded. It is necessary to make a consistent and accurate diagnosis of dementia through the captured images.

On the other hand, among the test methods for obtaining such a brain tomography image, positron emission tomography is obtained by administering radiopharmaceuticals that emit positrons into the body to obtain information on the active state of biochemical and physiological metabolism occurring in various tissues inside the human body ( PET) technique is the most widely used test method.

However, in order to perform a PET scan, continuous imaging is required while maintaining a fine motion state for the time during which the radiopharmaceutical administered to the examiner is absorbed into the body, for example, an average of about 1 hour. There is a problem in that it is difficult to proceed with the examination because the burden on the shooting itself is large.

KR 10-2016-0053220 A KR 10-2170968 B1

An object of the present invention is to solve the above problems, and by applying a neural network module that predicts near future information based on past information using time series data over time, the brain that calculates predictive data for the degree of response of an amyloid agent An object of the present invention is to provide a time series-based PET image analysis device and method for disease determination.

Time-series-based PET image analysis device for determining brain disease according to an aspect of the present invention for achieving the above object is the tomography for a predetermined time according to the PET technique of tomographically imaging the brain of a subject to whom an amyloid agent is administered. An image acquisition unit that acquires a PET image composed of a plurality of tomography images taken consecutively as imaging is performed, and time series information corresponding to time information obtained from each of the plurality of tomography images and the tomography image A time-series vector generation module that generates a predetermined image vector based on a score value corresponding to a pixel combination for each region for each region, and sequentially receives at least one image vector based on the time-series information and stores the image vectors separately for each input time point. An input layer unit including a plurality of input layers, and a predetermined state value based on an input value of the input layer corresponding to each input time point is generated, and the input value of the input layer corresponding to the current time point and the current A hidden layer unit including a plurality of hidden layers for generating a state value for the current time point based on the state value generated corresponding to a time point immediately before the point of view, and the input layer corresponding to the input layer having the last input time point An output layer unit generating a predetermined output value for the current time point based on the state value of the hidden layer, and a probability calculation unit calculating a probability value corresponding to at least one class among a plurality of preset classes based on the output value characterized by

In addition, a time-series-based PET image analysis method according to another aspect of the present invention for achieving the above object is a PET technique for taking tomography of the brain of a subject to whom an amyloid agent is administered, and the tomography is performed for a predetermined time. Acquiring a PET image composed of a plurality of tomography images sequentially taken as performed, time-series information corresponding to time information obtained from each of the plurality of tomography images, and pixels for each region of the tomography image generating a predetermined image vector based on a score value corresponding to the combination; receiving at least one image vector sequentially one by one based on the time-series information, classifying them according to input time points, and storing them as a plurality of input values; , Generates a predetermined state value based on an input value corresponding to each input time point, based on the input value corresponding to the current time point and the state value generated corresponding to a time point immediately before the current time point. generating a state value for a point in time; generating a predetermined output value for the current point in time based on the state value corresponding to the input value at which the input point in time is the last; and based on the output value, a predetermined output value It is characterized in that it includes calculating a probability value corresponding to at least one class among a plurality of classes.

According to the present invention, by analyzing time-series data over time, only PET images taken for a short time corresponding to the initial section derive predictive results with high accuracy and sensitivity comparable to dementia determination using the entire conventional PET image. can do.

In addition, according to the present invention, the total examination time according to the PET technique for brain disease determination can be shortened, thereby reducing the burden on the examiner of continuous imaging in a fine motion state.

In addition, according to the present invention, unlike conventional methods in which an optimal section representing an optimal metabolic state for an initial amyloid image is previously set or manually extracted in image frame units, feature vectors for each time point are analyzed through time-sequential data analysis. By collecting and acquiring information corresponding to the optimal section from the fields, it is possible to derive highly accurate prediction results in a short time.

1 is a diagram schematically showing the configuration of a time-series-based PET image analysis device for brain disease determination according to an embodiment of the present invention;
FIG. 2 is a block diagram showing in detail the configuration of a time-series-based PET image analysis device for brain disease determination according to an embodiment of the present invention of FIG. 1;
3 is a diagram showing an embodiment of the recurrent neural network module of FIG. 1;
4 is a diagram showing another embodiment of the recurrent neural network module of FIG. 1;
5 is a diagram showing an example of a result graph for each of a normal person and an Alzheimer's disease patient based on PET images acquired through the image acquisition unit of FIG. 1;
6 is a diagram showing an example of spatially normalized volume data in the spatial normalization unit of FIG. 2;
7 is a diagram showing an example of a pixel combination for each region detected by the region detection unit of FIG. 2;
FIG. 8 is a diagram showing an example of comparison data in which a tomographic image obtained by the image acquisition unit of FIG. 2 and a normalized image by the image preprocessing unit of FIG. 2 are classified by time period for a normal person and an Alzheimer's disease patient, respectively;
9 is a graph showing an example of a distribution state of feature vectors for output values generated by the output layer unit of FIG. 2 for each of a normal person and an Alzheimer's disease patient according to amyloid response;
10 is a flowchart schematically illustrating a time-series-based PET image analysis method for determining brain disease according to an embodiment of the present invention;
FIG. 11 is a flowchart illustrating the image vector generation step of FIG. 10 in more detail.

The specific details, including the problems to be solved, the solutions to the problems, and the effect of the invention for the present invention as described above are included in the embodiments and drawings to be described below. Advantages and features of the present invention, and methods of achieving them, will become clear with reference to the detailed description of the following embodiments taken in conjunction with the accompanying drawings. Like reference numbers designate like elements throughout the specification.

1 is a diagram schematically showing the configuration of a time-series-based PET image analysis device for determining brain disease according to an embodiment of the present invention, and FIG. 2 is a view showing brain disease determination according to an embodiment of the present invention in FIG. FIG. 3 is a diagram showing an embodiment of the recurrent neural network module of FIG. 1, and FIG. 4 is another embodiment of the recurrent neural network module of FIG. 1. FIG. 5 is a diagram showing an example of a result graph for each of a normal person and an Alzheimer's disease patient based on a PET image acquired through the image acquisition unit of FIG. 1 , and FIG. FIG. 7 is a diagram showing an example of a pixel combination for each region detected by the region detection unit of FIG. 2 .

Hereinafter, a time-series-based PET image analysis device for brain disease determination according to a preferred embodiment of the present invention will be described with reference to the above-described drawings.

1 and 2, the time-series-based PET image analysis device for brain disease determination according to the present invention largely includes an image acquisition unit 100, a time-series vector generation module 200, a recurrent neural network module, and a probability calculation unit ( 400).

The image acquisition unit 100 serves to acquire a plurality of tomographic images I ₀ according to tomography of the brain of a subject to be diagnosed.

Here, the image acquisition unit 100 is a position emission tomography (PET) technique of tomographically imaging the brain of a diagnosis subject to whom an amyloid agent based on a radioactive compound containing an amyloid PET radiotracer has been administered. According to this, as the tomography is performed for a predetermined time, a PET image composed of a plurality of tomography images (I ₀ ) continuously photographed may be obtained.

Such a tomography image (I ₀ ) is displayed with a border formed around a portion corresponding to a brain region as a boundary, and when a disease exists in a predetermined brain region, pixels of a cell or tissue portion belonging to the corresponding brain region are amyloid As a positive reaction by the deposition of (Brain Amyloid Plague Load), it may have a color of a pixel part appearing gray or white in contrast to a black part, that is, a color close to white.

At this time, the PET image is for measuring the metabolic amount or the absolute value of a specific substance through the accumulation and removal rate of the radiopharmaceutical over time in the brain region of the subject to whom the radiopharmaceutical was administered, and is generally brain disease. In order to determine, on average, based on the absorption time in the body, which is from immediately after the injection of the radiopharmaceutical to the point at which the radiopharmaceutical is removed, a dynamic image must be acquired through continuous shooting for about 60 to 90 minutes. prerequisites are required.

In this regard, as shown in FIG. 5, generally for an early period (eFBB; early phase FBB) from immediately after the injection of the radiopharmaceutical to a predetermined first time point, for example, for a period from the injection to 20 minutes. The resulting data 51 based on the PET image is used to find blood perfusion information, and is used to determine the later period (dFBB) from the second time point after the first time point to the time point at which the radiopharmaceutical is removed, e.g., after the injection. The resultant data 52 for the period from 60 minutes to 90 minutes has been used to determine whether or not a positive reaction is caused by the amyloid deposition described above.

Therefore, in the present invention, by using the characteristics of the time-series vector generation module 200 and the recurrent neural network module (Mn), which will be described later, the time-series data flow of some PET images captured in the initial section is identified, and dementia using the entire conventional PET image is used. It proposes a way to make predictions comparable to judgment.

The time-series vector generation module 200 is a score value corresponding to the pixel combination for each region of the time information and the tomographic image I ₀ obtained by the image acquisition unit 100, respectively, of the plurality of tomographic images I ₀ . Generates a predetermined image vector (V _I ) based on .

Here, the time-series vector generation module 200 may include a time-series information generation unit 220, an image pre-processing unit 240, and a vector conversion unit 260 as shown in FIG.

The time series information generator 220 generates predetermined time series information Ts for the plurality of tomographic images I ₀ .

Specifically, the time-series information generation unit 220 detects the time information Ti obtained from each of the plurality of tomographic images I ₀ by the image acquisition unit 100 and arranges them according to the order of time. Information (Oi) is generated, and time-series information ( about a plurality of tomographic images I ₀ ) based on the time information (Ti) and the sequence information (Oi) corresponding to each tomographic image (I ₀ ) Ts) can be created.

The image pre-processing unit 240 normalizes the plurality of tomographic images (I ₀ ) obtained by the image acquisition unit 100 according to a predetermined image processing technique and converts them into normalized images (I ₁ ) in a 2D array format. play a role

Here, the image pre-processing unit 240 may include a spatial normalization unit 242 and an intensity normalization unit 244 as shown in FIG. 2 .

The spatial normalization unit 242 spatially normalizes the plurality of tomographic images I ₀ with preset volume data Dv.

Here, the space normalizer 242 may spatially normalize a volume (width x length x height) of each of a plurality of brain regions included in the tomography image I ₀ to have a constant size.

The intensity normalizer 244 scales each value of a plurality of voxels constituting the volume data Dv spatially normalized by the spatial normalizer 242 at a predetermined ratio according to a preset scale factor.

At this time, the scale factor is a value representing the average image intensity of the cerebellum, which is one of a plurality of brain regions included in the tomographic image I ₀ , and can be summarized as in Equation 1 below. there is.

Here, C _i,ref represents a voxel value in the cerebellar region, and N _i,ref represents the number of voxels.

Although not shown in FIG. 2 described above, in the case of the time-series-based PET image analysis device for brain disease determination according to the present invention, based on the PET image from the user, it is a subject of judgment as to whether or not the patient is diagnosed with a brain disease. It may further include a configuration of a region selection unit (not shown) for selecting and receiving brain regions. In this case, the intensity normalization unit 244 allows the user to select a plurality of brain regions included in the tomographic image I ₀ . When any one of the regions is selected, a relative value based on the reference region is obtained by dividing a plurality of voxel values constituting the volume data Dv corresponding to the selected region by a preset scale factor to correspond to the predetermined reference region. You can perform scaling to convert to .

For example, if the reference region is the 'cerebellum', the scaling result value by the intensity normalizer 244 is '1' when the region selected by the user is the cerebellum region, and is based on the cerebellum region for other regions. converted to a relative value.

The vector converter 260 generates a predetermined image vector V based on the normalized image I ₁ converted by the image preprocessor 240 and the time series information Ts generated by the time series information generator 220. _I ) plays a role in generating.

Specifically, the vector converter 260 calculates a score value corresponding to a pixel combination for each region of the normalized image I ₁ , and then calculates a score value based on the calculated score value and the generated time series information Ts. Can be converted into an image vector (V _I ) of

Here, the vector converter 260 may include a region detector 262 , a representative value calculator 264 , a score calculator 266 and a matrix generator 268 as shown in FIG. 2 .

The region detection unit 262 classifies each normalized image I ₁ into a plurality of regions corresponding to a region of interest based on a preset SUV technique and detects a pixel combination for each region.

Here, the region detection unit 262 has a standardized uptake coefficient (SUV) based on a plurality of regions of interest (ROIs) preset to correspond to biological or physical factors with respect to the normalized image I ₁ . values), pixel combinations for each region of a plurality of regions corresponding to each of a plurality of regions of interest (ROIs) may be detected.

At this time, the standard uptake coefficient (SUV) is a kind of method for evaluating how much higher than the average uptake is in the tumor under the assumption that the radiopharmaceutical injected into the human body is evenly spread. represents a semi-quantitative indicator for

The representative value calculating unit 264 calculates a representative value for each pixel combination detected by the area detecting unit 262 .

Here, the representative value calculation unit 264 may calculate an average value as the representative value by adding all measured values based on the standard consumption coefficient (SUV) for each of a plurality of pixels constituting the pixel combination corresponding to each region. there is.

The score calculation unit 266 calculates a score value corresponding to the brightness coordinate value based on a result of applying the representative value calculated by the representative value calculation unit 264 to a preset color model.

In this case, the color model may include an HSV color model having spatial coordinate values (h, s, v) for three components of hue, saturation, and value.

The matrix generator 268 generates matrix data for a predetermined image vector V _I based on the score value calculated by the score calculator 266 and the time series information Ts generated by the time series information generator 220. generate

Here, the matrix generator 268 forms a 2D array having the score value as a value for a row and the time information (Ti) obtained for the corresponding image included in the time series information as a value for a column. Matrix data corresponding to a vector having an array of scalar structure may be generated.

The recurrent neural network module Mn is an artificial neural network for learning time-series data that changes over time. As shown in FIG. 1, the input layer unit 300, the hidden layer unit 400, and the output layer unit 500 ) may be included.

The input layer unit 300 sequentially receives at least one image vector (V _I ) generated by the matrix generator 268 one by one based on the time-series information (Ts), and classifies and stores the plurality of inputs for each input time point. layers (X ₁ ,X ₂ ,...,X _t ).

The hidden layer unit 400 generates a predetermined state value (h) based on the input value of the input layer corresponding to each input time point, and the input value of the input layer (X _t ) corresponding to the current time point (t) A plurality of hidden layers (A) generating a state value (h t ) for the current time point (t) based on a state value (h _t-1 ) generated corresponding to a point in time (t- ₁ ) immediately preceding the current point in time. includes

Here, as shown in FIG. 3, the hidden layer unit 400 has a cell structure including a hidden state therein, and generates a predetermined state value (h) according to a pre-stored activation function and stores it in the hidden state. It may be a recurrent neural network structure including two hidden layers (A), for example, a recurrent neural network (RNN).

In this case, in the hidden layer unit 400, a plurality of hidden layers (A) and a plurality of input layers (X) are respectively connected through input nodes provided to correspond to the number of the plurality of input layers, and mutually connect through a predetermined connection node. As at least one pair of adjacent hidden layers (A) are interconnected in series, corresponding to the input value of the input layer (X _t ) corresponding to the current time point (t) and the time point (t-1) immediately preceding the current time point Based on a result of applying the previously stored state value (h _t-1 ) to the activation function, a state value (h _t ) corresponding to the current point in time (t) may be generated.

Specifically, the hidden layer unit 400 generates a first value obtained by multiplying an input value of the input layer (X _t ) corresponding to the current time point (t) by a preset input weight (W _x ), and a time point immediately preceding the current time point ( After obtaining a second value obtained by multiplying a pre-stored state value (h _t-1 ) and a preset hidden weight (W _h ) corresponding to t-1), the sum of the first value and the second value Based on the result of applying the activation function, a state value (h _t ) corresponding to the current point in time (t) can be generated, and the input weight (W _x ) and hidden weight (W _h ) are all the same parameters for each layer ( weight) may have the same value depending on the characteristics of the recurrent neural network that shares it.

At this time, the activation function may include a hyperbolic tangent (tanh) function, which is a nonlinear function, and in this case, the formula for the state value (h _t ) corresponding to the current point in time (t) is summarized in Equation 2 below Same as

Here, W _x is an input weight, W _h is a hidden weight, and b is a predetermined bias value.

On the other hand, as shown in FIG. 4, the hidden layer unit 400 is provided with a cell structure further including a cell state S2 therein in addition to the aforementioned hidden state S1, and a predetermined state value according to the activation function ( h _t ) may be a recurrent neural network structure including a plurality of hidden layers (A) generating and storing the hidden state (S1), for example, LSTM (Long Short-Term Memory).

In this case, the hidden layer unit 400 generates input data based on the input value of the input layer (X _t ) corresponding to the current time point (t) and a state value (h corresponding to a point in time (t-1) immediately before the current point in time. Based on the result of applying the preset reflection factor corresponding to each of the past data based on _t-1 ), the cell state value (C _t ) of the hidden layer (A) corresponding to the current time point (t) is generated to generate the cell state (S2) , and a previously stored cell state value (C _t ) is applied to the activation function to generate a state value (h _t ) corresponding to the current point in time (t).

In addition, as shown in FIG. 4, the hidden layer unit 400 forms a connection relationship between the hidden state S1 and the cell state S2 inside the hidden layer A, and the input layer corresponding to the current time point t After generating a predetermined result value based on the input value of (X _t ) and the pre-stored state value (h _t- 1 ) corresponding to the time point (t-1) immediately preceding the current point in time, A deletion gate 410, an input gate 420, and an output gate 430 that are reflected in the cell state value (C _t ) or state value (h _t ) may be further included.

The deletion gate 410 generates a first result value by applying a preset first reflection factor based on the degree of deletion of the past data h _{t-1 and} transfers it to the cell state S2.

Here, the deletion gate 410, as _shown in Equation 3 below, has a weight (W After reflecting _f ), a first result value (f _t ) between '0' and '1' can be generated by applying the sum of all of them to a sigmoid function corresponding to the first reflection factor. .

At this time, the closer the first result value (f _t ) is to '0', the more past information is deleted and forgotten, and the closer to '1', the past information may be completely remembered.

The input gate 420 generates a gate value (i _t ) by applying a predetermined second reflection factor based on the storage degree of the input data (X _t ), the input data (X _t ), and the past data (h _{t )} . _-1 ) is applied to the activation function, and a second result value is generated by performing a preset operation on the candidate state value (g _t ) and transmitted to the cell state (S2).

Here, the input gate 420, as _shown in Equation 4 below, has a predetermined weight ( A gate value (i _t ) between '0' and '1' may be generated by applying the sum of W _i ) to a sigmoid function corresponding to the second reflection factor.

In addition, the input gate 420 reflects a preset weight (W _g ) to each of the input value at the current time point (t) and the state value (h t-1) at the time point ( _t-1 ) immediately before the current time point, and then A candidate state value (g _t ) may be generated by applying the sum of all values to the activation function.

In this case, the activation function may include a hyperbolic tangent (tanh) function, which is a nonlinear function, and the formula for the candidate state value (g _t ) in this case is summarized as Equation 5 below.

In addition, as shown in FIG. 4, the input gate 420 generates a second value obtained as a result of element-by-element multiplication (Hadamard Product) of the gate value (i _t ) and the candidate state value (g _t ). 2 The resulting value (i _t · g _t ) is finally transferred to the cell state (S2).

In this case, the cell state value Ct at the current point in time t is generated and stored in the cell state S2 based on the values transferred from each of the above-described deletion gate 410 and input gate 420 .

Specifically, the first result value (f _t ) delivered from the deletion gate 410 is multiplied ( ) by the cell state value (C _t -1 ) of the time point (t-1) immediately before the current point in time, and then , The value calculated as the result of adding the value (i _t g _t ) transferred from the input gate 420 can be stored as the cell state value (C _t ) at the current point in time (t). It is the same as Equation 6 of

The output gate 430 generates a third result value (o _t ) by applying a preset third reflection factor based on the output level of the cell state value (C _t ) at the current point in time (t).

Here, the output gate 430, as _shown in Equation 7 below, has a predetermined weight ( W _o ) is reflected and then the sum of all of them is applied to the sigmoid function corresponding to the third reflection factor to generate a third result value (o _t ) between '0' and '1' there is.

Accordingly, the hidden layer unit 400 calculates the cell state value (C _t ) and the third resultant value (o _t ) at the current time point (t), based on the result of applying the activation function to the current time point (t). It may be to generate a corresponding state value (h _t ).

At this time, the activation function may include a hyperbolic tangent (tanh) function, which is a nonlinear function. In this case, the formula for the state value (h _t ) corresponding to the current point in time (t) is summarized in Equation 8 below and same.

The output layer unit 500 generates a predetermined output value (y _t ) for the current time point (t) based on the state value (h t ) of the hidden layer (A) corresponding to the input layer (X _t ) at which the input time point is _the last. generate

Here, the output layer unit 500 is connected to each of the plurality of hidden layers (A) through a predetermined output node, and the hidden _layer ( A value in which a predetermined output weight (Wy) is reflected in the state value (h _t ) of A) may be generated and stored as an output value (y _t ).

The probability calculation unit 400 calculates a probability value corresponding to at least one class among a plurality of preset classes based on the output value of the output layer unit 500 .

At this time, the plurality of classes are composed of two types of Alzheimer's Disease (AD) and Cognitively Normal (CN), or amyloid-positive Alzheimer's disease group (Aβ(+)AD), amyloid-negative normal group ( Aβ(+)CN), amyloid-negative Alzheimer's disease group (Aβ(-)AD), and amyloid-positive normal group (Aβ(+)CN).

As described above, according to the time-series-based PET image analysis device for brain disease determination according to the present invention, it has the feature of being able to analyze based on the time-sequential data flow of PET images taken for a predetermined time. The result of comparing the result of performing analysis on a part of the whole PET image of the diagnosis subject using this method and the conventional analysis result is as follows.

In this regard, FIG. 8 shows tomographic images acquired by the image acquisition unit 100 and normalized images by the image pre-processing unit 240 for a normal person and an Alzheimer's disease patient, respectively, in an initial section (eFBB) and a later section ( dFBB) is a diagram showing an example of comparison data, and FIG. 9 is a distribution state of feature vectors for output values generated by the output layer unit 500 corresponding to the initial section (eFBB) and the later section (dFBB), respectively. is a graph showing

In addition, Table 1 below shows probability values calculated by the probability calculation unit 400 according to the present invention described above for PET images corresponding to the initial period (eFBB) and the late period (dFBB), respectively, as shown in FIG. 8 . It summarizes the accuracy, specificity, sensitivity, and ROC curve (AUROC) based on .

At this time, the initial interval is an interval from immediately after the injection to a predetermined first time point, for example, from the injection to 20 minutes, and the later interval is the radiopharmaceutical from the second time point after the first time point. As an interval up to the point of removal, for example, an interval from 60 minutes to 90 minutes after the injection may be indicated.

Ac. Phase Accuracy (%) Specificity (%) Sensitivity (%) AUROC (%) eFBB 78.182 67.872 84.191 0.846 dFBB 81.364 72.250 86.138 0.846

For example, referring to Table 1, the accuracy, sensitivity, and specificity of the analysis results using data corresponding to the early period (eFBB) compared to the analysis results by the conventional late period (dFBB) are '3.182 (%)', respectively. ','1.947 (%)','4.378 (%)', in that it has a small error range of around 5%, and in the case of the ROC curve based on sensitivity and specificity, it has the same value as the conventional one, the initial interval It can be confirmed that prediction comparable to dementia determination using the entire conventional PET image can be performed by data analysis based on a time-sequential data flow according to the present invention using only some of the PET images captured in the present invention.

In addition, Table 2 below collects the input data based on the PET image into the normal group (CN) and the Alzheimer's disease group (AD) according to the amyloid distribution (Aβ), and the early period (eFBB-LSTM) and late period (dFBB-SVM) ) shows the results of classifying the entire normal group and the Alzheimer's disease group using the amyloid PET image shown in ), which has results corresponding to some extent with the graph shown in FIG. 9.

Pop./Ac.Phase (Model) AUROC (%) Aβ(-)CN vs Aβ(-)AD Early FBB (LSTM) 0.843 (0.048) Delay FBB (SVM) 0.790 (0.055) Aβ(+)CN vs Aβ(+)AD Early FBB (LSTM) 0.867 (0.037) Delay FBB (SVM) 0.765 (0.075)

For example, referring to Table 2, in the case of ROC curve values for each of the comparative analysis results of the normal group and Alzheimer's disease group showing amyloid-negative reaction and the comparison analysis result of normal group and Alzheimer's disease group showing amyloid-positive reaction, Compared to the conventional SVM (Support Vector Machine) technique using PET image data of the interval (dFBB), the analysis method (LSTM) according to the present invention using the PET image data of the initial interval (eFBB) shows '0.843 (%)' and '0.867' (%)' to confirm that it is higher.

Accordingly, amyloid PET of the later phase, which was used to observe the deposition of existing amyloid plaques, shows better classification performance under the condition that it cannot classify Alzheimer's disease when similar amyloid uptake is shown. can prove that it can.

On the other hand, Table 3 below shows prediction results by various conventional regression classification techniques (LR, SVM) using data corresponding to the initial interval (eFBB) and time series-based PET image analysis for brain disease determination according to the present invention. Based on the prediction result by the method (LSTM), each of the accuracy, specificity, sensitivity, and ROC curve (AUROC) is compared and summarized.

In this regard, in general, in the case of PET image data for an initial period (eFBB) for a predetermined time from immediately after administration of an amyloid agent, for example, a period from the injection to 20 minutes, an important indicator for finding blood perfusion information has been used as

In the conventional regression classification techniques (LR, SVM) described in Table 3 below, the initial period (eFBB) is taken and the optimal period (2 to 7 minutes) representing the optimal metabolic state that most reflects cerebral blood flow is defined in advance. and analysis using images averaged on the time axis within a given time frame range, whereas the time-series-based PET image analysis technique (LSTM) for brain disease determination according to the present invention is an initial amyloid PET image ( 0 to 20 minutes), there is a difference in the method of obtaining results even though the same input data is used in that it is a method of deriving information corresponding to the optimal section by collecting feature vectors for each time point through time-series analysis. there is.

In this case, the LR represents Logistic Regression (LR), and the SVM represents Support Vector Machine (SVM).

Model Accuracy (%) Specificity (%) Sensitivity (%) AUROC (%) LR 74.546 61.278 82.066 0.843 SVM 76.364 59.108 85.647 0.828 LSTM 78.182 67.872 84.191 0.846

For example, referring to Table 3, the accuracy and specificity of the analysis result by the analysis technique (LSTM) according to the present invention compared to the analysis result by the conventional regression classification technique (LR, SVM) are '78.182 (%), respectively. )' and '67.872 (%)', and even in the case of the ROC curve based on sensitivity and specificity, the highest at '0.846 (%)', the technique according to the present invention is further improved compared to the conventional technique. It can be demonstrated that it is an early amyloid PET image quantification method that shows performance.

10 is a flowchart schematically illustrating a time-series-based PET image analysis method for determining a brain disease according to an embodiment of the present invention, and FIG. 11 is a flowchart illustrating the image vector generation step of FIG. 10 in more detail.

Hereinafter, a PET image analysis method based on time series for brain disease determination according to an embodiment of the present invention will be described with reference to the above-described drawings and FIGS. 10 and 11 .

First, a PET image consisting of a plurality of tomographic images (I ₀ ) taken sequentially is obtained by performing the tomography for a predetermined time according to a PET technique for taking tomography of the brain of a subject to whom an amyloid agent has been administered. (S100).

Here, the step S100 is performed according to a position emission tomography (PET) technique of tomography of the brain of a subject to whom an amyloid preparation based on a radioactive compound containing an amyloid PET radiotracer is administered. As the tomography is performed for a predetermined time, a PET image composed of a plurality of tomography images (I ₀ ) continuously photographed may be obtained.

Such a tomography image (I ₀ ) is displayed with a border formed around a portion corresponding to a brain region as a boundary, and when a disease exists in a predetermined brain region, pixels of a cell or tissue portion belonging to the corresponding brain region are amyloid As a positive reaction by the deposition of (Brain Amyloid Plague Load), it may have a color of a pixel part appearing gray or white in contrast to a black part, that is, a color close to white.

Next, a predetermined image vector based on time-series information corresponding to the time information obtained by each of the plurality of tomographic images in step 100 and a score value corresponding to a pixel combination for each region of the tomographic image I ₀ . V _I ) is generated (S200).

Here, the step S200 may include time series information generation step S210, normalization conversion step S220, and image vector conversion step S230, as shown in FIG.

In step S210 of generating time series information, predetermined time series information Ts for the plurality of tomographic images I ₀ is generated.

Specifically, in step S210, order information Oi for each image is generated by detecting time information Ti obtained from each of the plurality of tomographic images I ₀ in step S100 and arranging them in chronological order, , Based on the time information Ti and the sequence information Oi corresponding to each tomographic image I ₀ , time series information Ts for a plurality of tomographic images I ₀ may be generated. .

The normalization conversion step (S220) normalizes the plurality of tomographic images (I ₀ ) obtained in step S100 according to a preset image processing technique and converts them into normalized images (I ₁ ) in a 2D array format. and includes a spatial normalization step (S222) and an intensity normalization step (S224).

In the spatial normalization step ( S222 ), the plurality of tomographic images I ₀ are spatially normalized with preset volume data Dv.

Here, in step S222, spatial normalization may be performed so that the volume (width x length x height) of each of the plurality of brain regions included in the tomography image I ₀ has a constant size.

In the intensity normalization step (S224), each value of a plurality of voxels constituting the volume data Dv spatially normalized by the step S222 is scaled at a predetermined ratio according to a preset scale factor.

At this time, the scale factor is a value representing the average image intensity of the cerebellum, which is one of a plurality of brain regions included in the tomographic image I ₀ , and can be summarized as in Equation 1 above. .

Although not shown in FIG. 11, in the case of the time-series-based PET image analysis method for brain disease determination according to the present invention, based on the PET image from the user, the subject of diagnosis A step (not shown) of receiving a selection input of a brain region may be further included between the spatial normalization step (S222) and the intensity normalization step (S224), and in the added selection input step, the user selects the tomography image (I ₀ ) When any one of the plurality of brain regions included in is selected, the plurality of voxel values constituting the volume data Dv corresponding to the selected region are divided by a preset scale factor to correspond to the predetermined reference region, and the reference region It is possible to perform scaling that converts to a relative value based on .

For example, if the reference region is the 'cerebellum', the scaling result value by the intensity normalization step (S224) is '1' when the user-selected region is the cerebellum region, and other regions are based on the cerebellum region. converted to a relative value.

The image vector conversion step (S230) is a predetermined image vector (based on the normalized image (I ₁ ) converted by the normalization conversion step (S220) and the time series information (Ts) generated by the time series information generation step (S210). It plays a role in generating V _I ).

Specifically, the image vector conversion step (S230) calculates a score value corresponding to a pixel combination for each region of the normalized image I ₁ , and then calculates a score value based on the calculated score value and the generated time series information Ts. It can be converted into a predetermined image vector (V _I ).

Here, the image vector conversion step (S230) may include a region detection step (S262), a representative value calculation step (S234), a score calculation step (S236), and a matrix generation step (S238) as shown in FIG. 11. .

In the region detection step S262, each of the normalized images I ₁ is divided into a plurality of regions corresponding to a region of interest based on a preset SUV technique, and a pixel combination for each region is detected.

Here, the region detection step (S262) is a standardized uptake coefficient (SUV) based on a plurality of regions of interest (ROI) preset to correspond to biological or physical factors with respect to the normalized image (I ₁ ). A pixel combination for each region of a plurality of regions corresponding to each of a plurality of regions of interest (ROIs) may be detected according to a technique for measuring uptake values.

In the representative value calculation step (S234), a representative value for each pixel combination detected by the region detection step (S262) is calculated.

Here, in the representative value calculation step (S234), the average value obtained by adding all measured values based on the standard consumption coefficient (SUV) for each of a plurality of pixels constituting the pixel combination corresponding to each region may be calculated as the representative value. there is.

In the score calculation step (S236), a score value corresponding to the lightness coordinate value is calculated based on a result of applying the representative value calculated in the representative value calculation step (S234) to a preset color model.

In this case, the color model may include an HSV color model having spatial coordinate values (h, s, v) for three components of hue, saturation, and value.

The matrix generation step (S238) is the matrix data for a predetermined image vector (V _I ) based on the score value calculated by the score calculation step (S236) and the time series information (Ts) generated by the time series information generation step (S210). generate

Here, in the matrix generation step (S238), a 2D array having the score value as a value for a row and the time information (Ti) obtained for the image included in the time series information as a value for a column Matrix data corresponding to a vector having an array of scalar structure may be generated.

Next, at least one image vector (V _I ) is sequentially input one by one based on the time series information (Ts), and divided by input time point into a plurality of input values (X ₁ ,X ₂ ,...,X _t ) Save (S300).

Next, a predetermined state value (h) is generated based on the input value corresponding to each of the input time points, but the input value (X _t ) corresponding to the current time point (t) and the time point immediately preceding the current time point (t- Based on the state value (h _t-1 ) generated corresponding to 1), a state value (h _t ) for the current point in time (t) is generated (S400).

Here, in the step S400, the input value is received by the step S300, and a predetermined state value according to a pre-stored activation function is generated and stored, and the input value (X _t ) corresponding to the current point in time (t) and the current The state value (h _t ) corresponding to the current time point (t) is generated based on the result of applying the previously stored state value (h _t-1 ) corresponding to the time point (t-1) immediately before the time point to the activation function it may be

Specifically, the step S400 corresponds to a first value obtained by multiplying an input value (X _t ) corresponding to the current time point (t) by a preset input weight value (W _x ), and a time point (t-1) immediately preceding the current time point After obtaining the second value obtained by multiplying the pre-stored state value (h _t-1 ) and the preset hidden weight (W _h ), the result of applying the sum of the first value and the second value to the activation function Based on this, it is possible to generate a state value (h _t ) corresponding to the current point in time (t), and the input weight (W _x ) and hidden weight (W _h ) share the same parameter (weight) for each layer. It may have the same value according to the characteristics of the neural network.

At this time, the activation function may include a hyperbolic tangent (tanh) function, which is a nonlinear function, and in this case, the formula for the state value (h _t ) corresponding to the current point in time (t) is summarized to obtain Equation 2 and same.

On the other hand, the step S400 is based on the input data based on the input value (X _t ) corresponding to the current time point (t) and the state value (h _t- 1 ) corresponding to the time point (t-1) immediately preceding the current time point. A cell state value (C t ) corresponding to the current point in time (t) is generated and stored based _on the result of applying a predetermined reflection rate corresponding to each of the past data, and the previously stored cell state value (C _t ) is applied to the activation function It may be applied to generate a state value (C _t ) corresponding to the current time point (t).

Specifically, the step S400 includes generating a first result value by applying a preset first reflection rate based on the degree of deletion of the past data (h _t-1 ) (S410), and the input data (X A gate value by applying a second reflection factor preset based on the memory degree of _t ) and a candidate state value by applying the past data (h _t-1 ) and the input data (X _t ) to the activation function Generating a second result value by performing a predetermined operation for (S420), and generating a cell state value (C _t ) corresponding to the current time point based on the first result value and the second result value (S430), generating a third result value by applying a preset third reflection factor based on the output degree of the cell state value (C _t ) (S440), and It may include generating a state value (h _t ) corresponding to the current point in time ( _t ) based on a result of applying the cell state value (C t ) and the third result value to the activation function (S450).

In the generating of the first result value ( S410 ), a first result value is generated by applying a predetermined first reflection rate based on the degree of deletion of the past data (h _t-1 ).

Here, in the step S410, as shown in Equation 3 above, the input value of the current time point (t) and the state value (h t-1) of the time point (t _-1 ) immediately before the current time point have a preset weight (W _f ) A first resultant value (f _t ) between '0' and '1' may be generated by applying the sum of all values to a sigmoid function corresponding to the first reflection factor after reflecting .

At this time, the closer the first result value (f _t ) is to '0', the more past information is deleted and forgotten, and the closer to '1', the past information may be completely remembered.

In the step of generating the second result value (S420), the gate value (i _t ) according to applying the second reflection factor preset based on the storage degree of the input data (X _t ) and the input data (X _t ) and a second result value is generated by performing a preset operation on a candidate state value (g _t ) by applying the past data (h _t-1 ) to the activation function.

Here, in the step S420, as shown in Equation 4 above, the input value of the current time point (t) and the state value (h _{t-1) of the time point (t-1} ) immediately preceding the current time point have a preset weight (W _i ) A gate value (i _t ) between '0' and '1' can be generated by applying the sum of all values to a sigmoid function corresponding to the second reflection factor after reflecting .

In addition, in the step S420, as shown in Equation 5 above, the input value of the current time point (t) and the state value (h t-1) of the time point ( _t-1 ) immediately before the current time point have a predetermined weight (W _g ) A candidate state value (g _t ) can be generated by applying the sum of all values after reflecting t to the activation function.

In this case, the activation function may include a hyperbolic tangent (tanh) function, which is a nonlinear function, and the formula for the candidate state value (g _t ) in this case is summarized as Equation 5 above.

In addition, in the step S420, a second result calculated as a result of performing element-by-element multiplication (Hadamard Product) on the gate value (i _t ) and the candidate state value (g _t ) generated in the above-described step A value (i _t · g _t ) can be created.

In the step of generating the state value (C _t ) (S430), the first result value (f _t ) generated by the step S410 and the second result value (i _t g _t ) generated by the step S420 are Based on this, a cell state value (C _t ) corresponding to the current time point (t) is generated.

Specifically, in the step S430, the first result value (f _t ) is multiplied (·) by the cell state value (C t-1 ) of the time point (t _- 1) immediately preceding the current point in time, and then the second result is obtained. A value calculated as a result of adding the value (i _t ·g _t ) may be stored as a cell state value (C _t ) at the current point in time (t).

In the step of generating the third result value (S440), a third result value (o _t ) is obtained by applying a preset third reflection factor based on the output level of the cell state value (C _t ) at the current point in time (t). generate

Here, in the step S440, as shown in Equation 7, the input value of the current time point (t) and the state value (h _{t-1) of the time point (t-1} ) immediately preceding the current time point have a preset weight (W _o ) A third resultant value (o _t ) between '0' and '1' may be generated by applying the sum of all values to a sigmoid function corresponding to the third reflection factor after reflecting .

In the step of generating the state value (h _t ) (S450), the current time point (S450) is based on a result of applying the cell state value (C _t ) corresponding to the current time point (t) and the third result value to the activation function ( A state value (h _t ) corresponding to t) is generated.

In this case, the activation function may include a hyperbolic tangent (tanh) function, which is a nonlinear function, and in this case, the equation for the state value (h _t ) corresponding to the current time point (t) is as shown in Equation 8 above.

Next, a predetermined output value (y _t ) for the current time point (t) is generated based on the state value (h _t ) corresponding to the input value (X _t ) at which the input time point is the last (S500).

Here, in the step S500, a predetermined output weight (Wy) is reflected in the state value (h _t ) generated by the step S400 corresponding to the input layer (X _t ) in which the input time point based on the time series information (Ts) is the last. It may be to generate and store a value as an output value (y _t ), which is summarized as Equation 9 above.

Next, a probability value corresponding to at least one class among a plurality of preset classes is calculated based on the output value (y _t ) generated in step S500 (S600).

At this time, the plurality of classes are composed of two types of Alzheimer's Disease (AD) and Cognitively Normal (CN), or amyloid-positive Alzheimer's disease group (Aβ(+)AD), amyloid-negative normal group ( Aβ(+)CN), amyloid-negative Alzheimer's disease group (Aβ(-)AD), and amyloid-positive normal group (Aβ(+)CN).

Above, the present invention has been described in detail through preferred embodiments, but the present invention is not limited thereto and may be variously practiced within the scope of the claims.

In particular, the foregoing has outlined rather broadly the features and technical strengths of the present invention so that the claims of the invention to be described later may be better understood, so that the concept and specific embodiments of the present invention described above are intended to serve similar purposes to the present invention. It should be recognized by those skilled in the art that it can be readily used as a basis for designing or modifying other shapes for the purpose.

In addition, the embodiment described above is only one embodiment according to the present invention, and can be implemented in various modified and changed forms within the scope of the technical idea of the present invention by those skilled in the art. You will understand. Therefore, the disclosed embodiments should be considered from an explanatory point of view rather than a limiting point of view, and these various modifications and changes are also shown in the claims of the present invention described above as belonging to the scope of the technical idea of the present invention, and the scope equivalent thereto. Any differences within them should be construed as being included in the present invention.

100: image acquisition unit 200: time series vector generation module
220: time series information generation unit 240: image pre-processing unit
242: spatial normalization unit 244: intensity normalization unit
260: vector conversion unit 262: region detection unit
264: representative value calculation unit 266: score calculation unit
268: matrix generator 300; input layer
400: hidden layer unit 410: deletion gate
420: input gate 430: output gate
500: output layer unit 600: probability calculation unit

Claims (18)

an image acquiring unit configured to obtain a PET image consisting of a plurality of tomographic images sequentially taken by performing tomography for a predetermined time according to a PET technique for taking tomography of the brain of a subject to whom an amyloid agent is administered;
a time-series vector generation module that generates a predetermined image vector based on time-series information corresponding to time information obtained from each of the plurality of tomographic images and a score value corresponding to a pixel combination of each region of the tomography image;
an input layer unit including a plurality of input layers which sequentially receive at least one image vector one by one based on the time-series information and classify and store the image vectors according to input time points;
A predetermined state value is generated based on the input value of the input layer corresponding to each input time point, and the state generated corresponding to the input value of the input layer corresponding to the current time point and a time point immediately before the current time point a hidden layer unit including a plurality of hidden layers that generate a state value for the current view based on a value;
an output layer unit generating a predetermined output value for the current point in time based on a state value of the hidden layer corresponding to the input layer at which the input point in time is the last; and
A probability calculation unit for calculating a probability value corresponding to at least one class among a plurality of predetermined classes based on the output value; includes,
The plurality of tomographic images,
Time-series based PET image analysis device for determining brain disease, characterized in that obtained in correspondence with the initial section of the absorption time in the body of the diagnosed subject from the time the amyloid agent is administered to the time it is removed. According to claim 1,
The hidden layer unit,
A plurality of hidden layers provided in a cell structure including a hidden state therein to generate a predetermined state value according to a pre-stored activation function and store it in the hidden state,
The plurality of hidden layers and the plurality of input layers are connected to each other through input nodes corresponding to the number of the plurality of input layers, and at least one pair of the hidden layers adjacent to each other are connected in series to each other through a predetermined connection node. Accordingly, the state value corresponding to the current time point is generated based on a result of applying the activation function to the input value of the input layer corresponding to the current time point and the state value previously stored corresponding to a time point immediately before the current time point Time series-based PET image analysis device for brain disease determination, characterized in that. According to claim 2,
The hidden layer unit,
The hidden layer is provided with a cell structure further including a cell state therein to generate a predetermined state value according to the activation function and store it in the hidden state,
Based on a result of applying a preset reflectance corresponding to each of input data based on the input value of the input layer corresponding to the current view and past data based on the state value corresponding to a time immediately before the current view, the current view A cell state value of the hidden layer corresponding to is generated and stored in the cell state;
Time-series based PET image analysis device for determining brain disease, characterized in that the state value corresponding to the current time point is generated by applying the previously stored cell state value to the activation function. According to claim 3,
The hidden layer unit,
a deletion gate generating a first result value by applying a predetermined first reflection factor based on the degree of deletion of the past data and transmitting the first result value to the cell state;
A gate value by applying a second reflectance preset based on the storage degree of the input data and a candidate state value by applying the past data and the input data to the activation function. an input gate generating 2 result values and transferring them to the cell state; and
An output gate generating a third resultant value by applying a preset third reflection factor based on the output degree of the cell state value;
generating the cell state value corresponding to the current time point based on the first result value and the second result value;
Based on a result of applying the cell state value corresponding to the current time point and the third result value to the activation function, the state value corresponding to the current time point is generated. PET image analysis device. According to claim 2,
The output layer unit,
It is connected to each of the plurality of hidden layers through a predetermined output node, and a value in which a predetermined output weight is reflected in the state value of the hidden layer connected to the input layer corresponding to the input layer at which the input time based on the time series information is the last is generated as an output value A time-series-based PET image analysis device for determining brain disease, characterized in that for storing. According to claim 1,
The time series vector generation module,
Detecting time information obtained from each of the plurality of tomographic images and arranging them in chronological order to generate order information for each image, and based on the time information and the order information corresponding to each of the tomographic images, a time-series information generating unit generating time-series information for a plurality of tomographic images;
an image pre-processing unit normalizing the plurality of tomographic images according to a predetermined image processing technique and converting them into normalized images in a 2D array format; and
and a vector converter for calculating a score value corresponding to a pixel combination for each region of the normalized image and then converting it into a predetermined image vector based on the calculated score value and the generated time-series information. A time-series-based PET image analysis device for brain disease determination. According to claim 6,
The vector converter,
a region detection unit dividing each of the normalized images into a plurality of regions corresponding to a region of interest based on a predetermined SUV technique and detecting a pixel combination for each region;
a representative value calculating unit calculating a representative value for the pixel combination for each detected region;
a score calculation unit that calculates a score value corresponding to a lightness coordinate value based on a result of applying the calculated representative value to a preset color model; and
A time-series-based PET image analysis device for determining a brain disease, comprising: a matrix generator for generating matrix data for a predetermined image vector based on the calculated score value and the generated time-series information. According to claim 6,
The image pre-processing unit,
a spatial normalization unit for spatially normalizing the plurality of tomographic images with predetermined volume data; and
A time-series-based PET image analysis device for determining a brain disease, comprising: an intensity normalizer that scales each value of a plurality of voxels constituting the volume data at a predetermined ratio according to a preset scale factor. Obtaining a PET image consisting of a plurality of tomographic images sequentially taken by performing tomography for a predetermined time according to a PET technique for taking tomography of the brain of a subject to whom an amyloid agent has been administered;
generating a predetermined image vector based on time-series information corresponding to time information obtained for each of the plurality of tomographic images and a score value corresponding to a pixel combination for each region of the tomographic image;
receiving at least one image vector sequentially one by one based on the time-series information, classifying the image vectors according to input time points, and storing them as a plurality of input values;
A predetermined state value is generated based on an input value corresponding to each input time point, and the current time point is based on the input value corresponding to the current time point and the state value generated corresponding to a time point immediately before the current time point. generating a state value for;
generating a predetermined output value for the current point in time based on the state value corresponding to the input value at which the input point in time is the last; and
Calculating a probability value corresponding to at least one class among a plurality of predetermined classes based on the output value; includes,
The plurality of tomographic images,
Time series-based PET image analysis method for determining brain disease, characterized in that obtained in correspondence with the initial section of the absorption time in the body of the diagnosed subject from the time the amyloid agent is administered to the time it is removed. According to claim 9,
The step of generating the state value is,
Upon receiving the input value, a predetermined state value according to a previously stored activation function is generated and stored, and the input value corresponding to the current time point and the state value previously stored corresponding to the time point immediately before the current time point are converted into the activation function. Time-series-based PET image analysis method for determining brain disease, characterized in that for generating the state value corresponding to the current time point based on a result applied to. According to claim 10,
The step of generating the state value is,
Corresponding to the current time point based on a result of applying a predetermined reflection rate corresponding to each of input data based on the input value corresponding to the current point in time and past data based on the state value corresponding to a time point immediately before the current point in time Time-series-based PET image analysis method for determining brain disease, characterized by generating and storing a cell state value, and generating the state value corresponding to the current time point by applying the previously stored cell state value to the activation function. . According to claim 11,
The step of generating the state value is,
generating a first result value by applying a predetermined first reflection rate based on the degree of deletion of the past data and reflecting the first result value to the cell state value;
A gate value by applying a second reflectance preset based on the storage degree of the input data and a candidate state value by applying the past data and the input data to the activation function. 2 generating a result value and reflecting it to the cell state value;
generating the cell state value corresponding to the current time point based on the first result value and the second result value;
generating a third result value by applying a predetermined third reflectance based on the output degree of the cell state value; and
and generating the state value corresponding to the current time point based on a result of applying the cell state value corresponding to the current time point and the third result value to the activation function. Time series-based PET image analysis method for According to claim 9,
The step of generating the output value is,
A time-series-based PET image analysis method for determining brain disease, characterized in that a value obtained by reflecting a predetermined output weight in the state value corresponding to the input value at which the input time point based on the time-series information is the last is generated as an output value and stored. . According to claim 9,
The step of generating the image vector,
Detecting time information obtained from each of the plurality of tomographic images and arranging them in chronological order to generate order information for each image, and based on the time information and the order information corresponding to each of the tomographic images, generating time-series information for a plurality of tomographic images;
normalizing the plurality of tomographic images according to a predetermined image processing technique and converting them into normalized images in a 2D array format; and
Calculating a score value corresponding to a pixel combination for each region of the normalized image and then converting it into a predetermined image vector based on the calculated score value and the generated time-series information; Time series-based PET image analysis method for judgment. According to claim 14,
The step of converting the image vector,
dividing each of the normalized images into a plurality of regions corresponding to a region of interest based on a predetermined SUV technique and detecting a pixel combination for each region;
calculating a representative value of the pixel combination for each detected region;
calculating a score value corresponding to a lightness coordinate value based on a result of applying the calculated representative value to a preset color model; and
Generating matrix data for a predetermined image vector based on the calculated score value and the generated time-series information; time-series-based PET image analysis method for determining brain disease, comprising: According to claim 14,
The step of converting the normalized image,
spatially normalizing the plurality of tomographic images with preset volume data; and
Scaling each value of a plurality of voxels constituting the volume data at a predetermined ratio according to a predetermined scale factor; time-series-based PET image analysis method for determining a brain disease, comprising: delete delete

Images