JP7343145B2 - Information processing device, information processing method, and program - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program.

In various treatments such as glaucoma, it is necessary to understand the progress of a patient's symptoms. Specifically, in the treatment of glaucoma, it is necessary to understand the condition of the patient's visual field, and currently, visual field testing using optical stimulation is performed to understand the condition of this visual field. However, in order to obtain sufficient information for treatment, it is necessary to examine the patient's vision at many points that can be used as visual fields, and the burden of examination is heavy.

Furthermore, if measurement information on visual field sensitivity measured over a long period of time is required in order to know the state of progress of glaucoma, it will take a long time to make a decision regarding treatment.

On the other hand, in recent years, research has been conducted on whether there is a relationship between the partial area average of papillary retinal layer thickness and visual field sensitivity (Non-Patent Document 1).

Donald C. Hood, et.al., “A Framework for comparing structural and functional measures of glaucomatous damage”, Progress in Retinal and Eye Research, 2007 Nov. 26(6), pp. 688-710

As described above, in conventional inspections, etc., the burden of inspection may become large due to the measurement of the current state. On the other hand, by obtaining new knowledge as in Non-Patent Document 1, it is possible to reduce the burden of testing by referring to the results of various tests conducted in the past, and it is also possible to speed up the determination of the state of progression. We can expect that there may be cases.

However, even if a correlation as described in Non-Patent Document 1 is suggested between the results of multiple different types of tests, different types of tests that are considered to be correlated with each other may not be performed on the same day, for example. Generally speaking, it is rare that the progress is considered to be the same. For this reason, even if new knowledge is obtained that there is a correlation between multiple types of tests, the correlation between these test results cannot be verified, and in the end, information from past test results cannot be used effectively. The current situation is that it is not available.

The present invention has been made in view of the above-mentioned circumstances, and is an information processing device and an information processing method that can effectively utilize information on past test results, reduce the test load, and quickly provide decision materials that contribute to treatment, etc. One of its purposes is to provide , and programs.

One aspect of the present invention that solves the problems of the conventional example described above is an information processing device that includes a holding unit that holds information on a latent space defined by a parameter group including a parameter related to time information; an accepting means for receiving an input of a measurement information group including measurement information of different types, each of which is measured at an arbitrary time; and measurements included in the accepted measurement information group. a setting means for updating and setting the information of the latent space held by the holding means using the measurement information of the type in a predetermined manner for each type of information; and a setting means updated by the setting means. and output information generating means for generating predetermined output information using the information of the latent space held by the holding means.

According to the present invention, information on past test results can be effectively used.

1 is a block diagram illustrating a configuration example of an information processing device according to an embodiment of the present invention. FIG. FIG. 1 is a functional block diagram illustrating an example of an information processing device according to an embodiment of the present invention. FIG. 2 is an explanatory diagram conceptually representing the operation of the information processing device according to the embodiment of the present invention. FIG. 2 is a flowchart diagram illustrating an example of the operation of the information processing device according to the embodiment of the present invention.

Embodiments of the present invention will be described with reference to the drawings. As illustrated in FIG. 1, the information processing device 1 according to the embodiment of the present invention includes a control section 11, a storage section 12, an operation section 13, a display section 14, and an interface section 15. be done.

The control unit 11 is a program control device such as a CPU, and operates according to a program stored in the storage unit 12. In the example of the present embodiment, the control unit 11 stores and holds information on a latent space defined by a parameter group including parameters related to time information in the storage unit 12, and also stores information on a latent space defined by a parameter group including parameters related to time information, and Accepts input of a measurement information group including measurement information.

Here, the measurement information group includes a plurality of types of measurement information each measured at an arbitrary time. The control unit 11 uses the measurement information of the type included in the received measurement information group in a predetermined method for each type of measurement information to generate information on the latent space held in the storage unit 12. Update and configure.

Then, the control unit 11 generates predetermined output information using the updated latent space information held by the storage unit 12. The detailed operation of this control section 11 will be described later.

The storage unit 12 is a memory device or the like, and holds programs executed by the control unit 11. This program may be provided stored in a computer-readable, non-temporary recording medium, and may be copied to the storage unit 12. This storage section 12 also operates as a work memory for the control section 11 and also functions as a holding means.

The operation unit 13 is a device such as a mouse or a keyboard, and accepts the content of a user's instruction operation and outputs it to the control unit 11. The display unit 14 is a display or the like, and outputs information according to instructions input from the control unit 11.

The interface section 15 is a USB (Universal Serial Bus), a network interface, or the like, and receives measurement information from an external memory, inspection device, server, etc., and outputs it to the control section 11 . In one example of this embodiment, this measurement information includes information on test results related to the thickness of the patient's retinal layer, test results related to visual sensitivity at multiple points within the visual field (results of tests using optical stimulation, etc.). ). The interface unit 15 may also send information to be output to an external device or the like in accordance with instructions input from the control unit 11.

Next, an example of the operation of the control section 11 of this embodiment will be described. In the following example, it will be explained that one of the purposes is to predict the state of the visual field of a specific person (assuming that the person is undergoing a visual field test). By executing the program stored in the storage unit 12, the control unit 11 of the information processing device 1 according to the present embodiment controls the information receiving unit 21, the setting unit 22, and the output unit as illustrated in FIG. A configuration functionally including a generation section 23 and an output section 24 is realized.

In the following description, it is assumed that the storage unit 12 stores information on a latent space defined by a parameter group including parameters related to time information. Specifically, the latent space is
At+B
It is assumed that the function is described as a function that changes linearly with respect to time t (with a parameter group including parameters of a linear function of time information). Here, A and B are R-dimensional vectors, and below,
W=[AB]
is defined as a matrix with R rows and 2 columns. Initially, each element of A and B is randomly determined. In this example, the storage unit 12 holds the values of each element of vectors A and B (that is, matrix W) as latent space information. In this example, each element of vector A is a parameter related to time information, and each element of A and B corresponds to a parameter group.

Furthermore, in this embodiment, it is assumed that information on this latent space is obtained for each object of measurement information, which will be described later. As an example, if the information processing device 1 of this embodiment is to predict the state of a certain person's visual field, the object of the measurement information here is each eye of the person who has undergone a visual field test, etc. , the information of the latent space is determined for each person and each person's eyes. Hereinafter, the information on the latent space for each eye i (i=0, 1, 2, . . . ) will be described separately as Wi, etc., with i as a subscript.

The information receiving unit 21 receives an input of a measurement information group including a plurality of different types of measurement information, each of which is measured at an arbitrary time. Specifically, the measurement information here includes a retinal layer thickness test result (information regarding the structure of the retina) and a visual field test result by optical coherence tomography (OCT). In the present embodiment, the information receiving unit 21 obtains a plurality of different types of retinal layer thickness test results and visual field test results at arbitrary times (times that cannot necessarily be evaluated simultaneously). Accept the results of the test. Furthermore, the test results do not necessarily have to be the test results for a specific person, and may include test results for a plurality of people.

Note that in the following example, the results of the visual field test represent the lowest brightness visually recognized by the eyes of the test subject at D visual field test points set within the visual field of the human eye. In other words, the result of the visual field test is a D-dimensional vector.

The setting unit 22 associates the measurement information included in the received measurement information group with the information in the latent space using a predetermined method for each type of measurement information.

In one example of the present embodiment, the measurement information includes retinal layer thickness test results that are image data (GCIPL (Ganglion Cell and Inner Plexiform Layer), RNFL (Retinal Nerve Fiber Layer), RCL (Rod and Cone Layer), (obtained as a set) and the visual field test results, which are vector data. The setting unit 22 converts the image data, which is the retinal layer thickness test result, into vector data related to information in the latent space using a neural network that has undergone predetermined machine learning processing. The neural network used here will be described later. Furthermore, the setting unit 22 associates the visual field test results with latent space information using the following method.

That is, the setting unit 22 generates information W0 of the latent space corresponding to the visual field test result (here, the subscript "0" means the eyes of the person whose visual field is to be predicted (hereinafter referred to as the prediction target)) of,
G0W0P ^T 0+ε
The result of this calculation is to minimize the error between the actual visual field test results performed (T0 - 1) times in the past and the visual field test result matrix F0 representing the current (point of time when visual field is to be predicted). Determine G0 and W0 to be reduced.

Here, the visual field test result matrix F0 arranges the column vectors of the results of the visual field test of the prediction target at past time points τk (k=1, 2,..., T0-1) in the k column, and the current ( This is an array of column vectors that are provisional values (updated to predicted values by performing the following processing) of the results of the visual field test at the time when the visual field is to be predicted. Therefore, this visual field test result matrix is an array of ((T0-1)+1)=T0 D-dimensional vectors, and is a matrix with D rows and T0 columns.

Further, G0 is a projection matrix with D rows and R columns (R is the dimension of the latent space), and P0 is a matrix [t0, 1T0]. Here, t0 is a T0-dimensional vector representing the time when (T0-1) visual field tests were performed in the past (a relative time with a predetermined time in the past as the reference time Tb) and the desired time to predict. , 1T0 means a T0-dimensional vector in which all elements are "1". Also, the T on the shoulder of P0 means transposition. Note that ε represents an error.

を演算することとなる。 The setting unit 22 of this embodiment operates to obtain G0 and W0 that minimize the square of the Frobenius norm of F0 - G0 W0 P ^T 0. In other words, the setting section 22

will be calculated.

は、要素ごとの積を求めることを意味する。また、

は、Ｘのフロベニウスノルムである。 Here, M0 is a mask matrix with only the last column set to "0" and the other elements set to "1",

means to find the product of each element. Also,

is the Frobenius norm of X.

That is, in this example, G0W0P ^T0 is information that defines the conversion relationship between the visual field test result and the information on the latent space, and G0 corresponds to an example of the conversion parameter.

Note that if the current visual field test result is predicted only from the past visual field test result, the setting unit 22 obtains G0, W0 that minimizes the square of the Frobenius norm of F0 - G0W0P ^T 0. That's fine. Since there are various widely known methods for this method, a detailed explanation thereof will be omitted here.

Next, an example of how the setting unit 22 associates image data with latent space information will be described. In the example of this embodiment, the transformation relationship between image data and latent space information is defined using a multilayer neural network including a CNN (convolution) network.

The information input to this convolutional network is an array of image data (retinal layer thickness test results) obtained each time a retinal layer thickness test is performed at time t0_k (k=1, 2,...). The time dimension (1 dimension) is added to the data dimension (2 dimensions), and 4-dimensional data in which three image data sets are arranged (image width w x image height h x inspection time T0 x the number of image data for each examination (here, three: GCIPL, RNFL, and RCL)). Note that the time t0_k here is also expressed as a relative time with respect to the reference time tb that is used as a reference for each time t of the visual field test results (this is to align with the same time series as the visual field test results).

Here, the setting unit 22 concatenates the three image data obtained at each time point of the examination to form three-dimensional data, and arranges this in the order of the time of the examination to form four-dimensional data (image series data). The setting unit 22 then inputs this four-dimensional image series data to a neural network that defines the transformation relationship between the image series data and the information of the latent space.

This neural network is a three-dimensional convolutional network (multilayer neural network) on the input layer side, and the information is reshaped (dimensional conversion) in the intermediate layer (output layer of the three-dimensional convolutional network), and the two-dimensional The reshaped information is further input to a convolution network (multilayer neural network), and a matrix having R×2 elements is obtained as an output.

と表すことができる。ここでＣ0は、入力データ（複数の画像データを配列した４次元の画像系列データ）であり、ΘEは、ニューラルネットワークの各層間の重みやバイアス等であり、変換パラメータの一例に相当する。 The setting unit 22 uses the output of this two-dimensional convolutional network as latent space information W0. In other words, this neural network

It can be expressed as. Here, C0 is input data (four-dimensional image series data in which a plurality of image data are arranged), and ΘE is the weight, bias, etc. between each layer of the neural network, and corresponds to an example of a conversion parameter.

Therefore, the setting unit 22 may perform machine learning on the neural network so that the information W0 of the latent space that represents the current visual field test result is output, but such information W0 of the latent space is not necessarily obvious. do not have.

Therefore, in the present embodiment, the setting unit 22 uses the information W0 of the latent space, the conversion parameter G0 of the visual field test result, and the conversion parameter of the image series data (retinal layer thickness test result) as follows. Perform machine learning on a certain ΘE.

Specifically, the setting unit 22 sets an objective function LDLR defined as follows.

Note that λ1 is a hyperparameter indicating how much the retinal layer thickness test result influences the parameters of the latent space, and is determined experimentally or empirically.

Then, the setting unit 22 uses the information W0 of the latent space, the conversion parameter G0 of the visual field test results, and the image series data (four-dimensional ΘE, which is a conversion parameter of data), are optimized all at once. As the optimization method here, widely known methods such as the Adam method (D.P. Kingma et.al., Adam: A method for stochastic optimization, arXiv:1412.6980(2014)) can be adopted, so Detailed explanation will be omitted. The setting unit 22 stores G0, which is the conversion parameter of the visual field test result obtained through optimization, and the latent space information W0 in the storage unit 12 as updated latent space information W0.

As described above, in this embodiment, as conceptually shown in FIG. The information is related to the information (W0) in the latent space W by projecting or transforming the information onto the latent space. The information processing device 1 converts the conversion parameters and parameters representing the information of the latent space so that the output information (F0) based on the information of the latent space matches at least one of the measurement information (objective function). ) is determined through machine learning. Further, the latent space W is expressed as a space of a function (W0) that changes with time, as illustrated in FIG.

The output generation unit 23 uses the updated latent space information W0 stored in the storage unit 12 to generate output information desired by the user. In the example of the present embodiment, since the user wants information on the visual field test result at the current time (time of processing), the output generation unit 23 generates the latent space information W0 stored in the storage unit 12. and G_0, and P (parameter representing the current time),
F=G0W0P ^T
seek. The output generation unit 23 outputs this F to the output unit 24 as a vector representing information on the visual field test result at the current time (time of processing).

The output unit 24 presents the vector information output by the output generation unit 23 to the user by displaying it on the display unit 14 or the like.

[Example of including information about other eyes]
In addition, in the explanation up to this point, the setting unit 22 optimizes the information of the latent space using the measurement information regarding the eye that is the target of prediction, and based on the information of the optimized latent space, the setting unit 22 However, in this embodiment, measurements related not only to the eye targeted for prediction but also to other eyes (another eye of the person targeted for prediction, or the eyes of another person) are performed. Information may also be used to optimize the information in the latent space.

In this example, the setting unit 22 uses the following equation (2) instead of equation (1), and minimizes equation (2).

と定義しておく。 In this equation (2), λ0 is a hyperparameter, N is the number of eyes for which measurement information is obtained plus 1, and z0,j is a parameter representing the degree of eye similarity. It is. This z0,j is here,

Let's define it as

Here, F*j is the eye at the time τk (k=1, 2,...T0-1) when the visual field test results included in F0 are obtained for the eye to be measured other than the eye corresponding to F0. Predicted values (column vectors) of visual field test results for j (j=1, 2, . . . ) are arranged in corresponding columns (k-th column). Here, the predicted value of the visual field test result is a predicted value obtained by linear regression based on the actual visual field test result for eye j. The calculation method is detailed in Kenji Yamanishi, et.al., Predicting glaucoma progression using multi-task learning with heterogeneous features, In Big Data. 2014, IEEE International Conference on. IEEE 261-270, so I will not explain it here. The explanation will be omitted.

を、可能なｉ，ｊの組み合わせについて求めたときの中央値（median）である。 Also, σ ² is

is the median value obtained for possible combinations of i and j.

となるようニューラルネットワークｆの機械学習を行う。 Further, the setting unit 22 also sets the image series data obtained based on the retinal layer thickness test results to other eyes (another eye of the person who is the target of prediction, or an eye of another person) as follows. Processing including That is, the setting section 22

Perform machine learning on the neural network f so that

Specifically, in this example, the setting unit 22 also uses the information W0 of the latent space, the conversion parameter G0 of the visual field test result, and the conversion parameter of the image series data (four-dimensional data based on the retinal layer thickness test result). In order to perform machine learning on a certain ΘE at once, set the objective function LDLLR as follows:

Then, the setting unit 22 sets the latent space information Wi (including i=0), the conversion parameter Gi of the visual field test result, and the image series data (retinal layer thickness ΘE, which is a conversion parameter of four-dimensional data based on the inspection results, are collectively optimized.

としてもよい。ここで、ＬDLLRは、（４）式のＬDLLRである。また、ｇ（・，ΘD）は、ｆ（Ｃi，ΘE）のデコーダであり、ｆ（Ｃi，ΘE）と対称的な構造を有するニューラルネットワークである。従って、

は、オートエンコーダを意味し、その出力は、入力情報と同じサイズの画像系列データとなる。（５）式の目的関数の第２項ではオートエンコーダの出力と、入力情報との差のフロベニウスノルムの二乗を求めていることとなる。 Further, here, an L2 regularization term and an autoencoder regularization term may be added to the objective function as regularization terms. That is,

You can also use it as Here, LDLLR is LDLLR of equation (4). Furthermore, g(·, ΘD) is a decoder of f(Ci, ΘE), and is a neural network having a structure symmetrical to f(Ci, ΘE). Therefore,

means an autoencoder, and its output is image series data of the same size as the input information. The second term of the objective function in equation (5) calculates the square of the Frobenius norm of the difference between the output of the autoencoder and the input information.

Further, the third term in equation (5) is the L2 regularization term. Note that λ2 and λ3 are hyperparameters that determine the importance of each term. These λ2 and λ3 are also determined empirically or experimentally.

In this example as well, the setting unit 22 sets the information W0 of the latent space, the transformation parameter G0 of the visual field test results, and the image series data (retinal layer thickness test ΘE, which is a transformation parameter of 4-dimensional data based on the results, are optimized together.

Here again, widely known methods such as the adam method can be used as the optimization method. The setting unit 22 stores G0, which is the conversion parameter of the visual field test result obtained by optimizing LDLLR or L′DLLR, and the latent space information W0 as updated latent space information W0 in the storage unit. 12.

[motion]
The information processing device 1 according to the embodiment of the present invention has, for example, the above configuration and operates as follows. In addition, in the following explanation, the results of a visual field test conducted at any point in the past for multiple eyes (each eye of multiple measurement subjects) (information on the sensitivity of the measurement subject at each point in the visual field) will be used. ) or retinal layer thickness test results (image data) conducted at another time, one of the eyes of a specific measurement subject is selected as the prediction target. An example of predicting the results of a visual field test for an eye will be described.

In this example, as illustrated in FIG. 4, the information processing device 1 uses the results of a retinal layer thickness test (information regarding the structure of the retina) and the visual field test conducted in the past using optical coherence tomography (OCT). A group of measurement information including each result as measurement information is accepted (S11). As mentioned above, all of these tests were performed individually at arbitrary points in the past, and even for the same eye, it is not necessarily the case that both tests were performed at the same time.

The information processing device 1 associates each piece of measurement information included in the measurement information group accepted here with information in the latent space using a predetermined method for each type of measurement information. Specifically, the information processing device 1 sequentially selects the eye to be measured as the eye to be predicted, and repeats the following process (S12).

The information processing device 1 arranges the actual visual field test results performed (T0-1) times in the past on the eye of the prediction target as column vectors in the row direction, and in the last column, the current (field of view to be predicted) A visual field test result matrix F0 of D rows and T0 columns (D is the number of test points of the visual field test) is generated, in which provisional values of the visual field test results at the time point) are arranged as column vectors (S13).

In addition, the information processing device 1 calculates linear regression parameters of the visual field test results for each eye of the measurement target other than the prediction target, and calculates the linear regression parameters of the visual field test results for each eye included in the first (T0-1) column of F0 generated in process S13. The predicted value of the visual field test result for each eye to be measured at the time point corresponding to the time point of the test is obtained using the parameters of this linear regression, and the predicted values at the corresponding time point are arranged as a column vector and F *j is obtained (S14). The information processing device 1 uses F*j obtained in process S15 to obtain a parameter z0,j representing the degree of similarity between the prediction target and eye j according to equation (3) (S15).

Next, regarding the image data of the retinal layer thickness test results, the information processing device 1 concatenates the image data for each time point of the test to obtain four-dimensional image series data to be input into the neural network (S16). Note that if there is image data of retinal layer thickness test results more than a predetermined number of times, the information processing device 1 selects image data of retinal layer thickness test results less than or equal to the predetermined number. May be used.

を、予備的に機械学習する（Ｓ１８）。この機械学習は、処理Ｓ１６で生成した画像系列データのうちから任意に（ランダムに）選択した複数の画像系列データをオートエンコーダに入力し、オートエンコーダの出力と、入力した画像系列データとの誤差が小さくなるよう機械学習することで行えばよい。具体的に情報処理装置１は、

を最小化するように、adam法などを用いて

を定める。そして、情報処理装置１は、その後ＬDLLR，Ｌ′DLLRを用いた演算を行う際には、ＬDLLR，Ｌ′DLLRに含まれるΘE，ΘDを、

で初期化して演算を行うこととする。 Here, the information processing device 1 initializes information Wi (i=0, 1, 2, . . . ) that defines a latent space for each eye that can be a prediction target. In this example, the latent space of eye i is
Ait+Bi
Assuming that it is described as a function that changes linearly with respect to time t,
Wi=[AiBi]
It is defined as a matrix with R rows and 2 columns. The information processing device 1 determines each element of Ai and Bi, which are parameters representing information of this latent space, randomly or based on prior knowledge (S17). The information processing device 1 also includes an autoencoder

is subjected to preliminary machine learning (S18). This machine learning is performed by inputting a plurality of image series data arbitrarily (randomly) selected from among the image series data generated in process S16 into an autoencoder, and comparing the error between the output of the autoencoder and the input image series data. This can be done by using machine learning to reduce the value. Specifically, the information processing device 1

Using the adam method etc. to minimize

Establish. Then, when the information processing device 1 subsequently performs calculations using LDLLR and L'DLLR, ΘE and ΘD included in LDLLR and L'DLLR are

Let's initialize it with and perform the calculation.

In addition, in process S18, the information processing device 1 also performs preliminary machine learning on Gi, which is a conversion parameter of the visual field test result, to initialize it. This initialization can also be performed by performing machine learning to minimize Fi-GiWiP ^T i using the adam method or the like using Wi determined at random in advance.

を最小化するＧをadam法などにより得て、Ｇiを用いた演算を行う際には、いずれのＧiについても（任意のｉについて）、Ｇi＝Ｇと初期化する。 In this case as well, the information processing device 1

When G that minimizes G is obtained by the adam method or the like and an operation using Gi is performed, any Gi (for any i) is initialized as Gi=G.

Then, the information processing device 1 repeatedly performs calculations using the adam method or the like so as to minimize equation (5) (S19). This repeated calculation is performed until L'DLLR converges (until the difference between L'DLLR obtained last time and L'DLLR obtained by updating therefrom falls below a predetermined threshold).

When the repeated calculation in process S19 is completed (when it is determined that L'DLLR has been minimized), the information processing device 1 stores Gi and Wi at that time in the storage unit 12 (S20), and performs the prediction. If there is an eye that is not targeted, return to step S13 and continue the process (S21), and if all eyes have been processed as prediction targets, the process ends.

The information processing device 1 then selects Wi and Gi for the eye to be predicted from among the updated latent space information Wi and the transformation parameters Gi for the visual field test results stored in the storage unit 12. and set it as W0 and G0, and using P (a parameter representing the current time),
F=G0W0P ^T
seek. The information processing device 1 presents this F to the user as a vector representing information on the visual field test result at the current time (time of processing). Specifically, the value of each element of F may be displayed at the position of the visual field test point represented by each element of F.

Note that when calculating Gi and Wi, the information processing device 1 normalizes the elements of F (the maximum possible value at each inspection point so that the maximum value is "1"). The value of the test result at each test point may be divided by qmax). When normalization is performed in this way, the information processing device 1 adds the value of each element of F obtained as a vector representing the information of the visual field test result at the current time (point of processing) to the value of each of the above tests. It may be output after being multiplied by the maximum possible value qmax at the point.

As described above, according to the present embodiment, even if different types of tests are performed at different times, the results of each test are handled in association with the parameters of a common latent space, and the corresponding By generating results corresponding to the necessary information in the latent space and inversely converting them into the necessary information, it is possible to obtain a prediction that integrates the results of various tests at each point in time, and to obtain information on past test results. can be used effectively.

Note that in this example of the present embodiment, an example has been described in which information on a latent space related to the visual field is obtained and the result of a visual field test is predicted. Regarding cases where multiple types of measurement values that are assumed to be related to information that changes over time (for example, deterioration of some kind of structure over time) are obtained sporadically, not limited to information that is processed. Can be widely applied.

Next, an example will be shown in which the error (RMSE: root mean squared error) between the prediction of the visual field test result by the information processing device 1 of the present embodiment and the actually measured result is calculated.

[Table 1] shows an example (LR) predicted by linear regression for each measurement subject as Comparative Example 1 for comparison with the method of this embodiment (DLLR), and a prediction example (LR) by Zenghan Liang et al. et.al., Quantitative prediction of glaucomatous visual field loss from few measurements., In Data Mining(ICDM), 2013, IEEE 13th International Conference on. IEEE,. 1121-1126) Method using clustering (TSLR) and Shigeru (Shigeru Maya, et. al., Predicting glaucoma progression using multi-task learning with heterogeneous features. In Big Data(Big Data), 2014, IEEE International Conference on. IEEE, 261-270), Multitask Matrix Completion The method used (MTMC) is also shown.

Note that in [Table 1], S on the horizontal axis is the number of times the retinal layer thickness test results are used (the number used in process S16 in FIG. 4).

As shown in Table 1, compared to other methods, the method of this embodiment allows the visual field test results to be predicted with a relatively low error even if the number of tests used is relatively small.

1 information processing device, 11 control unit, 12 storage unit, 13 operation unit, 14 display unit, 15 interface unit, 21 information reception unit, 22 setting unit, 23 output generation unit, 24 output unit.

Claims (6)

holding means for holding information on a latent space defined by a parameter group including parameters of a time information function ;
Accepting means for receiving input of a measurement information group including a plurality of different types of measurement information, each of which is measured at an arbitrary time;
Setting means for updating and setting the information of the latent space held by the holding means using the measurement information of the type in a predetermined manner for each type of measurement information included in the accepted measurement information group; and,
output information generating means for generating predetermined output information using the information of the latent space held by the holding means updated by the setting means;
including;
An information processing device in which the information of the latent space is defined by a parameter group including a parameter of a function of time information . The information processing device according to claim 1,
For each type of measurement information, a conversion relationship with the information of the latent space is set in advance using a predetermined conversion parameter,
The setting means sets the conversion parameters for each type of measurement information and the information on the latent space held by the holding means so that the output information based on the information on the latent space is at least one of the measurement information. An information processing device that performs machine learning to reduce errors. The information processing device according to claim 1 or 2,
The measurement information includes image series data and visual field test results,
The information of the latent space is determined for each target of measurement information,
The setting means minimizes an objective function including a term based on a transformation parameter of the image series data and information on the latent space, and a term based on the transformation parameter of the visual field test result and information on the latent space. The conversion parameters for each type of measurement information and the information on the latent space held by the holding means are set such that the output information based on the information on the latent space has an error with each of the measurement information. An information processing device that performs machine learning all at once to make it smaller . The information processing device according to any one of claims 1 to 3,
The measurement information includes image series data,
The information processing device is configured such that the setting means defines a conversion relationship between the image series data and the information of the latent space using a neural network. Using a computer equipped with a holding means for holding information of a latent space defined by a parameter group including parameters of a time information function ,
the receiving means receives an input of a measurement information group including a plurality of different types of measurement information, each of which is measured at an arbitrary time;
The setting means updates the information of the latent space held by the holding means using the measurement information of the type in a predetermined manner for each type of measurement information included in the accepted measurement information group. Set,
An information processing method, wherein an output information generating means generates predetermined output information using information of the latent space held by the holding means, updated by the setting means. computer,
holding means for holding information on a latent space defined by a parameter group including parameters of a time information function ;
Accepting means for receiving input of a measurement information group including a plurality of different types of measurement information, each of which is measured at an arbitrary time;
Setting means for updating and setting the information of the latent space held by the holding means using the measurement information of the type in a predetermined manner for each type of measurement information included in the accepted measurement information group; and,
output information generation means for generating predetermined output information using the information of the latent space held by the holding means updated by the setting means;
A program that functions as

Images