WO2018083853A1 - Visual field sensitivity estimation device, method for controlling visual field sensitivity estimation device, and program - Google Patents

Abstract

In the present invention, provided information including information pertaining to visual field sensitivity and/or information pertaining to the thickness of a retina layer of an eye for each information provider, and information pertaining to the thickness of the retina layer of a prediction object person are received, and retina layer thickness feature value information for each information provider and prediction object person and visual field sensitivity feature value information for each information provider are extracted. A parameter is estimated for reconstituting the retina layer thickness feature value information of the prediction object person and feature value information based on information included by object pair information among the provided information including both the information pertaining to retina layer thickness and the information pertaining to visual field sensitivity, and the information pertaining to visual field sensitivity of the prediction object person is estimated on the basis of information reconstituted using the parameter.

Description

Field-of-view sensitivity estimation apparatus, method of controlling field-of-view sensitivity estimation apparatus, and program

The present invention relates to a visual field sensitivity estimation apparatus, a visual field sensitivity estimation apparatus control method, and a program.

There are treatments that require grasping the patient's visual field, such as treatment for glaucoma. Currently, visual field conditions include visual field inspection methods using light stimulation, but in order to obtain sufficient information for treatment, it is necessary to examine how the patient looks at many points that can be the visual field. There is a problem that the burden of inspection is large.

On the other hand, in recent years, research has been conducted on whether or not there is a relationship between the average partial area of the retinal layer thickness of the nipple and visual field sensitivity (Non-patent Document 1).

Retinal layer thickness information can be measured relatively easily with an optical coherence tomography (OCT). Therefore, if visual field sensitivity can be estimated from retinal layer thickness information obtained by OCT, the examination burden can be reduced and the treatment efficiency can be reduced. Can be expected to do. However, the present situation is that the method for estimating the visual field sensitivity has not been studied even at the research stage, including the conventional research described above.

The present invention has been made in view of the above circumstances, and provides a visual field sensitivity estimation device, a visual field sensitivity estimation device control method, and a program capable of estimating visual field sensitivity information from retinal layer thickness information. One of its purposes.

One aspect of the present invention that solves the problems of the above-described conventional example is a visual field sensitivity estimation device, which includes at least one of information related to the thickness of the eye retinal layer and information related to visual field sensitivity for each information provider. The provided information, means for accepting information related to the thickness of the retinal layer of the prediction target person who is the target of visual field prediction, information related to the thickness of the retinal layer included in the accepted provided information, and the prediction target Retinal layer thickness feature amount information for each information provider and prediction target person based on information on the thickness of the person's retinal layer, and information on the visual field sensitivity included in the received provided information And a means for extracting visual field sensitivity feature amount information for each information provider, and information relating to the thickness of the retinal layer and information relating to visual field sensitivity among the information provided for each of the accepted information providers Including at least information provided Pair feature amount information matrix including retinal layer thickness feature amount information based on information related to the thickness of the retinal layer included in the target pair information and information related to visual field sensitivity, and visual field sensitivity feature amount information And the visual field sensitivity feature quantity information of the prediction target person who is the target of the visual field prediction are set as the defect information, and the retinal layer thickness characteristic quantity information of the prediction target person and the visual field sensitivity feature quantity set as the defect information Obtaining a feature quantity information sequence including information and connecting to the pair feature quantity information matrix to generate a combined feature quantity information matrix; and a parameter of a latent variable model for reconstructing the combined pair feature quantity information matrix Estimating information using the combined pair feature quantity information matrix reconstructed based on the parameters of the estimated latent variable model. In which it was decided to output as information.

According to the present invention, visual field sensitivity information can be estimated from retinal layer thickness information.

It is a configuration block diagram showing an example of a visual field sensitivity estimation apparatus according to an embodiment of the present invention. It is a functional block diagram showing the example of the visual field sensitivity estimation apparatus which concerns on embodiment of this invention. It is explanatory drawing showing the example of the data input into the visual field sensitivity estimation apparatus which concerns on embodiment of this invention. It is explanatory drawing showing the example of the content of the process in the visual field sensitivity estimation apparatus which concerns on embodiment of this invention. It is a flowchart figure showing an example of the flow of the process in the visual field sensitivity estimation apparatus which concerns on embodiment of this invention. It is explanatory drawing showing the distribution example of the true value with respect to the predicted value in the visual field sensitivity estimation apparatus which concerns on embodiment of this invention. It is a functional block diagram showing the example of the visual field sensitivity estimation apparatus which concerns on another example of embodiment of this invention.

Embodiments of the present invention will be described with reference to the drawings. The visual field sensitivity estimation apparatus 1 according to the embodiment of the present invention includes a control unit 11, a storage unit 12, an operation unit 13, an output unit 14, and an interface unit 15, as illustrated in FIG. Composed.

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, from a plurality of patients (referred to as information providers), information on the thickness of the retinal layer (examination results), or information on the visual field sensitivity (examination results), The information (age, race, etc.) regarding the attributes of each information provider is provided and input to the visual field sensitivity estimation apparatus 1. The control unit 11 receives provided information including at least one of information related to the thickness of the retinal layer and information related to visual field sensitivity for each information provider.

Further, the control unit 11 accepts information (test result) related to the thickness of the retinal layer for the patient (predictor) who is the target of visual field sensitivity prediction. Then, the control unit 11 approximately reproduces the information relating to the thickness of the retinal layer and the information relating to the thickness of the retinal layer of the prediction target person among the received provision information, for each information provider and prediction target person Retinal layer thickness feature quantity information is extracted by calculation. In addition, the control unit 11 extracts, by calculation, visual field sensitivity feature amount information for each information provider that approximately reproduces information related to visual field sensitivity among the provided information. Here, various methods such as non-negative matrix decomposition (NMF), principal component analysis, and auto-encoder can be used for calculation of retinal layer thickness feature amount information and visual field sensitivity feature amount information.

The control unit 11 uses, as the target pair information, at least part of the provided information including both information related to the thickness of the retinal layer and information related to the visual field sensitivity among the provided information for each information provider. A pair feature amount information matrix including retinal layer thickness feature amount information based on information related to the thickness of the retinal layer included in the image and information related to the visual field sensitivity, and visual field sensitivity feature amount information is generated.

Further, the control unit 11 sets the visual field sensitivity feature amount information of the prediction target person as defect information, and includes feature amount information including the retinal layer thickness feature amount information of the prediction target person and the visual field sensitivity feature amount information set as the defect information. A sequence is generated, and this feature amount information sequence is connected to a pair feature amount information matrix to generate a combined feature amount information sequence. The control unit 11 estimates a parameter of the latent variable model that approximately reproduces the combined feature amount information sequence obtained by the connection, and learns the latent variable model. As the latent variable model, a general latent variable model, structural non-negative matrix decomposition (Structured NMF), Gaussian Bernoulli · restricted Boltzmann machine (Gaussian Bernoulli RBM) can be used.

The control unit 11 uses the estimated parameter to obtain a reconstructed combined feature amount information sequence that is approximately reproduced by the latent variable model. At this time, information of a portion corresponding to the visual field sensitivity feature amount information of the prediction target person in the reproduced combined feature amount information sequence becomes information different from the missing information depending on an approximate reproduction result. The control unit 11 approximately reproduces and outputs the prediction information of the visual field sensitivity as the prediction result by using information of a portion corresponding to the visual field sensitivity characteristic amount information of the prediction target person in the reproduced combined feature value information sequence. . The specific processing contents of the control unit 11 will be described in detail later.

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

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

The interface unit 15 is a USB (Universal Serial Bus), a network interface, or the like, and relates to information (examination result) on the thickness of a patient's retinal layer or visual field sensitivity from an external memory, an examination apparatus, a server, or the like. The information is accepted and output to the control unit 11. The interface unit 15 may send information to be output to an external device or the like in accordance with an instruction input from the control unit 11.

Here, the operation of the control unit 11 will be described. In the example of the present embodiment, as illustrated in FIG. 2, the control unit 11 includes a receiving unit 21, a retinal layer thickness feature quantity information extraction unit 22, a visual field sensitivity feature quantity information extraction unit 23, and a pair feature quantity. The information matrix generation unit 24, the combined feature amount information matrix generation unit 25, the parameter estimation unit 26, the estimation processing unit 27, and the output unit 28 are configured.

In an example of this embodiment, at least one of information related to the thickness of the retinal layer (test result) and information related to visual field sensitivity (test result) is received from a plurality of patients who are information providers. The receiving unit 21 accepts input of provided information including at least one of information related to the thickness of the retinal layer and information related to visual field sensitivity for each information provider.

Specifically, in this description, as information related to the thickness of the retinal layer, the thickness of the nerve fiber layer (NFL) in the macular region of the information provider and the thickness of the ganglion cell layer in the macular region (GCL + IPL) As information relating to visual field sensitivity, visual field sensitivity information (TH) at a plurality of observation points within a predetermined angle (for example, 10 degrees or 30 degrees) of the central visual field is used.

For each information provider, the receiving unit 21 in this example has a nerve fiber layer thickness (NFL) in the macula, a ganglion cell layer thickness (GCL + IPL) in the macula, and a central visual field predetermined angle (here, 10 degrees). And vector data (TH) (hereinafter simply referred to as visual field sensitivity information) having information on visual field sensitivity at a plurality of observation points. The NFL, GCL + IPL, and TH data of each information provider are exemplified in FIGS. 3A, 3N, 3B, GCL + IPL, and FIG. 3C, TH. Can be expressed as two-dimensional image data. The receiving unit 21 according to the present embodiment calculates a pixel value (such as a luminance value) of each pixel included in each two-dimensional image data or a value represented by each pixel (measurement value of retinal layer thickness or visual field sensitivity itself). Each of NFL, GCL + IPL, and TH is one-dimensional vector data arranged in a predetermined order (for example, so-called scan line order) and having a predetermined number of data elements (here, the number of data elements is the number of pixels). Accept data.

The receiving unit 21 is obtained by issuing identification information (which may be a unique identification number or the like) for identifying the information provider for each information provider, and is identified by the identification information with respect to the identification information. Provided by an information provider, the thickness of the nerve fiber layer in the macula (NFL), the thickness of the ganglion cell layer in the macula (GCL + IPL), and the visual field sensitivity at a plurality of observation points within the central visual field at a predetermined angle. The information (TH) is stored in the storage unit 12 in association with it. For information that has not been input, vector data in which arbitrary values (for example, “0”) are arranged as many as a predetermined number of elements is recorded as missing information.

In addition, the receiving unit 21 receives information (examination result) related to the thickness of the retinal layer for a patient (prediction target person) whose visual field sensitivity is to be predicted. The receiving unit 21 is obtained by issuing identification information for identifying the prediction target person (may be a unique identification number or the like as with the information provider), and is provided from the prediction target person for the identification information. The nerve fiber layer thickness (NFL) in the macula and the ganglion cell layer thickness (GCL + IPL) in the macula are stored in the storage unit 12 in association with each other. Regarding the field sensitivity information that is not input, as the missing information, an arbitrary value (hereinafter, this arbitrary value is set to “0” as an example) is arranged by the number of elements determined in advance as the field sensitivity information. Record vector data.

The retinal layer thickness feature quantity information extraction unit 22 includes information (vector data) related to the thickness of the retinal layer that is not missing information among the information (provided information) stored in the storage unit 12 and provided from the information provider. Then, a matrix in which information (vector data) related to the thickness of the retinal layer of the prediction target person is connected is generated. Then, the retinal layer thickness feature amount information extraction unit 22 extracts retinal layer feature amount information that approximately reproduces the generated matrix. In the example of the present embodiment, the retinal layer thickness feature amount information extraction unit 22 includes the NFL vector data (referred to as column vectors) of each information provider that is not missing information and the NFL vector of the prediction target person. Data (column vectors) are connected and arranged in the row direction to obtain a matrix X (NFL) having a size of NFL data elements × (number of information providers + 1). Similarly, the GCL + IPL vector data (column vector) of each information provider is not missing information and the GCL + IPL vector data (column vector) of the prediction target person is connected in the row direction. Arrangement is performed to obtain a matrix X (GCL + IPL) having a size of the number of data elements of GCL + IPL × (the number of information providers + 1).

At this time, the retinal layer thickness feature quantity information extraction unit 22 determines information i (i = 1, 2,...) Specifying a column and a vector corresponding to the i-th column for each of X (NFL) and X (GCL + IPL). The identification information of the data information provider or the person to be predicted is associated and recorded in the storage unit 12.

Then, the retinal layer thickness feature amount information extraction unit 22 performs non-negative matrix decomposition (NMF) on the obtained matrix X (D) (D = NFL, GCL + IPL), and mixes the basis W (D) and the basis. Decompose into product with coefficient H (D).

The retinal layer thickness feature amount information extraction unit 22 outputs the base mixing coefficient H (D) as retinal layer thickness feature amount information. In this equation (1), W (D) has the size of the number of data elements of domain D × the number of feature quantities (number of elements of feature quantities), and H (D) is the number of elements of domain D feature quantities. × (Number of information providers + 1). In addition, the retinal layer thickness feature amount information corresponding to the information provider or the prediction target person in the i-th column of X (D) is a column vector in the i-th column of H (D).

The field-of-view sensitivity feature amount information extraction unit 23 arranges the TH vector data (column vector) of each information provider that is not missing information in the row direction, and the number of TH data elements × the number of information providers A matrix X (TH) of the size At this time, the field-of-view sensitivity feature quantity information extraction unit 23 identifies the information provider (X = 1) that identifies the column i (i = 1, 2,...) And the vector data information provider corresponding to the i-th column. In association with information to be recorded, it is recorded in the storage unit 12.

The visual field sensitivity feature amount information extraction unit 23 performs non-negative matrix decomposition (NMF) on the matrix X (TH) using the equation (1) to obtain a base W (D) and a base mixing coefficient H (D) ( Here, it is decomposed into a product of D = TH). The visual field sensitivity feature amount information extraction unit 23 outputs the base mixing coefficient H (D) as visual field sensitivity feature amount information. Here, the visual field sensitivity feature amount information corresponding to the information provider in the i-th column of X (D) is a column vector in the i-th column of H (D).

The pair feature amount information matrix generation unit 24 is included in each of retinal layer thickness feature amount information H (NFL), H (GCL + IPL), and visual field sensitivity feature amount information H (TH) as schematically illustrated in FIG. The column associated with the identification information of the information provider (referred to as both providers) that provides both the retinal layer thickness information and the visual field sensitivity information is extracted (S1).

The pair feature amount information matrix generation unit 24 sequentially selects the identification information of both providers as attention identification information, retinal layer thickness feature amount information H (NFL), H (GCL + IPL), and visual field sensitivity feature amount information H (TH). The vector data associated with the selected attention identification information extracted from each of the columns is connected and arranged to generate a feature matrix (S2). Thus, the pair feature quantity information matrix generation unit 24 uses a certain mutual provider (a set of identification information {Pk} (k is an integer equal to or less than the number of information providers + the number of prediction subjects (1)). In FIG. 4, feature matrixes Hp (NFL), Hp in which retinal layer thickness feature quantity information and visual field sensitivity feature quantity information relating to Pa, Pb, Pc. (GCL + IPL), Hp (TH) is obtained. Further, the pair feature quantity information matrix generation unit 24 obtains a pair feature quantity information matrix by connecting and arranging these pair feature quantity matrices in the row direction (S3).

The combined feature quantity information matrix generation unit 25 extracts the retinal layer thickness feature quantity information h (NFL) and h (GCL + IPL) of the prediction target person from the retinal layer thickness feature quantity information H (NFL) and H (GCL + IPL). The combined feature amount information matrix generation unit 25 also sets the visual field sensitivity feature amount information h ′ (TH) of the prediction target person as missing information (vector data in which all components are “0”) (( S4)). Then, the combined feature amount information matrix generation unit 25 connects h ′ (NFL), h ′ (GCL + IPL) and h ′ (TH) set as missing information in the column direction to generate an additional information sequence, and FIG. As schematically illustrated, this additional information string is used as additional information h ′ and connected to the pair feature amount information matrix in the row direction to generate a combined feature amount information matrix χ (S5).

The parameter estimation unit 26 estimates the parameters of the latent variable model that approximately reproduces the combined feature amount information matrix χ by calculation. Specifically, the parameter estimation unit 26 approximately decomposes the combined feature amount information matrix χ into a product of a relational parameter matrix F representing a latent variable model and a latent variable matrix B. That is,

Find F and B. This decomposition may also be performed using non-negative matrix decomposition. Here, M is a matrix having the same size as χ, and is a matrix of “0” for an element corresponding to an element that is a missing value of χ, and “1” for the other elements. A symbol with a dot in it means an element product of a matrix (a product is calculated for each corresponding element).

The estimation processing unit 27 calculates a product of the relational parameter matrix F obtained by the parameter estimation unit 26 and the latent variable matrix B to obtain a reconstructed feature quantity matrix χ ″, and is in a range corresponding to the additional information h ′. Extract column vector h ″. Unlike the original combined feature quantity information matrix χ, χ ″ obtained here is changed to the additional information h ′ of the combined feature quantity information matrix χ due to the influence of components other than the additional information h ′ of the combined feature quantity information matrix χ. Of the corresponding column vector h ″, the component of the partial column vector h ″ (TH) that is originally in the range corresponding to h ′ (TH) set as missing information is generally “0” (the above arbitrary value). ) Is a different value.

In addition, the estimation processing unit 27 includes a base W (NFL) related to X (NFL) obtained by the visual field sensitivity feature amount information extraction unit 23, a base W (GCL + IPL) related to X (GCL + IPL), and a matrix X (TH). A base ω is generated by concatenating the base W (TH) related to in the column direction. The estimation processing unit 27 multiplies the base ω by the column vector h ″ extracted from the reconstructed feature quantity matrix χ ″ to obtain predicted visual field sensitivity information x ″ (column vector) (Equation (3)).

The estimation processing unit 27 outputs the column vector x ″ as a prediction result of the visual field sensitivity of the target patient to the output unit 28. The output unit 28 displays the prediction result of the visual field sensitivity of the target patient on the display unit 14. As an example, the output unit 28 obtains the size of the image data (vertical Pv pixel × horizontal Ph pixel) as input visual field sensitivity information, and an image having the same size as this. As shown in FIG. 3C, each pixel of the initialized image data is set based on the value of the corresponding component in the column vector output by the estimation processing unit 27. The image data representing the visual field sensitivity is generated and displayed.

In addition, this output part 28 is the value within the value range nearest to the said value about the value of each component in column vector x "which is a prediction result of the visual field sensitivity of the target patient that exceeds a predetermined value range. For example, when the value range is set to 0 or more and 40 or less, and there is a value of 45 in the column vector x ″, the value is set to a value closest to 45 in the value range. Reset to 40. Similarly, when there is a negative value in the column vector x ″ (when the feature amount information is generated by another method, which is impossible because non-negative matrix decomposition is used here), Reset the value to 0, which is the closest value in the range.

[Basic operation]
The visual field sensitivity estimation apparatus 1 according to the present embodiment basically has the above-described configuration and operates as follows. The user of the visual field sensitivity estimation apparatus 1 in advance receives at least one of information (examination result) relating to the thickness of the retinal layer and information (examination result) relating to visual field sensitivity from a plurality of patients who are information providers. The information is received and input to the visual field sensitivity estimation apparatus 1. The visual field sensitivity estimation apparatus 1 accepts input of provided information including at least one of information related to the thickness of the retinal layer and information related to visual field sensitivity for each information provider.

In the following description, as in the above example, the information relating to the thickness of the retinal layer includes the thickness (NFL) of the nerve fiber layer in the macular region of the information provider, and the ganglion cell layer in the macular region as well. Thickness (GCL + IPL) is used, and information on visual field sensitivity (TH) at a plurality of observation points within a predetermined angle (10 degrees) of the central visual field is used as information on visual field sensitivity, each of which is a one-dimensional vector. Keep it as data.

The visual field sensitivity estimation device 1 issues unique identification information for identifying an information provider for each information provider, and the macular provided from the information provider identified by the identification information to the identification information. A storage unit that correlates the nerve fiber layer thickness (NFL) in the head part, the ganglion cell layer thickness (GCL + IPL) in the macula part, and field sensitivity information (TH) at a plurality of observation points within a predetermined angle of the central field of view. 12. For information that has not been input, vector data in which “0” is arranged for the predetermined number of elements is recorded as missing information.

The user also obtains information (examination result) related to the thickness of the retinal layer for the patient (prediction target person) for which the visual field sensitivity is predicted, and inputs the information to the visual field sensitivity estimation apparatus 1. The visual field sensitivity estimation device 1 issues unique identification information for specifying the prediction target person, and for the identification information, the nerve fiber layer thickness (NFL) in the macular part of the prediction target person and the nerve in the macular part. The node cell layer thickness (GCL + IPL) is stored in the storage unit 12 in association with each other. Note that NFL and GCL + IPL are also set as one-dimensional vector data for the prediction target person, and the field sensitivity information that is not input is set as missing information, and “0” is arranged as the number of elements determined in advance as field sensitivity information. Record vector data.

The field-of-view sensitivity estimation apparatus 1 receives the instruction to predict the field-of-view sensitivity of the person to be predicted, starts the process illustrated in FIG. 5, and first extracts feature amount information that reproduces each vector data (S11). ). Specifically, the field-of-view sensitivity estimation apparatus 1 includes NFL vector data (column vector) of each information provider that is not missing information and NFL vector data (column vector) of the prediction target person. Are connected in the row direction to obtain a matrix X (NFL) having a size of NFL data dimension × (number of information providers + 1). Similarly, the GCL + IPL vector data (column vector) of each information provider is not missing information and the GCL + IPL vector data (column vector) of the prediction target person is connected in the row direction. The matrix X (GCL + IPL) is obtained by arranging the data dimensions of GCL + IPL × (number of information providers + 1).

The field-of-view sensitivity estimation device 1 also provides information i (i = 1, 2,...) For specifying a column and vector data information provider corresponding to the i-th column for each of X (NFL) and X (GCL + IPL). Alternatively, it is recorded in the storage unit 12 in association with the identification information of the person to be predicted.

The field-of-view sensitivity estimation apparatus 1 then performs non-negative matrix decomposition (NMF) on the obtained matrix X (D) (D = NFL, GCL + IPL), respectively, to obtain a base W (D) and a base mixing coefficient H (D And the base mixing coefficient H (D) is output as retinal layer thickness feature quantity information.

The field-of-view sensitivity estimation apparatus 1 arranges TH vector data (column vector) of each information provider that is not missing information in the row direction, and the TH data dimension × the number of information providers. To obtain a matrix X (TH). At this time, the field-of-view sensitivity estimation apparatus 1 includes information i (i = 1, 2,...) For specifying the column of X (TH), and information for identifying an information provider of vector data corresponding to the i-th column. Are stored in the storage unit 12 in association with each other.

The visual field sensitivity estimation device 1 performs non-negative matrix decomposition (NMF) on this matrix X (TH) to obtain a product of the base W (D) and the base mixing coefficient H (D) (here, D = TH). The base mixing coefficient H (D) is output as visual field sensitivity characteristic amount information.

The visual field sensitivity estimation apparatus 1 also includes retinal layer thickness information and visual field sensitivity information among columns included in the retinal layer thickness characteristic amount information H (NFL), H (GCL + IPL), and visual field sensitivity feature amount information H (TH). And the vector data of the column associated with the identification information of the information provider (referred to as the both providers) that provide both, and a predetermined latent variable model (retinal layer) using the extracted vector data A learning process related to a latent variable model related to both the thickness feature amount information and the visual field sensitivity feature amount information is executed (S12).

Specifically, the visual field sensitivity estimation apparatus 1 sequentially selects the identification information of both providers as attention identification information, retinal layer thickness feature amount information H (NFL), H (GCL + IPL), and visual field sensitivity feature amount information H ( Vector data associated with the selected attention identification information extracted from the TH) column are connected and arranged. Thereby, the retinal layer thickness feature amount information and the visual field sensitivity feature amount information related to a certain provider are feature amount matrices Hp (NFL), Hp (GCL + IPL), Hp (TH) arranged in the respective i columns. ) Is obtained. The field-of-view sensitivity estimation apparatus 1 connects these feature quantity matrices Hp (NFL), Hp (GCL + IPL), and Hp (TH) in the column direction to generate a pair feature quantity information matrix.

The visual field sensitivity estimation apparatus 1 also extracts the retinal layer thickness feature amount information h (NFL) and h (GCL + IPL) of the prediction target person from the retinal layer thickness feature amount information H (NFL) and H (GCL + IPL). Further, the visual field sensitivity estimation apparatus 1 sets the visual field sensitivity characteristic amount information h ′ (TH) of the prediction target person as missing information (vector data in which all components are “0”), and h ′ (NFL), h ′. The additional information h ′ is generated by connecting (GCL + IPL) and h ′ (TH) set as missing information in the column direction. The field-of-view sensitivity estimation apparatus 1 generates a combined feature amount information matrix χ by combining additional information h ′ with the pair feature amount information matrix in the row direction.

The visual field sensitivity estimation apparatus 1 calculates model parameters of a latent variable model that approximately reproduces the combined feature amount information matrix χ. Specifically, here, the combined feature amount information matrix χ is approximately decomposed into a product of the relational parameter matrix F representing the latent variable model and the latent variable matrix B using non-negative matrix decomposition.

The visual field sensitivity estimation apparatus 1 obtains visual field sensitivity feature amount information of the prediction target person using the learned latent variable model (S13). Specifically, the visual field sensitivity estimation apparatus 1 calculates a product of the relational parameter matrix F obtained by the above processing and the latent variable matrix B to obtain a reconstructed feature matrix χ ″, and corresponds to the additional information h ′. A column vector h ″ in the range to be extracted is extracted.

The visual field sensitivity estimation apparatus 1 obtains predicted visual field sensitivity information based on the obtained visual field sensitivity feature amount information (S14). Specifically, the base W (NFL) related to X (NFL) obtained in step S11, the base W (GCL + IPL) related to X (GCL + IPL), and the base W (TH) related to the matrix X (TH) are obtained. A base ω is generated by connecting in the column direction. The field-of-view sensitivity estimation apparatus 1 obtains predicted field-of-view sensitivity information x ″ (column vector) by multiplying this base ω by the column vector h ″ extracted from the reconstructed feature quantity matrix χ ″, and this column vector x ″ is the target. It outputs as a prediction result of a patient's visual field sensitivity (S15).

Specifically, the visual field sensitivity estimation device 1 arranges pixels in which pixel values are set based on the values of the respective components of the column vector obtained here, generates image data representing visual field sensitivity information, and outputs the display data. To do.

[Processing example of parameter estimation unit]
In the above example, the domain D (D = NFL, GCL + IPL, TH) is used to approximately decompose the combined feature quantity information matrix χ into the product of the relational parameter matrix F representing the latent variable model and the latent variable matrix B. ), An example in which non-negative matrix decomposition is simply performed without considering a specific factor has been described, but the present embodiment is not limited to this.

For example, considering the existence of a unique factor for each domain D (D = NFL, GCL + IPL, TH), a structural non-negative matrix decomposition (Structured NMF: For example, Hans Laurberg, et. Al., “Structured Non-Negative Matrix Factorization with Sparsity Patterns ”, Proc. Of Asilomar Conference on Signals, Systems, and Computers. (2008)) may be used.

in this case,

The following matrix U (NFL), U (GCL + IPL), U (TH), V (NFL), V (GCL + IPL), and feature column vectors Z (NFL), Z (GCL + IPL), Z (TH) Will be asked. Note that any element of U, V, and Z is non-negative (that is, 0 or a positive number). Χ (NFL) is a partial matrix corresponding to Hp (NFL) or h (NFL) in the combined feature quantity information matrix, and χ (GCL + IPL) is Hp in the combined feature quantity information matrix. (GCL + IPL) or a partial matrix corresponding to h (GCL + IPL), and χ (TH) is a partial matrix corresponding to Hp (TH) or h (TH) in the combined feature quantity information matrix. . And || α || _F ² means computing the Frobenius norm of α.

Also in this example, the matrix obtained as a result of the operation

The column vector

To obtain a reconstructed feature matrix χ ″, and extract a column vector h ″ in a range corresponding to the additional information h ′. Χ ″ obtained here is different from the original combined feature information matrix χ, and is added to the additional information h ′ of the combined feature information matrix χ by the influence of components other than the additional information h ′ of the combined feature information matrix χ. Of the corresponding column vector h ″, the component of the partial column vector h ″ (TH) that is originally in the range corresponding to h ′ (TH) set as missing information is generally a numerical value different from “0”. Become.

In Equation (4), M is a matrix having the same size as χ (TH), and “0” is assigned to elements corresponding to elements that are missing values in χ (TH), and other elements. Is a matrix of “1”, and a symbol with a dot in a circle means an element product of the matrix (operating the product for each corresponding element).

Also in this example, the estimation processing unit 27 uses the basis W (NFL) related to X (NFL) obtained by the visual field sensitivity feature amount information extraction unit 23, the basis W (GCL + IPL) related to X (GCL + IPL), and the matrix X ( TH) is connected to the base W (TH) in the column direction to generate a base ω, and this base ω is multiplied by a column vector h ″ extracted from the reconstructed feature matrix χ ″ to predict visual field sensitivity information x. ″ (Column vector) is obtained.

[Selection of data for model parameter generation]
Furthermore, in an example of the present embodiment, pair feature amount information used for generating model parameters may be selected based on a predetermined condition. This selection is performed as follows, for example.

That is, in an example of this embodiment, the control unit 11 of the visual field sensitivity estimation apparatus 1 performs a clustering process on at least one of NFL and GCL + IPL provided by the information provider and the prediction target person. For this clustering process, a widely known method such as principal component analysis or k-means method can be employed.

The control unit 11 acquires identification information that identifies an information provider who has provided NFL or GCL + IPL, which is determined to belong to the same cluster as at least one of NFL or GCL + IPL of the prediction target person.

The control unit 11 of the visual field sensitivity estimation apparatus 1 operates as the pair characteristic amount information matrix generation unit 24 in the retinal layer thickness feature amount information H (NFL), H (GCL + IPL), and the visual field sensitivity feature amount information H (TH). Among the columns included in each of the above, the identification information of the information provider (both providers) providing both the retinal layer thickness information and the visual field sensitivity information, the column being associated with the acquired identification information Is extracted as target pair information.

Then, the control unit 11 concatenates and arranges the vector data included in the extracted target pair information, and obtains feature quantity matrices Hp (NFL), Hp (GCL + IPL), and Hp (TH). Further, the control unit 11 concatenates these feature amount matrices in the column direction to generate a pair feature amount information sequence, and further concatenates each pair feature amount information sequence in the row direction to arrange the pair feature amount information matrix. obtain. The following processing is executed in the same manner as in the above example.

Of the provision information for each information provider received by the control unit 11 in this way, the provision information includes both information relating to the thickness of the retinal layer and information relating to the visual field sensitivity, and is included in the provision information. Feature information based on the target pair information, with the information related to the thickness of the retinal layer being provided as the target pair information by a predetermined clustering process and belonging to the same cluster as the information related to the thickness of the retinal layer of the prediction target person The model parameter is calculated using the model parameter by selecting the feature amount based on the information provided by the information provider having the same tendency as the prediction target person for the heterogeneous retinal layer thickness information. Thus, the prediction accuracy can be improved.

Further, the clustering method is not limited to the above example. For example, the control unit 11 of the field-of-view sensitivity estimation apparatus 1 is a point on the space represented by the information provider vector using at least one vector of NFL or GCL + IPL provided from the information provider and the prediction target person. Then, a point where a distance (Mahalanobis distance or the like) from a point on the space represented by the vector of the prediction target person is less than a predetermined threshold (or less than a distance exceeding the predetermined number of information providers) is extracted. Then, the control unit 11 extracts identification information that identifies an information provider who provides at least one of NFL and GCL + IPL, which is a vector related to the extracted point, and performs an operation as the pair feature amount information matrix generation unit 24. May be.

[Use attribute information]
In addition, as described above, the conditions for selecting the pair feature information used when generating the model parameters are related to the attributes of the information provider and the prediction target person in addition to the conditions using the result of clustering the information of the retinal layer thickness. Information may be used. Here, attributes are age (may be by age group), gender, race (for example, European, African or Asian), retinal blood vessel position, etc. The parameters of the latent variable model are estimated using the feature amount based on the information provided by the information provider who provides both the retinal layer thickness information and the visual field sensitivity information.

[Modification]
In the above description, examples of using the thickness of the nerve fiber layer (NFL) in the macula and the thickness of the ganglion cell layer (GCL + IPL) in the macula as the information of the retinal layer thickness have been described. In this embodiment, information on the thickness of the photoreceptor layer may be included instead of or in addition to at least one of them.

Further, the visual field sensitivity information (TH) may be data converted to decibels, or data before decibel conversion (linear data) may be used.

[Correction of results]
Further, in the description so far, the visual field sensitivity estimation apparatus 1 according to the present embodiment uses the column vector x ″, which is a prediction result of the visual field sensitivity of the target patient, as information based on the error distribution (for example, an error distribution such as an average error distribution). And the column vector after the correction may be output as a prediction result of the visual field sensitivity of the target patient.

Specifically, in an example of the embodiment of the present invention, the information (true value) of the visual field sensitivity of the target patient actually measured by the examination and the visual field sensitivity calculated on the assumption that the information is unknown. A plurality of sets of prediction results (column vector x ″) are used to obtain a distribution of true values r with respect to the prediction value p (FIG. 6).

The visual field sensitivity estimation apparatus 1 approximates this distribution with a predetermined function (may be approximated by multiple regression such as polynomial regression) to obtain an average distribution r = f (p).

The field-of-view sensitivity estimation apparatus 1 in this example corrects the column vector x ″ after correcting each component ξi (i = 1, 2,...) To ξi ′ = f (ξi) (each component is ξ ′ (vector i) is calculated, and the calculated column vector after correction is output as a prediction result of the visual field sensitivity of the target patient.

According to this example, when there is an average error tendency, it can be corrected by a simple method.

[Example using machine learning model]
In another example of the embodiment of the present invention, the control unit 11 operates according to a program stored in the storage unit 12 and relates to the thickness of the retinal layer from a plurality of patients (referred to as information providers). Information (examination result) and information related to visual field sensitivity (inspection result) and information (age, race, etc.) related to attributes of each information provider are received and input to the visual field sensitivity estimation device 1. The control unit 11 accepts provided information including information related to the thickness of the retinal layer for each information provider and information related to visual field sensitivity.

The control unit 11 accepts provision information including information related to the thickness of the retinal layer of the eye and information related to the visual field sensitivity for each information provider, and uses the information related to the thickness of the retinal layer as learning data. The learning model object in the machine-learned state is held in the storage unit 12 using the information related to the visual field sensitivity of the information provider corresponding to as teacher data. Further, the control unit 11 accepts information related to the thickness of the retinal layer of the prediction target person who is the target of the visual field prediction, and uses the retinal layer thickness feature amount information of the prediction target person as input to the learning model object. Generate sensitivity estimation data.

Specifically, as illustrated in FIG. 7, the control unit 11 according to this example is functionally configured including an acceptance unit 21, a training processing unit 31, an estimation processing unit 27 ′, and an output unit 28. The

Here, the accepting unit 21 accepts input of provision information including information on the thickness of the retinal layer for each information provider and information on the visual field sensitivity.

As in this example of the present embodiment, when using machine learning processing, as information related to the thickness of the retinal layer, the thickness (NFL) of the nerve fiber layer in the macular portion of the information provider, and the macular similarly In addition to the thickness of the ganglion cell layer (GCL + IPL) in the part, the thickness of the rod-shaped cone layer (RCL) is used. Similarly to the example described above, the information related to the visual field sensitivity includes a plurality of observation points (in the following example, arranged in an M × N matrix) within a predetermined angle (for example, 10 degrees or 30 degrees) of the central visual field. The field sensitivity information (TH) at the observation point) is used.

The receiving unit 21 in this example, for each information provider, the thickness of the nerve fiber layer (NFL) in the macula, the thickness of the ganglion cell layer in the macula (GCL + IPL), and the thickness of the rod-shaped cone layer ( RCL) and matrix data (TH) whose components are visual field sensitivity information at a plurality of observation points within a central visual field predetermined angle (here, 10 degrees) (hereinafter simply referred to as visual field sensitivity information). And receive. Note that the NFL, GCL + IPL, and TH data of each information provider are as illustrated in FIG. 3A (NFL), FIG. 3B (GCL + IPL), FIG. 3C, and TH. Furthermore, each can be expressed as two-dimensional image data (here, the number of pixels in the width direction is determined in advance). The receiving unit 21 according to the present embodiment calculates a pixel value (such as a luminance value) of each pixel included in each two-dimensional image data or a value represented by each pixel (measurement value of retinal layer thickness or visual field sensitivity itself). NFL, GCL + IPL, RCL, TH as one-dimensional vector data in which the number of data elements arranged in a predetermined order (for example, so-called scan line order) is predetermined (here, the number of data elements is the number of pixels). Each data is accepted.

The receiving unit 21 is obtained by issuing identification information (which may be a unique identification number or the like) for identifying the information provider for each information provider, and is identified by the identification information with respect to the identification information. Provided by the information provider, the thickness of the nerve fiber layer in the macula (NFL), the thickness of the ganglion cell layer in the macula (GCL + IPL), the thickness of the rod cone layer (RCL), and the center Field sensitivity information (TH) at a plurality of observation points within a predetermined field of view is associated with each other and stored in the storage unit 12, and the training processing unit 31 is instructed to perform a learning process based on the stored information. For information that has not been input, vector data in which arbitrary values (for example, “0”) are arranged as many as a predetermined number of elements may be recorded as missing information.

In addition, the receiving unit 21 receives information (examination result) related to the thickness of the retinal layer for a patient (prediction target person) whose visual field sensitivity is to be predicted. The receiving unit 21 is obtained by issuing identification information for identifying the prediction target person (may be a unique identification number or the like as with the information provider), and is provided from the prediction target person for the identification information. In addition, the nerve fiber layer thickness (NFL) in the macula part, the ganglion cell layer thickness (GCL + IPL) in the macula part, and the rod-shaped cone layer thickness (RCL) are stored in the storage unit 12 in association with each other. . Then, the receiving unit 21 instructs the estimation processing unit 27 ′ to perform the estimation process. Regarding the field sensitivity information that is not input, as the missing information, an arbitrary value (hereinafter, this arbitrary value is set to “0” as an example) is arranged by the number of elements determined in advance as the field sensitivity information. Record vector data.

The training processing unit 31 receives an instruction from the receiving unit 21 that learning processing should be performed, and stores in advance, based on the learning data, information received by the receiving unit 21 and stored in the storage unit 12 as learning data. Machine learning processing of the learning model object stored in the unit 12 is performed. Here, the learning model object is multi-layer neural network data.

In one example of the present embodiment, in this learning model object, at least the first layer to which learning data is directly input is a convolutional layer (CNN: Convolutional Neural Network). Specifically, the learning model object that is subjected to learning processing by the training processing unit 31 of the present embodiment is similar to the learning model object known as VGG16, for example, a convolutional layer that extracts a characteristic distribution of data, and a distribution image And a pooling layer for reducing the size (for example, max pooling is performed here). However, the layer on the output side of the learning model object according to the present embodiment is preferably a loosely coupled layer instead of the fully coupled layer. An example of this output layer will be described later.

Further, it is assumed that the learning data of the learning model object is data for three channels arranged in an X × Y matrix (that is, a data string arranged in 3 × X × Y).

In addition, in consideration of the fact that the number of patients who can be information providers is generally small in the example of the present embodiment, the learning model object has a thickness of the nerve fiber layer (NFL) in the macular region or a ganglion cell in the macular region in advance. General image data (for example, image net (J.Deng, et.al., ImageNet: A large-scale hierarchical image database) different from the layer thickness (GCL + IPL) and the thickness of the rod cone layer (RCL) Learning based on .In (IEEE) Conference (on) Computer (Vision) and Pattern (Recognition, 2009. CVPR (2009, 248-255)) may be performed. At this time, channels of the respective color components of the three primary colors (red R, green G, and blue B) representing the pixels of the image data are assigned to each channel of the learning data. For example, the channel information of the red component of the image data is input to the first channel, the channel information of the green component of the image data is input to the second channel, and so on. Let it be processed. Since such a learning processing method is widely known as a machine learning method for image data, a detailed description thereof will be omitted.

In the learning model object according to the present embodiment, the layer immediately before the output layer has a data string having the same size as the output data (in this example, a two-dimensional array of field sensitivity information (TH) at the time of output) In order to match these, a plurality of data (for example, L) is output (assuming that the data strings are real numbers arranged in an M × N matrix). In the following description, the value of data to be arranged at the position (m, n) of the l-th data string in this data string is denoted as dlmn.

The output layer of the learning model object here is parameters A and B (A is a L × M × N real data string that can be changed by learning, and the position (m, The data value corresponding to n) is expressed as almn, and B is an M × N real data string, and the data value corresponding to the position (m, n) is expressed as bmn), Using the data value dlmn, a data string xmn to be output is generated as follows.

That is, the sum of the product of amn and dmn for l = 1, 2,... L-th data string plus the value of bmn is added to the output data string (m, n ) Is the component xmn corresponding to the position of. As described above, the output layer of the learning model object used in the present embodiment represents a sparse coupling in which the calculation related to only the corresponding component is performed with the previous layer (that is, the position related to the position) Information) layer.

As described above, in the present embodiment, unlike general CNN, the spatial correspondence (OCT-VF) between the optical coherence tomography and the visual field is obtained without fully coupling the final layer and the previous layer. Therefore, the model is suppressed in complexity, overlearning can be prevented, and the prediction rate is improved.

The values of A and B are Xavier's method (Xavier et.al., Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the International Conference on Artificial Intelligence and Statistics ((AISTATSty) Intelligence に よ っ て and Statistics.) Etc.

The training processing unit 31 uses the information received by the receiving unit 21 as learning data, and among the learning data, the nerve fiber layer thickness (NFL) in the macula and the ganglion cell layer thickness (GCL + IPL) in the macula The information about the thickness (RCL) of the rod-like cone layer is input to the learning model object. Here, the training processing unit 31 sets the thickness (NFL) of the nerve fiber layer in the macula to the first channel, the thickness of the ganglion cell layer (GCL + IPL) in the macula to the second channel, and the rod-shaped cone layer. Thickness (RCL) information is input separately for each channel of data to be input, such as the third channel.

In addition, the training processing unit 31 has a real data string xmn arranged in an M × N matrix output from the output layer of the learning model object as a result of the input, and the visual field sensitivity that is correct data included in the learning data. The difference between the information (TH) and the corresponding position data (when the data at the position (m, n) of the field sensitivity information (TH) is expressed as THmn, the sum of the absolute values of THmn−xmn) or its square Using the general backpropagation method with the sum as a loss, the parameters A and B of the output layer and the convolution layer included in the learning model object are updated (generally, the pooling layer has Although there is no parameter to be updated, it is not described here, but it includes layers other than those described here, and the layer includes parameters to be updated by backpropagation. If this is the case, the parameters of the layer are also updated here).

Note that various methods, for example, a so-called momentum method can be adopted as the method of updating (back propagation), and thus detailed description thereof is omitted here.

The estimation processing unit 27 ′ receives an instruction from the receiving unit 21 that the estimation process should be performed, and receives the information received by the receiving unit 21 and stored in the storage unit 12 (the prediction target person's nerve fiber layer in the macular region). After performing the training process in the training processing unit 31 using the thickness (NFL), the thickness of the ganglion cell layer in the macula (GCL + IPL), and the thickness of the rod-shaped cone layer (RCL) as input data Input for the learning model object.

Similar to the training processing unit 31, the estimation processing unit 27 ′ uses the nerve fiber layer thickness (NFL) in the macular region as the first channel, and the ganglion cell layer thickness (GCL + IPL) in the macula region as the second channel. Information on the thickness (RCL) of the body cone layer is input separately for each channel of data to be input, such as the third channel.

In addition, the estimation processing unit 27 ′ uses the real data string xmn arranged in an M × N matrix output from the output layer of the learning model object as a result of the input as it is, and the field of view of the target patient (prediction target person). Output as a prediction result of sensitivity information (TH). That is, the output unit 28 according to this example of the present embodiment outputs the data THmn at the position (m, n) of the visual field sensitivity information (TH) as the prediction result to be output. The position sensitivity value xmn is set, and information on the visual field sensitivity prediction result is output as M × N image data.

In the description so far, the learning model object having the convolutional layer is obtained by providing the thickness of the nerve fiber layer in the macula (NFL) and the thickness of the ganglion cell layer in the macula (GCL + IPL) provided by the information provider. Learning based on the thickness (RCL) of the rod-shaped cone layer and the field sensitivity information (TH) at a plurality of observation points within a predetermined angle of the central field of view. The model object is not limited to this example.

For example, learning may be performed as follows using a learning model object including two to three layers of all connected layers (referred to here as a small object for distinction). That is, the nerve fiber layer thickness (NFL) in the macula, the ganglion cell layer thickness (GCL + IPL), and the rod cone layer thickness (RCL) provided by the information provider An output obtained by inputting information to a learning model object having a convolutional layer learned by the above-described method is obtained as teacher data. In addition, the nerve fiber layer thickness (NFL) in the macula, the ganglion cell layer thickness (GCL + IPL), and the rod cone layer thickness (RCL) provided by the information provider to the small object ) And the output obtained by inputting the information and the difference between the teacher data and the small object may be subjected to learning processing. Since such a method is a process widely known as a so-called “distillation method”, a detailed description thereof is omitted here.

In an example of the present embodiment, the estimation processing by the estimation processing unit 27 ′ may be executed using such a small object.

DESCRIPTION OF SYMBOLS 1 Field-of-view sensitivity estimation apparatus, 11 Control part, 12 Memory | storage part, 13 Operation part, 14 Output part, 15 Interface part, 21 Accepting part, 22 Retina layer thickness feature-value information extraction part, 23 Field-of-view sensitivity feature-value information extraction part, 24 Pair feature quantity information matrix generation section, 25 combined feature quantity information matrix generation section, 26 parameter estimation section, 27, 27 'estimation processing section, 28 output section, 31 training processing section.

Claims (7)

1. Provided information including at least one of information relating to the thickness of the retinal layer of the eye and information relating to the visual field sensitivity for each information provider, and information relating to the thickness of the retinal layer of the prediction target person to be subject to visual field prediction Means to accept,
   Based on the information related to the thickness of the retinal layer included in the received provision information and the information related to the thickness of the retinal layer of the prediction target person, the retinal layer thickness feature amount for each information provider and prediction target person A means of extracting information;
   Means for extracting visual field sensitivity feature amount information for each information provider based on information relating to the visual field sensitivity included in the received provision information;
   The retina included in the target pair information with at least a part of the provided information including both the information related to the thickness of the retinal layer and the information related to the visual field sensitivity as the target pair information. Means for generating a pair feature amount information matrix including retinal layer thickness feature amount information based on information relating to layer thickness and information relating to visual field sensitivity, and visual field sensitivity feature amount information;
   Feature quantity information including field sensitivity feature amount information of a prediction target person to be subject to visual field prediction is set as defect information, and includes retinal layer thickness feature quantity information of the prediction target person and field sensitivity feature amount information set as the defect information Means for obtaining an information string and concatenating the pair feature quantity information matrix to generate a combined pair feature quantity information matrix;
   Means for estimating parameters of a latent variable model for reconstructing the combined pair feature information matrix;
   Including
   Field-of-view sensitivity estimation that estimates information related to the field-of-view sensitivity that is the missing information using a combined pair feature amount information matrix reconstructed based on the parameters of the estimated latent variable model, and outputs the information as field-of-view sensitivity information apparatus.
2. The visual field sensitivity estimation apparatus according to claim 1,
   The means for generating the pair feature amount information matrix is provided information including both information related to the thickness of the retinal layer and information related to visual field sensitivity among the provided information for each received information provider, Visual field sensitivity estimation in which the information related to the thickness of the retinal layer included in the provided information is subject to the provided information belonging to the same cluster as the information related to the thickness of the retinal layer of the prediction target by the predetermined clustering process apparatus.
3. The visual field sensitivity estimation apparatus according to claim 1 or 2,
   The visual field sensitivity estimation apparatus, wherein the means for estimating the parameters of the latent variable model estimates the parameters of the latent variable model for reconstructing the combined pair feature quantity information matrix by using the operation of structural non-negative matrix decomposition.
4. Information corresponding to learning data by accepting provided information including information related to the thickness of the retinal layer of the eye and information related to visual field sensitivity for each information provider, and using the information related to the thickness of the retinal layer as learning data Means for holding a learning model object in a machine-learned state using information related to the visual field sensitivity of the provider as teacher data,
   Means for accepting information relating to the thickness of the retinal layer of the prediction target subject to visual field prediction;
   Generating means for generating visual field sensitivity estimation data for the input using the prediction target person's retinal layer thickness feature amount information as an input to the learning model object;
   Including
   A visual field sensitivity estimation device that outputs the estimation data as visual field sensitivity information.
5. The visual field sensitivity estimation apparatus according to claim 4,
   The visual field sensitivity estimation apparatus, wherein the learning model object includes a multilayer neural network, and at least a first layer to which the learning data is directly input is a convolutional layer.
6. For the visual field sensitivity estimation device, provided information including at least one of information related to the thickness of the retinal layer of the eye for each information provider and information related to visual field sensitivity, and a prediction target person to be subject to visual field prediction Inputting information relating to the thickness of the retinal layer;
   The visual field sensitivity estimation device is configured to provide an information provider and a prediction target person based on the information related to the thickness of the retinal layer included in the received provision information and the information related to the thickness of the retinal layer of the prediction target person. Extracting retinal layer thickness feature information for each,
   The visual field sensitivity estimation device extracts visual field sensitivity feature amount information for each information provider based on information related to the visual field sensitivity included in the received provision information;
   Of the provided information for each of the accepted information providers, the visual field sensitivity estimation device sets at least a part of the provided information including both information related to the thickness of the retinal layer and information related to visual field sensitivity as target pair information. Generating a pair feature amount information matrix including retinal layer thickness feature amount information based on information related to the thickness of the retinal layer included in the target pair information and information related to visual field sensitivity, and visual field sensitivity feature amount information;
   The visual field sensitivity estimation device sets the visual field sensitivity feature amount information of the prediction target person who is the target of visual field prediction as defect information, the retinal layer thickness characteristic amount information of the prediction target person, and the visual field sensitivity feature set as the defect information Obtaining a feature quantity information sequence including quantity information, and connecting the pair feature quantity information matrix to generate a combined pair feature quantity information matrix;
   The visual field sensitivity estimation device estimating a parameter of a latent variable model for reconstructing the combined pair feature information matrix;
   Including
   Field-of-view sensitivity estimation device for estimating information relating to visual field sensitivity, which is the missing information, using parameters of the latent variable model reconstructed based on the parameters of the estimated latent variable model, and outputting the information as visual field sensitivity information Control method.
7. Computer
   Provided information including at least one of information relating to the thickness of the retinal layer of the eye and information relating to the visual field sensitivity for each information provider, and information relating to the thickness of the retinal layer of the prediction target person to be subject to visual field prediction Means to accept,
   Based on the information related to the thickness of the retinal layer included in the received provision information and the information related to the thickness of the retinal layer of the prediction target person, the retinal layer thickness feature amount for each information provider and prediction target person A means of extracting information;
   Means for extracting visual field sensitivity feature amount information for each information provider based on information relating to the visual field sensitivity included in the received provision information;
   The retina included in the target pair information with at least a part of the provided information including both the information related to the thickness of the retinal layer and the information related to the visual field sensitivity as the target pair information. Means for generating a pair feature amount information matrix including retinal layer thickness feature amount information based on information relating to layer thickness and information relating to visual field sensitivity, and visual field sensitivity feature amount information;
   Feature quantity information including field sensitivity feature amount information of a prediction target person to be subject to visual field prediction is set as defect information, and includes retinal layer thickness feature quantity information of the prediction target person and field sensitivity feature amount information set as the defect information Means for obtaining an information string and concatenating the pair feature quantity information matrix to generate a combined pair feature quantity information matrix;
   Means for estimating parameters of a latent variable model for reconstructing the combined pair feature information matrix;
   Function as
   A program for estimating information relating to visual field sensitivity, which is the defect information, using a combined pair feature amount information matrix reconstructed based on the parameters of the estimated latent variable model, and outputting the information as visual field sensitivity information.
   

Images