JP2019152977A - Information processing apparatus, method for producing machine learning data, and program - Google Patents

Abstract

To provide an information processing apparatus, a method for producing machine learning data, and a program which allow machine learning to be improved in efficiency and be useful for processing such as diagnosis by performing machine learning processing while effectively utilizing even information not paired with diagnostic information relating to an eye.SOLUTION: Information representing a pattern of information about a symptom of an eye is acquired, and an output of a neural network resulting from input of diagnostic information concerning the eye, information about difference from information relating to the symptom of the eye corresponding to the input diagnostic information, and information concerning the acquired pattern of the information about the symptom of the eye are used to control a parameter of the neural network, thereby generating machine learning data.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an information processing apparatus that processes ophthalmic medical information, a method for manufacturing machine learning data, and a program.

  In recent years, machine learning has been used for various information processing, and for medical information, information processing using the results of machine learning has been proposed (for example, Patent Document 1).

  On the other hand, early detection is required for diseases associated with irreversible loss of visual function such as glaucoma. However, since glaucoma is generally time-consuming and requires a large amount of time for a definitive diagnosis, there is a need for a method for simply detecting the possibility of glaucoma by screening in advance.

JP, 2015-116319, A

  In the above-described apparatus using conventional machine learning, so-called supervised learning processing is performed using data in which input data and output data are paired. Regardless of the data to be output, there are many examples in which the output data is stored as information of the inspection result, and the input data and the output data are not necessarily in pairs.

  For example, in the diagnosis of glaucoma, there have been many accumulations of visual field examination results from the past, but information such as the retinal structure is not necessarily held in pairs with the results of visual field examinations.

  The present invention has been made in view of the above circumstances, and performs machine learning processing while effectively utilizing unpaired information, and improves the efficiency of the machine learning, and is useful for processing such as diagnosis. It is an object of the present invention to provide an information processing apparatus, a machine learning data manufacturing method, and a program that can be used.

  The present invention for solving the problems of the conventional example is an information processing apparatus, an acquisition unit for acquiring information representing a pattern of output information, such as information on a patient's symptom such as an eye symptom, and eye diagnosis information A means for generating machine learning data for associating input information such as patient diagnosis information with corresponding output information, and corresponding to the input of the learning device when the input information is input Learning processing means for generating machine learning data by controlling parameters of the learning device using information relating to the difference from the output information and information relating to the pattern of the received input information, the learning processing means The machine learning data generated by the above is used for processing for estimating output information such as information related to symptoms corresponding to input information such as diagnostic information to be processed.

  According to the present invention, it is possible to perform machine learning processing while effectively using information in which input information and output information are not paired, and to improve the efficiency of the machine learning. An information processing apparatus that can be used for processing such as diagnosis is provided.

1 is a configuration block diagram of an information processing apparatus according to an embodiment of the present invention. It is a functional block diagram concerning an example of an information processor concerning an embodiment of the invention. It is a flowchart figure showing the operation example of the information processing apparatus which concerns on embodiment of this invention. It is explanatory drawing showing the example of the information which the information processing apparatus which concerns on embodiment of this invention outputs.

  Embodiments of the present invention will be described with reference to the drawings. As illustrated in FIG. 1, the information processing apparatus 1 according to the embodiment of the present invention includes a control unit 11, a storage unit 12, an operation unit 13, a display unit 14, and an input / output unit 15. Composed.

  Here, 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 acquires information representing a pattern of output information, and inputs the input information when generating machine learning data that associates the input information with the corresponding output information. Machine learning data by controlling parameters of the learning device using information on the difference between the information output by the learning device and the output information corresponding to the input information, and the acquired information on the pattern of the output information Is generated. Then, the control unit 11 provides the generated machine learning data to a process for estimating output information corresponding to the input information to be processed.

  Specifically, in an example of the present embodiment, the control unit 11 acquires information representing an information pattern related to eye symptoms as output information. Further, at this time, the control unit 11 generates machine learning data associating the diagnostic information about the eyes as the input information with the information about the corresponding eye symptoms. Specifically, the control unit 11 corresponds to an output (output information) of a learning device that uses a nonlinear function, such as a neural network when diagnostic information relating to the eyes is input as input information, and diagnostic information that is the input information that has been input. The machine learning data is generated by controlling the parameters of the learning device using the information relating to the difference between the information relating to the eye symptom and the information relating to the information pattern relating to the eye symptom acquired.

  And this control part 11 may perform the process which estimates the information regarding the symptom of eyes corresponding to the diagnostic information which concerns on the machine learning data produced | generated by the learning process which is the object of a process. Details of the operation of the control unit 11 will be described later.

  In the following description of the present embodiment, the input information and the output information are described as those in the above example, but this is an example, and the same applies to other types of input information and output information. Processing can be applied. For example, the input information may be patient diagnosis information related to a part other than the eyes. Further, the output information may be information regarding symptoms related to parts other than the eyes. At this time, the part of the patient related to the input information may be different from the part of the patient related to the output information. Furthermore, examples of input information and output information according to the present embodiment are not limited to information related to the human body, but various fields to which a learning device using a nonlinear function is applied, such as language analysis, management information analysis, and weather information. However, any information can be used as long as it can extract pattern information about output information. Here, the output information may be multidimensional information.

  The storage unit 12 is a disk device, a memory device, or the like, and holds a program executed by the control unit 11. The program may be provided by being stored in a computer-readable non-transitory recording medium and stored in the storage unit 12.

  The operation unit 13 is a keyboard, a mouse, or the like, and accepts a user instruction and outputs it to the control unit 11. The display unit 14 is a display or the like, and displays and outputs information in accordance with an instruction input from the control unit 11.

  The input / output unit 15 is an input / output interface such as a USB (Universal Serial Bus), for example, and in the example of the present embodiment, various types of information to be processed are transferred to external devices (for example, various inspection devices and card readers). Etc.) and output to the control unit 11.

  Next, the operation of the control unit 11 will be described. In the present embodiment, the control unit 11 includes a learning processing unit 20 and an estimation processing unit 30 as illustrated in FIG. The learning processing unit 20 includes a pattern acquisition unit 21 and a learning data generation unit 22.

  In an example of the present embodiment, the pattern acquisition unit 21 accepts a plurality of pieces of information related to eye symptoms and represents a pattern of information related to eye symptoms obtained by unsupervised learning based on information related to the plurality of eye symptoms. Generate and obtain information.

  Specifically, the pattern acquisition unit 21 accepts a plurality of pieces of information related to the visual field as information related to eye symptoms. Here, the information on the visual field is information on the result of visual field inspection, that is, information on visual field sensitivity at a plurality of measurement points P1, P2,... Preset within a predetermined angle (for example, 10 degrees or 30 degrees) from the visual field center. Good. In the following example, it is assumed that each measurement point can be identified by unique identification information (serial number or the like).

  The pattern acquisition unit 21 extracts a symptom progression pattern by unsupervised learning based on information about the accepted visual field. For example, in the case of glaucoma, it is known that cases can be classified into several patterns, such as cases in which visual field defects progress from the upper side and cases from the lower side. In any case, it is known that relatively similar visual field sensitivity information can be obtained between visual field measurement points located relatively close to each other (having co-occurrence).

  In view of this, the pattern acquisition unit 21 uses, for example, non-negative matrix factorization (NMF), principal component analysis (PCA), an auto encoder, a method using a mixture distribution, and the like as unsupervised learning. Extract patterns of symptom progression from information. As an example, when non-negative matrix factorization is used, information on visual field sensitivity at each measurement point included in each piece of information on visual field is set in a predetermined order. That is, the pattern acquisition unit 21 converts the field sensitivity information included in the information about a plurality of fields (hereinafter, N (N is an integer of 2 or more)) to the corresponding measurement points Pi (i = 1, 2,...). ) In the order of vector information Yj (j = 1, 2,..., N).

The pattern acquisition unit 21 is a matrix in which column vectors of the vector information Yj (j = 1, 2,..., N) are connected and arranged.
This matrix Y is decomposed into a product of a pair of non-negative matrix WY and ΘY. That is,
And Here, ΘY is a base matrix obtained by transposing a matrix in which K column vectors θY are connected and arranged.
It is. Each row (each column θYi (i = 1, 2,... K) before transposition) included in the basis matrix ΘY represents a progress pattern.

  A specific method of this non-negative matrix factorization is described in DD Lee and HS Seung, Algorithms for Non-negative Matrix Factorization, In NIPS, Vol. 13, pp. 556-562, 2000, etc. Detailed explanation in is omitted.

  The pattern acquisition unit 21 according to this example of the present embodiment outputs the base matrix ΘY obtained here as information representing a pattern of information regarding eye symptoms.

  The learning data generation unit 22 learns a predetermined neural network as a learning device. The neural network here is, for example, a CNN (convolutional neural network), which is machine learning data that associates diagnostic information relating to the eyes with information relating to eye symptoms corresponding thereto. In this example of the present embodiment, the information related to the eye symptom may be a result of visual field examination expressed in the form of a vector as described above.

  The diagnostic information on the eye is, for example, any of fundus measurement results (information on the structure of the retina) or image data of fundus photographs by an optical coherence tomography (OCT). It is assumed that it is a fundus measurement result by a total meter (OCT). In the following description, a case where CNN is used as a neural network will be described as an example. However, the neural network according to the present embodiment is not limited to CNN, and a residual network (ResNet: Kaiming He, et.al. , Deep Residual Learning for Image Recognition, https://arxiv.org/pdf/1512.03385v1.pdf), etc. However, the output of the neural network is set to be a vector quantity having the same dimension as the dimension of the vector representing the visual field inspection result.

  The learning data generation unit 22 accepts an input of a set of diagnostic information relating to the eye and information relating to an eye symptom when the diagnostic information is obtained. This set of information is obtained from a specific subject who has performed an examination by OCT and a visual field examination substantially simultaneously (for example, within a period of several days). Although the number of such information sets is smaller than the total number of information related to eye symptoms (what is not paired with diagnostic information) used by the pattern acquisition unit 21, it is sufficient for generating learning data. There are as many as there are.

The learning data generation unit 22 sequentially selects a set of diagnostic information related to the accepted eye and information related to the eye symptom that is correct information when the diagnostic information is obtained, and diagnostic information included in the selected set Is the input to the neural network, the information output by the neural network (information that can be calculated from the information about the eye symptom) ypred and the information about the eye symptom included in the selected set (correct information) Find the squared residual average (L2 penalty term) l ₂ (ypred, ytrue) with ytrue.

The learning data generation unit 22 calculates the pattern regularization term R (ypred) as follows. In other words, this pattern regularization term
It is defined as Where ypredNMF is
It is prescribed by. However,
And

  This corresponds to the projection of ypred, which is the prediction result of the neural network, orthographically projected onto the subspace created by the symptom progression pattern (base), and the pattern regularization term as a whole is the result of this orthographic projection and the neural network. The square of the Euclidean distance from ypred, which is the net prediction result.

  As a result, as the prediction result deviates from the symptom progression pattern, the Euclidean distance between the orthographic projection of the prediction result to the partial space and the prediction result itself becomes larger, and larger regularization is performed.

The learning data generation unit 22 sums the L2 penalty term l ₂ (ypred, ytrue) and the pattern regularization term R (ypred) based on the set of the diagnostic information and the information about the eye symptoms (weighted sum l ₂ ( ypred, ytrue) + λ · R (ypred) (where λ is a real number, which is an example of a weight)), the parameters of the neural network are controlled and reset.

  As such a setting method, various known methods can be appropriately selected and adopted as a gradient descent method in a general neural network, and detailed description thereof is omitted here.

  The learning data generation unit 22 repeats the above-described processing for each set of the diagnostic information related to the accepted eye and the information related to the eye symptom that is correct information when the diagnostic information is obtained, Machine learning data is generated by controlling network parameters.

  In addition, the estimation processing unit 30 of the control unit 11 uses the machine learning data generated by the learning processing unit 20 and estimates information related to eye symptoms using the same neural network as the learning processing unit 20. Specifically, in this example, machine learning data is obtained by using fundus measurement results (information on the structure of the retina) obtained by an optical coherence tomography (OCT) as diagnostic information about the eyes. Also, the fundus measurement result obtained by the optical coherence tomography (OCT) of the subject to be diagnosed is input to a neural network in which parameters are set by machine learning data. The information output by the neural network (here, each element is information on a vector representing the visual field sensitivity at each measurement point in the field of view) is used to estimate the visual field sensitivity at each measurement point in the field of view (that is, regarding eye symptoms). Information estimation result).

  In the example of the present embodiment, the control unit 11 includes both the learning processing unit 20 and the estimation processing unit 30, but the control unit 11 of the present embodiment includes only the learning processing unit 20. Thus, machine learning data may be output.

  In this case, the process as the estimation processing unit 30 using the machine learning data output from the learning processing unit 20 is executed in another computer device.

  Here, the neural network used by the estimation processing unit 30 and the learning processing unit 20 is the same, but is not limited to this example. For example, the learning processing unit 20 is a CNN, and the estimation processing unit 30 is a two-layer full connection. A network may be used. In this case, the neural network used by the estimation processing unit 30 is learned using the result of estimation processing using the machine learning data output from the learning processing unit 20 and the neural network used by the learning processing unit 20. Distillation is performed to obtain machine learning data which is a parameter of the neural network used by the estimation processing unit 30 (referred to as distillation data for distinction from the machine learning data output by the learning processing unit 20). Then, the estimation processing unit 30 uses the distillation data and the neural network used during the distillation, and inputs the fundus measurement result obtained by the optical coherence tomography (OCT) of the subject to be diagnosed to the neural network. Obtain and output visual field sensitivity estimation results (that is, estimation results of information on eye symptoms) at each measurement point in the visual field.

[Operation]
This embodiment has the above-described configuration and operates as follows. When the information processing apparatus 1 according to the embodiment of the present invention receives an instruction to perform learning processing, the information processing apparatus 1 substantially performs inspection by an optical coherence tomography (OCT) and visual field inspection as illustrated in FIG. A set of fundus measurement results (information on the structure of the retina) and results of visual field examination by optical coherence tomography (OCT) obtained from a specific subject performed at the same time (for example within a period of several days) ( In the following description, a plurality of inputs of a pair of data (hereinafter referred to as “pair data”) and information on the results of visual field inspection (hereinafter referred to as “non-pair data”) that are not paired with the results of inspection by optical coherence tomography (OCT) are accepted. (S1).

  Then, the information processing apparatus 1 performs unsupervised learning processing such as non-negative matrix factorization (NMF) based on the non-pair data received in step S1, and generates and obtains information representing a pattern of information regarding eye symptoms. (S2). As an example, in this process S2, by using non-negative matrix factorization (NMF), several different representative patterns are obtained as a basis as information representing the pattern of information on the result of visual field examination, which is information on eye symptoms. .

  Next, the information processing apparatus 1 sets a CNN neural network (S3). In this setting, if there is machine learning data obtained by learning in the past, the parameter of the machine learning data is set in the neural network of the CNN (the learned model is loaded) and obtained by learning in the past. If there is no machine learning data, the CNN neural network is initialized. A widely known method may be adopted as the initialization method.

  Then, the information processing apparatus 1 selects a plurality of sets of pair data received in step S1 one by one and repeatedly executes the next loop (S4). That is, the information processing apparatus 1 obtains an output when the fundus measurement result (information on the structure of the retina) included in the selected pair data is input to the CNN neural network (S5).

The information processing apparatus 1 obtains a square residual average (l ₂ (ypred, ytrue)) between the output obtained in the processing S5 and information on the result of visual field inspection that is correct information included in the selected pair data, Further, a value (R (ypred)) for evaluating the compatibility between the information representing the pattern obtained in the process S2 and the information obtained from the output obtained in the process S5 is obtained. For example, this value is obtained by orthogonally projecting the output of the neural network into a partial space created by the information (base) representing the pattern obtained in the process S2, and the square of the Euclidean distance between the result of the orthogonal projection and the output of the neural network, etc. A value that increases as the output of the neural network deviates from the pattern represented by the information obtained in step S2 is used.

The information processing apparatus 1 calculates a loss function using the squared residual average l ₂ (ypred, ytrue) obtained here and a value R (ypred) indicating suitability to the pattern, and calculates the loss function. Based on the result, the parameters (machine learning data) of the neural network are updated (S6). Here, the loss function may be a weighted sum of the above-mentioned square residual average and a value representing suitability to the pattern. The update process itself based on the loss function can be selected by appropriately selecting various known methods as a gradient descent method in a general neural network.

  The information processing apparatus 1 repeatedly executes the processes S5 and S6 for each of a plurality of pairs of data received in the process S1, and when there is no unselected pair data, the machine learning data ( The learned model) is stored in the storage unit 12, and the process is terminated.

  This machine learning data may be used for processing for estimating information related to eye symptoms corresponding to diagnostic information relating to eyes to be processed by being output to an external computer, or the information processing apparatus 1 itself. May be used to perform processing (estimation processing) for estimating information regarding eye symptoms corresponding to diagnostic information relating to the eye to be processed.

  When receiving the instruction to perform the estimation process, the information processing apparatus 1 reads the machine learning data (learned model) stored in the storage unit 12 and applies the same CNN neural network as that used in the learning process. Set the machine learning data parameters (load the trained model).

  The information processing apparatus 1 receives information on the result of the inspection by the optical coherence tomography (OCT) and inputs the information on the result of the inspection by the optical coherence tomography (OCT) to the set neural network. The information processing apparatus 1 obtains information (information corresponding to the estimated visual field inspection result) output from the neural network with respect to the input, and outputs the information by displaying the information on the display unit 14.

  As described above, according to the present embodiment, it becomes possible to perform learning processing of a neural network that contributes to symptom estimation by effectively using non-pair data accumulated in the past, and can improve the efficiency of machine learning. The result of the machine learning can be used for processing such as diagnosis.

[Other examples of pattern extraction]
Further, the method of extracting the symptom progression pattern by unsupervised learning is not limited to the non-negative matrix factorization (NMF) method.

  The unsupervised learning method used by the control unit 11 of the information processing apparatus 1 may be principal component analysis (PCA) or an auto encoder.

Specifically, when the principal component analysis is used, the control unit 11 of the information processing apparatus 1 learns the principal component vector ΘY from the result of the visual field inspection. Then, the control unit 11 uses this principal component vector to calculate a value R (ypred) representing the degree of coincidence between the estimation result ypred output from the neural network and the pattern obtained by the principal component analysis.
However,
It prescribes as In addition,
Is an average for each element of visual field inspection information obtained from the visual field inspection result (a vector in which averages of information on visual field sensitivity at each measurement point are arranged).

When the auto encoder is used, the control unit 11 of the information processing apparatus 1 uses a value R (ypred) representing the degree of coincidence between the estimation result ypred output from the neural network and the pattern obtained by the auto encoder.
It prescribes as

  Here, the function fY (X) represents the output of the auto encoder when the vector corresponding to the result of the visual field inspection is input to the learned auto encoder. The neural network used in the auto-encoder here is, for example, a hidden layer having only one layer, a fully connected layer having a lower dimension than the input layer, and an activation function in the input layer, for example, a sigmoid function, The activation function in the hidden layer is set as the tanh function. Since it is an auto encoder, the output layer has the same dimensions as the input layer.

  The value R (ypred) representing the degree of coincidence with the pattern obtained by the auto encoder in this example is that the output of the neural network that is to learn the pair data is input to the neural network of the auto encoder that has been learned with the non-pair data. Then, the input information and the output of the neural network of the auto encoder (projected to match the pattern, that is, the output of the neural network trying to learn the pair data into the subspace learned from the visual field inspection result by the auto encoder) The greater the difference from the information (corresponding to the information), the more the pattern does not match the pattern.

  Furthermore, in an example of the present embodiment, a Gaussian model (GM) may be used as a method for unsupervised learning. In this example, the control unit 11 of the information processing apparatus 1 obtains an average vector μY for each element of the vector from vector information of a plurality of visual field inspection results included in the non-pair data, and estimates a variance covariance matrix ΣY. . A widely known method can be adopted as the method for estimating the variance-covariance matrix, and a detailed description thereof will be omitted here.

The control unit 11 is characterized by the average vector μY and the variance-covariance matrix ΣY of the output of the neural network trying to learn the pair data assuming that the visual field inspection result is generated according to the Gaussian distribution. Using the Mahalanobis distance, the value R (ypred) representing the degree of coincidence with the Gaussian model is
It prescribes as

As described above, the control unit 11 uses the pattern obtained by any one of the methods using any of non-negative matrix factorization (NMF), principal component analysis (PCA), auto encoder, or Gaussian model (GM). And the value R (ypred) of the pattern regularization term, which is a value representing the degree of coincidence between the data and the output of the neural network trying to learn the pair data, and the L2 penalty based on the set of diagnostic information and information on eye symptoms In order to reduce the sum of the term l ₂ (ypred, ytrue) and this pattern regularization term R (ypred) (the weighted sum l ₂ (ypred, ytrue) + λ · R (ypred) may be sufficient) The neural network learning process is performed by controlling and resetting the network parameters.

[Example of specifying information representing the pattern in advance]
In the description so far, the pattern related to the output of the neural network represented by the non-pair data is estimated based on the non-pair data, and the neural network learning process is performed so that the pattern matches the output of the neural network. However, the present embodiment is not limited to this.

  That is, the information representing the pattern (pattern normalization term R (ypred)) may be arbitrarily set by the user based on prior knowledge instead of non-pair data.

  As an example, in the visual field inspection result, taking into account that the measurement results at adjacent measurement points tend to correlate, control the parameters of the neural network in a direction to reduce the difference between the values of adjacent measurement points. It is good also as performing the learning process of a neural network by resetting. In this example, for each measurement point, for example, the difference between the estimation result of the value of the measurement point (output of the neural network) and the estimation result of the value of the adjacent measurement point is obtained, and the square of the difference obtained at each measurement point is obtained. The sum may be R (ypred).

[Visualization processing]
Using the machine learning results described above, in addition to diagnostic processing, input information relating to the patient's condition (in the above example, fundus measurement results (retinal retina) using optical coherence tomography (OCT) in the above example) Processing for visualizing the relationship between information relating to the structure) (diagnostic information relating to the eye such as image data of fundus photographs) and output information relating to the patient's state (in the above example, information corresponding to the result of visual field examination) Can also be done.

  That is, the information processing apparatus 1 according to the present embodiment uses undifferentiated input / output relationship differentiation by a non-linear learning device such as a neural network machine-learned by the above-described method, and performs unsupervised learning on input information related to a patient's state. A plurality of patterns (input information patterns) Pini (i = 1, 2,...) And a plurality of patterns (output information patterns) Poutj (j = j =) obtained by unsupervised learning on the output information related to the patient's condition. , 1, 2,...) Is determined as a weight Dij (a weight between Pini and Poutj).

  The information processing apparatus 1 may present this weight information as information indicating what pattern in the input information contributes more greatly to which pattern in the output information.

  Specifically, in an example of the present embodiment, a non-linear learning device such as a neural network machine-learned by the method described above is used. In the following example, it is assumed that the output of the estimation result θ by the learning device when x is input to the learning device is ypred (x; θ).

  In the following, it is assumed that the change amount Δx of the input information x and the change amount Δypred of the output information ypred with respect to the change amount Δx have a linear relationship Δypred = AΔx.

At this time, A is a Jacobian matrix in which a plurality of vectors a1 to aM are arranged in the row direction (because Δx and Δy are vectors) (the component of the mth row and lth column is _aml ). ,
Given in. The information processing apparatus 1 obtains the Jacobian matrix A using the learning device that has been machine-learned by the above-described method. Ypred_m represents the m-th component in the output of the estimation result θ. Further, ∂ypred_m / ∂xl represents a change amount (gradient) of ypred_m when xl is changed.

  In addition, the information processing apparatus 1 is paired with information XVF on the results of a plurality of non-pair visual field inspections and information on the results of visual field inspections that are not paired with the results of the optical coherence tomography (OCT) inspection. A plurality of non-pair optical coherence tomography (OCT) inspection result information XRT is obtained, and the following processing is performed.

  That is, the information processing apparatus 1 performs non-negative matrix factorization on each of the information XVF of the plurality of non-pair visual field inspection results and the information XRT of the plurality of non-pair optical coherence tomography (OCT) inspection results. Using an unsupervised learning method such as NMF), the respective base matrices BVF = ΘXVF and BRT = ΘXRT are obtained.

At this time, the minute change ΔXVF of the information XVF of the visual field inspection result is
In addition, the minute change ΔXRT of the information XRT of the result of the inspection by the optical coherence tomography (OCT) is also
Can be described.

By applying this relationship to the relationship of Δypred = AΔx assumed above, it can be described as ΔXVF = AΔXRT, and when the relationship between zVF and zRT is determined in consideration of the generalized inverse matrix,
It becomes. However,
It is.

  The information processing apparatus 1 obtains D using equation (16) using the basis matrices BVF and BRT obtained previously.

  In the case where none of the components of D should be negative, such as when non-negative matrix factorization is used for the calculation of the base matrix, the information processing device 1 is the component of D calculated by equation (16). , Negative ones are ignored (not used).

  Here, the vector of each row (each column before transposition) included in each base matrix BVF, BRT is the pattern PVF1, PVF2,..., PVFi (i = 1, 2,. Represents the patterns PRT1, PRT2,... PRTj (j = 1, 2,...) (Corresponding to the values of the respective inspection results at the corresponding visual field positions). Each obtained component Dij of D represents the weight of the relationship between each pattern PVFi and PRTj.

  Therefore, as illustrated in FIG. 4, the information processing apparatus 1 displays an image (P) of a pattern represented by each row included in each base matrix BVF, BRT, and the relationship weight (except for neglected ones) between them. Is displayed as a bipartite graph. At this time, the line segment corresponding to the link of the graph is displayed as a thicker line as the size of D is larger.

  According to this visualization processing, a learning device determines what kind of contribution relationship exists between a representative pattern of input information related to a patient's condition and a representative pattern of output information related to a patient's condition. It can be constructed using the differential information, and can be shown visually, and can be used for understanding the progression of symptoms.

  DESCRIPTION OF SYMBOLS 1 Information processing apparatus, 11 Control part, 12 Storage part, 13 Operation part, 14 Display part, 15 Input / output part, 20 Learning processing part, 21 Pattern acquisition part, 22 Learning data generation part, 30 Estimation processing part

Claims (7)

Acquisition means for acquiring information representing a pattern of output information;
A means for generating machine learning data for associating input information with corresponding output information, the difference between the information output by the learning device when the input information is input and the output information corresponding to the input information Learning processing means for generating machine learning data by controlling the parameters of the learning device using information on the acquired information on the pattern of the output information, and
Including
An information processing apparatus for subjecting machine learning data generated by the learning processing means to processing for estimating the output information corresponding to the input information to be processed. An acquisition means for acquiring information representing a pattern of information on eye symptoms;
A means for generating machine learning data for associating diagnostic information relating to the eye with information relating to a corresponding eye symptom, the output of the learning device when the diagnostic information relating to the eye is input, and the inputted diagnostic information Learning processing means for generating machine learning data by controlling parameters of the learning device using information regarding a difference from information regarding eye symptoms corresponding to the information and information regarding information patterns regarding the accepted eye symptoms When,
Including
An information processing apparatus for subjecting machine learning data generated by the learning processing means to processing for estimating information relating to eye symptoms corresponding to diagnostic information relating to eyes to be processed. An information processing apparatus according to claim 2,
The acquisition unit receives a plurality of pieces of information relating to eye symptoms, and generates and acquires information representing a pattern of information relating to eye symptoms obtained by unsupervised learning based on the information relating to the plurality of eye symptoms. . The information processing apparatus according to claim 2 or 3,
The information processing apparatus, wherein the eye diagnostic information is information representing a retinal structure, and the information about the eye symptom is information about a visual field. The information processing apparatus according to claim 3,
The information processing apparatus, wherein the information regarding the eye symptom accepted by the acquisition means is information including at least information indicating a visual field. Receiving a plurality of information related to eye symptoms and obtaining information representing a pattern of diagnostic information obtained by unsupervised learning based on information related to the plurality of eye symptoms;
A step of generating machine learning data for associating eye diagnostic information with corresponding eye symptom information, the output of the learning device when the eye diagnostic information is input, and the input diagnostic information Controlling the parameters of the learner using information relating to the difference between the information relating to the eye symptom corresponding to and information relating to the pattern of information relating to the accepted eye symptom;
including,
Machine learning data manufacturing method. Computer
An acquisition means for acquiring information representing a pattern of information on eye symptoms;
A means for generating machine learning data for associating diagnostic information relating to the eye with information relating to a corresponding eye symptom, the output of the learning device when the diagnostic information relating to the eye is input, and the inputted diagnostic information Learning processing means for generating machine learning data by controlling parameters of the learning device using information regarding a difference from information regarding eye symptoms corresponding to the information and information regarding information patterns regarding the accepted eye symptoms When,
Function as
A program used for processing for estimating information relating to eye symptoms corresponding to diagnostic information relating to eyes to be processed, wherein the machine learning data generated by the learning processing means.