WO2024190161A1 - Training data generation system, machine learning system, oct system, training data generation method, and training data generation program - Google Patents

Abstract

[Problem] To provide a technology for efficiently generating training data. [Solution] A training data generation system according to the present invention comprises: a measurement information acquisition unit that acquires measurement information that indicates measurement results for a measurement target by polarization-sensitive OCT, a polarization OCT image acquisition unit that, on the basis of the measurement information, acquires a polarization OCT image that is an image that corresponds to the polarization characteristics of the measurement target, a conventional OCT image acquisition unit that, on the basis of the measurement information, acquires a plurality of conventional OCT images that are intensity images from when the measurement target has been measured by OCT that does not consider polarization sensitivity, and a training data generation unit that generates a plurality of training data pairs that pair the polarization OCT image and each of the plurality of conventional OCT images.

Description

Teacher data generation system, machine learning system, OCT system, teacher data generation method, and teacher data generation program

The present invention relates to a teacher data generation system, a machine learning system, an OCT system, a teacher data generation method, and a teacher data generation program.

Polarization-sensitive OCT (Polarization-Sensitive Optical Coherence Tomography, hereafter referred to as polarization OCT) has been known so far. In addition to the functions of OCT (hereafter referred to as conventional OCT) that does not take into account the polarization sensitivity that OCT inherently has in principle, polarization OCT has the function of measuring the polarization characteristics of the object being measured. Polarization-sensitive OCT is more expensive than conventional OCT. Therefore, technology has been developed that uses relatively inexpensive conventional OCT to measure polarization characteristics equivalent to those of polarization-sensitive OCT.

For example, Patent Document 1 and Non-Patent Documents 1 and 2 disclose a technology in which a conventional OCT image is input, a machine learning model is trained to output a pseudo polarized OCT image corresponding to the polarized OCT image, and a pseudo polarized OCT image is generated from the conventional OCT image using the machine learning model.

International Publication No. WO2022/234828 Y. Sun, J. Wang, J. Shi, and S. A. Boppart, "Deep-learning-enabled polarization-sensitive optical coherence tomography (OCT)," in Biophotonics Congress 2021 (Optica Publishing Group, 2021), p. OF2E.3. Y. Sun, J. Wang, J. Shi, and S. A. Boppart, "Synthetic polarization-sensitive optical coherence tomography by deep learning," NPJ Digit Med 4(1), 105 (2021).

In order to learn a machine learning model, a sufficient amount of training data is generally required to obtain a desired learning rate. Patent Document 1 and Non-Patent Documents 1 and 2 do not mention a technique for efficiently preparing a large amount of training data. For example, in Patent Document 1, a conventional OCT image and a polarized OCT image are measured by the same polarized OCT and paired to generate training data (Patent Document 1,0070, etc.). However, in order to generate training data by such measurements, it is necessary to perform an extremely large number of measurements, which is very labor-intensive.
The present invention has been made in consideration of the above-mentioned problems, and has an object to efficiently generate training data.

To achieve the above objective, the teacher data generation system includes a measurement information acquisition unit that acquires measurement information indicating the measurement results of the object to be measured using polarization-sensitive OCT, a polarized OCT image acquisition unit that acquires a polarized OCT image that is an image corresponding to the polarization characteristics of the object to be measured based on the measurement information, a conventional OCT image acquisition unit that acquires multiple conventional OCT images that are intensity images when the object to be measured is measured using OCT that does not take polarization sensitivity into account based on the measurement information, and a teacher data generation unit that generates multiple sets of teacher data each consisting of a polarized OCT image and each of the multiple conventional OCT images.

In other words, when measurement information is acquired by polarization-sensitive OCT, a polarized OCT image and a conventional OCT image can be generated from the measurement information. Furthermore, a conventional OCT image is an intensity image that can be measured by conventional OCT, which does not actively utilize polarization information as a contrast source/information source, but it is also possible to generate a conventional OCT image captured with measurement light of any polarization state from the measurement information of the polarization-sensitive OCT. A conventional OCT image generated from the measurement information of the polarization-sensitive OCT can be considered to be an image obtained when the measurement object of the polarization-sensitive OCT is measured with the conventional OCT. Therefore, any number of conventional OCT images corresponding to one polarization OCT image can be generated.

In the teacher data generation system, a number of conventional OCT images corresponding to one polarized OCT image are generated, and multiple sets of teacher data are generated by pairing each of the multiple conventional OCT images with the polarized OCT image. Therefore, multiple sets of teacher data can be generated from the measurement information that is the basis of one polarized OCT image. As a result, teacher data can be generated efficiently.

FIG. 1 is a diagram illustrating a teacher data generation system and a machine learning system according to one embodiment of the present invention. 1 is a schematic diagram illustrating the configuration of an optical system of a polarization-sensitive OCT according to an embodiment of the present invention. FIG. 2 is a block diagram showing in detail the configuration of a sampling trigger/clock generator according to the present embodiment. FIG. 2 is a diagram for explaining the subject's eye and the direction of a B-scan. 13 is a flowchart of a teacher data creation process. FIG. 1 is a diagram for explaining a learning model. 1 is a flowchart of a machine learning process. FIG. 1 illustrates a conventional OCT system according to an embodiment of the present invention. 13 is a flowchart of a polarization OCT image generating process. FIG. 1 is a diagram showing a schematic configuration of an apparatus for measuring a Jones vector.

Here, the embodiments of the present invention will be described in the following order.
(1) Data processing system configuration:
(1-1) Configuration of polarization-sensitive OCT:
(1-2) Training data creation process:
(2) Machine learning processing:
(3) Configuration of conventional OCT system:
(4) Other embodiments:

(1) Data processing system configuration:
FIG. 1 is a diagram showing a data processing system 1000 that functions as a teacher data generation system and a machine learning system according to an embodiment of the present invention. A polarization-sensitive OCT 2000, a storage medium 3000, an input unit 4000, and a display unit 5000 are connected to the data processing system 1000 according to this embodiment. The polarization-sensitive OCT 2000 is an apparatus that performs optical coherence tomography by splitting light from a light source into reference light and measurement light, and measuring the interference between the reference light returned by a mirror or the like and the measurement light returned by the test eye. However, the polarization-sensitive OCT 2000 is capable of acquiring measurement information including the polarization characteristics of the measurement object. The polarization-sensitive OCT 2000 measures the test eye to generate measurement information, and outputs it to the data processing system 1000.

The storage medium 3000 is a non-volatile storage medium such as a hard disk drive (HDD) or a solid state drive (SSD). Various data can be stored in the storage medium 3000, and the data processing system 1000 can store data in the storage medium 3000 at any time and can also read out data stored in the storage medium 3000. In this embodiment, the storage medium 3000 stores measurement information 3000a obtained by photographing the subject's eye and teacher data 3000b generated by the data processing system 1000. Furthermore, the storage medium 3000 stores learning model data 3000c generated by machine learning processing performed by the data processing system 1000 based on the teacher data 3000b.

The input unit 4000 is a device operated by a user such as an examiner, and is a device for making various inputs to the data processing system 1000. The input unit 4000 may be in any form, but examples include a keyboard and a mouse. The display unit 5000 is a display device that displays various types of information, and displays various images such as characters and images generated by the data processing system 1000.

The data processing system 1000 includes a control unit 1100, a communication interface (I/F) 1200, and a display I/F 1300. The communication I/F 1200 is an interface that connects the data processing system 1000 to an external device. In this embodiment, the polarization-sensitive OCT 2000, the storage medium 3000, and the input unit 4000 are connected to the data processing system 1000 via the communication I/F 1200. The communication I/F 1200 may be of various standards, such as the USB standard, PCI-Express standard, Thunderbolt standard, and Ethernet standard (which are registered trademarks).

The display I/F 1300 is an interface that connects the data processing system 1000 and the display unit 5000. The display I/F 1300 may be of various standards, such as HDMI (High-Definition Multimedia Interface: registered trademark) and DVI (Digital Visual Interface). Of course, the communication I/F 1200 and the display I/F 1300 are not limited to these and may be wireless communication interfaces, etc.

The control unit 1100 includes a CPU, RAM, and ROM (not shown) and can execute various programs stored in the storage medium 3000, ROM, etc. The control unit 1100 may include a GPU (Graphics Processing Unit) to efficiently perform machine learning, which will be described later. The programs executed by the control unit 1100 include various programs. In this embodiment, the programs include a teacher data generation program 1110 for generating teacher data 3000b and a machine learning program 1120 for generating learning model data 3000c from the teacher data 3000b. The functions executed by the control unit 1100 using the teacher data generation program 1110 include a measurement information acquisition unit 1110a, a polarized OCT image acquisition unit 1110b, a conventional OCT image acquisition unit 1110c, and a teacher data generation unit 1110d. The functions executed by the control unit 1100 using the machine learning program 1120 include a machine learning unit 1120a.

The measurement information acquisition unit 1110a is a function for acquiring measurement information 3000a indicating the measurement results of the object to be measured by the polarization-sensitive OCT 2000. That is, the control unit 1100 controls the polarization-sensitive OCT 20, acquires the measurement information 3000a obtained by measuring the subject's eye, and stores it in the storage medium 3000. In this embodiment, the control unit 1100 measures the subject's eyes of multiple subjects (assumed to be L people) and acquires the measurement information 3000a for the L people. Note that when the left and right eyes of the same person are measured, this may be counted as measurement information 3000a for two people. Furthermore, the measurement information 3000a is a Jones matrix indicating the polarization characteristics of the subject's eye, which is the object to be measured, or a Jones vector indicating the polarization state of light reflecting the polarization characteristics.

(1-1) Configuration of polarization-sensitive OCT:
Here, a configuration example of a polarization-sensitive OCT 2000 for measuring the Jones matrix will be described. Here, an example will be described in which the polarization-sensitive OCT 2000 is a wavelength-swept Fourier domain type (so-called SS-OCT type) device using a wavelength-swept light source. However, the OCT type is not limited to SS-OCT, and other types using the Fourier main type, such as SD-OCT (spectral domain OCT) and types other than the Fourier domain type (such as the time domain type) can also be used.

As shown in FIG. 2, the polarization-sensitive OCT 2000 of this embodiment includes a light source 11, a measurement light generation unit (21-29, 31, 32) that generates measurement light from the light of the light source 11, a reference light generation unit (41-46, 51) that generates reference light from the light of the light source 11, interference light generation units 60, 70 that combine the reflected light from the measurement object 30 generated by the measurement light generation unit with the reference light generated by the reference light generation unit to generate interference light, and interference light detection units 80, 90 that detect the interference light generated by the interference light generation unit.

(light source)
The light source 11 is a wavelength sweep type light source, and the wavelength (wave number) of the emitted light changes at a predetermined period. Since the wavelength of the light irradiated to the measurement object 30 changes (sweeps), the intensity distribution of the light reflected from each part in the depth direction of the measurement object 30 can be obtained by performing Fourier analysis on the signal obtained from the interference light between the reflected light from the measurement object 30 and the reference light.

The light source 11 is connected to a polarization control device 12 and a fiber coupler 13, and the fiber coupler 13 is connected to a PM (Polarization Maintaining) coupler 14 and a sampling trigger/clock generator 100. Therefore, the light output from the light source 11 is input to the PM coupler 14 and the sampling trigger/clock generator 100 via the polarization control device 12 and the fiber coupler 13. The sampling trigger/clock generator 100 uses the light from the light source 11 to generate a sampling trigger and a sampling clock for each of the signal processors 83 and 93, which will be described later.

(Measurement light generating unit)
The measurement light generating unit (21 to 29, 31, 32) includes a PM coupler 21 connected to the PM coupler 14, two measurement light paths S1, S2 branching from the PM coupler 21, a polarized beam combiner/splitter 25 connecting the two measurement light paths S1, S2, and a light path extension unit 306, a collimator lens 26, galvanometer mirrors 27, 28, and a lens 29 connected to the polarized beam combiner/splitter 25. The measurement light path S1 includes an optical path length difference generating unit 22 and a circulator 23. The measurement light path S2 includes a circulator 24. Therefore, the optical path length difference Δl between the measurement light path S1 and the measurement light path S2 is generated by the optical path length difference generating unit 22. The optical path length difference Δl may be set to be longer than the measurement range in the depth direction of the measurement object. This makes it possible to prevent interference light having different optical path length differences from overlapping. The optical path length difference generating unit 22 may be, for example, an optical fiber or an optical system such as a mirror or a prism. In this embodiment, a 1 m PM fiber is used for the optical path length difference generating unit 22. The measurement light generating unit further includes PM couplers 31 and 32. The PM coupler 31 is connected to the circulator 23. The PM coupler 32 is connected to the circulator 24.

One of the beams (i.e., the measurement beam) split by the PM coupler 14 is input to the measurement beam generating unit (21-29, 31, 32). The PM coupler 21 splits the measurement beam input from the PM coupler 14 into a first measurement beam and a second measurement beam. The first measurement beam split by the PM coupler 21 is guided to the measurement beam path S1, and the second measurement beam is guided to the measurement beam path S2. The first measurement beam guided to the measurement beam path S1 passes through the optical path length difference generating unit 22 and the circulator 23 and is input to the polarized beam combiner/splitter 25. The second measurement beam guided to the measurement beam path S2 passes through the circulator 24 and is input to the polarized beam combiner/splitter 25.

The PM fiber 304 is connected to the polarized beam combiner/splitter 25 in a state rotated 90 degrees in the circumferential direction relative to the PM fiber 302. As a result, the second measurement light input to the polarized beam combiner/splitter 25 becomes light having a polarization component perpendicular to the first measurement light. Since the optical path length difference generating unit 22 is provided in the measurement light path S1, the first measurement light is delayed relative to the second measurement light by the distance of the optical path length difference generating unit 22 (i.e., an optical path length difference Δl is generated). The polarized beam combiner/splitter 25 superimposes the input first measurement light and second measurement light.

The light output from the polarized beam combiner/splitter 25 (the superimposed light of the first and second measurement light) is irradiated onto the measurement object 30 via the collimator lens 26, the galvanometer mirrors 27 and 28, and the lens 29. An optical path extension 306 may be disposed between the collimator lens 26 and the galvanometer mirror 27. For example, a PM fiber of about 60 m may be used for the optical path extension 306, as shown in FIG. 2. This makes it possible to suppress the occurrence of crosstalk between the two modes of the PM fiber. The light irradiated onto the measurement object 30 is scanned in the x-y directions by the galvanometer mirrors 27 and 28. The x- and y-directions are orthogonal to each other in a plane perpendicular to the optical axis.

The light irradiated to the object to be measured 30 is reflected by the object to be measured 30. Here, the light reflected by the object to be measured 30 is scattered on the surface of the object to be measured 30 and inside the object to be measured. The reflected light and scattered light from the object to be measured 30 are input to the polarized beam combiner/splitter 25 through the lens 29, the galvanometer mirrors 28 and 27, and the collimator lens 26 in the opposite direction to the incident path. The polarized beam combiner/splitter 25 splits the input reflected light into horizontally polarized reflected light (horizontally polarized component) and vertically polarized reflected light (vertically polarized component), which are orthogonal polarization components, and the horizontally polarized reflected light is guided to the measurement optical path S1, and the vertically polarized reflected light is guided to the measurement optical path S2, respectively.

The horizontally polarized reflected light has its optical path changed by the circulator 23 and is input to the PM coupler 31. The PM coupler 31 splits the input horizontally polarized reflected light and inputs it to each of the PM couplers 61 and 71. Therefore, the horizontally polarized reflected light input to the PM couplers 61 and 71 contains a reflected light component due to the first measurement light and a reflected light component due to the second measurement light.

The vertically polarized reflected light has its optical path changed by the circulator 24 and is input to the PM coupler 32. The PM coupler 32 splits the input vertically polarized reflected light and inputs it to each of the PM couplers 62 and 72. Therefore, the vertically polarized reflected light input to the PM couplers 62 and 72 contains a reflected light component due to the first measurement light and a reflected light component due to the second measurement light.

(Reference light generating unit)
The reference light generating unit (41 to 46, 51) includes a circulator 41 connected to the PM coupler 14, a reference delay line (42, 43) connected to the circulator 41, a PM coupler 44 connected to the circulator 41, two reference light paths R1, R2 branched from the PM coupler 44, a PM coupler 46 connected to the reference light path R1, and a PM coupler 51 connected to the reference light path R2. An optical path length difference generating unit 45 is disposed in the reference light path R1. No optical path length difference generating unit is provided in the reference light path R2. Therefore, the optical path length difference Δl' between the reference light path R1 and the reference light path R2 is generated by the optical path length difference generating unit 45. For example, an optical fiber is used for the optical path length difference generating unit 45. The optical path length difference Δl' of the optical path length difference generating unit 45 may be the same as the optical path length difference Δl of the optical path length difference generating unit 22. By making the optical path length differences Δl and Δl′ equal, the depth positions of a plurality of interference lights, which will be described later, relative to the measurement object become the same, which means that it is not necessary to align the positions of a plurality of tomographic images to be acquired.

The other light (i.e., reference light) branched by the PM coupler 14 is input to the reference light generating unit (41-46, 51). The reference light input from the PM coupler 14 is input to the reference delay line (42, 43) through the circulator 41. An optical path length extension unit 308 may be arranged between the optical circulator 41 and the reference delay line (42, 43). A PM fiber of about 60 m may be used for the optical path length extension unit 308, as shown in FIG. 2. The reference delay line (42, 43) is composed of a collimator lens 42 and a reference mirror 43. The reference light input to the reference delay line (42, 43) is irradiated to the reference mirror 43 via the collimator lens 42. The reference light reflected by the reference mirror 43 is input to the circulator 41 via the collimator lens 42. Here, the reference mirror 43 is movable in a direction approaching or moving away from the collimator lens 42. In this embodiment, before starting measurement, the position of the reference mirror 43 is adjusted so that the optical path length (measurement optical path length) from the PM coupler 14 to the measurement object 30 via the second measurement optical path S2 matches the optical path length (reference optical path length) from the PM coupler 14 to the reference mirror 43.

The reference light reflected by the reference mirror 43 has its optical path changed by the circulator 41 and is input to the PM coupler 44. The PM coupler 44 splits the input reference light into a first reference light and a second reference light. The first reference light is guided to the reference light path R1, and the second reference light is guided to the reference light path R2. The first reference light is input to the PM coupler 46 through the optical path length difference generating unit 45. The reference light input to the PM coupler 46 is split into a first branched reference light and a second branched reference light. The first branched reference light is input to the PM coupler 61 through the collimator lens 47 and the lens 48. The second branched reference light is input to the PM coupler 62 through the collimator lens 49 and the lens 50. The second reference light is input to the PM coupler 51 and split into a third branched reference light and a fourth branched reference light. The third branched reference light passes through collimator lens 52 and lens 53 and is input to PM coupler 71. The fourth branched reference light passes through collimator lens 54 and lens 55 and is input to PM coupler 72.

(Interference light generating unit)
The interference light generating units 60 and 70 include a first interference light generating unit 60 and a second interference light generating unit 70. The first interference light generating unit 60 includes PM couplers 61 and 62. As described above, the PM coupler 61 receives the horizontally polarized reflected light from the measurement light generating unit, and the first branched reference light (light having an optical path length difference Δl) from the reference light generating unit. Here, the horizontally polarized reflected light includes a reflected light component due to the first measurement light (light having an optical path length difference Δl) and a reflected light component due to the second measurement light. Therefore, in the PM coupler 61, the reflected light component due to the first measurement light (light having an optical path length difference Δl) of the horizontally polarized reflected light and the first branched reference light are combined to generate the first interference light (horizontally polarized component).

In addition, the PM coupler 62 receives vertically polarized reflected light from the measurement light generation unit and receives the second branched reference light (light having optical path length difference Δl) from the reference light generation unit. Here, the vertically polarized reflected light contains a reflected light component due to the first measurement light (light having optical path length difference Δl) and a reflected light component due to the second measurement light. Therefore, in the PM coupler 62, the reflected light component due to the first measurement light (light having optical path length difference Δl) of the vertically polarized reflected light is combined with the second branched reference light to generate a second interference light (vertically polarized component).

The second interference light generating unit 70 has PM couplers 71 and 72. As described above, the horizontally polarized reflected light is input to the PM coupler 71 from the measurement light generating unit, and the third branched reference light (light that does not have an optical path length difference Δl) is input from the reference light generating unit. Therefore, in the PM coupler 71, the reflected light component of the horizontally polarized reflected light due to the second measurement light (light that does not have an optical path length difference Δl) and the third branched reference light are combined to generate the third interference light (horizontally polarized component).

In addition, the PM coupler 72 receives vertically polarized reflected light from the measurement light generation unit and receives the fourth branched reference light (light without optical path length difference Δl) from the reference light generation unit. Therefore, in the PM coupler 72, the reflected light component (light without optical path length difference Δl) by the second measurement light of the vertically polarized reflected light is combined with the fourth branched reference light to generate a fourth interference light (vertically polarized component). The first interference light and the second interference light correspond to the measurement light that has passed through the measurement light path S1, and the third interference light and the fourth interference light correspond to the measurement light that has passed through the measurement light path S2.

(Interference light detection unit)
The interference light detection units 80, 90 include a first interference light detection unit 80 that detects the interference light (first interference light and second interference light) generated by the first interference light generation unit 60, and a second interference light detector 90 that detects the interference light (third interference light and fourth interference light) generated by the second interference light generation unit 70.

The first interference light detection unit 80 includes balanced photodetectors 81 and 82 and a signal processor 83 connected to the balanced photodetectors 81 and 82. The PM coupler 61 is connected to the balanced photodetector 81, and the signal processor 83 is connected to the output terminal of the balanced photodetector 81. The PM coupler 61 splits the first interference light into two interference lights with a phase difference of 180 degrees and inputs them to the balanced photodetector 81. The balanced photodetector 81 performs differential amplification and noise reduction processing on the two interference lights with a phase difference of 180 degrees input from the PM coupler 61, converts them into an electrical signal (first interference signal), and outputs the first interference signal to the signal processor 83. In other words, the first interference signal is an interference signal HH of the horizontally polarized reflected light from the measurement object by the horizontally polarized measurement light and the reference light.

Similarly, the PM coupler 62 is connected to the balanced photodetector 82, and the signal processor 83 is connected to the output terminal of the balanced photodetector 82. The PM coupler 62 splits the second interference light into two interference lights with a phase difference of 180 degrees and inputs them to the balanced photodetector 82. The balanced photodetector 82 performs differential amplification and noise reduction processing on the two interference lights with a phase difference of 180 degrees, converts them into an electrical signal (second interference signal), and outputs the second interference signal to the signal processor 83. That is, the second interference signal is an interference signal HV of the vertically polarized reflected light from the measurement object and the reference light due to the horizontally polarized measurement light. The signal processor 83 samples the first interference signal (signal due to the first interference light) and the second interference signal (signal due to the second interference light) based on the sampling trigger and sampling clock input from the sampling trigger/clock generator 100. The first interference signal and the second interference signal sampled by the signal processor 83 are input to the calculation unit 202, which will be described later. A known data acquisition device (so-called DAQ) can be used as the signal processor 83.

The second interference light detector 90, like the first interference light detector 80, includes balanced photodetectors 91 and 92 and a signal processor 93 connected to the balanced photodetectors 91 and 92. The PM coupler 71 is connected to the balanced photodetector 91, and the signal processor 93 is connected to the output terminal of the balanced photodetector 91. The PM coupler 71 splits the third interference light into two interference lights with a phase difference of 180 degrees and inputs them to the balanced photodetector 91. The balanced photodetector 91 performs differential amplification and noise reduction processing on the two interference lights with a phase difference of 180 degrees, converts them into an electrical signal (third interference signal), and outputs the third interference signal to the signal processor 93. In other words, the third interference signal is an interference signal VH of the horizontally polarized reflected light from the measurement object and the reference light due to the vertically polarized measurement light.

Similarly, the PM coupler 72 is connected to the balanced photodetector 92, and the signal processor 93 is connected to the output terminal of the balanced photodetector 92. The PM coupler 72 splits the fourth interference light into two interference lights with a phase difference of 180 degrees and inputs them to the balanced photodetector 91. The balanced photodetector 92 performs differential amplification and noise reduction processing on the two interference lights with a phase difference of 180 degrees, converts them into an electrical signal (fourth interference signal), and outputs the fourth interference signal to the signal processor 93. That is, the fourth interference signal is an interference signal VV of the vertically polarized reflected light of the measurement object and the reference light from the vertically polarized measurement light. The signal processor 93 samples the third interference signal (signal due to the third interference light) and the fourth interference signal (signal due to the fourth interference light) based on the sampling trigger and sampling clock input from the sampling trigger/clock generator 100. The third and fourth interference signals sampled by the signal processor 93 are input to the calculation unit 202, which will be described later. A known data acquisition device (so-called DAQ) can also be used for the signal processor 93. With this configuration, it is possible to obtain interference signals that represent the four polarization characteristics of the measurement object 30.

Next, the configuration of the control system of the polarization-sensitive OCT 2000 according to this embodiment will be described. As shown in FIG. 3, the polarization-sensitive OCT 2000 is controlled by a calculation device 200. The calculation device 200 is composed of a calculation unit 202, a first interference light detection unit 80, and a second interference light detector 90. The first interference light detection unit 80, the second interference light detector 90, and the calculation unit 202 are connected to the measurement unit 10. The calculation unit 202 outputs a control signal to the measurement unit 10, and drives the light source 11 and the galvanometer mirrors 27 and 28 to scan the incident position of the measurement light on the measurement object 30.

The first interference light detector 80 acquires first sampling data based on the sampling clock 1 input from the measurement unit 10 for the interference signal (interference signal HH and interference signal HV) input from the measurement unit 10, using the sampling trigger 1 as a trigger, and outputs the first sampling data to the calculation unit 202. The calculation unit 202 performs a predetermined calculation process on the first sampling data to generate the HH component and the HV component of the Jones matrix. The second interference light detector 90 acquires second sampling data based on the sampling clock 2 input from the measurement unit 10 for the interference signal (interference signal VH and interference signal VV) input from the measurement unit 10, using the sampling trigger 2 as a trigger, and outputs the second sampling data to the calculation unit 202. The calculation unit 202 performs a predetermined calculation process on the second sampling data to generate the VH component and the VV component of the Jones matrix. The HH, HV, VH, and VV components of the Jones matrix are the first row, first column, the first row, second column, the second row, first column, and the second row, second column of the 2-row, 2-column Jones matrix, respectively. If the first row, first column is denoted as a, the first row, second column as b, the second row, first column as c, and the second row, second column as d, the measured Jones matrix Jmeasured can be expressed by the following formula (1).

In the Jones matrix Jmeasured measured in the above manner, each of the components a, b, c, and d is a complex number. Furthermore, the Jones matrix depends on the position of the object to be measured in the depth direction (optical axis direction) and the position in the direction perpendicular to the optical axis. Here, if the two orthogonal directions in the plane perpendicular to the optical axis are the x direction and the y direction, and the optical axis direction is the z direction, then the Jones matrix is defined for each coordinate (x, y, z) in three-dimensional space.

In this embodiment, the Jones matrix defined for each position in three-dimensional space is the measurement information 3000a. When the polarization-sensitive OCT 2000 measures the test eye, the polarization-sensitive OCT 2000 transmits the measurement information 3000a to the data processing system 1000. The control unit 1100 acquires the measurement information 3000a via the communication I/F 1200 and stores it in the storage medium 3000. The data processing system 1000 performs the above processing for L test eyes and stores the measurement information 3000a for L people in the storage medium 3000.

Continuing the explanation by returning to FIG. 1, the polarization OCT image acquisition unit 1110b has a function of acquiring a polarization OCT image, which is an image corresponding to the polarization characteristics of the object to be measured, based on the measurement information 3000a. The polarization OCT image may be an image for evaluating the polarization characteristics. Here, an example will be described in which a PD (polarization-diverse) image is the polarization OCT image. In this embodiment, the polarization OCT image is an image on a two-dimensional plane obtained by cutting a three-dimensional space with a cutting plane including the optical axis. The two-dimensional plane may be defined by various methods, but in this embodiment, the control unit 1100 considers a B-scan image to be the polarization OCT image.

FIG. 4 is a diagram for explaining the test eye Ey and the direction of the B-scan. In this example, the optical axis Lx passes through the corneal apex of the test eye Ey. In FIG. 4, among the cut surfaces including the optical axis Lx, the angle of the cut surface parallel to the horizontal direction is shown as 0°, and the angle of the cut surface parallel to the vertical direction is shown as 90°. The angle of the cut surface including the optical axis Lx is not limited to 0° or 90°, and the cut surface can be defined for any rotation angle centered on the optical axis Lx. In this way, extracting information on a cut surface including the optical axis Lx and cut at any angle centered on the optical axis Lx is called extracting information by a B-scan.

The control unit 1100 generates a polarization OCT image based on the measurement information 3000a on the cut surface obtained by the B-scan. For this purpose, the control unit 1100 specifies the cut surface on which the polarization OCT image is to be generated, and acquires the measurement information 3000a on the cut surface. The control unit 1100 regards the measurement information 3000a as information for each coordinate on the cut surface, specifies the intensity I _PD of each coordinate by the following formula (2), and generates a polarization OCT image. When visualizing an image based on intensity, it is preferable to perform processing such as expressing the intensity logarithmically and expressing a specific range by a gradation value (same below). With the above configuration, it is possible to generate teacher data 3000b for a two-dimensional image that can visualize the three-dimensional structure of the measurement object.

The conventional OCT image acquisition unit 1110c has a function of acquiring multiple conventional OCT images, which are intensity images obtained when the object to be measured is measured using OCT that does not take polarization sensitivity into account, based on the measurement information. The conventional OCT image is an intensity image obtained when the object to be measured is measured using conventional OCT that does not take polarization sensitivity into account, but in this embodiment, it is generated based on the measurement information 3000a of the polarization-sensitive OCT 2000. In other words, the conventional OCT image is generated without performing measurement using conventional OCT that does not take polarization sensitivity into account.

By using the Jones matrix indicated by the measurement information 3000a, it is possible to calculate the Jones vector of the measurement light obtained as a response to input light of an arbitrary polarization state that is incident on the measurement object. Specifically, when the Jones vector Ein of the incident light is expressed by equation (3) and the Jones vector Eout of the measurement light is expressed by equation (4), the control unit 1100 obtains the Jones vector Eout of the measurement light by equation (5) in which the Jones matrix is applied to the Jones vector Ein of the incident light.

By multiplying the Jones vector Eout in equation (5), which represents the polarization state of the measurement light, by the Jones matrix representing an arbitrary polarization acting element, another arbitrary Jones vector can be calculated by numerical calculation. In this embodiment, a half-wave plate with an angle θ is applied to the Jones vector Eout of the measurement light shown in equation (5). Furthermore, consider a situation in which a horizontal linear polarizer is applied to cause the horizontal linear polarization component of the measurement light to interfere with the horizontal linear polarization reference light for measurement.

Equation (6) is the equation when a 1/2 wavelength plate with an angle θ is applied to the Jones vector Eout of the measurement light. Equation (6) corresponds to rotating the polarization state of the Jones vector Eout of the measurement light shown in equation (5) by 2θ. That is, the polarization state can be changed by equation (6). Therefore, by setting an arbitrary θ in equation (6), a polarization operation that changes the polarization state at an arbitrary angle can be performed. In this embodiment, a 1/2 wavelength plate with an angle θ is used for explanation, but it may be realized by applying various other polarization acting elements instead of a wavelength plate. For example, it is generally known that a Faraday rotator also has a Jones matrix that is mathematically equivalent to a wavelength plate, and similar calculations are possible by applying a Faraday rotator. In addition to the rotation angle of the wavelength plate, the phase delay amount may be changed. Here, the case of applying a 1/2 wavelength plate with an angle θ is explained.

Conventional OCT images measured in conventional OCT that does not consider polarization sensitivity are generated based on the result of measuring the interference light between the measurement light and the reference light, but it is known that the measurement light interferes only with the same polarization component as the polarization component of the reference light. This phenomenon is generally known as the Fresnel-Arago law. Therefore, the conventional OCT image is a measurement result of only a specific polarization component. Therefore, in this embodiment, the state in which a horizontal linear polarizer is applied based on formula (6) is obtained by formula (7). The polarizer is not limited to a horizontal linear polarizer, but since the polarization operation of any angle is performed in formula (6), generality is not lost even if a horizontal linear polarizer is used.

そこで、制御部１１００は、偏光ＯＣＴ画像の生成対象となった切断面と同一の切断面についてのジョーンズ行列を取得し、複数のθについて式（７）を演算することで、変化後の偏光状態が異なる複数回の偏光操作を実施する。そして、制御部１１００は、得られた結果に基づいて式（８）を計算することで、複数枚の従来型ＯＣＴ画像を計算する。なお、１つの切断面において式（２）に基づいて生成される偏光ＯＣＴ画像は１枚であるが、当該切断面において式（８）に基づいて生成される従来型ＯＣＴ画像は複数枚である。 Equation (7) shows the information measured in conventional OCT that does not take polarization sensitivity into account. Here, consider the case where horizontally linearly polarized reference light is interfered with equation (7) and the intensity signal is calculated. That is, the intensity I single is calculated by equation (8) based on the horizontally linearly polarized component of equation (7). In this way, a conventional OCT image can be generated in a pseudo manner.

Therefore, the control unit 1100 obtains the Jones matrix for the same cut surface as the cut surface for which the polarized OCT image was generated, and performs multiple polarization operations with different polarization states after the change by calculating equation (7) for multiple θ. The control unit 1100 then calculates equation (8) based on the obtained results to calculate multiple conventional OCT images. Note that, while the polarized OCT image generated based on equation (2) for one cut surface is one, the conventional OCT images generated based on equation (8) for the same cut surface are multiple.

The teacher data generating unit 1110d has a function of generating multiple sets of teacher data, each of which pairs a polarized OCT image with multiple conventional OCT images. That is, the control unit 1100 generates multiple sets of teacher data 3000b by associating each conventional OCT image generated based on formula (8) with a polarized OCT image generated based on formula (2) to form a pair. The generated teacher data 3000b is stored in the storage medium 3000.

In the above configuration, multiple conventional OCT images are generated from the measurement information 3000a at one cut surface, and multiple sets of teacher data 3000b can be generated. In addition, information on multiple cut surfaces can be extracted from the measurement information 3000a, and multiple sets of teacher data 3000b can be generated from the measurement information 3000a of each cut surface. Therefore, an extremely large amount of teacher data 3000b can be generated based on a single measurement, making it possible to generate teacher data efficiently.

Furthermore, in this embodiment, since the Jones matrix is obtained as the measurement information 3000a by the polarization-sensitive OCT 2000, it is possible to easily obtain a polarized OCT image and a conventional OCT image based on equations (2) and (8). Furthermore, in this embodiment, it is possible to generate multiple conventional OCT images from the same information by performing a polarization operation that changes the polarization state. Therefore, it is possible to easily and efficiently generate training data 3000b.

(1-2) Training data creation process:
Next, the teacher data creation process performed by the control unit 1100 in the above-mentioned configuration will be described in detail. Fig. 5 is a flowchart showing the teacher data creation process. Before the teacher data creation process is executed, the control unit 1100 controls the polarization-sensitive OCT 2000 using the function of the measurement information acquisition unit 1110a to measure measurement information 3000a for L people and store it in the storage medium 3000.

When the teacher data creation process is started, the control unit 1100 initializes variables l and n to 0 using the function of the polarization OCT image acquisition unit 1110b (steps S100, S105). The variable l is a variable for counting how many people's measurement information 3000a has been processed, with a minimum value of 0 and a maximum value of L-1. The variable n is a variable for identifying the B-scan, with a minimum value of 0 and a maximum value of N-1. In this embodiment, it is assumed that B-scan data for cut surfaces at angles from 0° to 180/N around the optical axis Lx shown in FIG. 4 is used. In this example, it is assumed that 0° shown in FIG. 4 is the 0th B-scan.

Then, the control unit 1100 acquires Jones matrix data corresponding to the nth B-scan using the function of the polarization OCT image acquisition unit 1110b (step S110). That is, the control unit 1100 refers to the measurement information 3000a, identifies the cut surface of the nth B-scan, and acquires Jones matrix data at each coordinate on the cut surface.

Next, the control unit 1100 acquires a polarization OCT image based on the Jones matrix using the function of the polarization OCT image acquisition unit 1110b (step S115). That is, the control unit 1100 acquires the intensity I _PD of each coordinate based on the formula (2) and generates a polarization OCT image. The generated polarization OCT image is stored in the RAM or storage medium 3000 (not shown).

Next, the control unit 1100 uses the function of the conventional OCT image acquisition unit 1110c to apply the Jones vector Ein of the input light to the Jones matrix to obtain the Jones vector Eout of the measurement light (step S120). That is, the control unit 1100 defines the Jones vector Ein of the light assumed to be the input light of the polarization-sensitive OCT 2000 by equation (3), and obtains the Jones vector Eout of the measurement light corresponding to each coordinate based on equation (5).

Then, the control unit 1100 initializes the variable m to 0 by using the function of the conventional OCT image acquisition unit 1110c (step S125). Note that the variable m is a variable for counting the number of conventional OCT images generated for one cut surface, and has a minimum value of 0 and a maximum value of M-1.

Next, the control unit 1100 applies the Jones matrix of the half-wave plate for angle θ = (m/M)π (radians) to Eout to obtain a conventional OCT image (step S130). That is, in this example, each angle obtained by dividing 180° into M is set as the mth angle, and polarization operation is performed at every certain angle. Therefore, in step S130, the control unit 1100 obtains the Jones vector after polarization operation using the mth angle by equation (6). Furthermore, the control unit 1100 applies a horizontal linear polarizer to the Jones vector obtained by equation (7). Then, the control unit 1100 obtains a conventional OCT image by equation (8).

Then, the control unit 1100 uses the function of the teacher data generation unit 1110d to generate teacher data 3000b that pairs the conventional OCT image and the polarized OCT image (step S135). That is, the control unit 1100 generates a set of teacher data 3000b by pairing the polarized OCT image acquired in step S115 with the conventional OCT image acquired in step S130, and stores the set in the storage medium 3000.

The control unit 1100 then determines whether the variable m is equal to the maximum value M-1, and if it is not equal, increments the variable m (step S145) and repeats the processes from step S130 onwards. That is, M conventional OCT images corresponding to one polarized OCT image are generated. With the above configuration, conventional OCT images can be easily generated based on θ at regular angles. In addition, because polarization operations are performed at regular angles that evenly divide the rotation angle range of the waveplate over the entire range from 0° to 180°, conventional OCT images can be generated under a variety of conditions.

If it is determined in step S140 that the variable m matches the maximum value M-1, the control unit 1100 determines whether the variable n matches the maximum value N-1 (step S150). If it is not determined in step S150 that the variable n matches the maximum value N-1, the control unit 1100 increments the variable n (step S155) and repeats the processing from step S110 onwards. That is, processing is performed to generate M sets of teacher data 3000b for each of the N B-scans.

If it is determined in step S150 that the variable n matches the maximum value N-1, the control unit 1100 determines whether the variable l matches the maximum value L-1 (step S160). If it is determined in step S160 that the variable l does not match the maximum value L-1, the control unit 1100 increments the variable l (step S165) and repeats the processing from step S105 onwards. That is, a process is performed to generate M x N sets of teacher data 3000b for each of the measurement information 3000a for L individuals. If it is determined in step S160 that the variable l matches the maximum value L-1, the control unit 1100 ends the teacher data creation process. As a result, L x M x N sets of teacher data 3000b are generated.

(2) Machine learning system configuration:
Next, machine learning using the teacher data 3000b will be described. In this embodiment, machine learning is performed by the control unit 1100 executing the machine learning program 1120 in the data processing system 1000. In this embodiment, the machine learning process is a process of optimizing a training model that forms a neural network. In this embodiment, a model that generates a polarization OCT image from a conventional OCT image is machine-learned. Therefore, in this embodiment, the model is information indicating an equation that derives a correspondence relationship between both data when the grayscale value of each pixel of the input conventional OCT image is used as an input value and the grayscale value of each pixel of the generated polarization OCT image is used as an output value.

FIG. 6 is a diagram showing a schematic diagram of a learning model in this embodiment. In this embodiment, a generative model G and a discriminative model D are prepared as training models. The generative model G is a model that generates one image from another. The discriminative model D is a model that compares a first image with a second image and discriminates whether the second image is a fake image generated from the first image.

In this embodiment, the conventional OCT image Int is input to the generation model G, and machine learning is performed to generate a polarized OCT image Ipf. Also, in this embodiment, the conventional OCT image Int is set as a first image, and the polarized OCT image Ipf generated by the generation model G, or the polarized OCT image Ipt associated with the conventional OCT image Int in the teacher data 3000b, is input as a second image to the discrimination model D. The discrimination model D is machine-learned to output false when the second image is the polarized OCT image Ipf generated by the generation model G, and true when the second image is the polarized OCT image Ipt associated with the conventional OCT image Int in the teacher data 3000b. The model structure and loss function settings when performing machine learning may be performed using various methods, and for example, a well-known GAN (Generative Adversarial Network) can be used.

FIG. 7 is a flowchart showing the machine learning process. When the machine learning process is started, the control unit 1100 acquires a training model using the function of the machine learning unit 1120a (step S200). In this embodiment, the control unit 1100 acquires a training model (information such as a filter, activation function, loss function, etc. indicating the model) of a generation model G in which the conventional OCT image Int is an input value and the polarized OCT image Ipf is an output value. The control unit 1100 also acquires a training model (information such as a filter, activation function, loss function, etc. indicating the model) of a discrimination model D in which the conventional OCT image and the polarized OCT image are input values and the output value indicates whether the input polarized OCT image is true or false.

Next, the control unit 1100 acquires the teacher data 3000b stored in the storage medium 3000 using the function of the machine learning unit 1120a (step S205). Next, the control unit 1100 acquires test data using the function of the machine learning unit 1120a (step S210). In this embodiment, a portion of the teacher data 3000b is extracted and used as test data for checking whether learning has been generalized. Note that the test data is not used for machine learning.

Next, the control unit 1100 determines initial values using the function of the machine learning unit 1120a (step S215). That is, the control unit 1100 assigns initial values to the variable parameters (filter weights, biases, etc.) to be learned in the training model acquired in step S200. The initial values may be determined by various methods. Of course, the initial values may be adjusted so that the parameters are optimized during the learning process, or learned parameters may be acquired and used from various databases, etc.

Next, the control unit 1100 performs learning using the function of the machine learning unit 1120a (step S220). That is, the control unit 1100 inputs the conventional OCT image indicated by the teacher data 3000b acquired in step S205 to the generation model G of the training model acquired in step S200, and outputs a polarized OCT image. Next, the control unit 1100 inputs the conventional OCT image and the comparison image to the discrimination model D, and outputs either true or false. The comparison image is either the polarized OCT image Ipt indicated by the teacher data 3000b (i.e., the true image) or the polarized OCT image Ipf generated by the generation model G (i.e., the false image). Which one is used is statistically determined in advance.

When the final output value by the training model, i.e., the stained image generated by the generative model G and the output of true or false identified by the discriminative model D, is obtained, the control unit 1100 identifies an error based on the output value and the correct answer value (information indicating the true or false of the conventional OCT image Int and the polarized OCT image Ipt indicated by the teacher data 3000b, and the comparison image) based on a loss function that evaluates the error between the output value and the correct answer value. When the loss function E is obtained, the control unit 1100 updates the parameters using a predetermined optimization algorithm, such as stochastic gradient descent. In other words, the control unit 1100 repeats the process of updating the parameters based on the derivative of the loss function E by the parameters a predetermined number of times.

When the parameters have been updated a preset number of times in this manner, the control unit 1100 determines whether generalization of the training model has been completed (step S225). That is, the control unit 1100 inputs the conventional OCT image of the test data acquired in step S210 into a generation model G of the training model to generate a polarization OCT image. The control unit 1100 also inputs the generated polarization OCT image and the conventional OCT image of the test data into a discrimination model D of the training model to determine true or false. The control unit 1100 then acquires the number of test data for which true was output, and divides this by the total number of samples in the test data to obtain the estimation accuracy. In this embodiment, the control unit 1100 determines that generalization has been completed if the estimation accuracy is equal to or greater than a threshold value.

In addition to evaluating the generalization performance, the validity of the hyperparameters may be verified. That is, in a configuration in which a hyperparameter that is a variable amount other than the variable parameters to be learned, such as the number of nodes, is tuned, the control unit 1100 may verify the validity of the hyperparameters based on the validation data. The validation data may be obtained by extracting validation data in advance from the teacher data 3000b using a process similar to that of step S210, and saving it as data not used for training.

If it is not determined in step S225 that generalization of the training model is complete, the control unit 1100 repeats step S220. That is, it performs a process of updating the variable parameters to be learned. On the other hand, if it is determined in step S225 that generalization of the training model is complete, the control unit 1100 records the machine-learned training model as learning model data 3000c in the storage medium 3000 (step S230). With the above configuration, it is possible to generate a learning model that generates polarized OCT images from conventional OCT images based on the efficiently generated teacher data 3000b. Therefore, it is possible to greatly simplify the generation of teacher data 3000b, which is a very burdensome task among the tasks required to train the learning model.

(3) Configuration of conventional OCT system:
8 is a diagram showing a data processing system 1500 functioning as a conventional OCT system according to an embodiment of the present invention. A conventional OCT 2500, a storage medium 3500, an input unit 4500, and a display unit 5500 are connected to the data processing system 1500 according to this embodiment. The conventional OCT 2500 is a device that performs optical coherence tomography by interfering reference light and measurement light without taking polarization sensitivity into consideration. The conventional OCT 2500 captures an image of the subject's eye to generate measurement information and outputs it to the data processing system 1500.

The storage medium 3500, input unit 4500, display unit 5500, and their connection interfaces, communication I/F 1700 and display I/F 1800, can be realized in the same configuration as the storage medium 3000, input unit 4000, display unit 5000, communication I/F 1200, and display I/F 1300 shown in FIG. 1. However, the storage medium 3500 stores measurement information 3500a measured by the conventional OCT 2500 and learning model data 3000c generated by the data processing system 1000. In addition, conventional OCT image data 3500b and polarized OCT image data 3500c are stored in the storage medium 3500 during the operation of the conventional OCT system.

The control unit 1600 includes a CPU, RAM, and ROM (not shown), and can execute various programs stored in the storage medium 3500, ROM, etc. The programs executed by the control unit 1600 include various programs. In this embodiment, the programs include a measurement program 1610 for generating a conventional OCT image and a polarized OCT image based on measurement information 3500a measured by the conventional OCT 2500. Functions executed by the control unit 1600 according to the measurement program 1610 include a measurement unit 1610a, a conventional OCT image generation unit 1610b, and a polarized OCT image generation unit 1610c.

The measurement unit 1610a is a function of measuring the measurement object using an OCT optical system that is not sensitive to polarization. That is, the control unit 1600 controls the conventional OCT 2500 by the function of the measurement unit 1610a, and acquires measurement information 3500a obtained by measuring the subject's test eye. The measurement information 3500a is stored in the storage medium 3500. The conventional OCT image generation unit 1610b is a function of generating a conventional OCT image based on the measurement results. That is, the control unit 1600, by the function of the conventional OCT image generation unit 1610b, accepts a cut surface input by the examiner by operating the input unit 4500, and acquires measurement information 3500a on the cut surface by referring to the measurement information 3500a. Then, the control unit 1600 acquires intensity information for each coordinate on the cut surface based on the measurement information 3500a, and generates a conventional OCT image. In addition, the control unit 1600 stores conventional OCT image data 3500b representing the generated conventional OCT image in the storage medium 3500.

The polarized OCT image generating unit 1610c has a function of inputting the generated conventional OCT image into a learning model and generating a polarized OCT image of the object measured by the conventional OCT 2500. FIG. 9 is a flowchart of the polarized OCT image generating process by the polarized OCT image generating unit 1610c. The polarized OCT image generating process shown in FIG. 9 is a process for collectively converting conventional OCT image data 3500b stored in the storage medium 3500.

When the polarized OCT image generation process is started, the control unit 1600 acquires the conventional OCT image data 3500b stored in the storage medium 3500 (step S300). That is, the control unit 1600 selects and acquires one of the conventional OCT image data 3500b stored in the storage medium 3000 that has not been subject to processing in steps S300 to S315.

Next, the control unit 1600 inputs the conventional OCT image into the learning model (step S305). That is, the control unit 1600 inputs the conventional OCT image indicated by the conventional OCT image data 3500b acquired in step S300 into the generation model G indicated by the learning model data 3000c. As a result, a polarized OCT image is generated.

Then, the control unit 1600 stores the generated polarized OCT image (step S310). That is, the control unit 1600 stores the polarized OCT image data 3500c indicating the generated polarized OCT image in the storage medium 3000. Next, the control unit 1600 determines whether processing of all conventional OCT images has been completed, and repeats the processing from step S300 onwards until it is determined that processing has been completed. Of course, the above processing is one example, and a polarized OCT image may be generated each time a measurement is performed by the conventional OCT 2500, or the captured conventional OCT image or polarized OCT image may be displayed on the display unit 5500.

The above process makes it possible to generate a polarized OCT image measured with a polarization insensitive OCT from a conventional OCT image taken with the conventional OCT 2500, which is a polarization insensitive OCT. Therefore, it is possible to measure and use a polarized OCT image that reflects the polarization characteristics without using the polarization sensitive OCT 2000, which is more expensive than the conventional OCT 2500.

(4) Other embodiments:
The above embodiment is an example for implementing the present invention, and various other embodiments can be adopted. For example, the device configuration of the data processing system is not limited to the configuration shown in FIG. 1. The devices shown in FIG. 1 may be fewer devices that share functions, or may be more devices. Specifically, the data processing system 1000 may be an integrated device with at least one of the polarization-sensitive OCT 2000, the storage medium 3000, the input unit 4000, and the display unit 5000. In addition, the data processing system 1000 may be distributed to more devices, such as at least a part of the data processing system 1000 being configured by a server. Of course, the data processing system 1500 may also be distributed and consolidated in the same manner. Of course, the fact that various configurations are adopted is similar to the configuration shown in FIG. 8.

The measurement information acquisition unit only needs to be able to acquire measurement information that indicates the measurement results of the object to be measured using polarization-sensitive OCT. In other words, the measurement information acquisition unit acquires measurement information that is for acquiring a polarized OCT image and is also capable of creating a conventional OCT image. The object to be measured is not limited to the subject's eye, and any object that can be measured using OCT by interference between reference light and measurement light can be used as the measurement object.

In addition, the measurement information may be adjusted in various ways depending on the application of the learning model. For example, if a learning model for general use is to be generated, it is preferable to prepare measurement information measured under various conditions so as not to cause bias in terms of race, sex, age, etc. On the other hand, if the subject is of a specific race or measurements are planned for a specific case, measurement information for the specific race or measurement information for the specific case may be collected.

Furthermore, the measurement information may be information indicating a measurement result reflecting the polarization characteristics of the measurement object, and is not limited to the Jones matrix as in the above embodiment. For example, the measurement information may be a Jones vector of interference light measured by a polarization-sensitive OCT. That is, when the measurement information 3000a for each coordinate on a cut surface obtained by a certain B-scan is obtained as a Jones vector, it is also possible to specify the intensity I _PD of each coordinate by the following formula (9) and generate a polarization OCT image.

In this case, the conventional OCT image is generated by calculating the intensity as shown in equation (11) based on equation (10) obtained by substituting equation (4) into the left side of equation (7).

Therefore, a configuration using Jones vectors can be realized, for example, by applying a device for measuring Jones vectors as the polarization-sensitive OCT 2000 shown in FIG. 1, generating teacher data 3000b using the data processing system 1000, and generating learning model data 3000c. In other words, if the control unit 1100 calculates a polarization OCT image using equation (9) and calculates a conventional OCT image using equation (11), the other configurations can be realized in the same manner as in FIG. 1.

The device assuming the Jones vector may be various devices, and can be realized, for example, by the configuration shown in FIG. 10. FIG. 10 is a diagram showing a schematic configuration example of a polarization-sensitive OCT. In this example, the polarization-sensitive OCT is equipped with a light source LS, and the light output from the light source LS passes through a polarizer PL (e.g., 90°) and is incident on a beam splitter BS. The light incident on the beam splitter BS is separated into reference light and measurement light. The reference light output from the beam splitter BS passes through a quarter-wave plate QW1 (e.g., 22.5°) and a lens LS1 and is incident on a mirror MR. The reference light reflected by the mirror MR returns to the beam splitter BS via lens LS1 and the quarter-wave plate QW1.

Meanwhile, the measurement light output from the beam splitter BS passes through the galvanometer mirror GM, the quarter-wave plate QW2 (for example, 45°), and the lens LS2 and is incident on the measurement object Sa. The measurement light returning from the measurement object Sa passes through the lens LS2, the quarter-wave plate QW2, and the galvanometer mirror GM and returns to the beam splitter BS.

The reference light and measurement light that return to the beam splitter BS are made into interference light by the beam splitter BS and enter the polarizing beam splitter PBS. The light is then separated by the polarizing beam splitter PBS into interference light of orthogonal polarization components, which are measured by the vertical polarization component detector DV and the horizontal polarization component detector DH. Polarization-sensitive OCT is not limited to this configuration, but in any case, even when a device that measures Jones vectors is used, the above configuration makes it possible to generate an extremely large amount of training data 3000b based on a single measurement, making it possible to generate training data efficiently.

The polarized OCT image acquisition unit only needs to be able to acquire a polarized OCT image that is an image according to the polarization characteristics of the object to be measured based on the measurement information. In other words, the polarized OCT image acquisition unit only needs to be able to acquire a polarized OCT image that visualizes the measurement results that reflect the polarization characteristics. The polarization characteristics only need to indicate the response for each polarization state when the response (scattering, reflection, etc.) from the object to light irradiated onto the object varies depending on the polarization state of the light.

Therefore, a polarized OCT image may be various images other than the above-mentioned PD (polarization-diverse) image. For example, it may be an image that shows cumulative phase retardation due to birefringence, local phase retardation due to birefringence, optic axis of birefringence, diattenuation, degree of polarization uniformity, and polarimetric entropy. In other words, any image that can be generated from the measurement results of polarization-sensitive OCT and shows polarization characteristics can be a polarized OCT image.

The conventional OCT image acquisition unit only needs to be able to acquire multiple conventional OCT images, which are intensity images obtained when the object to be measured is measured using conventional OCT that does not take polarization sensitivity into account, based on the measurement information. In other words, the conventional OCT image acquisition unit only needs to be able to acquire, from the measurement information, conventional OCT images obtained when the object to be measured is photographed using OCT that does not take polarization sensitivity into account, based on the measurement information. There may be various methods for acquiring multiple conventional OCT images from the measurement information, and for example, the angle used as θ is not limited to the above-mentioned embodiment.

Furthermore, the polarization operation using the half-wave plate and horizontal linear polarizer shown in formula (7) is one example, and a conventional OCT image may be generated based on any other polarization operation. The polarization operation may be, for example, a polarization operation using a wave plate with any rotation angle and phase delay amount, or a linear polarizer with any rotation angle. A conventional OCT image can be generated by calculating the Jones vectors of various polarization states using the polarization operation and calculating the intensities of the horizontal polarization component and the vertical polarization component. Note that the polarization operation may be a simpler calculation. For example, any polarization operation may be performed by linearly combining the Jones matrix or each component of the Jones vector indicated by the measurement information with an arbitrary weighting coefficient. Of course, in this case, it is preferable that the magnitude of the vector is maintained before and after the operation.

The teacher data generation unit only needs to be able to generate multiple sets of teacher data in which a polarized OCT image is paired with each of multiple conventional OCT images. In other words, the teacher data generation unit only needs to be able to generate multiple conventional OCT images from measurement information and generate multiple sets of teacher data by associating them with one polarized OCT image. Of course, the measurement information may be the measurement results for multiple test eyes, and teacher data may be generated from each of the measurement results. Furthermore, there is no limit to the number of B-scans, etc., used when acquiring polarized OCT images.

Furthermore, the technique of generating multiple conventional OCT images based on measurement information can also be applied as a method or program invention. In addition, the above-mentioned system, method, and program can be realized as a stand-alone device or as part of a device with multiple functions, and include various aspects.

10... measurement unit, 11... light source, 12... polarization control device, 13... fiber coupler, 14... PM coupler, 21... PM coupler, 22... optical path length difference generating unit, 23... circulator, 24... circulator, 25... splitter, 26... collimator lens, 27... galvanometer mirror, 28... galvanometer mirror, 29... lens, 30... measurement object, 31... PM coupler, 32... PM coupler, 41... optical circulator, 42... collimator lens, 43... reference mirror, 44... PM coupler, 45... optical path length difference generating unit, 46... PM coupler, 47... collimator lens lens, 48...lens, 49...collimator lens, 50...lens, 51...PM coupler, 52...collimator lens, 53...lens, 54...collimator lens, 55...lens, 60...first interference light generation unit, 61...PM coupler, 62...PM coupler, 70...second interference light generation unit, 71...PM coupler, 72...PM coupler, 80...first interference light detection unit, 81...balanced type photodetector, 82...balanced type photodetector, 83...signal processor, 90...second interference light detector, 91...balanced type photodetector, 92...balanced type photodetector, 93...signal processor, 100...clock A pulse generator, 200, a calculation device, 202, a calculation unit, 302, a PM fiber, 304, a PM fiber, 306, an optical path extension unit, 308, an optical path extension unit, 1000, a data processing system, 1100, a control unit, 1110, a teacher data generation program, 1110a, a measurement information acquisition unit, 1110b, a polarization OCT image acquisition unit, 1110c, a conventional OCT image acquisition unit, 1110d, a teacher data generation unit, 1120, a machine learning program, 1120a, a machine learning unit, 1200, a communication I/F, 1300, a display I/F, 1500, a data processing system system, 1600...control unit, 1610...measurement program, 1610a...measurement unit, 1610b...conventional OCT image generation unit, 1610c...polarized OCT image generation unit, 1700...communication I/F, 1800...display I/F, 3000...storage medium, 3000a...measurement information, 3000b...teaching data, 3000c...learning model data, 3500...storage medium, 3500a...measurement information, 3500b...conventional OCT image data, 3500c...polarized OCT image data, 4000...input unit, 4500...input unit, 5000...display unit, 5500...display unit

Claims (12)

a measurement information acquisition unit that acquires measurement information indicating a measurement result of the measurement object by the polarization-sensitive OCT;
a polarization OCT image acquisition unit that acquires a polarization OCT image that is an image according to the polarization characteristics of the measurement object based on the measurement information;
a conventional OCT image acquisition unit that acquires a plurality of conventional OCT images, which are intensity images obtained when the object to be measured is measured by OCT that does not take polarization sensitivity into consideration, based on the measurement information;
A training data generating unit that generates a plurality of sets of training data each including the polarization OCT image and each of the plurality of conventional OCT images;
A teacher data generation system comprising: The measurement information is
A Jones matrix indicating the polarization characteristics of the measurement object.
The teacher data generation system according to claim 1 . The measurement information is
The Jones vector of the interference light measured by the polarization-sensitive OCT.
The teacher data generating system according to claim 1 or 2. The conventional OCT image acquisition unit includes:
a polarization operation for changing a polarization state of the Jones vector of the measurement light output from the measurement object based on the measurement information, the polarization operation being different from each other, and a plurality of the conventional OCT images being obtained based on the components after each polarization operation;
The teacher data generating system according to claim 1 or 2. The polarization manipulation is
This is an operation that changes the polarization state.
The different polarization operations are:
The polarization state after the change is different.
The teacher data generating system according to claim 4 . The polarization manipulation is
This is done by a polarization effect element,
The different polarization operations are:
An operation in which the polarization effect element having a different rotation angle set at every constant angle is applied to the Jones vector of the measurement light.
The rotation angle of the polarization effect element is changed at regular intervals.
The teacher data generating system according to claim 4 . The polarization manipulation is
This is done by a polarization effect element,
The different polarization operations are:
An operation in which the polarization effect element having a different rotation angle obtained by equally dividing a predetermined angle range is applied to the Jones vector of the measurement light.
The teacher data generating system according to claim 4 . The measurement information is information for each coordinate in a three-dimensional space of the measurement object,
the polarized OCT image and the conventional OCT image are images on a two-dimensional plane obtained by cutting the three-dimensional space with a cutting plane including an optical axis,
The polarized OCT images and the conventional OCT images are acquired for a plurality of the cross sections.
The teacher data generating system according to claim 1 or 2. Based on the teacher data generated by the teacher data generation system according to claim 1 or 2,
A machine learning system including a machine learning unit that inputs the conventional OCT image and machine-learns a learning model that outputs the polarization OCT image. A measurement unit that measures a measurement object using an OCT optical system that does not take polarization sensitivity into consideration;
a conventional OCT image generating unit that generates a conventional OCT image based on the measurement result;
a polarization OCT image generating unit that inputs the generated conventional OCT image into the learning model according to claim 9 and generates a polarization OCT image of the measurement object measured by the OCT optical system;
An OCT system comprising: a measurement information acquisition step of acquiring measurement information indicating a measurement result of the measurement object by the polarization-sensitive OCT;
a polarization OCT image acquisition step of acquiring a polarization OCT image, which is an image corresponding to the polarization characteristics of the measurement object, based on the measurement information;
a conventional OCT image acquisition step of acquiring a plurality of conventional OCT images, which are intensity images obtained when the object to be measured is measured by OCT that does not take polarization sensitivity into consideration, based on the measurement information;
A training data generating step of generating a plurality of sets of training data each including the polarization OCT image and each of the plurality of conventional OCT images;
A method for generating teacher data, comprising: Computer,
A measurement information acquisition unit that acquires measurement information indicating a measurement result of the measurement object by the polarization-sensitive OCT;
a polarization OCT image acquisition unit that acquires a polarization OCT image that is an image according to the polarization characteristics of the measurement object based on the measurement information;
a conventional OCT image acquisition unit that acquires a plurality of conventional OCT images, which are intensity images obtained when the object to be measured is measured by OCT that does not take polarization sensitivity into consideration, based on the measurement information;
a training data generating unit that generates a plurality of sets of training data each including the polarization OCT image and each of the plurality of conventional OCT images;
A teacher data generation program that functions as a training data generator.

Images