Measurement signal processing device, measurement signal processing method, and program

Abstract

In the present invention, a model learning unit defines, for each learning set that contains a measurement signal and at least one type of prescribed characteristic value indicative of a characteristic of the measurement signal as a target value, a model parameter used to calculate an estimate value obtained by performing a calculation on the measurement signal using a prescribed mathematical model so that the difference between the estimate value and the target value becomes small, defines a characteristic value set, which is a set of multiple types of characteristic values indicative of characteristics of the measurement signal including the target value, such that each type of the target value becomes common between a plurality of characteristic value sets, and generates, for each of the characteristic sets, the learning set that contains the target value and a measurement signal having the characteristics represented by the aforementioned multiple types of characteristic values.

Description

[Technical field]

[0001] The present invention relates to a measurement signal processing device, a measurement signal processing method, and a program. The present invention relates to a technology for estimating various characteristics of a measured sample by analyzing a signal acquired by, for example, an optical coherence tomography, and a technology for implementing the technology in software. This application claims priority to Japanese Patent Application No. 2020-070913, filed on April 10, 2020, the contents of which are incorporated herein by reference. [Background technology]

[0002] Optical coherence tomography (OCT) is a technology that uses the coherence of light to obtain cross-sectional images of a sample (mainly a living body). OCT can obtain images that show not only the surface of a sample but also its internal structure with high spatial resolution. OCT is applied to medical imaging diagnosis and tissue evaluation of living bodies, and its interest is sometimes directed to changes in tissue and changes in the density of scatterers in tissues such as cell nuclei. On the other hand, the intensity of the OCT signal is affected not only by the scatterer density, but also by the morphology of the tissue itself, the depth of the measurement site, the probe light intensity from the light source and sample, the signal-to-noise ratio, and the image resolution. Therefore, it was not practical to estimate the scatterer density from the OCT signal. As a result, practical diagnosis and evaluation using the scatterer density was not possible. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Timothy R. Hillman, et al., “Correlation of static speckle with sample properties in optical coherence tomography”, OPTICS LETTERS Vol.31, No.2, p.190-192, January 15, 2006 [Non-Patent Document 2] Kazuhiro Kurokawa, et al., “In-plane and out-of-plane tissue micro-displacement measurement by correlation coefficient of optical coherence tomography”, OPTICS LETTERS Vol.40, No.9, p.2153-2156, May 1, 2015 Summary of the Invention [Problem to be solved by the invention]

[0004] On the other hand, a speckle pattern inevitably occurs in the OCT signal. Speckle is an irregularly distributed spot-like pattern formed by light reflected or diffused from the sample. The speckle pattern refers to the spatial distribution of these patterns. The speckle contrast depends on the number of scatterers within the range of the coherence volume of the OCT probe light, as described in Non-Patent Document 1. The coherence volume is a volume in three-dimensional space determined by the spatial resolution of the OCT. This suggests that if the spatial resolution of the OCT is known, the scatterer density can be estimated from the speckle contrast. In other words, it is also possible to relatively and subjectively estimate the scatterer density from the OCT signal under the assumption that the stronger the OCT signal, the higher the scatterer density. However, the spatial resolution of OCT also changes depending on the structure of the sample itself, the aberration caused by the OCT device, and the aberration occurring between the OCT device and the sample, so the resolution for the sample cannot be known in advance. In addition, the intensity of the OCT signal is also affected by factors other than the scatterer density.

[0005] On the other hand, Non-Patent Document 2 describes a method of estimating the resolution in tissue used as a sample by analyzing the speckle pattern. This also suggests that the scatterer density in a sample can be estimated by analyzing the speckle pattern. However, since the speckle pattern is affected by various factors such as the signal-to-noise ratio and aberration of the OCT signal, it is not possible to obtain the scatterer density with sufficient accuracy. For the same reason, it is difficult or impossible to accurately estimate the characteristics of the measurement signal itself or the characteristics of the sample to be measured reflected in the measurement signal from the measurement signal such as the OCT signal in addition to the scatterer density.

[0006] The present invention has been made in consideration of the above-mentioned points, and one object of the present invention is to provide a measurement signal processing device, a measurement signal processing method, and a program that can more accurately estimate a characteristic value that indicates the characteristics of a measurement signal. [Means for solving the problem]

[0007] (1) The present invention has been made to solve the above-mentioned problems, and one aspect of the present invention includes a model learning unit that determines model parameters for calculating an estimated value for each learning set including a measurement signal and at least one predetermined characteristic value that indicates a characteristic of the measurement signal as a target value, such that a difference between the estimated value calculated using a predetermined mathematical model for the measurement signal and the target value becomes small; and a characteristic value set that is a set of multiple types of characteristic values ​​that indicate a characteristic of the measurement signal including the target value, for each of the target values. , the type of the target value is Common across multiple sets of characteristic values To be included in and generating, for each of the characteristic value sets, the learning set including a measurement signal having a characteristic represented by the plurality of types of characteristic values ​​and the target value. The learning data generating unit determines a tissue structure of a sample to be measured based on at least one of the plurality of types of characteristic values, and generates the measurement signal for a wave incident on the sample based on the tissue structure. A measurement signal processing device.

[0009] (2) Another aspect of the present invention is a method for producing a semiconductor device comprising the steps of: )of In the measurement signal processing device, the learning data generating unit may determine characteristic values ​​that fall within a predetermined value range for each of the plurality of types by using random numbers.

[0010] (3) Another aspect of the present invention is (1) or (2) In the measurement signal processing device, the target value may include a scatterer density of a sample, and the multiple types of characteristic values ​​may include an intensity or a signal-to-noise ratio of the measurement signal.

[0011] (4) Another aspect of the present invention is (3) In the measurement signal processing device of claim 1, the learning data generating unit may determine a tissue structure of the sample in which scatterers are distributed at the scatterer density.

[0012] (5) Another aspect of the present invention is (3) or (4) In the measurement signal processing device of claim 1, the target value may further include a spatial resolution in a measurement system.

[0013] (6) Another aspect of the present invention is the (5) In any one of the measurement signal processing devices described above, the plurality of types of characteristic values ​​may include a spatial resolution, an aberration coefficient, or a dispersion coefficient related to chromatic dispersion in a measurement system.

[0014] (7) Another aspect of the present invention is the (6) wherein the mathematical model comprises a neural network including an input layer that outputs a signal value for each observation point of the input measurement signal to a first intermediate layer, a plurality of intermediate layers, and an output layer that outputs an estimated value based on an output value input from a last intermediate layer, the plurality of intermediate layers being configured by repeatedly stacking one or more convolutional layers that output to a next layer a calculated value obtained by performing a convolution process on an input value input from a previous layer, and at least one or more final layers including the output layer may be fully connected layers that output a calculated value obtained by performing a convolution process on all of a plurality of input values ​​input from the previous layer, each using a number of parameter sets that is less than the number of the input values.

[0015] (8) Another aspect of the present invention is the (7) The measurement signal processing device may further include a characteristic estimating unit that calculates an estimate of the target value for the measurement signal by using the model parameters based on the mathematical model.

[0016] (9) Another aspect of the present invention is (8) The measurement signal processing device may further include an output processing unit, wherein the measurement signal indicates a signal value for each observation point, the characteristic estimation unit calculates the estimated value for each block including a plurality of observation points, and the output processing unit converts the estimated value into a pixel value and outputs display data indicating the pixel value to a display unit.

[0017] (10) Another aspect of the present invention is a measurement signal processing device according to any one of (1) to (9), wherein the measurement signal may be any one of an optical coherence tomography measurement signal, a speckle interferometry measurement signal, a photoacoustic measurement signal, and an ultrasonic tomography signal.

[0018] (11) Another aspect of the present invention is a measurement signal processing method in a measurement signal processing device, comprising: a model learning step of determining model parameters for calculating an estimated value calculated for the measurement signal using a predetermined mathematical model for each learning set including a measurement signal and at least one predetermined characteristic value indicating a characteristic of the measurement signal as a target value such that a difference between the estimated value and the target value is small; and a characteristic value set, which is a set of multiple types of characteristic values ​​indicating a characteristic of the measurement signal including the target value, for each of the target values. , the type of the target value is Common across multiple sets of characteristic values To be included in a learning data generating step of generating, for each of the characteristic value sets, the learning set including a measurement signal having a characteristic represented by the plurality of types of characteristic values ​​and the target value; In the learning data generating step, a tissue structure of a sample to be measured is determined based on at least one of the plurality of types of characteristic values, and the measurement signal for a wave incident on the sample is generated based on the tissue structure. A measurement signal processing method.

[0019] (12) Another aspect of the present invention is a method for controlling a computer of a measurement signal processing device to perform a model learning procedure for determining model parameters for calculating an estimated value calculated for a measurement signal using a predetermined mathematical model, for each learning set including a measurement signal and at least one predetermined characteristic value that indicates a characteristic of the measurement signal as a target value, so that a difference between the estimated value and the target value becomes small; and a characteristic value set, which is a set of multiple types of characteristic values ​​that indicate a characteristic of the measurement signal including the target value, for each of the target values. , the type of the target value is Common across multiple sets of characteristic values Included in a learning data generation step of generating, for each of the characteristic value sets, the learning set including a measurement signal having characteristics represented by the plurality of types of characteristic values ​​and the target value; wherein, in the learning data generation step, a tissue structure of a sample to be measured is determined based on at least one of the plurality of types of characteristic values, and the measurement signal for a wave incident on the sample is generated based on the tissue structure. It is a program. Effect of the Invention

[0020] According to the present invention, it is possible to estimate the characteristics of a measurement signal more accurately. For example, it is possible to estimate the scattering density of a tissue as a measurement sample from an OCT signal with higher accuracy than in the past. [Brief description of the drawings]

[0021] [Figure 1] FIG. 2 is an explanatory diagram for explaining an overview of the main measurement signal processing according to the present embodiment. [Diagram 2] 1 is a block diagram showing an example of the configuration of a measurement signal processing device according to an embodiment of the present invention; [Diagram 3] FIG. 1 is a configuration diagram showing an example of an OCT system according to an embodiment of the present invention. [Figure 4] FIG. 2 is an explanatory diagram illustrating an example of a mathematical model according to the present embodiment. [Diagram 5] 1 is a table showing an example of a configuration of a mathematical model according to the present embodiment. [Figure 6] FIG. 11 is a diagram illustrating a first calculation example of a target value. [Figure 7] FIG. 11 is a diagram illustrating a second calculation example of a target value. [Figure 8] FIG. 13 is a diagram illustrating a third calculation example of a target value. [Figure 9]FIG. 13 is a diagram illustrating a fourth calculation example of a target value. [Figure 10] FIG. 13 is a diagram illustrating a fifth calculation example of a target value. [Figure 11] FIG. 13 is a diagram showing an example of observation of a biological sample. [Figure 12] FIG. 11 is a diagram illustrating an example of changes in each index value over time. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0022] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. First, an outline of this embodiment will be described. FIG. 1 is an explanatory diagram for explaining an outline of measurement signal processing according to this embodiment. A measurement signal processing device 100 according to this embodiment is capable of executing a learning step (S01) and an estimation step (S02). The learning step S01 is a process of determining model parameters for estimating a predetermined target value among characteristic values ​​indicating the characteristics of a measurement signal from an input measurement signal according to a predetermined mathematical model. The predetermined target value may be one type, or may be multiple types. In the learning step S01, the measurement signal processing device 100 performs supervised learning. Supervised learning means using a set including known input values ​​and output values ​​as learning data (also called supervised data, training data, teacher data, etc.), and recursively calculating model parameters so that an estimated value calculated according to a mathematical model for an input value approximates an output value.

[0023] In a learning step S01, the measurement signal processing device 100 uses a measurement signal (e.g., an OCT signal) and a target value as an input value and an output value, respectively. At least one predetermined characteristic value of the measurement signal (in the example shown in FIG. 1, the scatterer density (SPUV: Number of Scatterers Per Unit Volume)) is used as the target value. In the learning step S01, the measurement signal processing device 100 generates a plurality of sets of input values ​​and output values ​​(hereinafter, learning sets), and calculates model parameters so that the difference between the estimated value and the output value (hereinafter, model error) is minimized across the plurality of learning sets. For example, a convolutional neural network (CNN) can be used as the mathematical model. Here, minimization does not only mean absolute minimization, but also means performing calculations to estimate or search for model parameters with the aim of minimizing as much as possible. Therefore, the model error does not necessarily decrease monotonically during the minimization process, and may temporarily increase.

[0024] The learning step S01 according to the present embodiment includes, for example, the following steps S102 to S108. (Step S102) The measurement signal processing device 100 sets a plurality of sets of a plurality of types of characteristic values ​​indicating the characteristics of the measurement signal as characteristic value sets. In the measurement signal processing device 100, one characteristic value set includes a plurality of types of characteristic values, and at least one predetermined characteristic value among the plurality of types of characteristic values ​​is set in advance as a target value. That is, one characteristic value set may include a plurality of types of target values. Also, one target value is determined to be common among the plurality of characteristic value sets. The number of types of characteristic values ​​included in one characteristic value set is set to be greater than the number of types of target values. As illustrated in FIG. 1, when SPUV is adopted as one type of characteristic value as a target value, the characteristic value set may include any one of the following types of characteristic values ​​as other types of characteristic values: intensity, SNR (Signal-to-Noise Ratio), spatial resolution, ENS (Effective Number of Scatterers), or a combination consisting of some or all of them.

[0025] (Step S104) The measurement signal processing device 100 determines, for each characteristic value set, the distribution of scatterers in the sample located in a predetermined observation region as the tissue composition of the sample using SPUV as an example of the target value set in step S102. (Step S106) The measurement signal processing device 100 generates (simulates) a measurement signal in the observation region according to a predetermined signal model using a tissue structure determined for each characteristic value set and a characteristic value of a type other than the SPUV, which is the target value. The measurement signal processing device 100 may generate an OCT signal as an example of the measurement signal, or may generate OCT image data for visualizing the state of the sample based on the OCT signal (OCT imaging).

[0026] (Step S108) The measurement signal processing device 100 sets a set of measurement signals and target values ​​(SPUV in the example shown in FIG. 1) generated for each characteristic value set as a learning set including the sets as input values ​​and output values, respectively. The measurement signal processing device 100 calculates (learns) model parameters across multiple learning sets so as to minimize the model error between the estimated values ​​calculated for the measurement signals according to a mathematical model and the output values.

[0027] On the other hand, in the estimation step S02, the measurement signal processing device 100 executes, for example, the processes of the following steps S112 and S114. (Step S112) A measurement signal indicating the state of the sample in a predetermined observation area where the sample is placed is obtained by OCT measurement or the like. (Step S114) The measurement signal processing device 100 calculates an estimate of the target value (in this example, SPUV) using the model parameters obtained in the learning step S01 according to a predetermined mathematical model with the acquired measurement signal as an input value (target value estimation).

[0028] In this way, multiple sets of characteristic values ​​are determined for one target value, and measurement signals generated based on each of the characteristic value sets consisting of multiple types of characteristic values ​​are used as input values, and model parameters are determined so that estimated values ​​for the input values ​​across the multiple characteristic value sets approximate the target value. Using the determined model parameters, values ​​approximating the target value are calculated as estimated values ​​for measurement signals acquired separately as input values. Thus, the measurement signal processing device 100 can determine characteristic values ​​that are robust or invariant against types of characteristic values ​​(e.g., intensity, resolution, etc.) other than the target value as fluctuation factors for the measurement signal.

[0029] (Measurement signal processing device) Next, a description will be given of a configuration example of the measurement signal processing device 100 according to this embodiment. Fig. 2 is a block diagram showing a configuration example of the measurement signal processing device 100 according to this embodiment. The measurement signal processing device 100 includes a control unit 110 , a storage unit 130 , an input / output unit 140 , a display unit 150 , and an operation unit 160 .

[0030] A part or all of the functions of the control unit 110 are realized as a computer including a processor such as a CPU (Central Processing Unit). The processor reads a program previously stored in the storage unit 130, and performs processing instructed by commands written in the read program to fulfill its functions. In this application, performing processing instructed by commands written in a program may be referred to as executing a program, running a program, or the like. A part or all of the control unit 110 is not limited to general-purpose hardware such as a processor, and may be configured to include dedicated hardware such as an LSI (Large Scale Integration) or an ASIC (Application Specific Integrated Circuit). The control unit 110 includes a measurement system control unit 112 , a measurement signal acquisition unit 114 , a learning data generation unit 116 , a model learning unit 118 , a characteristic estimation unit 120 , and an output processing unit 122 .

[0031] The measurement system control unit 112 controls the range of the observation area of ​​the object to be measured used as a sample installed in the measurement system. For example, the measurement system control unit 112 sets the position or size of the observation area based on an operation signal input from the operation unit 160, and controls the position of the observation point within the set observation area. For example, the measurement system control unit 112 drives a driving mechanism that changes the positions of galvanometer mirrors 60a and 60b in the OCT system 1 (FIG. 3) described later, and scans the observation point of the sample Sm. The observation point of the sample Sm is scanned in a direction intersecting the depth direction of the sample Sm (for example, a direction parallel to the front surface of the sample Sm).

[0032] The measurement signal acquisition unit 114 acquires a measurement signal indicating the state of the measured object from the measurement system. For example, the measurement signal acquisition unit 114 sequentially acquires measurement signals from the spectrometer 70 in the OCT system 1. The measurement signal acquisition unit 114 acquires a signal value indicating the intensity distribution of reflected light in the depth direction of the sample Sm for each observation point arranged at a predetermined interval based on the acquired measurement signal. The measurement signal acquisition unit 114 repeats a process of acquiring the intensity distribution of reflected light for each observation point changed by scanning. The acquired intensity distribution of reflected light is based on the distribution of the refractive index of the sample Sm in the depth direction. The measurement signal acquisition unit 114 can acquire an OCT signal indicating the state of the sample Sm in the observation area as a measurement signal. The OCT signal has a signal value indicating the complex amplitude of the intensity of interference light resulting from interference between the reference light incident on the sample Sm and the reflected light from the sample Sm for each observation point. The measurement signal acquisition unit 114 stores the acquired measurement signal in the storage unit 130.

[0033] The learning data generating unit 116 generates a learning set for each set of characteristic values. The learning data generating unit 116 stores a series of data groups including the learning sets generated for each set of characteristic values ​​in the storage unit 130 as learning data. One learning set includes a measurement signal as an input value and at least one type of predetermined characteristic value as a target value. First, the learning data generating unit 116 generates a plurality of characteristic value sets. Each characteristic value set is configured to include a plurality of types of characteristic values, and at least one of the plurality of types of characteristic values ​​is adopted as a target value. However, the learning data generating unit 116 determines each characteristic value set such that one target value is common to two or more characteristic value sets. The combinations of the plurality of types of characteristic values ​​constituting each characteristic value set are different between the individual characteristic value sets. In the present application, a characteristic value set that has some types of characteristic values ​​in common with a certain characteristic value set but different characteristic values ​​for the remaining types is distinguished as a separate characteristic value set. For example, in two characteristic value sets 1 and 2 that include SPUV and SNR as characteristic values ​​and have SPUV as a target value, the learning data generating unit 116 determines a characteristic value set that has an SPUV of 0.01 particles / μm 3 and the SNR is 10 dB.3 The characteristic value set 2 with an SNR of 20 dB is the SPUV of 0.01 particles / μm 3 The training data generating unit 116 generates the characteristic value sets that are common to each other but different from each other in that they are different in the characteristic value set. Even when each of the target values ​​is a set of characteristic values ​​including a plurality of types of characteristic values, the training data generating unit 116 distinguishes the target values ​​that differ in at least one type of characteristic value as different target values. In general, the training data generating unit 116 sets a plurality of target values. That is, the total number of generated characteristic value sets is the sum of the number of characteristic value sets set for each of the target values. Therefore, the number of training sets (the number of sets) used to generate one set of model parameters is equal to the number of characteristic value sets, and is the sum of the number of characteristic value sets common to each of the target values ​​(this number is two or more as described above) across the plurality of target values.

[0034] Also, each characteristic value can be either a scalar quantity or a vector quantity. For example, SPUV, ENS, intensity, SNR, etc. are each scalar quantities. On the other hand, the resolution may be treated as a vector quantity having the same number of dimensions as the observation area, or may be treated as an independent scalar quantity. For example, the resolution in a three-dimensional observation area may be treated as a three-dimensional vector including the x-direction resolution, y-direction resolution, and z-direction resolution as elements, or the x-direction resolution, y-direction resolution, and z-direction resolution may be treated as separate scalar quantities. However, the x-direction, y-direction, and z-direction are directions of coordinate axes that are mutually orthogonal in a three-dimensional Cartesian coordinate system. The z-direction is the direction of the depth of the sample, and the x-direction and y-direction are mutually orthogonal directions in a plane perpendicular to the z-direction. The process of determining this series of characteristic value sets corresponds to the process of step S102 above.

[0035] In setting the characteristic value set, the learning data generating unit 116 determines a plurality of possible values ​​within a predetermined range for each type of characteristic value. For example, the learning data generating unit 116 determines three or more characteristic values ​​that are distributed at equal intervals within a predetermined range for that type. The learning data generating unit 116 may also generate pseudo-random numbers (hereinafter, uniform random numbers) that are uniformly distributed within a predetermined range using a predetermined algebraic method, and set values ​​obtained by scaling the generated pseudo-random numbers within a range corresponding to the type of characteristic value as the characteristic value. As the algebraic method, for example, a linear congruential method, an M-sequence method, or the like can be used. For a characteristic value that is a vector quantity, the learning data generating unit 116 determines the value of each element of the vector individually. In general, the value range of the characteristic value may differ depending on the type of characteristic. However, a common value range may be set for element values ​​or characteristic values ​​of the same type, such as x-direction resolution and y-direction resolution. The learning data generating unit 116 may be able to set the type and range of characteristic values ​​to be processed based on an operation signal input from an operation unit 160, for example.

[0036] The training data generator 116 uses some or all of the characteristic values ​​of the characteristic value set to generate a measurement signal having characteristics expressed by the characteristic values ​​in accordance with a predetermined propagation model for the incident wave. The process of determining the scatterer distribution illustrated below corresponds to the process of step S104 above, and the process of determining an OCT speckle pattern as an example of a measurement signal from the scatterer distribution corresponds to the process of step S106 above.

[0037] Taking the case where the characteristic values ​​constituting the characteristic value set include SPUVs as an example, the training data generation unit 116 generates a tissue structure in which scatterers are stochastically distributed in a predetermined observation region with the determined SPUV. More specifically, when the observation region is a three-dimensional region having a rectangular parallelepiped shape, the training data generation unit 116 calculates the number of scatterers by multiplying the determined SPUV by the number of observation points in the occupied region of the sample located in the observation region. The occupied region of the sample is set in advance. When assuming that the sample occupies the entire observation region, the occupied region of the sample is not taken into consideration, and the training data generation unit 116 regards the number of observation points in the observation region as the number of observation points in the occupied region.

[0038] The training data generating unit 116 then sequentially generates uniform random numbers and scales every third random number generated for the predetermined width, depth, and depth of the observation region to determine the coordinate values ​​(x, y, z) of the scatterers in the observation region. The training data generating unit 116 repeats the process of determining the positions of the scatterers a number of times corresponding to the number of scatterers. A distribution of the scatterers (hereinafter, scatterer distribution f(x, y, z)) in the observation region is generated as a tissue structure.

[0039] Taking the case where the characteristic value set includes spatial resolution as an example, the learning data generating unit 116 determines a point spread function (PSF) corresponding to the spatial resolution. For example, as shown in formula (1), the learning data generating unit 116 determines a three-dimensional normal distribution having x, y, and z directional resolutions as standard deviations in the x, y, and z directions, respectively, as PSF(x, y, z). Then, as shown in formula (2), the learning data generating unit 116 performs a convolution operation of the point spread function PSF(x, y, z) with the scatterer distribution f(x, y, z) to determine an OCT speckle pattern O(x, y, z). The OCT speckle pattern is an OCT signal that can be observed when incident light is scattered by a scatterer. Note that in formula (1), σ x ,σ y ,σ z indicate the standard deviation in the x, y, and z directions, respectively.

[0040]

number

[0041]

number

[0042] For example, when the characteristic value set includes intensity, the learning data generating unit 116 scales the measurement signal by multiplying the signal value for each observation point in the above OCT speckle pattern O(x, y, z) by a constant proportional to the intensity. The learning data generating unit 116 updates the data indicating the signal value for each observation point after scaling as a new OCT speckle pattern (x, y, z).

[0043] For example, when the characteristic value set includes SNR, the learning data generation unit 116 adjusts the measurement signal by adding a noise component value to the signal value of each observation point in the above-mentioned OCT speckle pattern O(x, y, z). When determining the noise component value, the learning data generation unit 116 sequentially generates pseudo-random numbers whose distribution is a complex normal distribution, and determines the noise component value for each observation point by multiplying each generated random number by a gain. The learning data generation unit 116 determines a gain by which the noise component value is multiplied so that the ratio of the absolute value sum of squares of the signal value to the absolute value sum of squares of the noise component value in the observation area becomes the SNR. Note that, when scaling is performed on the measurement signal, the learning data generation unit 116 adds the noise component value to the scaled measurement signal.

[0044] Note that a plurality of types of characteristic values ​​may have mutual dependencies. In that case, even if a certain type of characteristic value is not set, the learning data generation unit 116 may derive the characteristic value of that type from the characteristic values ​​of other types. Also, the learning data generation unit 116 may or may not use the derived characteristic value as a target value. For example, the learning data generation unit 116 may determine the SPUV by dividing the set ENS by a coherence volume which is the product of the spatial resolutions in the x, y, and z directions. Conversely, the learning data generation unit 116 may determine the ENS by multiplying the SPUV by a coherence volume which is the product of the spatial resolutions in the x, y, and z directions. Also, the learning data generation unit 116 may determine the cube root of the volume obtained by dividing the ENS by the SPUV by the ratio of the spatial resolutions in the z direction and the xy direction as the x-direction spatial resolution and the y-direction spatial resolution, respectively, and may multiply the x-direction spatial resolution by the ratio of the spatial resolutions to determine the z-direction spatial resolution. The learning data generating unit 116 may also assume that the noise level, i.e., the sum of squares of the noise component values ​​within the tactile region, is a constant value and determine the SNR to be proportional to the intensity of the signal component, or may calculate the signal value for each observation point by adding the signal component value and the noise component value scaled by the intensity. As described above, the SNR is given by the ratio of the sum of squares of the absolute values ​​of the signal value to the sum of squares of the absolute values ​​of the noise component value.

[0045] In the above explanation, an example was given in which an OCT speckle pattern in an observation area in a three-dimensional space was defined as a measurement signal, but an OCT speckle pattern in an observation area in a two-dimensional space can be defined as a measurement signal using a similar method. In that case, it is sufficient to omit one of the coordinates in the direction of the coordinate axis among the x, y, and z coordinates. For example, the point spread function PSF(x, y) and OCT speckle pattern O(x, y) in a two-dimensional space such as a sample surface can be given by omitting the z coordinate, as shown in equations (3) and (4), respectively.

[0046]

number

[0047]

number

[0048] The model learning unit 118 reads out the learning data from the storage unit 130, and determines one set of model parameters using a series of multiple learning sets included in the read learning data. The process of determining the model parameters corresponds to the process of step S108 above. The model parameters refer to one or more parameters used in a calculation according to a mathematical model, and are also called parameter sets or hyperparameters. The model learning unit 118 sets the measurement signal for each learning set included in the learning data as an input value of the mathematical model. The model learning unit 118 calculates the model parameters so as to minimize the model error between the target value and an estimated value calculated as an output value according to the mathematical model for the measurement signal as a whole series of learning data. For example, the mean squared error (MSE) can be used as a loss function, which is an index indicating the magnitude of the model error. The MSE is expressed as the mean squared error (MSE) of the target value c among N samples i as shown in Equation (5): i and the estimated value c i p The MSE is a simple average value of the squares of the differences between the estimated and target values. The smaller the MSE, the smaller the model error, that is, the higher the degree of approximation of the estimated value to the target value. The model learning unit 118 repeats a process of recursively updating the model parameters using, for example, the steepest descent method until the MSE becomes equal to or less than a predetermined judgment threshold. The model learning unit 118 stores model parameter data indicating the determined model parameters in the storage unit 130. An example of the mathematical model according to this embodiment will be described later.

[0049]

number

[0050] The characteristic estimation unit 120 reads out the model parameter data stored by the model learning unit 118 and the measurement signal stored by the measurement signal acquisition unit 114 from the memory unit 130. The process of acquiring the measurement signal corresponds to the process of step S112 described above. The characteristic estimation unit 120 calculates, as an output value, an estimate of a target value to be calculated, using model parameters indicated by the read model parameter data, in accordance with the mathematical model described above for the read measurement signal. The process of calculating the estimate corresponds to the process of step S114 described above. The characteristic estimation unit 120 stores the calculated estimate in the memory unit 130 in association with the measurement signal.

[0051] The output processing unit 122 controls the output of various data based on an operation signal input from the operation unit 160. The operation signal instructs, for example, whether or not to display an OCT image, whether or not to display a target value, etc. When an operation signal instructing the display of an OCT image is input, the output processing unit 122 reads out the measurement signal acquired by the measurement signal acquisition unit 114 or the learning data generation unit 116 from the storage unit 130, converts the absolute value of the complex amplitude, which is the signal value for each observation point indicated by the measurement signal, into a luminance value within a predetermined value range of the pixel corresponding to the observation point, and outputs output image data having the converted luminance value to the display unit 150 as display data. As a result, an OCT image based on the output image data is displayed on the display unit 150. The output processing unit 122 may select a measurement signal to be output based on the operation signal. When an operation signal instructing the display of the target value is input, the output processing unit 122 reads out the estimated value of the target value calculated by the characteristic estimation unit 120 from the storage unit 130, and outputs display data indicating the read estimated value to the display unit 150. The output processing unit 122 may cause the display unit 150 to display the estimated value by superimposing it on an OCT image based on the measurement signal used to calculate the estimated value.

[0052] The output processing unit 122 may read out setting screen data for guiding the setting of the above-mentioned setting values ​​(for example, the observation region, the target value, the type of characteristic value, the value range, etc.) from the storage unit 130, and output the read out setting screen data as display data to the display unit 150. In this case, a setting screen based on the setting screen data is displayed on the display unit 150.

[0053] The storage unit 130 stores the above programs, as well as various data used in the processes executed by the control unit 110 and various data acquired by the control unit 110. The storage unit 130 includes, for example, a non-volatile (non-transient) storage medium such as a ROM (Read Only memory), a flash memory, or a HDD (Hard Disk Drive). The storage unit 130 includes, for example, a volatile storage medium such as a RAM (Random Access memory) or a register.

[0054] The input / output unit 140 is connected wirelessly or by wire to other devices so as to input and output various data. The input / output unit 140 includes, for example, an input / output interface. The input / output unit 140 is connected to, for example, peripheral devices and a measurement system. Display unit 150 displays display information such as images, characters, and symbols based on display data input from control unit 110. Display unit 150 may include, for example, any of a liquid crystal display, an organic electroluminescence display, and the like.

[0055] The operation unit 160 may be configured to include members that receive a user's operation and generate an operation signal according to the received operation, such as a button, a knob, a dial, a mouse, or a joystick. The operation unit 160 outputs the acquired operation signal to the control unit 110. The operation unit 160 may be an input interface that receives an operation signal wirelessly or wired from another device (for example, a portable device such as a remote controller).

[0056] (OCT system) Next, an OCT system 1 according to this embodiment will be described. Fig. 3 is a configuration diagram showing an example of the OCT system 1 according to this embodiment. The OCT system 1 is an observation system for observing the state of a sample using OCT. The OCT system 1 irradiates the sample Sm with light, obtains interference light generated by interfering between the light reflected from the sample Sm and the reference light reflected by the reference mirror 40, and obtains an OCT signal indicating the state of the surface and the interior of the sample Sm as a measurement signal from the obtained interference light. The OCT system 1 converts the obtained measurement signal into an OCT image to visualize it.

[0057] The object to be observed as the sample Sm may be, for example, a living human or animal body, or a non-living body. The living body may be the fundus of the eye, blood vessels, teeth, subcutaneous tissue, etc. The non-living body may be an artificial structure such as an electronic part or a machine part, a natural structure such as stone or a mineral, or a substance without a specific shape.

[0058] The OCT system 1 includes a light source 10, a beam splitter 20, collimators 30a, 30b, 50a, and 50b, a reference mirror 40, galvanometer mirrors 60a and 60b, a spectrometer 70, and a measurement signal processing device 100. Among these components, the beam splitter 20, the collimators 30a, 30b, 50a, and 50b, the reference mirror 40, the galvanometer mirrors 60a and 60b, and the spectrometer 70 constitute an optical system called an interferometer. The interferometer illustrated in FIG. 1 is a Michelson interferometer. More specifically, the light source 10, the spectrometer 70, the collimator 30a, and the collimator 50a are each connected to the beam splitter using an optical fiber. The OCT system 1 is, for example, a Fourier-domain OCT (FD-OCT).

[0059] The light source 10 is a broadband light source such as an ultrashort pulse laser or an SLD (Superluminescent Diode). The light source 10 irradiates, for example, a probe light having a near-infrared wavelength (for example, 800 to 1000 nm) and low coherence. The light irradiated from the light source 10 is guided inside an optical fiber and enters the beam splitter 20. The beam splitter 20 splits the incident light into light guided toward the collimator 30a (hereinafter, reference light) and light guided toward the collimator 50a (hereinafter, measurement light). The beam splitter 20 is, for example, a cube beam splitter.

[0060] The collimator 30a converts the reference light guided from the beam splitter 20 into parallel light, and emits the parallel light toward the collimator 30b. The collimator 30b collects the parallel light incident from the collimator 30a and emits the collected reference light toward the reference mirror 40. The collimator 30b receives the reference light reflected by the reference mirror 40, converts it into parallel light, and emits the converted parallel light toward the collimator 30a. The collimator 30a collects the parallel light incident from the collimator 30b and guides it toward the beam splitter 20.

[0061] On the other hand, the collimator 50a converts the measurement light guided from the beam splitter 20 into parallel light and emits the converted parallel light toward the galvanometer mirror 60a. At the surfaces of the galvanometer mirrors 60a and 60b, the parallel light incident from the collimator 50a is reflected and emitted toward the collimator 50b. The collimator 50b collects the parallel light incident from the collimator 50a via the galvanometer mirrors 60a and 60b, and irradiates the collected measurement light onto the sample Sm. The measurement light irradiated onto the sample Sm is reflected by the reflecting surface of the sample Sm and enters the collimator 50b. The reflecting surface is not limited to, for example, the boundary surface between the sample Sm and the surrounding environment of the sample Sm (e.g., the atmosphere), but may be a boundary surface that divides materials or tissues with different refractive indices inside the sample Sm. Hereinafter, the light reflected by the reflecting surface of the sample Sm and incident onto the collimator 50b is called reflected light.

[0062] The collimator 50b outputs the incident reflected light toward the galvanometer mirror 60b. The light is reflected by the surfaces of the galvanometer mirrors 60b and 60a, respectively, and is output toward the collimator 50a. The collimator 50a collects the parallel light incident from the collimator 50a via the galvanometer mirrors 60a and 60b, and guides the collected reflected light toward the beam splitter 20.

[0063] The beam splitter 20 guides the reference light reflected by the reference mirror 40 and the light reflected by the sample Sm to the spectrometer 70 via optical fibers. The spectrometer 70 includes a diffraction grating and a light receiving element therein. The diffraction grating separates the reference light and the reflected light guided from the beam splitter 20. The separated reference light and the reflected light interfere with each other to become interference light. The light receiving element is disposed on an imaging surface onto which the interference light is irradiated. The light receiving element detects the irradiated interference light and generates a measurement signal according to the detected interference light. The light receiving element outputs the generated measurement signal to the measurement signal processing device 100. Note that it is desirable that the interval between adjacent observation points in the observation region is equal to or less than the spatial resolution of the optical system.

[0064] (Mathematical Model) Next, an example of a mathematical model according to this embodiment will be described. Fig. 4 is an explanatory diagram showing an example of a mathematical model according to this embodiment. Fig. 5 is a table showing the configuration of the mathematical model exemplified in Fig. 4. Fig. 5 shows the number, type, number of kernels, kernel size, stride, activation function and axis for each layer in each column. However, - indicates that there is no corresponding item due to the type of layer.

[0065] The mathematical model shown in Figures 4 and 5 is a CNN. In the example shown in Figures 4 and 5, the input value to the CNN is a three-dimensional OCT signal, and the output value from the CNN is used to calculate a one-dimensional target value (scalar). The one-dimensional characteristic value corresponds to, for example, the diffuser density, signal intensity, ENS, SNR, and resolution in each direction (i.e., each of the x, y, and z direction resolutions).

[0066] CNN is a type of artificial neural network, and includes one input layer, multiple intermediate layers, and an output layer. The CNN illustrated in FIG. 4 includes an input layer 118a, a feature extraction layer group 118b, and an estimation layer group 118c. The 0th layer illustrated in FIG. 5 corresponds to the input layer 118a, the 1st to 16th layers correspond to the feature extraction layer group 118b, and the 17th and 18th layers correspond to the estimation layer group 118c. Each layer belonging to the feature extraction layer group 118b and the first layer of the estimation layer group 118c correspond to the intermediate layer, and the last layer of the estimation layer group 118c corresponds to the output layer. Each layer has one or more nodes (also called nodes, neurons, etc.). Each node outputs a function value of a predetermined function for an input value as an output value. However, when a three-dimensional measurement signal is handled, the resolution takes a three-dimensional vector value. Therefore, when calculating the resolution as a three-dimensional output value, a CNN having an output layer with three nodes may be used.

[0067] The input layer 118a outputs to the next layer a signal value for each observation point indicated by a measurement signal input as an input value. A signal value of an observation point corresponding to each node of the input layer 118a is input to the node, and the input signal value is output to the corresponding node of the next layer. The number of kernels in FIG. 5 indicates the number of kernels used for processing (e.g., arithmetic) for each input value. A kernel refers to a processing unit for calculating one output value at a time. A kernel is also called a filter. A kernel size indicates the number of input values ​​used in one processing in a kernel. For example, the 0th layer in FIG. 5 has a kernel with 32×32×32 nodes constituting the input layer 118a, and indicates that one output value is given for each input value to each node.

[0068] The feature extraction layer group 118b calculates a feature extraction value indicating the feature from a multidimensional input value. In the example shown in FIG. 4, the feature extraction layer group 118b is formed by alternately stacking one or more convolutional layers and pooling layers two or more times. More specifically, the feature extraction layer group 118b is formed by stacking two adjacent convolutional layers Conv11 and Conv12, one pooling layer Pool13, two convolutional layers Conv21 and Conv22, one pooling layer Pool23, two convolutional layers Conv31 and Conv32, one pooling layer Pool33, and convolutional layers Conv41 and Conv42 in that order. However, one normalization layer is sandwiched between the layer after each of the two convolutional layers and the layer next thereto. The feature extraction layer group 118b includes a flattening layer as the last layer.

[0069] A convolutional layer is a layer that performs a convolution operation for each kernel on the input values ​​input from the previous layer to each of multiple nodes to calculate a convolution value, calculates the function value of a predetermined activation function for a correction value obtained by adding the calculated convolution value and a bias value as an output value, and outputs the calculated output value to the next layer. In the convolution operation, one or more input values ​​are input to each node from the previous layer, and an independent convolution coefficient is used for each input value. The parameters of the convolution coefficient, bias value, and activation function are part of one set of model parameters.

[0070] For example, in the first and second layers in FIG. 5, the kernel numbers are set to 1 and 20, respectively. The kernel size of both the first and second layers is 32×32×32. The number of kernels is 20, which means that there are 20 convolution coefficients, and output values ​​corresponding to each of the 20 kernels are obtained for each node, which are input to the corresponding nodes of the next layer. In addition, ReLU in the activation function section means that the activation function is a rectified linear unit. The rectified linear unit is a function that sets a threshold value (for example, 0) as the output value for an input value equal to or less than the threshold, and outputs an input value that exceeds the threshold as is. Therefore, this threshold value can also be part of a set of model parameters. In addition, for the convolution layer, whether or not it is necessary to refer to an input value from a node in the previous layer, and whether or not it is necessary to output an output value to a node in the next layer can also be part of a set of model parameters. Therefore, unlike the fully connected layer described below, each node in a convolutional layer is not necessarily connected to all nodes in the previous layer so that it receives input values, nor is it necessarily connected to all nodes in the next layer so that it outputs output values.

[0071] A normalization layer is a layer that normalizes input values ​​input to each of multiple nodes within a predetermined range using a common normalization parameter (batch normalization), and outputs the normalized value as an output value. Note that the axis item of the third layer in Figure 5 is set with a code indicating the type and range of parameters as the axis used for normalization. The -1 set in this item is a code indicating that the output values ​​from all nodes are normalized within the range from 0 to 1 by multiplication using a common multiplication value.

[0072] A pooling layer is a layer having a node that determines one representative value from input values ​​input from multiple nodes in the previous layer and outputs the determined representative value as an output value to the next layer. The representative value is, for example, a value that statistically represents multiple input values, such as the maximum value, average value, or mode. Note that the "2" set in the stride item of the fourth layer in FIG. 5 indicates the range of adjacent nodes in the previous layer that refer to the input value for one node. In the example shown in FIG. 5, 2x2x2 spatially adjacent previous nodes in three-dimensional space are referenced. Therefore, a pooling layer is a layer that downsamples the input value from the previous layer to a lower dimension and provides the output value to the next layer.

[0073] The estimation layer group 118c is a layer that calculates an estimated value as an output value from the feature extraction value input from the feature extraction layer group 118b. The estimation layer group 118c is formed by stacking a flattening layer (Flatten Layer, FIG. 5, but omitted in FIG. 4) and two fully connected layers FC01 and FC02. The flattening layer is a layer that unrolls the input value of a multidimensional sample of three or more dimensions into an output value of a two-dimensional sample. For example, the 16th layer in FIG. 5 receives as input values ​​the element values ​​of 80 three-dimensional 4×4×4 observation points normalized and output in the 15th layer, and linearly rearranges the three-dimensional array of 5120 input values ​​in total into one vector having 5120-dimensional element values. Therefore, the flattening layer does not perform substantial arithmetic processing. The fully connected layer is a layer that performs a convolution operation on input values ​​input from the previous layer to each of a plurality of nodes to calculate a convolution value, calculates an operation value obtained by adding the calculated convolution value and a bias value as an output value, and outputs the calculated output value to the next layer. In other words, the fully connected layer is a layer that outputs an operation value obtained by performing a convolution process on all of a plurality of input values ​​input from the previous layer using parameter sets (kernels) whose number of sets is less than the number of input values. A predetermined activation function may be set to each node of the fully connected layer, and a function value of the predetermined activation function for the calculated correction value may be output to the next layer as an output value. However, in the 17th layer of FIG. 5, a normalized linear unit is set as an activation function for each of 128 kernels, and one output value is calculated by referring to the input values ​​from 40 nodes constituting the 16th layer as input values. In the 18th layer of FIG. 5, one convolution value obtained by performing a convolution operation on input values ​​from 128 nodes constituting the 17th layer is calculated as an output value. Therefore, in the fully connected layer, the parameters of the convolution coefficients, bias values, and activation functions become part of one set of model parameters. In this way, by making one or more final layers, including the output layer, fully connected layers, it is possible to reduce the degrees of freedom while taking into account all components that significantly affect the characteristic values ​​given from the previous layer, and ultimately derive the characteristic values.

[0074] The number of layers of the CNN, the type of each layer, the number of nodes of each layer, and the like are not limited to those shown in FIG. 4 and FIG. 5. The CNN according to this embodiment may have a configuration capable of calculating an estimate of a predetermined target value as an output value for a measurement signal having signal values ​​for a plurality of observation points as an input value. The estimate value is not limited to a one-dimensional scalar value as described above, but may be a vector or matrix of two or more dimensions. For example, the characteristic value may be a three-dimensional resolution vector having x-direction resolution, y-direction resolution, and z-direction resolution as elements. In addition, the measurement signal is not limited to a three-dimensional signal in which observation points are distributed in a three-dimensional space, but may be a signal of a lower dimension than a two-dimensional signal, or a high-dimensional signal of four or more dimensions. However, as exemplified in FIG. 4 and FIG. 5, the CNN according to this embodiment is preferably provided with an intermediate layer configured by alternately repeating one or more convolution layers and pooling layers for two or more periods and sequentially stacking them. This is because the repetition of the convolution layers narrows down the components that significantly affect the characteristic values. In addition, the pooling layer may be omitted in this repetition of the convolution layers.

[0075] As described above, the characteristic estimation unit 120 uses a mathematical model such as CNN to calculate an estimate for each block consisting of a plurality of mutually adjacent observation points that form an OCT signal. Therefore, an OCT signal that gives one frame of an OCT image usually corresponds to a plurality of blocks. Therefore, when an operation signal indicating the display of a target value is input, the output processing unit 122 may convert the estimate calculated for each block into a luminance value within a predetermined value range of the pixel corresponding to that block. The output processing unit 122 outputs display data indicating the converted luminance value to the display unit 150. As a result, a target value image that expresses the distribution of the target value for each block using luminance is displayed on the display unit 150.

[0076] (Example of calculating characteristic values) Next, a calculation example of the characteristic values ​​calculated by the estimation step (S02) using the model parameters obtained by executing the learning step (S01) will be described. In the following calculation example, a two-dimensional sample surface (en-face surface) with 32 × 32 observation points was set as the observation area. Each characteristic value set includes x-direction resolution, y-direction resolution, SPUV, and SNR as characteristic values. As a measurement signal, an OCT speckle pattern in which scatterers are stochastically distributed at the SPUV and has the set x-direction resolution, y-direction resolution, and SNR as its characteristics was used. Then, the model learning unit 118 was caused to calculate CNN model parameters for calculating an estimate of the target value using the measurement signal and multiple learning data having any of the x-direction resolution, y-direction resolution, SPUV, and SNR as a target value. That is, a total of four sets of model parameters for calculating the estimates of the x-direction resolution, y-direction resolution, SPUV, and SNR were calculated from one common measurement signal given by each one characteristic value set. However, the x-direction resolution and y-direction resolution ranges are 3 to 20 μm, the SNR range is 0 to 50 dB, and the SPUV is 0 to 0.0497 particles / μm. 2 The characteristic values ​​of each type were determined using uniform random numbers so that they were randomly distributed within their ranges.

[0077] In addition, a learning set based on each of the 50,000 characteristic value sets was used to calculate one set of model parameters. The 50,000 characteristic value sets each had one target value (e.g., SPUV of 0.03 particles / μm 2) are included, and an OCT speckle pattern of 32 pixels x 32 pixels is generated from each of the characteristic value sets. Although it depends on the specifications of various devices and the scanning method of the OCT signal, the size of one pixel corresponds to approximately 1 μm to several μm. Then, the characteristic estimation unit 120 is caused to calculate, as a measurement signal, an OCT speckle pattern whose target value indicating the characteristic is known, as an estimate of each target value using each of the four sets of model parameters calculated by the model learning unit 118 according to the CNN illustrated in FIG. 5. However, the characteristic estimation unit 120 uses an OCT speckle pattern different from the OCT speckle pattern used by the model learning unit 118 to calculate the model parameters.

[0078] 6 to 10 are diagrams showing examples of calculation of target values. In each of Figs. 6 to 10, the vertical axis shows the estimated value of the characteristic value, and the horizontal axis shows the set value of the characteristic value. However, an OCT speckle pattern having signal values ​​of 32 x 32 x 32 points in width, depth, and depth in the x-, y-, and z-directions, respectively, totaling 32,768 points, was used as an input value. In Figs. 6 to 10, the estimated value refers to the estimated value of the target value as an output value calculated from the measurement signal using the model parameters, and the set value refers to the target value included in the characteristic value set used to calculate the model parameters. Figs. 6, 7, 8, 9, and 10 respectively show the x-direction resolution, y-direction resolution, SPUV, SNR, and ENS as target values. It is shown that the set value and the estimated value are almost equal for each of the x-direction resolution, y-direction resolution, SPUV, SNR, and ENS.

[0079] The coefficient of determination R obtained by performing linear regression analysis on the distribution of estimated values ​​for each individual setting for x-direction resolution, y-direction resolution, SPUV, SNR, and ENS. 2 The values ​​were 0.965, 0.961, 0.981, 0.957, and 0.991, which are close to 1. In linear regression analysis, it was assumed that the set value used as the input value was equal to the estimated value used as the output value. Generally, the coefficient of determination R 2indicates the contribution of the variance determined by the input value to the variance of the output value, and the closer it is to 1, the higher the estimation accuracy using the model parameters. The calculation examples shown in Figures 6 to 10 show that model parameters that enable highly accurate estimation can be obtained for x-direction resolution, y-direction resolution, SPUV, SNR, and ENS. However, there are some differences in estimation accuracy depending on the type of target value. For example, the coefficient of determination R for x-direction resolution and y-direction resolution 2 are almost the same, at 0.965 and 0.961, respectively, while the coefficients of determination R 2 are 0.981 and 0.991, respectively, and the coefficient of determination for resolution R 2 is higher. In particular, for ENS, the variance around the regression line is smaller than for other estimates, and the variance tends to decrease as the absolute value of ENS increases. This suggests that when ENS is high, the effect of fluctuations in the measurement signal due to an increase in scatterers is more significant than that due to the resolution. In contrast, for SNR, the variance tends to increase as the absolute value of SNR increases. This suggests that when SNR is high, factors other than noise, such as calculation errors, have a significant effect on the accuracy of SNR estimation.

[0080] In the examples shown in Figs. 4 to 10, the learning data generating unit 116 generates a learning set including one characteristic value from each of the one characteristic value sets as a target value, and the model learning unit 118 generates model parameters using the target value included in the learning set as an output value, but the present invention is not limited to this. The learning data generating unit 116 may generate a learning set including two or more characteristic values ​​from each of the one characteristic value sets as target values. This allows the model learning unit 118 to generate model parameters for simultaneously calculating two or more characteristic values ​​as estimated values ​​calculated from the measurement signal as an input value. For example, the x-direction resolution, the y-direction resolution, and the SPUV may be included in one learning set.

[0081] In the verification example described below, the model parameters learned under the above conditions were used to evaluate the SPUV estimated from the OCT signal acquired from a real biological sample using the OCT system 1 over a long period of time. Human breast cancer cell spheroids (MCF-7 cell line) were used as the biological sample. The diameter of the biological sample was approximately 250 μm. The biological sample was cultured in a culture chamber under an environment where carbon dioxide (CO2) was supplied from seeding. After that, the culture was terminated, the biological sample was removed from the culture chamber, and placed in a culture medium. No CO2 was supplied to the medium, and the temperature was set to 22°C. The OCT signal of the biological sample was acquired every two hours.

[0082] In acquiring the OCT signal, SS-OCT (Swept Source OCT) was used, and the central wavelength of the light source was set to 1.3 μm. SS-OCT corresponds to one form of FD-OCT. The A-scan speed of the OCT signal was set to 50,000 lines per second. In addition, the spatial resolution in observing the OCT signal was 19 μm in the lateral direction (corresponding to any direction in the xy plane mentioned above) and 14 μm in the depth direction (corresponding to the z direction mentioned above). In the evaluation, in addition to SPUV calculated from the acquired OCT signal, logarithmic intensity variance (LIV) was used. LIV corresponds to the time variance of the logarithmic intensity of the OCT signal within a predetermined time (e.g., 1 to 10 seconds). LIV is used as an index value indicating the degree of dynamics of biological tissue.

[0083] FIG. 11 shows an example of the observation of a biological sample. In FIG. 11, the first, second, and third rows show an OCT image, an SPUV image, and an LIV image, respectively. Each image is based on an OCT signal observed every 4 hours from the end of the culture. Each OCT image shows the OCT signal intensity at each observation point. The brighter the area, the higher the OCT signal intensity. The SPUV image shows the SPUV for each block (32 samples x 32 samples in this example). The LIV image shows the LIV for each observation point. The brighter the area, the higher the LIV. According to this, while the OCT image shows that the OCT intensity does not change significantly over time, the SPUV and LIV images show that the SPUV and LIV decrease significantly over time. In particular, the decrease in SPUV from immediately after the end of the culture (0 hr) to 4 hours later (4 hr) is remarkable. This reflects the phenomenon that intracellular motion, which was active immediately after the end of the culture, slows down over time.

[0084] FIG. 12 is a diagram showing an example of the change over time of each index value. In FIG. 12, the vertical axis shows the average OCT intensity, the average SPUV, and the average LIV in the first, second, and third columns, respectively, and the horizontal axis shows the time. However, in the first row, the central part of the biological sample was observed. In the second row, the peripheral part of the biological sample was observed. The average OCT intensity is stable in the central part of the biological sample, and in the peripheral part, although fluctuations are observed, a downward trend with time is observed. In contrast, for both the central part and the peripheral part of the biological sample, the average SPUV significantly decreases within 4 hours after the end of the culture, and the change thereafter becomes slow. The average LIV shows a relatively constant downward trend over the observation period of 28 hours. These facts show that the SPUV estimated from the OCT signal by this embodiment is more convenient for observing the activity of cells than the OCT signal. Furthermore, it is considered that the activity of cells due to different mechanisms can be observed with the SPUV and the LIV.

[0085] In the above example, the x-direction resolution, y-direction resolution, SPUV, SNR, and ENS are mainly used as characteristic values, but the present invention is not limited to this. The characteristic value set may include, for example, an aberration coefficient indicating the mode and degree of wavefront aberration of the optical system. Since the aberration brings about a phase change in the xy plane with respect to the measurement signal, it can be considered as a type of factor that causes fluctuations in the measurement signal. As the aberration coefficient, for example, the Zernike coefficient A nm For example, the 0th order Zernike coefficient A 00 is the constant term, the 1st and 0th order Zernike coefficients A 10 is the gradient component in the x direction, and the 1st-order Zernike coefficient A 11 is the gradient component in the y direction, and the 2nd and 0th order Zernike coefficients A 20 are the astigmatism in the 0° and 90° directions, and the 2nd and 1st order Zernike coefficients A 21 is the focus shift, and the 2nd and 2nd order Zernike coefficients A 22 is the astigmatism in the 45° direction, and the 3rd and 1st order Zernike coefficients A 31 is the x-component coma, and the third and second order Zernike coefficients A 32 are coefficients indicating the degree of y-component coma, etc.

[0086] When the incident direction of light is the z direction, the frequency (f x ,f y ) for aberration W(f x ,f y ) is expressed by the Zernike polynomial shown in equation (6). In equation (6), ρ represents the radial frequency from the optical axis oriented in the z direction, and θ represents the azimuthal frequency. R n n-2m (ρ) denotes a radial polynomial of the nth, nmth degree, and is defined by formula (7). In formula (7), ...! denotes the factorial of an integer .... n is an integer of 0 or more, and m is an integer of 0 or more and n or less. However, the maximum degree k is a predetermined integer of 0 or more. In other words, the aberration W(f x ,f y ) are the nth and mth Zernike coefficients Anm and the sum of the products of the n-th and n-2m-th degree radial polynomials at that position and the cosine value cos|n-2m|θ (where n-2m is 0 or greater) or the sine value sin|n-2m|θ (where n-2m is less than 0).

[0087]

number

[0088]

number

[0089] Therefore, when generating a measurement signal that includes the influence of aberration, the learning data generating unit 116 performs a two-dimensional Fourier transform on the three-dimensional point spread function PSF(x, y, z) on the xy plane to obtain a frequency domain (f x ,f y ) conversion coefficient p(f x ,f y Then, the learning data generating unit 116 converts the input data into a conversion coefficient p(f x ,f y ;z) Phase factor e, which indicates the phase change caused by aberration iaW(fx,fy ) for the conversion coefficient after phase change obtained by multiplying x -f y A two-dimensional inverse Fourier transform is performed on the plane to calculate the point spread function PSF'(x,y,z) including the phase change due to the aberration in the three-dimensional spatial domain (x,y,z). In equation (8), a is a predetermined real value indicating the degree of the phase change due to the aberration. fx,fz -1 […] is the f x -f y This shows the two-dimensional inverse Fourier transform on a plane.

[0090]

number

[0091] The learning data generation unit 116 performs a convolution operation of the above-mentioned scatterer distribution f(x, y, z) with the PSF'(x, y, z) containing aberration, and generates an OCT speckle pattern O'(x, y, z) that has been subjected to a phase change due to aberration as a measurement signal. Therefore, by using the measurement signal that has been subjected to a phase change due to aberration as an input value of the learning data, the learning data generation unit 116 can calculate model parameters for calculating the aberration coefficient as a target value, or model parameters for robustly calculating other types of characteristic values ​​as target values ​​against fluctuations in the aberration coefficient.

[0092] The characteristic value set may also include, for example, a dispersion coefficient related to the chromatic dispersion of the optical system. Since chromatic dispersion brings about a phase change in the measurement signal in the z direction according to the wavelength, it can be considered as a factor that causes fluctuations in the measurement signal. In general, chromatic dispersion occurs due to the frequency dependence of the propagation constant in the optical system. The propagation constant is a physical quantity that corresponds to an effective wave number. When the propagation constant β(ω) is expanded by Taylor expansion with respect to the frequency ω around the center frequency ω0 of the light source, it is expressed as shown in Equation (9).

[0093]

number

[0094] In equation (9), β(ω0) represents the propagation constant at the center frequency ω0 of the incident light. a1 and a2 represent the first and second derivatives at the center frequency ω0, respectively. The first derivative a1 corresponds to the inverse of the group velocity, and the second derivative a2 corresponds to the group velocity dispersion. Therefore, for example, the group velocity and group velocity dispersion can be used as the dispersion coefficient.

[0095] Therefore, when generating a measurement signal including the influence of chromatic dispersion, the learning data generating unit 116 performs a one-dimensional Fourier transform in the z direction on the three-dimensional point spread function PSF(x, y, z) to convert it into a conversion coefficient q(ω; x, y) in the frequency domain (ω). Then, the learning data generating unit 116 adds a phase factor e, which indicates a phase change due to chromatic dispersion, to the conversion coefficient q(ω; x, y) as shown in equation (10). ibβ(ω) A one-dimensional inverse Fourier transform is then performed in the ω direction on the transformation coefficients after the phase change, which are obtained by multiplying by F, to calculate the point spread function PSF″(x,y,z) that has undergone a phase change due to chromatic dispersion in the three-dimensional spatial domain (x,y,z). Note that in equation (10), b is a predetermined real value that indicates the degree of phase change due to chromatic dispersion. F ω -1 […] denotes the one-dimensional inverse Fourier transform of… in the ω direction.

[0096]

number

[0097] The learning data generation unit 116 performs a convolution operation of the above-mentioned scatterer distribution f(x, y, z) with the PSF''(x, y, z) containing the phase change due to chromatic dispersion, and can generate an OCT speckle pattern O''(x, y, z) subjected to the phase change due to chromatic dispersion as a measurement signal. Therefore, by using the measurement signal that has been subjected to a phase change due to chromatic dispersion as an input value of the learning data, the learning data generation unit 116 can calculate model parameters for calculating the dispersion coefficient as a target value, or model parameters for robustly calculating other types of characteristic values ​​as target values ​​against fluctuations in the dispersion coefficient. The characteristic value set may include, as the dispersion coefficient, either the aberration coefficient or the dispersion coefficient, or may include both.

[0098] In the above description, the measurement signal to be processed by the measurement signal processing device 100 is an OCT signal, but this is not limited to this. The measurement signal can be applied to this embodiment even if it is a signal that indicates the state of a sample using a measurement principle other than OCT or a signal for analyzing that state. The measurement signal is not limited to a raw signal obtained by measurement or a signal generated by simulating the raw signal, but may be a signal obtained by performing a predetermined post-processing on the signal, for example, output display data such as an OCT image, the main purpose of which is visualization.

[0099] The measurement signal may be, for example, any one of a speckle interferometry signal, a photoacoustic measurement signal, an ultrasonic tomography signal, and the like. The speckle interferometry signal is a measurement signal acquired by using speckle interferometry. The speckle interferometry method includes a step of irradiating a coherent wave such as a laser beam as an incident wave onto a sample in a predetermined reference state, detecting an interference wave generated by interference between the incident wave and a reflected wave from the sample as a reference wave, a step of further detecting an interference wave generated by irradiating a sample whose state such as position, direction, or shape has changed from the reference state by using a similar method, and a step of acquiring a speckle interferometry signal showing an interference fringe obtained by interfering the detected interference wave with the reference wave. The speckle interferometry signal is mainly used to analyze the amount of movement, change in orientation, or deformation of the sample as a characteristic value. The incident wave is not limited to a laser beam, which is a visible light, but may be an X-ray, an infrared ray, an ultrasonic wave, or the like.

[0100] The photoacoustic measurement signal is a measurement signal obtained by using the photoacoustic method. The photoacoustic method includes a step of irradiating a sample with a laser light, which is visible light or infrared light (including near-infrared light), as an incident wave, and detecting a sound wave emitted from the sample as a photoacoustic measurement signal. Here, the molecules constituting the sample absorb the incident wave and generate heat, and sound waves are generated by the expansion of the sample due to the generated heat. The photoacoustic measurement signal is used for the distribution of light absorbers contained in the sample. Therefore, as a characteristic value derived from the distribution of light absorbers, it can be used for analyzing, for example, the density of light absorbers and active sites of the sample. The generated sound waves are not limited to audible sounds, and may be ultrasonic waves.

[0101] The ultrasonic tomography signal is a measurement signal obtained by ultrasonic tomography. The ultrasonic tomography includes a step of irradiating a sample with ultrasonic waves as an incident wave and detecting an interference wave between the incident wave and a reflected wave emitted from the sample as an ultrasonic tomography signal. The ultrasonic tomography signal is used to analyze characteristic values ​​derived from the density, distribution, structure, etc. of tissue. The above tomography is not limited to a method using visible light or ultrasonic waves as an incident wave, but may use other types of waves such as infrared rays, ultraviolet rays, or X-rays.

[0102] As described above, the measurement signal processing device 100 according to the present embodiment includes a model learning unit 118 that determines model parameters for calculating an estimated value for each learning set including a measurement signal and at least one type of predetermined characteristic value indicating the characteristics of the measurement signal as a target value, so that the difference between the estimated value calculated using a predetermined mathematical model for the measurement signal and the target value is small. The measurement signal processing device 100 also includes a learning data generating unit 116 that determines a characteristic value set, which is a set of multiple types of characteristic values ​​indicating the characteristics of the measurement signal including the target value, so that the characteristic value set is common among multiple characteristic value sets for each type of target value, and generates a learning set including a measurement signal having characteristics indicated by multiple types of characteristic values ​​and the target value for each characteristic value set. The measurement signal may be any of an OCT signal, a speckle interferometric measurement signal, a photoacoustic measurement signal, an ultrasonic tomography signal, and the like. One type of target value may include two or more types of characteristic values. In that case, model parameters for estimating two or more types of characteristic values ​​from the measurement signal can be generated. By using the generated model parameters, it is possible to estimate two or more types of characteristic values ​​from the measurement signal with practical accuracy, which was previously difficult or impossible. The two or more types of characteristic values ​​may include, for example, the density of scatterers and the resolution of the measurement system.

[0103] With this configuration, measurement signals based on multiple types of characteristic values ​​are generated so that each characteristic value is common among multiple learning sets, and model parameters can be determined so that an estimated value calculated from the generated measurement signals is close to a target value. Therefore, by using the determined model parameters, a target value can be calculated from the measurement signals according to a mathematical model, robust against fluctuations in a type of characteristic value different from the target value. For example, when the characteristic value is the scatterer density of the sample and the parameters of the multiple types of characteristic values ​​include the intensity or signal-to-noise ratio of the measurement signal, the scatterer density of the sample can be estimated without depending on fluctuations in the intensity or signal-to-noise ratio. As an example of the scatterer density, the cell nucleus density of the biological tissue used as the sample can be more accurately estimated using the measurement signal. Therefore, the burden of artificial judgment and erroneous judgment in pathological research and diagnosis can be reduced.

[0104] The learning data generating unit 116 may determine a tissue structure of the sample (e.g., a tissue structure in which scatterers are distributed at a set scatterer density) based on at least one of a plurality of types of characteristic values ​​(e.g., scatterer density), and generate a measurement signal for a wave incident on the sample based on the determined tissue structure.

[0105] With this configuration, measurement signals are generated based on tissue structures determined using set characteristic values, making it possible to quantitatively indicate the characteristics of tissue structures under different conditions and generate a variety of measurement signals.

[0106] Furthermore, the learning data generating unit 116 may use random numbers to determine the characteristic values ​​so that each of the multiple types falls within a predetermined range of values. This configuration makes it possible to efficiently generate multiple types of characteristic values ​​that are randomly distributed across a large number of sets of characteristic values. In addition, it is possible to reduce the bias of the distribution of the multiple types of characteristic values ​​compared to artificially determined values. As a result, model parameters that enable the calculation of the target value with higher accuracy can be obtained.

[0107] The multiple types of characteristic values ​​may include a spatial resolution, an aberration coefficient, or a dispersion coefficient related to chromatic dispersion in a measurement system for the sample, thereby reducing or eliminating a decrease in the estimation accuracy of the characteristic values ​​caused by fluctuations in the spatial resolution, the aberration coefficient, or the chromatic dispersion that depend on the measurement system itself or on the sample to be measured.

[0108] The mathematical model also includes a neural network including an input layer that outputs a signal value for each observation point of the input measurement signal to the first intermediate layer, multiple intermediate layers, and an output layer that outputs an estimated value based on the output value input from the last intermediate layer. The multiple intermediate layers may be configured by repeatedly stacking one or more convolution layers that output a calculated value obtained by performing a convolution process on the input value input from the previous layer to the next layer. With this configuration, features that contribute to the fluctuation of the target value are extracted from the input measurement signal with each stack of layers, so that the estimation accuracy of the target value can be improved. Furthermore, at least one or more final layers including the output layer may be fully connected layers that output a calculated value obtained by performing a convolution process on all of the multiple input values ​​input from the previous layer using parameter sets whose number of sets is less than the number of the input values. In this way, the degree of freedom is reduced while taking into consideration all components that significantly affect the characteristic value given from the previous layer, and the characteristic value can finally be derived.

[0109] The measurement signal processing device 100 may include a characteristic estimating unit 120 that calculates an estimate of the target value for the measurement signal by using the model parameters based on the mathematical model. With this configuration, the measurement signal processing device 100 can calculate an estimate of the target value with high accuracy from the input measurement signal by using the model parameters acquired by the device itself.

[0110] Although the embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to the above, and various design changes, etc. are possible within the scope that does not deviate from the gist of the present invention.

[0111] For example, the measurement signal processing device 100 may be realized as a part of the OCT system 1, or may be a single device independent of the optical system. In that case, the measurement system control unit 112 may be omitted from the control unit 110 of the measurement signal processing device 100. The measurement signal acquisition unit 114 is not limited to the optical system, and may acquire the measurement signal from other devices such as a data storage device or a PC.

[0112] Furthermore, the measurement signal processing device 100 may include a display unit 150 and an operation unit 160, or one or both of them may be omitted. In the measurement signal processing device 100, one or both of the characteristic estimation unit 120 and the output processing unit 122 may be omitted. When the characteristic estimation unit 120 is omitted, the model learning unit 118 may output the calculated model parameters to other devices such as a data storage device, a PC, or other measurement signal processing devices. The device to which the data is output may have the same function as the characteristic estimation unit 120, that is, a function of calculating a characteristic value for the measurement signal according to a predetermined mathematical model using the model parameters acquired from the measurement signal processing device 100.

[0113] In the above description, the training data generating unit 116 generates, as information indicating the tissue structure, a scatterer distribution f(x, y, z) indicating a situation in which scatterers are stochastically randomly distributed in the observation area with a scatterer density given as a characteristic value, but the present invention is not limited to this. The training data generating unit 116 may acquire tissue structure information indicating the tissue structure of the sample from another device, and the tissue structure information may include a plurality of tissue structures and known characteristic values ​​indicating the characteristics (e.g., SPUV, ENS, etc.) possessed by each tissue structure. When acquiring tissue structure information, the training data generating unit 116 does not need to determine the tissue structure of the sample, and these tissue structures and characteristic values ​​may be used as part of the characteristic value set. The tissue structure may be one in which scatterers are regularly arranged in the observation area, for example, a lattice arranged spatially at a predetermined period. In addition, the scattering intensity or scattering cross section of each scatterer may be included in the characteristic value set as a characteristic value related to the tissue structure.

[0114] Although the mathematical model is exemplified as a CNN, the present invention is not limited to this. The mathematical model may be another type of neural network, such as a recurrent neural network or a probabilistic neural network. As long as the relationship between the measurement signal and the target value can be obtained, another type of mathematical model, such as a random forest or a multiple regression model, may be used. The multiple regression model is preferable if the relationship between the measurement signal as the input value and the target value as the output value is linear or can be approximated to linear. In addition, the activation function in CNN is not limited to the normalized linear unit, and any of a sigmoid function, a softmax function, etc. may be used. The index of model error is not limited to the MSE, and cross entropy, etc. may be used.

[0115] Also, a part or the whole of the measurement signal processing device 100 in the above-mentioned embodiment may be realized as an integrated circuit such as an LSI (Large Scale Integration). Each functional block of the measurement signal processing device 100 may be individually processed, or a part or the whole may be integrated and processed. The integrated circuit method is not limited to LSI, and may be realized by a dedicated circuit or a general-purpose processor. Also, when an integrated circuit technology that replaces LSI appears due to the progress of semiconductor technology, an integrated circuit based on that technology may be used. [Explanation of symbols]

[0116] 1...OCT system, 10...light source, 20...beam splitter, 30a, 30b, 50a, 50b...collimator, 40...reference mirror, 60a, 60b...galvanometer mirror, 70...spectroscope, 100...measurement signal processing device, 110...control unit, 112...measurement system control unit, 114...measurement signal acquisition unit, 116...learning data generation unit, 118...model learning unit, 120...characteristic estimation unit, 122...output processing unit, 130...storage unit, 140...input / output unit, 150...display unit, 160...operation unit

Claims

1. a model learning unit that determines model parameters for calculating an estimated value calculated for the measurement signal using a predetermined mathematical model, for each learning set including a measurement signal and at least one predetermined characteristic value indicating a characteristic of the measurement signal as a target value, so as to reduce a difference between the estimated value and the target value; defining a characteristic value set, which is a set of a plurality of types of characteristic values ​​indicating characteristics of the measurement signal including the target value, for each type of the target value such that the type of the target value is included in common among the plurality of characteristic value sets; a learning data generating unit that generates, for each of the characteristic value sets, the learning set including a measurement signal having a characteristic represented by the plurality of types of characteristic values ​​and the target value; The learning data generation unit determining a tissue structure of a sample to be measured based on at least one of the plurality of types of characteristic values; and generating the measurement signal for the wave incident on the sample based on the tissue structure. Measurement signal processing device.

2. The learning data generation unit Using random numbers, determine characteristic values ​​that fall within a predetermined range for each of the plurality of types. The measurement signal processing device according to claim 1 .

3. the target value includes a scatterer density of the sample; The plurality of types of characteristic values ​​include the intensity or the signal-to-noise ratio of the measurement signal.

3. The measurement signal processing device according to claim 1.

4. The learning data generation unit A structure of the sample in which scatterers are distributed is determined based on the scatterer density. The measurement signal processing device according to claim 3 .

5. The target value further includes the spatial resolution of the measurement system.

5. The measurement signal processing device according to claim 3.

6. The plurality of types of characteristic values ​​include a spatial resolution, an aberration coefficient, or a dispersion coefficient related to wavelength dispersion in a measurement system. The measurement signal processing device according to any one of claims 1 to 5.

7. The mathematical model is a neural network including an input layer that outputs a signal value for each observation point of the input measurement signal to a first intermediate layer, a plurality of intermediate layers, and an output layer that outputs an estimated value based on an output value input from a last intermediate layer; The intermediate layers are configured by repeatedly stacking one or more convolutional layers each of which outputs a calculated value obtained by performing a convolution process on an input value input from a previous layer to a next layer, At least one final layer including the output layer is A fully connected layer outputs a calculated value obtained by performing convolution processing on all of the multiple input values ​​input from the previous layer using parameter sets whose number is less than the number of the input values. The measurement signal processing device according to any one of claims 1 to 6.

8. A characteristic estimating unit is provided for calculating an estimated value of the target value for the measurement signal based on the mathematical model and using the model parameters. The measurement signal processing device according to any one of claims 1 to 7.

9. An output processing unit is provided, The measurement signal indicates a signal value for each observation point, The characteristic estimation unit calculates the estimated value for each block including a plurality of observation points; The output processing unit includes: The estimated value is converted into a pixel value, and display data indicating the pixel value is output to a display unit. The measurement signal processing device according to claim 8 .

10. The measurement signal is any one of an optical coherence tomography measurement signal, a speckle interferometry measurement signal, a photoacoustic measurement signal, and an ultrasonic tomography signal. The measurement signal processing device according to any one of claims 1 to 9.

11. A measurement signal processing method in a measurement signal processing device, comprising: a model learning step of determining model parameters for calculating an estimated value calculated for the measurement signal using a predetermined mathematical model, for each learning set including a measurement signal and at least one predetermined characteristic value indicating a characteristic of the measurement signal as a target value, so that a difference between the estimated value and the target value becomes small; defining a characteristic value set, which is a set of a plurality of types of characteristic values ​​indicating characteristics of the measurement signal including the target value, for each type of the target value such that the type of the target value is included in common among the plurality of characteristic value sets; a learning data generating step of generating, for each of the characteristic value sets, the learning set including a measurement signal having characteristics represented by the plurality of types of characteristic values ​​and the target value; In the learning data generating step, determining a tissue structure of the sample to be measured based on at least one of the plurality of types of characteristic values; generating a measurement signal for a wave incident on the sample based on the tissue structure; Measurement signal processing method.

12. The computer of the measurement signal processing device a model learning procedure for determining model parameters for calculating an estimated value calculated for the measurement signal using a predetermined mathematical model, for each learning set including a measurement signal and at least one predetermined characteristic value indicating a characteristic of the measurement signal as a target value, such that a difference between the estimated value and the target value becomes small; defining a characteristic value set, which is a set of a plurality of types of characteristic values ​​indicating characteristics of the measurement signal including the target value, for each type of the target value such that the type of the target value is included in common among the plurality of characteristic value sets; a learning data generating step for generating, for each of the characteristic value sets, the learning set including a measurement signal having characteristics represented by the plurality of types of characteristic values ​​and the target value, In the learning data generation step, determining a tissue structure of the sample to be measured based on at least one of the plurality of types of characteristic values; generating a measurement signal for a wave incident on the sample based on the tissue structure; program.

Images