JP7378063B2 - Temperature estimation device and temperature estimation method - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act (Publication 1) Publication date October 26, 2018 Publication Journal of the Japan Society of Computer Surgery Special issue of the 27th Annual Conference of the Japan Society of Computer Surgery, Page 272, Japan Society of Computer Surgery (Publication 2) Date: November 9, 2018 Name of the meeting: 27th Japan Society of Computerized Surgery Conference Venue: Nara Prefectural Cultural Center (6-2 Noborioji-cho, Nara City, Nara Prefecture)

The present invention relates to a temperature estimation device and a temperature estimation method.

Endoscopic laser therapy is known as a method for treating lesions within the body, in which the lesion is treated by directly irradiating laser light through an optical fiber inserted into the forceps port of an endoscope. Currently, the only information that the surgeon can grasp regarding the lesion during endoscopic laser treatment is a two-dimensional image of the lesion, but if the temperature of the lesion can be grasped in real time, it would be possible to Laser treatment can be performed more appropriately (without overheating or underheating).

Therefore, the development of a technique for estimating the temperature of a lesion from the optical characteristics of the lesion (see Non-Patent Document 1) is underway.

Takeshi Seki, Akito Takahashi, Kiyoshi Oka, Akita Naganawa, "Basic verification of temperature estimation method for endoscopic laser therapy", 310th Research Meeting of the Tohoku Branch of the Society of Instrument and Control Engineers (2017.7.21) , document number 310-1

There is a strong correlation between the temperature of the lesion and the optical characteristics of the lesion during endoscopic laser treatment. However, during endoscopic laser treatment, the lesion degenerates due to heating, and as a result, the optical characteristics and thermal characteristics of the lesion change. Therefore, it has been difficult to accurately estimate the temperature of the lesion from the optical characteristics of the lesion. Furthermore, in the industrial field, there has been a strong demand for technology that accurately estimates the temperature of a specific location from optical characteristics.

The present invention has been made in view of the above-mentioned problems, and an object of the present invention is to provide a technology that can accurately estimate the temperature of a sample that can be denatured by temperature rise during heating from its optical characteristics in real time.

A temperature estimating device according to one aspect of the present invention includes a measurement unit that periodically measures an optical characteristic value regarding reflectance of the sample during heating of the sample that can be denatured by temperature rise; The measurement is performed using a trained neural network including an output unit for outputting the output, and a plurality of input units into which the optical characteristic values measured in time series by the measurement unit are input at mutually different delay times. and a temperature estimating section that estimates the temperature of the sample at the time of the current measurement of the optical characteristic value, each time the optical characteristic value is measured by the section.

In other words, this temperature estimating device is configured to estimate the temperature at each point in time of a sample that can be denatured by temperature rise (heating), taking into account the optical characteristic values related to the reflectance of the sample measured up to that point. has. Since the optical characteristic values related to the reflectance of the sample measured up to each point in time are information indicating the degree of denaturation of the sample, according to this temperature estimation device, it is possible to The temperature of a warm medium can be estimated in real time and accurately from its optical properties.

Various values can be adopted as the optical characteristic values of the temperature estimation device. For example, the area under the curve of the reflected light spectrum of the sample may be employed as the optical characteristic value.

If the sample whose temperature is to be estimated is one whose temperature has been increased by irradiation with laser light, the neural network further includes at least one input unit for inputting a value indicating the intensity of the laser light. You may also adopt.

The present invention also provides a measurement unit that irradiates light during heating of a sample that can be denatured by heating and measures characteristic values obtained from the reflected light;
Estimating the temperature of the sample using a trained neural network including an output unit for outputting the temperature of the sample and a plurality of input units into which the characteristic values measured by the measurement unit are input. A temperature estimation section;
Equipped with
When training the neural network,
The output unit outputs a value of a specific function corresponding to the temperature of the sample in a temperature range to be measured,
Among the output values of the specific function, a value corresponding to the temperature of the learning target is given a label of 1, and other values are given a label of 0,
The temperature estimating device executes repeated calculations to update a coupling coefficient in the neural network so that a cross-entropy error based on the value of the specific function outputted by the output unit and the value of the label is reduced. It's okay.

According to this, the number of neural networks is only one regardless of the time related to the temperature information of the sample. Furthermore, it is not necessary to input past characteristic values into the neural network; it is sufficient to input only current information. Note that the above specific function may be any one of a sigmoid function, a softmax function, a linear function, a step function, a ramp function, and a hyperbolic tangent function.

Further, in the present invention, the characteristic value may include a wavelength distribution of the reflected light, biological information based on the reflected light, intensity of the irradiated light, and an image of the sample. Thereby, it is possible to improve the estimation accuracy of the temperature estimation device. Further, the characteristic value may include numerical data obtained by color space decomposition of image information of the sample. According to this, it is possible to improve estimation accuracy when estimating temperature using an image of a sample.

Further, a temperature estimation method according to one aspect of the present invention includes a measurement step of periodically measuring an optical characteristic value regarding the reflectance of the sample while increasing the temperature of the sample that can be denatured by temperature increase; Using a trained neural network including an output unit for outputting temperature and a plurality of input units into which the optical characteristic values measured in time series in the measurement step are input with mutually different delay times, Each time the optical characteristic value is measured in the measuring step, the method includes a temperature estimation step of estimating the temperature of the sample at the time of the current measurement of the optical characteristic value.

This temperature estimation method also makes it possible to accurately estimate in real time the temperature during heating of a sample, which can be denatured by heating, from its optical properties. Note that in the present invention, a "trained neural network" includes not only a neural network that has completed learning and no further learning is performed, but also a neural network that is continuously trained. For example, it is possible to retrain a trained neural network in real time using values indicating the intensity of laser light, optical property values, etc. to improve accuracy.

Furthermore, the temperature estimation method of the present invention includes a measurement step of irradiating a sample that can be denatured by temperature rise with light during temperature rise and measuring characteristic values obtained from the reflected light;
Estimating the temperature of the sample using a trained neural network including an output unit for outputting the temperature of the sample and a plurality of input units into which the characteristic values measured by the measurement unit are input. a temperature estimation step;
A temperature estimation method comprising:
When training the neural network,
The output unit outputs a value of a specific function corresponding to the temperature of the sample in a temperature range to be measured,
Among the output values of the specific function, a value corresponding to the temperature of the learning target is given a label of 1, and other values are given a label of 0,
The temperature estimation method includes performing repeated calculations for updating a coupling coefficient in the neural network so that a cross-entropy error based on the value of the specific function output by the output unit and the value of the label is reduced. It's okay. According to this, the number of neural networks is only one regardless of the time related to the temperature information of the sample. Furthermore, it is not necessary to input past characteristic values into the neural network; it is sufficient to input only current information. Here, the above specific function may be any one of a sigmoid function, a softmax function, a linear function, a step function, a ramp function, and a hyperbolic tangent function.

According to the present invention, the temperature during heating of a sample that can be denatured by heating can be estimated in real time and accurately from its optical properties.

FIG. 1 is a schematic configuration diagram of a temperature estimating device according to a first embodiment of the present invention. FIG. 2 is an explanatory diagram of how the temperature estimating device according to the embodiment is used. FIG. 3 is an explanatory diagram of an example of the configuration of an optical fiber that makes it easy to measure the intensity of reflected light. FIG. 4 is an explanatory diagram of AUC (area under the curve). FIG. 5 is an explanatory diagram of a neural network used by the temperature estimator for temperature estimation. FIG. 6 is an explanatory diagram of hidden units of a neural network. FIG. 7 is a diagram plotting the actually measured temperature of the target to be irradiated with laser light, the temperature estimated by the temperature estimation device (“estimated temperature”), and the temperature estimated by the linear model against the elapsed time after the start of laser irradiation. be. FIG. 8 is an explanatory diagram of a neural network used for temperature estimation by the temperature estimation unit according to the second embodiment of the present invention. FIG. 9 is a flowchart of the learning process of the temperature estimation device according to the second embodiment of the present invention. FIG. 10 is a diagram plotting the actually measured temperature of the laser beam irradiation target and the temperature estimated by the temperature estimation device (“estimated temperature”) against the elapsed time after the start of laser beam irradiation according to Example 2 of the present invention. It is.

Embodiments of the present invention will be described below with reference to the drawings.

<Example 1>
FIG. 1 shows a schematic configuration diagram of a temperature estimating device 10 according to Example 1 of the present invention, and FIG. 2 shows an explanatory diagram of a usage pattern of the temperature estimating device 10 according to the embodiment.

A temperature estimation device 10 (FIG. 1) according to this embodiment is a device developed to estimate a lesion 40 (FIG. 2) during endoscopic laser treatment. Here, as shown in FIG. 2, endoscopic laser treatment refers to the lesion area 40 being directly irradiated with laser light through the optical fiber 35 inserted into the forceps port 31 of the endoscope 30. It is a laser treatment method to treat. The temperature estimating device 10 is configured as a device that estimates the temperature of the lesion 40 based on the intensity of the measurement light and the white light from the ride guide 32 reflected by the lesion 40 (details will be described later). Therefore, the optical fiber 35 has a structure that makes it easy to measure the reflected light intensity, for example, as shown in FIG. A device having a plurality (six in the figure) of light-receiving optical fibers 37 arranged around it is used.

As shown in FIG. 1, the temperature estimation device 10 includes a spectrometer 11, an information processing device 12, and a display device 13.

The spectrometer 11 is a device that periodically measures the spectrum of reflected light from the lesion 40 (hereinafter referred to as reflected light spectrum). Hereinafter, the measurement period of the reflected light spectrum by this spectrometer 11 will be referred to as a sampling period.

The display device 13 is a device for presenting to the user the result of estimating the temperature of the lesion 40 by the information processing device 12 (temperature estimation unit 22 described later). As the display device 13, a liquid crystal display that can also display text information or a segment display that can display only temperature is used.

The information processing device 12 is a so-called computer (computer body). The information processing device 12 is configured (programmed) to function as an AUC calculation section 21, a temperature estimation section 22, a learning processing section 23, and a storage section 24. Further, the information processing device 12 is configured to be connectable to a laser light source (not shown) of the endoscope 30, and when estimating the temperature of the lesion 40, the information processing device 12 connects to the laser light source of the endoscope 30. It operates with the laser light intensity input from.

The AUC calculation unit 21 is a unit that calculates the AUC of the reflected light spectrum each time the spectroscope 11 measures the reflected light spectrum. Note that the AUC (Area Under the Curve) of the reflected light spectrum is the area below the curve of the reflected light spectrum, as shown in FIG.

The temperature estimation unit 22 estimates the temperature of the lesion 40 based on each AUC calculated by the AUC calculation unit 21 and the laser light intensity from the laser light source of the endoscope 30, and displays the estimation result on the display device 13. This is the unit to display. The temperature estimating unit 22 estimates the temperature of the lesion 40 using a neural network configured as shown in FIG. Note that Z ^-k in FIG. 5 is a discrete time operator that outputs an input delayed by k sampling periods.

That is, the input layer of the neural network used by the temperature estimator 22 to estimate the temperature of the lesion 40 includes q+1 input units 26 into which AUCs are input at different delay times, and laser light intensities are input at different delay times. q+1 input units 26 for input. The intermediate layer of the neural network also includes a plurality of hidden units 27 into which the output of each input unit 26 or the output from other hidden units 27 is input. The output layer of the neural network is composed of one output unit 28 that outputs the temperature of the lesion 40.

Each input unit 26 is a unit that outputs input (AUC or laser light intensity) to a plurality of predetermined hidden units 27.

Furthermore, as schematically shown in FIG. 6, each hidden unit 27 in this embodiment calculates the sum a of the multiplication results of each input x _i (i=1 to n) and the weight w _i , and calculates the sum a This is a unit (artificial neuron) that calculates y=1/(1+e ^−a ) and outputs the calculation result y. The output unit 28 is also a unit similar to each hidden unit 27. Note that the hidden unit 27 and the output unit 28 may take bias into consideration or may use an activation function different from 1/(1+e ^−a ).

The storage unit 24 (FIG. 1) is a unit (such as a hard disk) that stores a plurality of training data for educating the temperature estimation unit 22 (details will be described later). Here, the training data refers to time-series data of AUC (AUC for each sampling period) obtained in an experiment simulating endoscopic laser treatment, time-series data of laser light intensity during the experiment, and lesion area 40. (a portion corresponding to the lesion 40) and time-series temperature data. As the training data, a relatively large amount of data with different conditions (laser light intensity, etc.) is prepared.

The learning processing section 23 is a unit that educates the temperature estimating section 22 based on the teacher data group in the storage section 24. Note that the temperature estimation device 10 (temperature estimation unit 22) functions as a device that can accurately estimate the temperature of the lesion 40 in real time after the training of the temperature estimation unit 22 by the learning processing unit 23 is completed. be.

Specifically, the learning processing section 23 performs a process of determining each weight value (and bias value) of the neural network used by the temperature estimating section 22 for temperature estimation so that a predetermined evaluation function is minimized. This process performed by the learning processing unit 23 according to the present embodiment is a process of minimizing the evaluation function using the error backpropagation method. Furthermore, the evaluation function minimized by the learning processing unit 23 corresponds to the AUC in the teacher data #j (j is the identification number of the teacher data) and the output of the temperature estimation unit 22 at each time when the laser light intensity is input. If the sum of the squared values of the differences from the temperature in the teacher data #j at the time is expressed as C _j , it is ΣC _j (=C ₁ +C ₂ +...+C _m ; m is the total number of teacher data). However, methods other than the error backpropagation method may be used to educate the temperature estimator 22 (determining weight values, etc.), and an evaluation function different from the above evaluation function may be used.

As is clear from the above description, the temperature estimating device 10 calculates the temperature at each point in time of the sample (lesioned area 40) that can be denatured by temperature rise (heating), taking into account the AUC of the sample measured up to that point. It has a configuration that estimates by putting it in. The AUC of the sample measured up to each time point is information indicating the degree of denaturation of the sample. Therefore, according to the present temperature estimating device 10, as illustrated in FIG. 7, the temperature of a sample that can be denatured by heating during heating can be estimated accurately in real time from its optical characteristics. Note that FIG. 7 shows the relationship between the actually measured temperature of the laser beam irradiation target (“actual temperature”), the estimated temperature by the temperature estimation device 10 (“estimated temperature”), and the linear relationship with respect to the elapsed time after the start of laser beam irradiation. It is a figure which plots estimated temperature by a model (first-order lag model). Further, the error between the actual temperature and the temperature estimated by the temperature estimating device 10 in this experimental result is ±2°C.

《Modified example》
The temperature estimating device 10 according to the embodiment described above can be modified in various ways. For example, the time variation pattern of AUC changes depending on the intensity of the laser light. Therefore, the input unit 26 for inputting the laser light intensity may be removed from the temperature estimation neural network (FIG. 5). Further, if the target is only the case where the laser light intensity is constant (does not change over time), the number of input units 26 for inputting the laser light intensity of the neural network for temperature estimation may be set to one. .

The neural network for temperature estimation may be modified to input the reflected light intensity of light of a specific wavelength and the sum of the reflected light intensity of light of specific M (M≧2) wavelengths instead of AUC. Further, the temperature estimating device 10 may be modified into a device that measures the intensity of reflected light without using an optical fiber, or a device that estimates the temperature of a sample heated by a heat source other than laser light.

<Example 2>
Next, Example 2 of the present invention will be described. In the first embodiment, the learning processing unit 23 updates the weight (coupling coefficient) W so that the error between the estimated temperature output from the output unit (output layer) and the temperature of the sample (lesion part 40) becomes small. Although the temperature estimation device has been described, in the second embodiment, the characteristic values input to the input unit (input layer) are increased, and the values of the soft matrix function are output from a plurality of output units (output layer), and these values are An example will be described in which a correct label is given to , and the weight (coupling coefficient) W is updated so that the cross-entropy error becomes smaller. Note that the schematic configuration of the temperature estimating device 10 in this embodiment is equivalent to that shown in FIG. 1.

FIG. 8 is a diagram for explaining the temperature estimation device 10 in this embodiment. In this embodiment, the temperature estimating unit 22 estimates the temperature of the lesion 20 using the neural network shown in FIG. That is, in this embodiment, the lesion 40, which is a sample, is irradiated with laser light and measurement light. Then, at time k, (1) wavelength distribution obtained by dispersing reflected light with a spectrometer, (2) biological information represented by blood flow velocity obtained by inputting reflected light into a blood flow meter, (3) the amount of laser light (laser light intensity) obtained by irradiating the optical sensor with laser light; (4) the image obtained after processing the fiberscope image obtained by the CCD camera 33; layer) 26. Here, as the biological information, in addition to the blood flow rate, oxygen saturation concentration, fluorescence, drug reaction, etc. may be used. Further, here, an example has been described in which the CCD camera 33 is used as a camera for acquiring images, but it goes without saying that other types of cameras, such as a CMOS camera, may be used.

なお、このソフトマックス関数は、複数の出力ユニット（出力層）２８の値の総和が１になる関数と言える。 In the hidden unit (middle layer) ₂₇ of this embodiment, as in the first embodiment, as schematically shown in _FIG. The sum a is calculated, and from the sum a, y=1/(1+e ^−a ) is calculated and the calculation result y is output. Then, the output unit (output layer) 28 outputs a softmax function based on the following equation (1).

Note that this softmax function can be said to be a function in which the sum of the values of the plurality of output units (output layer) 28 is 1.

図８では、３０℃を学習させる場合について記載しており、３０℃に相当する出力に正
解ラベル“１”が付与されている。なお、本実施例においても、隠れユニット（中間層）２７、出力ユニット（出力層）２８は、バイアスを考慮するものであっても、１／（１＋ｅ^-a）とは異なる活性化関数を使用するものであっても良い。 Then, a correct label "1" is given to the output corresponding to the temperature to be learned, and "0" is given to the other outputs. Then, the weight (coupling coefficient) W is adjusted so that the cross-entropy error E based on the following equation (2) becomes small.

In FIG. 8, a case is described in which 30° C. is learned, and a correct answer label “1” is given to the output corresponding to 30° C. Note that in this embodiment as well, the hidden unit (middle layer) 27 and the output unit (output layer) 28 use an activation function different from 1/(1+e ^{- a} ) even if bias is taken into account. It may be something that you do.

FIG. 9 shows a flowchart of learning processing by the learning processing section 23 in this embodiment. When this process is executed, first, in step S101, a laser beam for temperature increase and measurement light for measuring wavelength distribution and blood flow are emitted from the irradiation optical fiber 36 at the center of the optical fiber 35 to the lesion area as a sample. 40 irradiation. In step S102, the reflected light of the two measurement lights is received by a plurality of surrounding optical fibers 37 for light reception. In step S103, the amount of laser light, the wavelength distribution of reflected light, biological information (in this example, tissue blood flow velocity), and a fiberscope image are acquired. In step S104, the signal acquired in step S103 is normalized and input to each input unit (input layer) 26.

In step S105, each signal and weight (coupling coefficient) W pass through the hidden unit (intermediate layer) 27. In step S106, a signal is output from the output unit (output layer) 28 as a softmax function. Here, the number of output units 28 is made to match the number of temperature values in the target temperature range. In this example, the target temperature range is 1° C. to 80° C., and the number of output units 27 is 80, which is the same number as the number of temperatures of the lesion 40 as teacher data. In step S107, a correct label is given to the output from each output unit (output layer) 28. Here, a label "1" is given to the temperature at which learning for estimation is desired (in this example, 30° C.), and a label "0" is given to the other temperatures.

In step S108, the weight (coupling coefficient) W is updated so that the cross-entropy error E shown in equation (2) becomes smaller. In step S109, the process returns to S105. Then, the processes of S105 to S108 are repeatedly executed. In step S110, the processes of S104 to S109 are repeatedly executed for each temperature in the target temperature range (1° C. to 80° C. in this example), and learning for each temperature is completed. When the learning for each temperature in the target temperature range is completed, this routine ends.

In the temperature estimation device 10 shown in Example 1, it was necessary to prepare neural networks for the number of times (q+1) going back in the past, whereas in the temperature estimation device 10 shown in this example, it was necessary to prepare neural networks for the number of times (q+1) going back in time. Regardless, it is sufficient to prepare one neural network. Furthermore, in the temperature estimating device 10 shown in Example 1, it was necessary to input past information on AUC and laser light intensity into the input unit 26; The estimation device 10 only needs to input current information such as wavelength distribution, laser light intensity, biological information, and fiberscope images.

FIG. 10 shows the actually measured temperature of the laser beam irradiation target (“actual temperature”) and the temperature estimated by the trained temperature estimation device 10 in this embodiment (“estimated temperature”) with respect to the elapsed time after the start of laser beam irradiation. FIG. As can be seen from FIG. 10, in this example, although the estimation is performed using one neural network, it is possible to estimate the temperature of the lesion 40 with the same accuracy as in Example 1. ing.

In this embodiment, a laser beam and measurement light are irradiated onto the lesion 40, which is a sample, and the fiberscope image obtained by the CCD camera 33 at time k is processed, and the image is sent to the input unit of the neural network. (Input layer) 26 was input. Regarding this image, instead of or in addition to the image itself, image information taken with an optical fiber or CCD camera 33 is used as RGB (Red, Green, Blue) or , numerical data obtained by color space decomposition such as HLS (Hue, Saturation, Lightness/Luminance),
Alternatively, changes in this numerical data may be input to the input unit (input layer) 26 of the neural network.

Furthermore, in this embodiment, an example has been described in which the signal from the output unit (output layer) 28 is output as a softmax function. However, the signal from the output unit (output layer) 28 may be output as another function, such as a sigmoid function, a linear function, a step function, a ramp function, or a hyperbolic tangent function.

Furthermore, in the above embodiment, a neural network with a multilayer structure including an input unit (input layer) 26, a hidden unit (middle layer) 27, and an output unit (output layer) 28 was used. A neural network including a convolutional neural network), an RNN (Recurrent Neural Network), reinforcement learning, etc. may be used.

In the embodiments described above, the present invention has been explained by taking as an example a medical scene in which an optical fiber is inserted into a channel of an endoscope, but the field to which the present invention is applied is not limited to this. For example, the present invention can be applied to other medical fields, such as estimating the temperature of a target to be incised when incising the target using a thermal method such as an electric scalpel, as well as to the industrial field, such as estimating the temperature of a welding target in piping. I do not care.

10 Temperature estimation device 11 Spectrometer 12 Information processing device 13 Display device 21 AUC calculation unit 22 Temperature estimation unit 23 Learning processing unit 24 Storage unit 26 Input unit 27 Unit 28 Output unit 30 Endoscope 31 Forceps port 32 Ride guide 33 CCD camera 35 Optical fiber 36 Optical fiber for irradiation 37 Optical fiber for light reception 40 Lesion area

Claims (10)

a measurement unit that periodically measures optical characteristic values regarding reflectance of the sample while the sample is being heated, which can be denatured by temperature rise;
A trained neural network including an output unit for outputting the temperature of the sample, and a plurality of input units into which the optical characteristic values measured in time series by the measurement unit are input at mutually different delay times. a temperature estimation unit that estimates the temperature of the sample at the time of the current measurement of the optical property value, each time the optical property value is measured by the measurement unit;
A temperature estimation device comprising: The optical characteristic value is the area under the curve of the reflected light spectrum of the sample,
The temperature estimating device according to claim 1, characterized in that: The measurement unit periodically measures the optical characteristic value of the sample whose temperature is being increased by irradiation with laser light,
The neural network further comprises at least one input unit for inputting a value indicating the intensity of the laser light.
The temperature estimating device according to claim 1 or 2, characterized in that: a measurement unit that irradiates light during heating of a sample that can be denatured by heating and measures characteristic values obtained from the reflected light;
Estimating the temperature of the sample using a trained neural network including an output unit for outputting the temperature of the sample and a plurality of input units into which the characteristic values measured by the measurement unit are input. A temperature estimation section;
Equipped with
When training the neural network,
The output unit outputs a value of a specific function corresponding to the temperature of the sample in a temperature range to be measured,
Among the output values of the specific function, a value corresponding to the temperature of the learning target is given a label of 1, and other values are given a label of 0,
It is characterized by performing repeated calculations to update coupling coefficients in the neural network so that a cross-entropy error based on the value of the specific function outputted by the output unit and the value of the label is reduced. Temperature estimation device. 5. The temperature estimating device according to claim 4, wherein the specific function is any one of a sigmoid function, a softmax function, a linear function, a step function, a ramp function, and a hyperbolic tangent function. The temperature estimation according to claim 4 or 5, wherein the characteristic value includes a wavelength distribution of the reflected light, biological information based on the reflected light, an intensity of the irradiated light, and an image of the sample. Device. 6. The temperature estimating device according to claim 4, wherein the characteristic value includes numerical data obtained by color space decomposition of image information of the sample. a measuring step of periodically measuring an optical characteristic value regarding reflectance of the sample while increasing the temperature of the sample that can be denatured by increasing the temperature;
a trained neural system including an output unit for outputting the temperature of the sample; and a plurality of input units into which the optical characteristic values measured in time series from the measurement step are input at mutually different delay times. a temperature estimation step of estimating the temperature of the sample at the time of the current measurement of the optical property value, using a network, each time the optical property value is measured in the measurement step;
A temperature estimation method, comprising: a measurement step of irradiating light during heating of a sample that can be denatured by heating and measuring characteristic values obtained from the reflected light;
Estimating the temperature of the sample using a trained neural network including an output unit for outputting the temperature of the sample and a plurality of input units into which the characteristic values measured in the measurement step are input. a temperature estimation step;
A temperature estimation method comprising:
When training the neural network,
The output unit outputs a value of a specific function corresponding to the temperature of the sample in a temperature range to be measured,
Among the output values of the specific function, a value corresponding to the temperature of the learning target is given a label of 1, and other values are given a label of 0,
It is characterized by performing repeated calculations for updating coupling coefficients in the neural network so that a cross-entropy error based on the value of the specific function output by the output unit and the value of the label is reduced. Temperature estimation method. 10. The temperature estimation method according to claim 9, wherein the specific function is any one of a sigmoid function, a softmax function, a linear function, a step function, a ramp function, and a hyperbolic tangent function.

Images