JP7798264B2 - Biological testing methods - Google Patents

Description

The present invention relates to a method for testing living organisms, such as plants and animals, for parasitic infection.

Generally, living organisms such as edible plants, ornamental plants, or edible animals (fish and meat) can become infected with parasites and become unusable, so they are tested for parasitic infections. For example, to explain the testing method for living organisms using nematodes as parasites, the following example of a living organism such as garlic can be given.

In recent years, garlic has become infected with the parasitic nematode, C. difficile, and various countermeasures have been implemented. In this regard, it is essential to first test harvested garlic for C. difficile infection. Conventional methods for testing for C. difficile are described, for example, in Patent Document 1 (JP 2017-219423 A). This method involves crushing C. difficile nematodes, centrifuging the resulting body fluid, and injecting it into a rabbit for immunization to obtain antiserum. Nonspecific reactants are then removed from this antibody serum to obtain anti-nematode serum. The test garlic is then crushed and contacted with the anti-nematode serum to induce an antigen-antibody reaction. The antibodies in the anti-nematode serum are then labeled. If a reaction complex is formed, it is detected. Labeling methods using enzymes, fluorescent substances, or other conventional methods are used to detect the reaction complex.

Another known method for extracting nematodes is the Berman method, in which infected garlic clove, the surface of which has been sterilized, is finely chopped and treated with chemicals on a funnel to obtain a nematode suspension (see Patent Document 1 (JP 2017-219423 A) and Patent Document 2 (JP 2010-100568 A), etc., mentioned above).

Japanese Patent Application Laid-Open No. 2017-219423 JP 2010-100568 A

However, this conventional method of testing live organisms requires that the organisms be crushed one by one and treated with anti-nematode serum or the required chemicals, which is time-consuming and makes the testing process extremely tedious. In addition, because the organisms are destroyed, it is not possible to test every single product, such as garlic, which is shipped in cloves, which is a problem.

The present invention was made in consideration of these problems, and aims to provide a living organism testing method that enables non-destructive testing of living organisms, even those infected with parasites, thereby improving testing efficiency and versatility.

In order to achieve such an object, the method for testing a living organism of the present invention is a method for testing a living organism for infection with a parasite, which comprises:
A calculation model is constructed in advance, which performs calculations relating to the presence or absence of parasitic infection in the living organism and the degree of damage, which is quantified based on the second derivative spectrum of the absorbance, by irradiating near-infrared light onto a sample living organism infected with a parasite and a sample living organism not infected with a parasite, receiving diffuse reflected light or diffuse transmitted light from the sample living organism, and measuring the absorbance of the received light, and quantifying the degree of onset due to infection, from a predetermined maximum of no onset to onset,
Near-infrared light is irradiated onto the living body of the subject, diffuse reflected light or diffuse transmitted light from the living body of the subject is received, the absorbance of the received light is measured, and the presence or absence of the infection and/or the degree of damage is determined based on the calculation results calculated from this absorbance and the above-mentioned calculation model.

Here, living organisms include plants or animals infected with parasites, and may be alive or dead. Examples include edible and ornamental plants such as vegetables and fruits, fish, and meat. Parasites include protozoa, sarcoflagellates, sporozoa, ciliates, helminths, nematodes, cestodes, trematodes, and thorny headed worms.

In addition, infection refers to the state in which a parasite has taken over the body, and onset refers to the appearance of some symptoms due to parasitic infection, such as changes in the body's tissues and the appearance of lesions in the body. The extent of damage can be determined not only on the surface, but also when there is internal disease. Because it can distinguish between the surface and the internal, it is possible to grasp even the earliest stages of disease, making it more suitable for diagnosing early stages of disease than visual examination.

This system irradiates a living organism under test with near-infrared light, receives diffusely reflected or transmitted light from the organism, measures the absorbance of the received light, and determines the presence or absence of parasitic infection and/or the extent of damage based on the results of calculations performed using this absorbance and a computational model. This allows for non-destructive testing of the living organism under test, improving testing efficiency and versatility. Furthermore, the use of a computational model ensures that even when parasites have invaded the organism or the infection is minimal, it is possible to reliably determine the presence or absence of parasitic infection in the organism and determine the extent of damage caused by the onset of symptoms, thereby improving testing accuracy. Furthermore, when determining the extent of damage, the extent of the onset of parasitic symptoms can be determined, facilitating quality assessment of the organism. Furthermore, the degree of damage is a quantified value of the onset of symptoms, allowing for a reliable understanding of the extent of the onset of symptoms. This allows for the quality of organisms to be classified, further improving versatility.

If necessary, the above damage level can also be used to rank non-infected test subjects. This makes it possible to determine whether or not a subject is infected by determining the damage level, resulting in extremely high efficiency.

In this case, it is effective to determine the degree of damage based on the degree of discoloration that appears on the subject. While this depends on the parasite, it is suitable for organisms in which lesions appear on the surface of the organism when the disease develops due to infection with the parasite.

In this configuration, the damage level can be configured as an index with 0 being no onset and 100 being the maximum with onset.
Furthermore, in this configuration, the damage level can be configured as a class in which the degree of onset is expressed as an index with no onset being 0 and the maximum onset being 100, with no onset being 0 and the onset side increasing in order for each predetermined index range. When configured as a class, it is easy to sort the quality of living organisms, and versatility can be further improved.

and, as necessary, the computational model is an analytical model constructed using statistical analysis or a trained model constructed using machine learning;
The statistical analysis is selected from discriminant analysis, multiple regression analysis, PLS regression analysis, logistic regression analysis, and support vector machine (SVM);
The machine learning is configured to be selected from tree analysis such as decision trees (CART), regression trees, random forests, and gradient boosting trees; neural networks such as perceptrons, convolutional neural networks (CNN), recurrent neural networks (RNN), and residual networks (ResNet); deep learning (DNN); and ensemble analysis consisting of a combination of these.

The computational model consists of an analytical model constructed using statistical analysis or a trained model constructed using machine learning, which can improve testing accuracy. In particular, when deep learning (DNN) is used, testing accuracy can be significantly improved. The wavelengths used in the computational model can be selected and combined from statistically significant wavelengths in the near-infrared region between 700 nm and 2500 nm. This makes it possible to accurately determine whether a living organism is infected with a parasite and the extent of the resulting disease.

If necessary, the above-mentioned computational model can use multiple computational models of different types for determining whether or not an infection is present, and if any one of the discrimination results using each computational model determines that an infection is present, that discrimination result is given priority. This makes it possible to accurately determine whether a living organism is infected with a parasite.

Alternatively, if necessary, three or more different types of computational models for determining the presence or absence of infection can be used as the computational model, and the infection determination results of each computational model can be determined by majority vote. This also makes it possible to accurately determine whether a living organism is infected with a parasite.

The parasites are nematodes that infect living plants. The living plants may be any plants that can be infected by nematodes, such as garlic, potato, sweet potato, or iris.
Examples of nematodes include potato lesion nematodes, root-knot nematodes (sweet potato root-knot nematode, Java root-knot nematode, arenaria root-knot nematode, northern root-knot nematode), root lesion nematodes (southern root lesion nematode, channel root lesion nematode, northern root lesion nematode, walnut root lesion nematode), peel nematodes (leaf nematode, peel nematode), stalk nematode, bursa nematode, cyst nematode, spiral nematode, spine nematode, pin nematode, ring nematode, and soybean nematode.

In this case, the living organism is selected from garlic, potato, sweet potato, and iris, and the parasite is a potato lesion nematode.

In this case, the test subject is made up of garlic bulbs, which are collections of garlic cloves. Since each garlic bulb is inspected, the inspection is efficient. Also, since garlic is mainly shipped in the form of garlic bulbs, the inspection is also efficient in this respect.

If the test subject is a garlic bulb with clusters of clove, the damage level is determined as necessary based on the degree of discoloration that appears on the test subject, and is calculated as follows using an index where 0 indicates no symptoms and 100 indicates symptoms.

Damage index = ((number of scales in t = 1) + (number of scales in t = 2 x 2) + (number of scales in t = 3 x 3) + (number of scales in t = 4 x 4)) / (total number of scales surveyed x 4) x 100
where:
t is a number corresponding to the size of the discolored area caused by parasitic infection per garlic bulb, and is set so that the larger the value, the larger the size of the discolored area, and is set to four categories: t=1, t=2, t=3, and t=4.

Furthermore, if the test subject is a garlic bulb with clusters of clove, the degree of damage is determined as necessary based on the degree of discoloration that appears on the test subject, and the degree of onset is expressed as an index with no onset being 0 and the maximum onset being 100. The degree of onset is expressed as follows, with no onset being 0 and the onset side being ranked in ascending order for each specified index range.

Severity of onset (index) = ((number of scales at t = 1) + (number of scales at t = 2 x 2) + (number of scales at t = 3 x 3) + (number of scales at t = 4 x 4)) / (total number of scales surveyed x 4) x 100
where:
t is a number corresponding to the size of the discolored area caused by parasitic infection per garlic bulb, and is set so that the larger the value, the larger the size of the discolored area, and is set to four categories: t=1, t=2, t=3, and t=4.

According to the present invention, near-infrared light is irradiated onto a living organism under test, diffusely reflected or transmitted light from the living organism is received, the absorbance of the received light is measured, and the results of calculations based on this absorbance and a computational model are used to determine the presence or absence of parasitic infection and/or the extent of damage in the living organism. This allows for non-destructive testing of the living organism under test, improving testing efficiency and versatility. Furthermore, the use of a computational model makes it possible to reliably determine the presence or absence of parasitic infection in a living organism and determine the extent of damage caused by the onset of disease, even when parasites have invaded the organism or the infection is slight, thereby improving testing accuracy. Furthermore, when determining the extent of damage, the extent of the onset of parasitic disease can be determined, facilitating quality assessment of the living organism. Furthermore, the degree of damage is a value that quantifies the extent of onset, allowing for a reliable understanding of the extent of onset. This allows for the quality of living organisms to be classified, further improving versatility.

1 is a perspective view showing an examination device for implementing a living body examination method according to an embodiment of the present invention; 1 is a diagram showing garlic as a living organism to be tested by a testing device that implements a method for testing a living organism according to an embodiment of the present invention, together with a holder. 1 is a front view showing an examination device for implementing a living body examination method according to an embodiment of the present invention; 1 is a plan view showing an examination device for implementing a living body examination method according to an embodiment of the present invention; 1 is a side view showing an examination device for implementing a living body examination method according to an embodiment of the present invention; 1 is a diagram showing the relationship between a gripping hand of a robot and an inspection unit, together with the operation process of the gripping hand, in an inspection device that realizes a living body inspection method according to an embodiment of the present invention. FIG. 1A and 1B show an inspection unit in an inspection device that realizes a living body inspection method according to an embodiment of the present invention, where FIG. 1A is a perspective view of the inspection unit and FIG. 1B is a view showing an inspection section of the inspection unit. 1 is a cross-sectional view showing a state in which a testing unit is testing garlic as a living organism in a testing device that realizes a living organism testing method according to an embodiment of the present invention. 1 is a block diagram showing the configuration of a control unit in an examination device that realizes a living body examination method according to an embodiment of the present invention. FIG. 10 is a diagram schematically illustrating a trained model used by a control unit in an examination device that realizes a biological examination method according to an embodiment of the present invention. 4 is a flowchart showing a control flow in a control unit of an examination device that realizes a living body examination method according to an embodiment of the present invention. 4 is a flowchart showing a control flow of a sample measurement and discrimination routine in a control unit of a testing device that realizes a living body testing method according to an embodiment of the present invention. 10 is a flowchart showing a control flow of a parasite discrimination routine, which is a control flow of a control unit of an inspection device that realizes a living body inspection method according to an embodiment of the present invention. FIG. 1A is a graph showing the correlation between actual measured values and estimated values, and FIG. 1B is a table showing the underlying numerical values, relating to the calculation model (multiple regression analysis (determining the presence or absence of infection)) of Example 1 of the present invention. 1A and 1B show a verification example of the calculation model (multiple regression analysis (determining whether or not there is infection)) of Example 1 of the present invention, in which (a) is a graph showing the correlation between actual measured values and estimated values, and (b) is a table showing the underlying numerical values. FIG. 10A is a graph showing the correlation between actual measured values and estimated values, and FIG. 10B is a graph showing the verification results of the calculation model (multiple regression analysis (damage level (index) determination)) of Example 2 of the present invention. 10A is a table showing the correlation between actual measured values and estimated values, and FIG. 10B is a table showing the verification results of the calculation model (logistic regression model (infection determination)) of Example 3 of the present invention. FIG. 10A is a table showing the correlation between actual measured values and estimated values, and FIG. 10B is a table showing the verification results of the calculation model (logistic regression model (damage level (6 levels) discrimination)) of Example 4 of the present invention. FIG. 10A is a table showing the correlation between actual measured values and estimated values, and FIG. 10B is a graph showing the distribution of estimated probabilities for the computational model (AI-trained model A (determining whether or not there is infection)) of Example 5 of the present invention. This is a table showing the verification results of the computational model (AI trained model A (infection determination)) of Example 5 of the present invention. FIG. 10A is a table showing the correlation between actual measured values and estimated values, and FIG. 10B is a table showing the verification results of the calculation model (AI-trained model B (damage level (6 levels) discrimination)) of Example 6 of the present invention. FIG. 10A is a graph showing the correlation between actual measured values and estimated values, and FIG. 10B is a graph showing the verification results of the calculation model (AI-trained model C (damage level (index) determination)) of Example 7 of the present invention. FIG. 10A is a table showing the correlation between actual measured values and estimated values, and FIG. 10B is a table showing the verification results of the calculation model (AI-trained model D (infection determination)) of Example 8 of the present invention. 9A is a table showing the correlation between actual measured values and estimated values, and FIG. 9B is a table showing the verification results of the calculation model (AI-trained model E (damage level (6 levels) discrimination)) of Example 9 of the present invention. FIG. 10A is a graph showing the correlation between actual measured values and estimated values, and FIG. 10B is a graph showing the verification results of the calculation model (AI-trained model F (damage level (index) determination)) of Example 10 of the present invention. FIG. 11 is a table showing the correlation between actual measured values and estimated values for the computational model (AI-trained model G (infection determination)) of Example 11 of the present invention; and FIG. 12 is a table showing the verification results. 12A is a table showing the correlation between actual measured values and estimated values, and FIG. 12B is a table showing the verification results of the calculation model (AI-trained model H (damage level (6 levels) discrimination)) of Example 12 of the present invention. FIG. 13A is a graph showing the correlation between actual measured values and estimated values, and FIG. 13B is a graph showing the verification results of the calculation model (AI-trained model I (damage level (index) determination)) of Example 13 of the present invention.

A living body testing method according to an embodiment of the present invention will be described below with reference to the accompanying drawings. This living body testing method is implemented using a living body testing device, and will therefore be explained in the description of this living body testing device.

As shown in Figures 1 to 9, a living organism testing device S according to an embodiment of the present invention tests for parasitic infection in a living organism, which is a test subject W. As shown in Figure 2, the living organism is harvested garlic, and the parasite is a potato lesion nematode. The test subject W is composed of a garlic bulb with a collection of scales Wa. The garlic bulb (test subject W) has a flower diameter Wb protruding several centimeters upward, a disk stalk Wc below, is covered with an exoskeleton, and is washed and dried. Infection by the parasite begins on the disk stalk Wc side. F indicates the parasitic lesion (shown by a dotted line). Generally, the lesion F is hidden by the exoskeleton and is difficult to see. In severe cases, the lesion extends to the exoskeleton and becomes prominent on the surface.

The biological testing device S according to the embodiment comprises a base 1, a transport unit 2 mounted on the base 1 and transporting test subjects W one by one at predetermined intervals, a robot 10 installed near the base 1, a gripping hand 12 at the tip of the robot arm 11 of the robot 10 that grips the test subject W during transport by the transport unit 2 and positions the test subject W at the testing position Y, an inspection unit 30 mounted at the tip of the robot arm 11 of the robot 10 near the gripping hand 12 and inspects the test subject W positioned at the testing position Y by the gripping hand 12, a receiving unit 50 that receives the test subject W moved by the robot 10 and gripped by the gripping hand 12 after inspection by the inspection unit 30, and a control unit 60 that performs the inspection function and controls the transport unit 2, robot 10, and testing unit 30. Reference numeral 70 in the figure denotes a display unit such as a CRT.

As shown in Figure 6, in the robot 10, the gripping hand 12 is attached to the tip of the robot arm 11 via an opening/closing actuator 13 that opens and closes the gripping hand 12. The opening/closing actuator 13 is swingable, and the swinging of the opening/closing actuator 13 moves the gripping hand 12 to two positions: a gripping position X where the object W on the transport unit 2 can be gripped, and an inspection position Y where inspection is performed in the inspection unit 30.

The conveying unit 2 comprises an endless belt conveyor 3 driven by a motor (not shown), and multiple holding units 20 arranged at predetermined intervals along the belt conveyor 3's movement to hold test subjects W. As shown in FIG. 2, each holding unit 20 supports and holds the test subjects W, i.e., garlic bulbs, with their stems Wc facing up. Each holding unit 20 comprises a cup-shaped holding body 23 with a cylindrical side wall 21 and top wall 22, the underside of which is fixed to the surface of the belt conveyor 3; an insertion hole 24 formed in the top wall 22 of the holding body 23, centered on the axis of the holding body 23, through which the garlic bulb with its flower diameter Wb passes; and a conical support recess 25 extending over the entire insertion hole 24 or above (the entire insertion hole 24 in this embodiment) to support the periphery of the garlic bulb with its flower diameter Wb. This positions the stems Wc side of the garlic bulb of the test subject W grasped by the grasping hand 12 facing the inspection unit 30.

The base 1 is provided with a position detection sensor (not shown) that detects when the test object W, transported by the belt conveyor 3 of the transport unit 2, is positioned at a predetermined position for grasping by the gripping hand 12 of the robot 10. Based on the detection by this position detection sensor, the test object W is grasped and removed by the gripping hand 12 of the robot 10. A camera 26 is provided above the belt conveyor 3 of the transport unit 2 to capture an image of the test object W being transported. The size of the test object W can be determined from the image captured by the camera 26.

As shown in Figures 6 to 8, the inspection section 30 is composed of a cylindrical inspection unit 31, a support rod 32 that supports the inspection unit 31 at the tip of the robot arm 11, near the gripping hand 12, and a cover 33 that is provided at the tip of the inspection unit 31 and covers the object W to be inspected that is positioned at the inspection position Y. Because the inspection section 30 and gripping hand 12 are provided on the robot arm 11, the object W to be inspected can be transported and inspected, and subsequent sorting can be performed extremely efficiently.

As shown in Figures 7 and 8, the inspection unit 31 comprises a cylindrical unit body 34, a light source unit 35 composed of the tips of multiple optical fibers that irradiates light in the near-infrared region onto the test object W, and a detection unit 37 equipped with a light receiving unit 36 composed of the tips of optical fibers that receives diffusely reflected or transmitted light from the test object W. The optical fibers are collected and held in a solid cylindrical holder 38, with the end face of each optical fiber exposed at one end face of the holder 38 facing the test object W, and this end face serves as the detection unit 37. The holder 38 is held in the solid cylindrical unit body 34 coaxially with the axis of the holder 38. The detection unit 37, which is one end face of the holder 38, is positioned rearward of the end face of the unit body 34 facing the test object W, and a recess 39 is formed in front of the detection unit 37.

The inspection unit 31 also includes a dustproof mechanism 40 that protects the detection unit 37 from dust. The dustproof mechanism 40 includes a protruding portion 41 that protrudes from a portion of the outer periphery of the end face of the unit body 34. This protruding portion 41 is configured with an air injection port 42 that injects air along one end face of the unit body 34 to form an air curtain that covers the recess 39. The unit body 34 is formed with an air passage 43 that is parallel to its axis and directs air to the air injection port 42. 44 is an air supply pipe connected to the air passage 43 at the other end of the unit body 34. The light source portion 35 and light receiving portion 36 of the detection unit 37 can be kept clean at all times, ensuring reliable inspections and making maintenance easier.

The receiving unit 50 is equipped with six chutes 51 that receive the test objects W. The chutes 51 are provided to correspond to the determined damage level (0-5), as described below, and the robot 10, upon receiving a command from the control unit 60, places the test objects W held by the gripping hands 12 into the corresponding chute 51 after inspection. In other words, this device is equipped with a sorting means 52 that sorts the test objects W according to the presence or absence of parasitic infection, and in this embodiment, according to the damage level. The sorting means 52 is composed of the robot 10 and the receiving unit 50, which are controlled as required by the control unit 60, and sorting is performed using these functions.

As shown in Figure 9, the control unit 60 is equipped with a calculation processing unit 61 that performs overall control. The calculation processing unit 61 is equipped with a calculation model storage unit 62 that stores a calculation model that is constructed based on the second derivative spectrum of absorbance at wavelengths in the near-infrared region of diffuse reflected light or diffuse transmitted light obtained from a sample organism infected with a parasite and a sample organism not infected with a parasite by irradiating the sample organism with light in the near-infrared region. The calculation model performs calculations related to the presence or absence of parasite infection in the organism and/or the degree of damage, which is quantified as a predetermined level of onset due to infection, ranging from no onset to a maximum of onset.

The calculation processing unit 61 also has a discrimination unit 63 that determines the presence or absence of infection and/or the degree of damage based on the calculation results calculated from the absorbance of light received by the light receiving unit 36 and the calculation model stored in the calculation model storage unit 62. That is, the discrimination unit 63 has an infection discrimination function that determines whether or not a living organism is infected with a parasite based on the calculation results calculated from the absorbance of light received by the light receiving unit 36 and the calculation model. The discrimination unit 63 also has a damage degree discrimination function that determines the degree of damage, which is a numerical value representing the degree of onset, ranging from a predetermined maximum of no onset to onset, due to infection of a living organism with a parasite, based on the calculation results calculated from the absorbance of light received by the light receiving unit 36 and the calculation model.

In this embodiment, the damage level is determined based on the degree of discoloration that appears on the subject W, as shown in Figure 2. Subjects W that are not infected can also be ranked in terms of damage level. In this embodiment, no onset includes no infection. If it is necessary to closely examine the damage level and the presence or absence of infection, each trained model can be used. Two types of damage level are used. One type of damage level is an index where no onset is set to 0 and the maximum value of onset is 100. In more detail, this is done using the damage level below.

Damage index = ((number of scales in t = 1) + (number of scales in t = 2 x 2) + (number of scales in t = 3 x 3) + (number of scales in t = 4 x 4)) / (total number of scales surveyed x 4) x 100
where:
t is a number corresponding to the size of the discolored area caused by parasitic infection per garlic bulb, and is set so that the larger the value, the larger the size of the discolored area, and is set to four categories: t=1, t=2, t=3, and t=4.

The other damage level is expressed as an index where the degree of onset is 0 for no onset and 100 for the maximum of onset, with no onset being 0 and onset being 100 being the maximum, and is composed of ranks that increase in order for each specified index range. In this embodiment, the damage level is set to the following six ranks.

Damage level 0 = severity of onset 0
Severity of onset 0 < Damage level 1 ≦ Severity of onset 20
Severity of onset 20 < Damage level 2 ≦ Severity of onset 40
Severity of onset 40 < Damage level 3 ≦ Severity of onset 60
Severity of onset 60 < Damage level 4 ≦ Severity of onset 80
Severity of onset 80 < Damage level 5 ≦ Severity of onset 100

The severity of the disease is determined as follows:
Severity of onset (index) = ((number of scales at t = 1) + (number of scales at t = 2 x 2) + (number of scales at t = 3 x 3) + (number of scales at t = 4 x 4)) / (total number of scales surveyed x 4) x 100
where:
t is a number corresponding to the size of the discolored area caused by parasitic infection per garlic bulb, and is set so that the larger the value, the larger the size of the discolored area, and is set to four categories: t=1, t=2, t=3, and t=4.

The computational model is an analytical model constructed using statistical analysis or a trained model constructed using machine learning,
The statistical analysis is selected from discriminant analysis, multiple regression analysis, PLS regression analysis, logistic regression analysis, and support vector machine (SVM);
The machine learning is selected from tree analysis such as decision trees (CART), regression trees, random forests, and gradient boosted trees, neural networks such as perceptrons, convolutional neural networks (CNN), recurrent neural networks (RNN), and residual networks (ResNet), or deep learning (DNN), and ensemble analysis consisting of a combination of these.

In more detail, in multiple regression analysis, PLS regression analysis, and logistic regression analysis, near-infrared light is irradiated onto a sample living organism infected with a parasite and a sample living organism not infected with a parasite, the diffuse reflected light or diffuse transmitted light from the sample living organism is received, the absorbance of the received light is measured, and an analytical model (computational model) relating to the wavelengths attributed to the parasite is identified through statistical analysis of the second-order derivative spectrum of the absorbance. Near-infrared light is then irradiated onto the object W to be examined, the diffuse reflected light or diffuse transmitted light from the living organism of the subject W is received, and the absorbance of the received light is measured. From the results calculated from these absorbances and the analytical model, the presence or absence of parasite infection in the living organism and/or the degree of damage, quantified as a predetermined level of onset due to infection ranging from no onset to a maximum of onset, is determined.

<Multiple regression analysis>
The analytical model is determined by multiple regression analysis, where the absorbance spectrum of the light received from the sample is subjected to MSC processing and then second-order differentiation. MSC (Multiplicative Scatter Correction) is a preprocessing method for removing multiplicative factors (such as light scattering) and additive factors that occur in the measured spectrum (hereinafter the same).

The above analytical model was identified using an equation (discriminant) that satisfies the relationship of the following general formula (B), which uses the absorbances of wavelengths 1 to n that have high correlation coefficients as variables.

In general formula (B), when selecting the first wavelength (λ1) to the nth wavelength (λn), the absorbance of the sample organism is first measured in advance, and then whether or not the organism is infected with a parasite is determined by another method, such as a culture method using the Bellman method. Using the absorbance of the sample organism infected with the parasite and the absorbance of the sample organism not infected with the parasite as an indicator, the near-infrared wavelength range of the first wavelength (λ1) with the highest correlation coefficient attributable to the parasite is selected by multiple regression analysis of the near-infrared wavelength range of the first wavelength (λ1) with the wavelength range of 700 nm to 2500 nm. A wavelength range of a second wavelength (λ2) is selected that has a correlation coefficient equal to or greater than the correlation coefficient of the near-infrared wavelength range of the wavelength (λ1). Then, a wavelength range of a third wavelength (λ3) is selected that is attributed to the parasite and has a correlation coefficient equal to or greater than the correlation coefficient of the near-infrared wavelength range of the first wavelength (λ1) and the second wavelength (λ2) through a multiple regression analysis of the near-infrared wavelength ranges of the first wavelength (λ1) to the second wavelength (λ2) and a wavelength range of 700 nm to 2500 nm. In this way, a wavelength range of an nth wavelength (λn) is selected that is attributed to the parasite and has a correlation coefficient equal to or greater than the correlation coefficient of the near-infrared wavelength range of the first wavelength (λ1).

<PLS regression analysis>
The analytical model was determined by calculating and determining the absorbance spectrum of the light received from the sample object using a partial least square projection to latent structure (PLS) analysis method, by subjecting the absorbance spectrum to MSC processing and then second-order differentiation.

<Logistic regression analysis method>
The trained model was identified by calculating and using logistic regression analysis after second-order differentiation of the absorbance spectrum of the light received from the sample object after MSC processing.

<Support Vector Machine (SVM)>
The trained model was identified by the SVM (Support Vector Machine) method by subjecting the absorbance spectrum of the light received from the sample object to MSC processing and then second-order differentiation.

Next, we will give an example of deep learning methods.
<Deep Learning Method (1)>
For example, the trained model is a trained model obtained by a so-called ensemble method using a decision tree of a deep learning method and a random forest method, after MSC processing of the absorbance spectrum of light received from the sample living body. The ensemble method is used to obtain a statistically superior trained model. The combination is not limited to this, and may be selected and combined as appropriate. These are types A to C of AI trained models described below.

<Deep Learning Method (2)>
As the trained model, a trained model using a deep learning method, DNN (Deep Neural Network), is used, in which the absorbance spectrum of light received from the sample living body is subjected to MSC processing and then second-order differentiation. These are types D to F of the AI trained models described below.

<Deep Learning Method (3)>
As the trained model, the absorbance spectrum of the light received from the sample living body is subjected to MSC processing, and the resulting waveform is imaged and trained using a deep learning CNN (Convolutional Neural Network) method. These are the AI trained models G to I described below.

In an embodiment, the trained model for the presence or absence of infection is a trained model for causing a computer to function to output a calculation result relating to the presence or absence of infection of a living organism (garlic) with a parasite (potato-lesion nematode),
The absorbance of a sample organism is measured in advance, and the presence or absence of parasites in the organism is then confirmed by another means such as a culture method. Using the confirmed presence or absence of parasites as an indicator, the absorbance data of the sample organism not infected with parasites and the absorbance data of the sample organism infected with parasites are matched to the presence or absence of parasites, respectively, and this matched data is used as training data.The computer is trained using this training data, and near-infrared light is irradiated onto the organism to be tested, diffuse reflected light or diffuse transmitted light from the organism to be tested is received, the absorbance of the received light is measured, and a value quantifying whether the organism is infected with a parasite is output from the absorbance data.

The trained model for the degree of damage related to onset is a trained model for causing a computer to function so as to output a calculation result relating to the degree of damage, which is a numerical value representing the degree of onset from a predetermined maximum of no onset to onset due to infection of a parasite (potato lesion nematode) of a living organism (garlic),
The absorbance of a sample organism is measured in advance, and the degree of damage caused by parasitic infection in the organism is then determined by another means, such as visual inspection. Using the determined degree of damage as an index, the absorbance data of a non-infected sample organism and the absorbance data of an infected sample organism are each associated with the degree of damage to the organism. This associated data is used as training data, and the computer is trained using this training data. The computer is then caused to irradiate the subject organism with near-infrared light, receive diffusely reflected or transmitted light from the subject organism, measure the absorbance of the received light, and output a value quantifying the degree of damage caused by parasitic infection in the organism from the absorbance data. Figure 10 shows an example configuration of the trained model. In this embodiment, for example, 550 absorbance data sets at wavelength intervals of 2 nm are input to the input layer.

In the embodiment, multiple different types of computational models are provided, which can be selectively used. Figures 9 and 13 show cases where a multiple regression analysis model, a logistic regression analysis model, a support vector machine (SVM), AI-trained models types A to F, and other AI-trained models types G to I are used. Specific details of these are provided in the examples below.

The calculation processing unit 61 is equipped with multiple calculation models of different types and a judgment unit 64 that judges whether to accept or reject the judgment results of each calculation model. When multiple calculation models of different types for determining the presence or absence of infection are selected, the judgment unit 64 can be configured to prioritize and adopt any one of the judgment results of the calculation models that determines the presence of infection. The judgment unit 64 can also be configured to use three or more calculation models of different types for determining the presence or absence of infection, and to decide and adopt the infection judgment results of each calculation model by majority vote. When determining the level of damage, any one of the calculation models related to the level of damage can be selected and used. In this case, calculations can be performed using all calculation models, and the results can be selectively used. The judgment unit 64 adopts the calculation result of the selected calculation model. The calculation results of all calculation models can be displayed on the display unit 70. In this embodiment, six levels of damage level are output as the final result.

As shown in FIG. 9, the control unit 60 includes an arithmetic processing unit 61 with an operation control unit 65. This unit has a processing history storage function, a result display function on the display unit 70, a robot 10 operation function, an internet communication function, and a light intensity control function. The control unit 60 also includes a communication unit 66 controlled by the internet communication function, an inspection control unit 67 that has a spectrometer and controls the inspection unit 30, such as controlling its spectroscopic control circuit, a drive control unit 68 that includes a transport unit control circuit that controls the transport unit 2 and a robot control circuit that controls the robot 10, and an operation command unit 69 that includes a keyboard (not shown) and a setting panel (not shown).

The processing history storage function in the operation control unit 65 accumulates and stores the cumulative number of measurements of the test object W, the discrimination results, and the cumulative results in chronological order. The internet communication function is used to understand the discrimination situation and system status when used in a remote location, and to perform maintenance such as application (discrimination method) updates remotely. By using this internet communication function, the test object W can be inspected regardless of location, thereby improving the versatility of the device. The light intensity control function adjusts the light intensity of the light source and the sensitivity of the sensor in the spectroscopic unit so that the light detection sensor in the spectroscopic unit can perform measurements at an appropriate level.

Various data are displayed on the display unit 70. The display on the display unit 70 is operated by the screen operation unit (operation command unit 69), and can be switched appropriately to display an input setting screen, a discrimination result display screen, an error display screen, etc. The discrimination results may also be displayed on an LCD panel. The discrimination results may also be output as audio. Furthermore, an external data output interface may be provided.

In this embodiment, when the discrimination unit 63 determines the degree of damage, the calculation processing unit 61 uses the classification means 52 to classify the test objects W according to the degree of damage. That is, when the discrimination unit 63 determines the degree of damage, the calculation processing unit 61 operates the robot 10 via the robot operation function of the operation control unit 65 and the robot control circuit of the drive control unit 68 to classify the test objects W according to the determined degree of damage, and causes the test objects W grasped by the grasping hand 12 to be thrown into the corresponding chute 51. The control unit 60 also performs various other control functions.

Therefore, when the living body testing apparatus S according to this embodiment is used to test the subject W, the procedure is as follows: This will be explained using the flow charts shown in FIGS.
As shown in Figure 11, when the main power supply switch (not shown) is turned on, the control unit 60 starts up and checks each mechanism, such as the spectroscope, sensors, and lamps (1-1). At this time, the display unit 70 displays a screen indicating that the system is initializing, and the indicator light turns yellow (1-2). After each mechanism has been checked, a reference waveform, which serves as the basis for calculating absorbance, is measured (1-3) and stored in the memory inside the control unit 60 (1-4). The main screen then appears, and the indicator light turns green (1-8). If a setting value is recalled from the setting panel (not shown) during this time (1-5), a confirmation screen for the accumulated lamp lighting time, date, and time, etc., is displayed (1-6). To perform an inspection, operate the setting panel to return to the main screen (1-7).

A garlic bulb serving as the test object W is placed on the holder 20 with its stem Wc facing up. When the indicator light on the main screen is green, pressing the start switch (1-9) on the control button executes a measurement and discrimination routine for the test object W (1-10). In this measurement and discrimination routine for the test object W, as shown in FIG. 12, the control unit 60 first issues an instruction to operate the belt conveyor 3 of the transport unit 2 (2-1), transporting the test object W. During this transport process, the position detection sensor detects that the test object W has reached the predetermined position for gripping by the gripping hand 12 of the robot 10 (2-2). An image is then captured by the camera (2-3). The test object W is then gripped and removed by the gripping hand 12 of the robot 10 at the predetermined position. The test object W is then positioned at the inspection position Y, and the inspection unit 30 measures the test object W. In the inspection unit 30, light is irradiated onto the subject W from the light source unit 35, and the diffusely reflected or diffusely transmitted light is received by the light receiving unit 36 and transmitted via optical fiber to the spectroscopic unit of the inspection control unit 67, where the intensity of the near-infrared wavelengths is converted into voltage for each wavelength range and extracted and transmitted to the calculation processing unit 61 (2-4). The intensity spectrum of the diffusely reflected or diffusely transmitted light is then obtained and used to perform a parasite discrimination using a separate routine (2-5).

As shown in Figure 13, the discrimination routine involves noise reduction (3-1, 3-2) of the diffused light intensity (diffuse light spectrum) transmitted from the spectroscopic unit, reading in the reference waveform stored during initialization (3-3), and calculating the near-infrared absorbance of the measurement object W (3-4). This absorbance is then standardized (3-5) and subjected to second-order differentiation (3-6), and the second-order derivative spectrum is then centered and scaled (3-7). This spectrum is then used to discriminate using a computation model pre-stored in the computation model storage unit 61. Discrimination is performed using multiple regression analysis (3-8), logistic regression analysis (3-9), AI-trained model (1) (2) type A to F calculation (3-10), AI-trained model (3) type G to I calculation (3-11), and support vector machine (SVM) calculation (3-12).

When multiple different types of computation models for determining the presence or absence of infection are selected as the computation model, if the determination unit 64 determines that one of the computation models indicates the presence of infection, it prioritizes and adopts that determination result (3-13) and calculates the final determination result (3-14). Alternatively, when three or more different types of computation models for determining the presence or absence of infection are used as the computation model, the determination unit 64 decides and adopts the infection determination results of each computation model by majority vote (3-13) and calculates the final determination result (3-14). In this embodiment, a computation model for determining the degree of damage is used as the computation model, and the determination unit 64 adopts the computation result of the selected computation model (3-13) and calculates the final determination result (3-14). In this embodiment, six levels of damage are output as the final result. All determination results can be viewed on the display unit 70.

Once the measurement of the test object W is complete, the process returns to Figure 12. Based on the damage level results, the robot 10 is operated and the test object W is placed into the corresponding chute 51 of the receiving unit 50 for sorting (2-6). The process also returns to Figure 11, where the discrimination results and the cumulative discrimination results up to that point are displayed (1-13). If an error occurs in any mechanism or the emergency stop button is pressed during this time (1-11), an error screen is displayed and the indicator light turns red (1-12). To recover, check the lamp, spectrometer, sensor, and other mechanisms. The discrimination results and the diffuse light spectrum during measurement of the test object W are stored in the memory function within the control unit 60. This process is repeated until the stop button is pressed (1-15).

Next, an example of a calculation model will be shown, along with the results of verification performed by actually inspecting and verifying the calculation model.
Example 1: Multiple regression analysis (determining whether or not an infection is present)
The garlic bulbs used as test subject W were produced in an experimental field in Rokunohe Town, Aomori Prefecture, and 12,480 garlic bulbs prepared for shipping were used.
14, the data for creating the multiple regression analysis model was obtained by subjecting all test subjects W to near-infrared spectroscopy, culturing the test subjects W with potato-lesion nematodes using the Bellman method, and then examining the test subjects W to confirm whether they were infected, and correlating the presence or absence of infection with the absorbance data of the near-infrared spectral wavelengths of the samples. Using this creation data, the wavelengths necessary to determine the presence or absence of infection were searched for in order of highest correlation, and a multiple regression analysis model for the presence or absence of infection was created.

Specific wavelengths required for the multiple regression analysis model to determine whether or not an infection is present were 191 wavelengths with a logarithmic value of 0.069 or greater. As shown in Figure 14, the correlation coefficient for the multiple regression model to determine whether or not an infection is present is 0.6898. While a correlation is observed, it cannot be said to be highly correlated. However, when the prediction value of the multiple regression analysis model was set to 0.5 or greater as the discriminant value for "infection," the probability of "not infected" was 99.6%, and even the probability of "infection" was 99.6%. This multiple regression analysis model can determine whether or not garlic is infected with potato-lesion nematodes, depending on the discriminant value set.

The computational model according to Example 1 (multiple regression model for infection) was used to verify the discrimination rate and accuracy. 200 garlic bulbs, produced in an experimental field in Rokunohe, Aomori Prefecture, and prepared for shipping, were used as test subjects W. The results are shown in Figure 15. The results show that the multiple regression model for infection had a multiple correlation coefficient of 0.6836 (Figure 15(a)). When the estimated discrimination value was set to 0.5 or higher, the accuracy of this multiple regression model for discriminating between the presence and absence of infection was 0.9422, with a precision of 97.1% for "no infection" and 90.7% for "infection" (Figure 15(b)).

Example 2: Multiple regression analysis (damage level (index) discrimination)
The garlic bulbs used for the test W were 12,480 garlic bulbs prepared for shipping from the same production area as above. After near-infrared spectroscopy testing of the garlic bulbs for test W, the test garlic bulbs were cultured for potato leaf nematodes using the Bellman method, and the damage level was calculated by measuring the extent of lesions on the garlic using the damage level (index) described above. A multiple regression model for the damage level (index) was created using 196 wavelengths (not shown) selected with a logarithmic value of 0.018 or higher. A multiple regression equation using these wavelengths was used as a damage level (index) estimation multiple regression model. As shown in Figure 16(a), the multiple regression model yielded a correlation coefficient of 0.7815, indicating that the damage level (index) could be estimated.

The discrimination rate and accuracy were verified using the calculation model (multiple regression model of damage level (index)) according to Example 2. 200 garlic bulbs prepared for shipping from the same production area as above were used. The results are shown in Figure 16(b). From these results, the correlation coefficient of the multiple regression model was 0.7965. The results showed that it was possible to estimate the damage level (index).

Example 3: Logistic regression model (determining whether or not someone is infected)
The garlic bulbs used for the test subject W were 12,480 garlic bulbs from the same production area as above, prepared for shipping. After near-infrared spectroscopy of the garlic bulbs from the test subject W, they were cultured for C. elegans using the Bellman method to confirm the presence or absence of infection. Using this test data, wavelengths necessary for determining the presence or absence of infection were searched for in order of highest correlation, and a logistic regression model for infection was created. Specifically, 164 wavelengths (not shown) with a logarithmic value of 0.039 or greater were used for the determination using the logistic regression model for infection. As shown in Figure 17(a), the discrimination accuracy of the logistic regression model for C. elegans infection was 95.4% for "no infection" and 96.0% for "infection." In other words, this is an extremely accurate model.

The discrimination rate and accuracy were verified using the computational model according to Example 3 (a logistic regression model for determining whether or not an infection is present). 200 garlic bulbs prepared for shipping and from the same production area as above were used. The results are shown in Figure 17(b). From these results, the accuracy of the logistic regression model for determining whether or not an infection is present was 97.7% for "no infection" and 92.9% for "infection," with an average precision of 0.9674.

Example 4: Logistic regression model (damage level (6 levels) discrimination)
For the garlic bulbs of test subject W, 3,120 garlic bulbs prepared for shipping from the same production area as above were used as the data for creation. Near-infrared spectroscopy testing was performed on all test subjects W, and then the test garlic bulbs were cultured for potato leaf nematodes using the Bellman method, and the level of damage was calculated based on the extent of garlic lesions using the above-mentioned damage level (6 levels). The logistic regression analysis model for the damage level (6 levels) was created by analyzing near-infrared spectroscopic absorbance data corresponding to the measured value data of the above-mentioned damage level (6 levels) using logistic regression analysis as training data, and creating a regression equation.

As shown in Figure 18(a), the logistic regression analysis model performed a statistical analysis of the measured and estimated damage levels (six levels), and showed an accuracy of 0.7583, an average precision of 0.7149, an average recall of 0.4044, and an average F-score of 0.4295. Although the recall tends to be low, the logistic regression analysis model is capable of classifying the damage level (six levels) caused by garlic potato-lesion nematodes.

The discrimination rate and accuracy were verified using the computational model according to Example 4 (a logistic regression model for discriminating between damage levels (six levels)). 870 garlic bulbs prepared for shipping from the same production area as above were used to verify the discrimination rate and accuracy of the logistic regression model for determining whether or not the damage level (six levels) is infected. The results are shown in Figure 18(b). These results show that the classification accuracy of the ordered logistic regression model for damage levels (six levels) varies depending on the degree of damage, ranging from 0.000 to 0.8709, but this is possible by changing the classification class.

<Example 5: AI trained model A (infection determination)>
For the garlic bulbs of the test subjects W, 15,075 garlic bulbs prepared for shipping from the same production area as above were used as supervised training data. This training data was prepared by culturing the test garlic bulbs for potato leaf nematodes using the Bellman method after near-infrared spectroscopy testing for all test subjects W, and then checking for the presence or absence of infection, and correlating this infection with the absorbance data of the sample at the near-infrared spectroscopic wavelength. The basic structure of the trained model in this example consists of an input layer (550 absorbance data sets at 2 nm wavelength intervals are input), an intermediate layer, and an output layer (the same applies to the following examples).

This trained model A uses near-infrared spectroscopic absorbance data corresponding to the aforementioned actual infection measurement data as training data, creates 3-5 types of models using the cross-5-fold validation method of deep learning (DNN method), and 3-10 types of models using the cross-5-fold method of gradient boosting trees, and is constructed using an ensemble method to select the optimal model based on the accuracy rate, error rate, recall rate, and F-value of each model.

As shown in Figure 19(a), statistical analysis has shown that trained model A has extremely high discrimination accuracy, with an AUC of 0.9999, an Accuracy of 0.9992, a Precision of 0.9992, a Recall of 0.9992, and an F-score of 0.9992. Furthermore, as shown in Figure 19(b), the distribution of the training data used to establish trained model A is distributed in a manner that allows for extremely clear separation.

The discrimination rate and accuracy were verified using the computational model according to Example 5 (AI-trained model A for determining the presence or absence of infection). 3,873 garlic bulbs prepared for shipping from the same production area as above were used. As shown in Figure 20, with this trained model A, the multiple correlation coefficient was 0.9979, and the discrimination accuracy for the presence or absence of infection was 100.0% for no infection and 99.9% for infection. This shows that the model can accurately determine the presence or absence of potato-lesion nematode infection in garlic.

<Example 6: AI trained model B (damage level (6 levels) discrimination)>
For the test subject W, 12,480 garlic bulbs prepared for shipping from the same production area as above were used as supervised training data. Near-infrared spectroscopy testing was performed on all test subjects W. The test garlic bulbs were then cultured for potato lesion nematodes using the Bellman method, and the damage level was calculated based on the severity of the garlic lesions using the six levels described above. The trained model B for the six levels of damage was constructed using near-infrared spectroscopic absorbance data corresponding to the measured damage level data (six levels) as training data. Three to five models were created using a cross-five-fold validation method with deep learning (DNN method) and three to ten models were created using a cross-five-fold method with gradient boosting trees. The optimal model was then selected using an ensemble method based on the accuracy rate, error rate, recall rate, and F-value of each model.

As shown in Figure 21(a), in a statistical analysis of the measured and estimated damage levels (six levels), trained Model B showed an Accuracy of 0.9965, an average Precision of 0.9986, an average Recall of 0.9902, and an average F-score of 0.9943, demonstrating extremely high discrimination accuracy, goodness of fit, and recall. In other words, this model can classify the damage level (six levels) of garlic potato-lesion nematodes with high accuracy.

The discrimination rate and accuracy were verified using the computational model (AI-trained model B for discriminating damage level (6 levels)) according to Example 6. 3,120 garlic bulbs prepared for shipping from the same production area as above were used. The results are shown in Figure 21(b). From these results, trained model B for damage level (6 levels) demonstrated extremely high discrimination accuracy and goodness of fit, with an accuracy of 0.9999 between actual and estimated values and an average precision of 0.9979. In other words, it can be seen that this model is capable of classifying the damage level (6 levels) of garlic caused by potato-lesion nematodes with high accuracy.

<Example 7: AI trained model C (damage level (index) determination)>
For the test subject W, 12,480 garlic bulbs prepared for shipping from the same production area as above were used as supervised training data. The actual damage index was calculated by near-infrared spectroscopy testing of all test subjects W, culturing the test garlic bulbs for potato lesion nematodes using the Bellman method, and calculating the extent of garlic lesions using the above damage index. To create the trained model C for the damage index, the aforementioned measured damage data was used as supervised data. Three to five models were created using deep learning (DNN) cross-validation (5-fold) and three to ten models were created using gradient boosting tree cross-validation (5-fold). For each model, the optimal model was selected based on the statistical results of the coefficient of determination, median error, mean error, and PMSE.

As shown in Figure 22(a), the coefficient of determination between the measured and predicted values for the damage level (index) trained model C was 0.9809. With a median error of 0.303, a mean error of 2.36, an RMSE of 5.98, and a coefficient of determination of 0.9809, extremely high prediction accuracy was achieved. Therefore, trained model C is a model that can quickly, accurately, and non-destructively predict the damage level (index) caused by garlic potato-lesion nematode infection.

The method and accuracy were verified using the computational model (damage level (index) discrimination AI trained model C) according to Example 7. 3,120 garlic bulbs prepared for shipping from the same production area as above were used. The results are shown in Figure 22(b). From these results, trained model C demonstrated extremely high prediction accuracy, with a coefficient of determination of 0.9936, median error of 0.1646, mean error of 0.9788, and RMSE of 3.5269. This means that the model can predict the damage level (index) of garlic caused by potato-lesion nematodes with high accuracy.

<Example 8: AI trained model D (infection determination)>
18,960 garlic cloves prepared for shipping from the same production area as above were used as training data for the test subjects W. The training data for this trained model D was prepared by testing all test subjects W using near-infrared spectroscopy, cultivating the test garlic for potato-lesion nematodes using the Berman method, and checking for infection, and correlating the presence or absence of infection with the absorbance data of the sample at the near-infrared spectroscopic wavelength.

The trained model D is based on the deep learning DNN (deep neural network) method, and is composed of an input layer, five intermediate layers (composed of activation functions such as drop and Relu functions), and an output layer. The optimization algorithm used is Adam (adaptive moment estimation), with a Learning Rate of 0.1, Momentum of 0.9, Weight Decay of 0.0001, a Learning Rate Scheduler of Exponential, Multipier of 0.1, and an Interval of 600540 iterations. This trained model was created with due consideration given to overfitting.

As shown in Figure 23(a), this trained model D has an AUC (Area Under the Curve) of 0.9998, Accuracy of 0.9994, Precision of 0.9996, Recall of 0.9972, and an F-score of 9.994, demonstrating extremely high discrimination accuracy and making it an extremely effective model for determining whether or not a patient is infected with potato-lesion nematodes.

The trained model D for determining the presence or absence of infection according to Example 8 was used to verify its discrimination rate and accuracy. 3,914 garlic bulbs prepared for shipping from the same production area as above were used. As shown in Figure 23(b), with Precision 1.0, the trained model D achieved an infection discrimination accuracy of 100.0%. This means that the model can accurately determine whether garlic is infected with potato-lesion nematodes.

<Example 9: AI trained model E (damage level (6 levels) discrimination)>
For the test subject W, 15,600 garlic bulbs prepared for shipping from the same production area as above were used as supervised training data. Near-infrared spectroscopy was performed on all test subjects W, and then the test garlic bulbs were cultured for potato lesion nematodes using the Bellman method. The damage level was calculated based on the extent of garlic lesions using the six levels described above. The trained model E was created using deep learning (DNN: deep neural network) with near-infrared spectroscopic absorbance data corresponding to the measured damage level (six levels) as training data. The trained model E is composed of an input layer, six intermediate layers (activation functions such as drop and Relu functions), and an output layer. The optimization algorithm used was Adams, with a weight decay of 0.0001, and consideration was given to overfitting.

As shown in Figure 24(a), trained model E has an accuracy of 0.9976 between measured and estimated values, an average precision of 0.9991, an average recall of 0.9945, and an average F-measure of 0.9968. This means that it has extremely high discrimination accuracy, goodness of fit, and recall, making it a trained model that can be used as an extremely effective method for discriminating and classifying damage levels (6 levels).

The discrimination rate and accuracy were verified using the computational model (AI-trained model E for discriminating damage level (6 levels)) according to Example 9. 3,873 garlic bulbs prepared for shipping from the same production area as above were used. The results are shown in Figure 24(b). From these results, the trained model E for damage level (6 levels) had an average precision of 0.9994, demonstrating extremely high fitness and can be used to classify the degree of damage. In other words, it can be seen that the trained model E is a model that can quickly, accurately, and non-destructively predict the degree of damage (6 levels) caused by garlic potato-lesion nematode infection.

<Example 10: AI trained model F (damage level (index) determination)>
For the garlic bulbs of test subject W, 12,480 garlic bulbs prepared for shipping from the same production area as above were used as supervised learning data. The actual measured values of the damage level (index) were determined by near-infrared spectroscopy testing for all test subjects W, followed by culturing the tested garlic bulbs for potato lesion nematodes using the Bellman method, and calculating the extent of garlic lesions using the above-mentioned damage level (index). The trained model F was created using deep learning (DNN: deep neural network) using near-infrared spectroscopy absorbance data corresponding to the actual measured value data of the damage level (index) as training data.

The trained model F uses the DNN method as its basic structure, and after repeated fundamental research and ingenuity to obtain optimal conditions for the input, intermediate, and output layers, is composed of an input layer, five intermediate layers (composed of activation functions such as drop and Relu functions), and an output layer. The optimization algorithm is Adam (Adaptive moment estimation), with a Learning Rate of 0.1, Momentum of 0.9, Weight Decay of 0.0001, a Learning Rate Scheduler of Exponential, Multipier of 0.1, and Interval of 600540 iterations. This model is characterized by being created with attention to overfitting.

As shown in Figure 25(a), the trained model F for damage level (index) is a highly accurate model, based on statistical data analysis of the estimated values and actual measured values, with a coefficient of determination of 0.9987, a median error of 0.2737, a mean error of 0.2593, and an RMSE of 1.6037. This model can quickly, accurately, and non-destructively estimate the level of damage caused by garlic lesion nematode infection.

The discrimination rate and accuracy were verified using the computational model (damage level (index) discrimination AI trained model F) according to Example 10. 2,851 garlic bulbs prepared for shipping from the same production area as above were used. The results are shown in Figure 25(b). From these results, trained model F obtained test results in a statistical data analysis of the predicted values and actual measured values, with a multiple correlation coefficient of 0.999, a mean error of 0.5260, a median error of 0.2743, and an RMSE of 1.3025. In other words, it can be seen that trained model F is a model that can quickly, accurately, and non-destructively predict the damage level (index) of garlic caused by potato-lesion nematode infection.

<Example 11: AI trained model G (infection determination)>
The training data used for the test subject W were 12,480 garlic bulbs prepared for shipping from the same production area as above. The training data for this trained model G was prepared by testing all test subjects W using near-infrared spectroscopy, cultivating the test garlic for potato-lesion nematodes using the Berman method, and then checking for the presence or absence of infection, and correlating the presence or absence of infection with the absorbance data of the sample at the near-infrared spectroscopic wavelength.

The trained model G was created using deep learning (CNN: convolutional neural network) using near-infrared spectroscopic absorbance data corresponding to actual measured values of infection as training data. The trained model G visualizes the waveform (1 row, 550 columns) of the near-infrared spectroscopic absorbance data and is composed of an input layer, six intermediate layers (activation functions such as drop and Relu functions), and an output layer. The model was created using the Adams optimization algorithm with a weight decay of 0.0001, taking into consideration overfitting.

As shown in Figure 26 (a), this trained model G has an accuracy of 1.0000 between the actual measured and predicted values, an average precision of 1.0000, an average recall of 1.0000, and an average F-measure of 1.0000. This means that it has extremely high discrimination accuracy, goodness of fit, and recall, making it a model that can be used as an extremely effective method for determining whether or not a patient is infected with potato lesion nematodes.

The trained model G for determining the presence or absence of infection according to Example 11 was used to verify its discrimination rate and accuracy. 3,121 garlic bulbs prepared for shipping from the same production area as above were used. As shown in Figure 26(b), the trained model G had an infection discrimination accuracy of 100.0%. This means that the model can accurately determine whether garlic is infected with potato-lesion nematodes.

<Example 12: AI trained model H (damage level (6 levels) discrimination)>
For the garlic bulbs of the test subject W, 12,480 garlic bulbs prepared for shipping from the same production area as above were used as supervised training data. Near-infrared spectroscopy testing was performed on all test subjects W, and then the test garlic bulbs were cultured for potato leaf nematodes using the Bellman method. The damage level was calculated based on the extent of garlic lesions using the above-mentioned damage level (6 levels). The trained model H was created using deep learning (CNN: convolutional neural network) using near-infrared spectroscopic absorbance data corresponding to the measured damage level (6 levels) as training data. The trained model H was created by imaging the waveform (1 row, 550 columns) of the near-infrared spectroscopic absorbance measurement data. It is composed of an input layer, six intermediate layers (activation functions such as drop and Relu functions), and an output layer. The optimization algorithm was Adams, with a weight decay of 0.0001, and consideration was given to overfitting.

As shown in Figure 27(a), the trained model H had an accuracy of 1.0000 between the actual measured and estimated values, an average precision of 1.0000, an average recall of 1.0000, and an average F-measure of 1.0000. This means that the model has extremely high discrimination accuracy, goodness of fit, and recall, making it a model that can be used as an extremely effective method for discriminating and classifying damage levels (6 levels).

The discrimination rate and accuracy were verified using the computational model (AI-trained model H for discriminating damage level (6 levels)) according to Example 12. 3,120 garlic bulbs prepared for shipping from the same production area as above were used. The results are shown in Figure 27(b). These results demonstrate that trained model H has a fitness of 1.000 and can be used to classify damage level (6 levels). In other words, it can be seen that trained model H is a model that can quickly, accurately, and non-destructively predict the damage level (6 levels) caused by garlic potato-lesion nematode infection.

<Example 13: AI trained model I (damage level (index) determination)>
For the garlic bulbs of test subject W, 12,480 garlic bulbs prepared for shipping from the same production area as above were used as supervised learning data. The actual measured values of the damage index were determined by near-infrared spectroscopy testing of all test subjects W, followed by culturing the tested garlic bulbs for potato lesion nematodes using the Bellman method, and calculating the extent of garlic lesions using the above-mentioned damage index. The trained model I was created using deep learning (CNN: convolutional neural network) using near-infrared spectroscopic absorbance data corresponding to the actual measured value data of the damage index as training data. Trained Model I is constructed using the CNN method by visualizing the waveform (1 row, 550 columns) of near-infrared spectroscopic absorbance data, and is composed of an input layer, six intermediate layers (activation functions such as drop and Relu functions), and an output layer.The optimization algorithm used is Adams, with Weight Decay set to 0.0001, and attention paid to preventing overfitting.

As shown in Figure 28(a), the trained model I is a highly accurate model, with a coefficient of determination of 0.9831, a median error of 1.2806, a mean error of 4.2906, and an RMSE of 5.8131 in statistical data analysis of the estimated values and actual measured values. This model can quickly, accurately, and non-destructively estimate the degree of damage (index) caused by garlic lesion nematode infection.

The discrimination rate and accuracy were verified using the computational model (damage level (index) discrimination AI trained model I) according to Example 13. 2,851 garlic bulbs prepared for shipping from the same production area as above were used. The results are shown in Figure 28(b). The results show that the statistical multiple correlation coefficient between the estimated values and the actual measured values was 0.999, with a mean error of 4.2590, a median error of 1.2513, and an RMSE of 5.72955. In other words, it can be seen that trained model I is a model that can quickly, accurately, and non-destructively predict the damage level (index) of garlic caused by potato-lesion nematode infection.

In the above embodiment, discrimination was performed using a combination of five discrimination methods: discriminant analysis, multiple regression analysis, PLS analysis, backpropagation neural network method, and SVM. However, this is not necessarily limited to these; any one discrimination method may be used, or several methods may be combined, and modifications may be made as appropriate. Also, in the above embodiment, classification was performed in six stages, but this is not necessarily limited to this; the number of stages may be determined as appropriate. Furthermore, in the embodiment, the test object was garlic bulbs, but this is not necessarily limited to this; individual cloves may also be inspected, and modifications may be made as appropriate.

Furthermore, while the embodiments have been described using garlic as the living organism and potato-lesion nematode as the parasite, the present invention is not limited to these and can be applied to various plants and the nematodes that parasitize them. Furthermore, it goes without saying that the present invention can be applied to various living organisms and various parasites, not just plants and nematodes. In short, those skilled in the art can easily make many modifications to these exemplary embodiments without substantially departing from the novel teachings and effects of the present invention, and these many modifications are within the scope of the present invention.

S: Living organism testing device W: Test subject (garlic bulb)
Wa: scale Wb: flower diameter Wc: disc stalk 1: base 2: conveyor 3: belt conveyor 10: robot 11: robot arm 12: gripping hand 13: opening/closing actuator X: gripping position Y: inspection position 20: holding section 21: side wall 22: upper wall 23: holding body 24: insertion hole 25: support recess 26: camera 30: inspection section 31: inspection unit 32: support rod 33: cover 34: unit body 35: light source section 36: light receiving section 37: detection section 38: holding body 39: recess 40: dustproof mechanism 41: protrusion section 42: air injection port 43: air passage 50: receiving section 51: chute 52: sorting means 60: control section 61: calculation processing section 62: calculation model storage section 63: discrimination section 64: judgment section 65: operation control section 66: communication section 67: inspection control section 68: drive control section 69: operation command section 70: display section

Claims (12)

A method for testing a living organism for infection with a parasite in a living organism to be tested, comprising:
A calculation model is constructed in advance, which performs calculations relating to the presence or absence of parasitic infection in the living organism and the degree of damage, which is quantified based on the second derivative spectrum of the absorbance, by irradiating near-infrared light onto a sample living organism infected with a parasite and a sample living organism not infected with a parasite, receiving diffuse reflected light or diffuse transmitted light from the sample living organism, and measuring the absorbance of the received light, and quantifying the degree of onset due to infection, from a predetermined maximum of no onset to onset,
The computational model is an analytical model constructed using statistical analysis or a trained model constructed using machine learning,
The statistical analysis is selected from discriminant analysis, multiple regression analysis, PLS regression analysis, logistic regression analysis, and support vector machine (SVM);
The machine learning is selected from tree analysis such as decision trees (CART), regression trees, random forests, and gradient boosting trees, neural networks such as perceptrons, convolutional neural networks (CNN), recurrent neural networks (RNN), and residual networks (ResNet), deep learning (DNN), and ensemble analysis consisting of a combination of these,
A method for examining a living organism, characterized by irradiating near-infrared light onto the living organism to be examined, receiving diffusely reflected light or diffusely transmitted light from the living organism to be examined, measuring the absorbance of the received light, and determining the presence or absence of the infection and/or the degree of damage based on the calculation results calculated from this absorbance and the above-mentioned calculation model. The method for examining a living body according to claim 1, characterized in that the degree of damage is also graded for non-infected test subjects. A living organism inspection method according to claim 1 or 2, characterized in that the degree of damage is determined based on the degree of discoloration appearing on the subject. The biological testing method described in claim 3, characterized in that the damage level is constructed as an index where 0 indicates no onset and 100 indicates onset. The biological testing method described in claim 3, characterized in that the damage level is expressed as an index where the degree of onset is 0 for no onset and 100 for onset, with no onset being 0 and onset being the maximum being 100, and the onset side being ranked in ascending order for each predetermined index range. A method for examining a living body as described in any one of claims 1 to 5, characterized in that as the computational model, a plurality of computational models of different types for determining the presence or absence of infection are used, and when any one of the discrimination results using each computational model determines the presence of infection, that discrimination result is given priority. A method for examining a living body as described in any one of claims 1 to 5, characterized in that three or more different types of computational models for determining the presence or absence of infection are used as the computational models, and the infection presence or absence determination results of each computational model are determined by majority vote. 8. The method for inspecting a living organism according to claim 1 , wherein the parasite is a nematode that infects a living plant. 9. The method for inspecting a living organism according to claim 8 , wherein the living organism is selected from the group consisting of garlic, potato, sweet potato, and iris, and the parasite is a potato-lesion nematode. 10. The method for examining a living body according to claim 9 , wherein the subject is a garlic bulb made up of a collection of garlic clove. A method for examining a living body as described in claim 10, characterized in that the degree of damage is determined based on the degree of discoloration appearing in the subject, and is composed of the following damage degree, which is an index where 0 indicates no onset and 100 indicates the maximum of onset.
Damage index = ((number of scales in t = 1) + (number of scales in t = 2 x 2) + (number of scales in t = 3 x 3) + (number of scales in t = 4 x 4)) / (total number of scales surveyed x 4) x 100
where:
t is a number corresponding to the size of the discolored area caused by parasitic infection per garlic bulb, and is set so that the larger the value, the larger the size of the discolored area, and is set to four categories: t=1, t=2, t=3, and t=4. A method for inspecting living organisms as described in claim 10, characterized in that the degree of damage is determined based on the degree of discoloration appearing in the subject, and the degree of onset is expressed as an index where 0 represents no onset and 100 represents the maximum of onset, and the degree of onset is expressed as the following degrees of onset, where 0 represents no onset and the onset side is composed of classes that increase in order for each specified index range.
Severity of onset (index) = ((number of scales at t = 1) + (number of scales at t = 2 x 2) + (number of scales at t = 3 x 3) + (number of scales at t = 4 x 4)) / (total number of scales surveyed x 4) x 100
where:
t is a number corresponding to the size of the discolored area caused by parasitic infection per garlic bulb, and is set so that the larger the value, the larger the size of the discolored area, and is set to four categories: t=1, t=2, t=3, and t=4.