JP6468576B1 - Image diagnostic system for fertilized egg, image diagnostic program for fertilized egg, and method for creating classifier for image diagnosis of fertilized egg. - Google Patents

Abstract

Provided is an image diagnostic system for a fertilized egg that can obtain a diagnosis result that can predict whether or not a normal child can be acquired by using only image data of a fertilized egg without subjective judgment.
[Solution]
The image data of a fertilized egg is image data composed of normal image data that has been found to have resulted in the acquisition of newborns and abnormal image data that has been found to have not resulted in the acquisition of newborns, or pregnancy. It is image data composed of normal image data that has been found to have been established, and abnormal image data that has been found to have not been established, and includes patient age information, The analysis means 6 divides the image data into a plurality of groups according to age, creates a plurality of discriminators for the analysis target group selected from the plurality of groups, and calculates the correct diagnosis rate of the test data. The optimum classifier for each age is determined from the plurality of classifiers based on the correct diagnosis rate.
[Selection] Figure 1

Description

  The present invention relates to an image diagnosis system for a fertilized egg, an image diagnosis program for a fertilized egg, and a method for creating a classifier for image diagnosis of a fertilized egg.

  While Japan is in danger of declining birthrate, many patients suffer from infertility. In vitro fertilization or microinsemination can be used as a treatment for establishing pregnancy in infertile patients. In vitro fertilization is a treatment that removes egg cells from the ovaries (egg collection) and gives sperm outside the body to make a fertilized egg. It is a treatment to make. In any treatment, fertilized eggs are cultured and grown, frozen and stored once, thawed when the patient's condition is good, and returned to the uterus to establish pregnancy.

  The selection of the fertilized egg to be returned to the uterus is made by a doctor, laboratory technician, or cultivator subjectively by microscopic observation before thawing and after thawing. Only 30% of people get pregnant. That is, the accuracy of subjective judgment is low, and even if the judgment method is devised, there is a certain limit to raising the precision.

  On the other hand, the cost of treatment per treatment is high and is a self-paying burden. For this reason, if the accuracy of the determination of fertilized egg selection is low, the number of treatments increases and time is wasted, and in addition to psychological stress, economic stress is also great. Under such circumstances, infertility treatment requires both patients and society to reach pregnancy as soon as possible and with less economic burden.

  In this regard, in the medical field, various techniques for analyzing and evaluating diagnostic images using a computer have been proposed. For example, Patent Literature 1 proposes a diagnosis support technology that improves the accuracy of diagnosis by causing a computer to search for similar cases using image feature amounts when interpreting a diagnostic image. Patent Document 2 proposes a fertilized egg quality evaluation technique that uses a computer to realize information useful for quality evaluation of a fertilized egg without identifying cells in the fertilized egg using only the captured image. Yes.

JP 2002-230518 A JP 2010-181402 A

  However, the technique described in Patent Document 1 does not target a fertilized egg image, but the target image is a CT image targeting an organ, and it is possible to search for similar cases to improve diagnostic accuracy. It was a thing. That is, from the description of Patent Document 1, it has been difficult to derive a method for determining whether or not pregnancy can be established for an image of a fertilized egg. On the other hand, the technique described in Patent Document 2 is intended for an image of a fertilized egg, but it is limited to a technique for evaluating the quality of a fertilized egg, and is not a technique for directly determining whether pregnancy is established.

  The present invention solves the conventional problems as described above, and can obtain a diagnosis result that can predict whether or not a normal child can be acquired or pregnancy can be established by using only image data of a fertilized egg without subjective judgment. An object of the present invention is to provide a fertilized egg image diagnostic system, a fertilized egg image diagnostic program, and a method for creating a fertilized egg image diagnostic discriminator.

  In order to achieve the above object, an image diagnostic system of the present invention is a fertilized egg image diagnostic system using a machine learning technique, and stores image data of a fertilized egg whose pregnancy prognosis is known. And receiving means for capturing image data of a fertilized egg into a storage means, analysis means for analyzing the image data, and output means for outputting an analysis result by the analysis means, Image data composed of normal image data that has been found to have resulted in the acquisition of births and abnormal image data that has been found to have not resulted in the acquisition of births, or found that pregnancy has been established Image data composed of normal image data and abnormal image data that has been found not to have become pregnant, and includes age information of the patient. The image data is divided into a plurality of groups according to age based on the age information of the patient, the analysis target group selected one from the plurality of groups is classified into test data and learning data, A plurality of discriminators are created from the learning data, the test data is diagnosed by each discriminator, a correct diagnosis rate of the test data is calculated, and the plurality of discriminators are calculated based on the correct diagnosis rate. Determining an optimum classifier, performing the processing from the classification to the determination of the optimum classifier by selecting another group as an analysis target group, and repeating this to determine the optimum classifier by age It is characterized by.

  The diagnostic imaging program of the present invention is a diagnostic image program for a fertilized egg for causing a computer to perform image diagnosis of a fertilized egg using a machine learning technique, and fetching image data of the fertilized egg into a storage means And an analysis step for analyzing the image data and an output step for outputting an analysis result by the analysis means, and the image data is a normal image that has been found to have acquired a baby. Image data composed of data and abnormal image data that has been found not to have acquired a baby, or normal image data that has been found to have achieved pregnancy, and pregnancy has not been achieved It is image data composed of abnormal image data that has been found, and includes age information of the patient, in the analysis step, The image data is divided into a plurality of groups according to age based on the age information of the patient, and the analysis target group selected one from the plurality of groups is classified into test data and learning data, A plurality of discriminators are created from the learning data, the test data is diagnosed by each discriminator, a correct diagnosis rate of the test data is calculated, and the plurality of discriminators are calculated based on the correct diagnosis rate The optimum classifier is determined from the above, and the processing from the classification to the determination of the optimum classifier is performed by selecting another group as the analysis target group, and this is repeated to determine the optimum classifier by age. It is characterized by that.

  The method for creating a classifier for image diagnosis of the present invention is a method for creating a classifier for image diagnosis of a fertilized egg using a machine learning technique, and includes a step of taking in image data of a fertilized egg into a storage means, An analysis step for analyzing image data; and an output step for outputting an analysis result by the analysis means, wherein the image data includes normal image data that has been found to have acquired a child, and acquisition of a child. Image data composed of abnormal image data that has been found not to have reached, or normal image data that has been found to have been established, and abnormalities that have been found not to have been established Image data composed of image data and includes patient age information, and in the analyzing step, the image data is classified according to age based on the age information of the patient. The analysis target group selected from the plurality of groups is classified into test data and learning data, and a plurality of classifiers are created from the learning data. Diagnosing the test data, calculating a correct diagnosis rate of the test data, determining an optimal classifier from the plurality of the classifiers based on the correct diagnosis rate, and determining the optimal classifier from the classification The above processes are executed by selecting another group as an analysis target group, and the process is repeated to determine the optimum classifier for each age.

  According to the present invention, it is possible to diagnose a fertilized egg image without subjective judgment by simply executing a computer program without using equipment such as photographing and measurement. The classifier that diagnoses the data is statistically created by machine learning based on the pregnancy prognosis image data whose pregnancy prognosis is known, and is also created by age, so the diagnosis results are normal. Whether or not acquisition or pregnancy can be established can be predicted with high accuracy.

  In the diagnostic imaging system of the present invention, the analysis means repeatedly creates the discriminator and calculates the correct diagnosis rate by changing the number of learning data to be selected before determining the optimum discriminator. Determining the optimum number of the learning data based on the correct diagnosis rate calculated for the learning data having different numbers, and setting the number of the learning data as the optimum number, The calculation of the correct diagnosis rate is repeated again while changing the learning data to be selected, and the plurality of discriminators obtained by the second repetition based on the plurality of correct diagnosis rates obtained by the second repeat. It is preferable to determine the optimum classifier from

  In the image diagnostic program and the method for creating an image diagnostic discriminator according to the present invention, in the analyzing step, the analyzing means creates the discriminator and calculates the correct diagnosis rate before determining the optimum discriminator. By changing the number of the learning data to be selected, and determining the optimum number of the learning data based on the correct diagnosis rate calculated for the learning data having different numbers. Using the number as the optimum number, the creation of the discriminator and the calculation of the correct diagnosis rate are repeated again by changing the learning data to be selected, and based on the correct check rates obtained by the repeat again It is preferable that an optimum classifier is determined from a plurality of the classifiers obtained by the second iteration.

  According to the preferred configuration of the present invention, since the optimum number of learning data is determined based on the correct diagnosis rate, it is possible to prevent a decrease in diagnostic accuracy due to overlearning of machine learning.

  The effects of the present invention are as described above, and according to the present invention, subjective judgment is made by simply executing a computer program on image data of a fertilized egg without using equipment such as photographing and measurement. Diagnosis is possible without any problems. The classifier that diagnoses the data is statistically created by machine learning based on the pregnancy prognosis image data whose pregnancy prognosis is known, and is also created by age, so the diagnosis results are normal. Whether or not acquisition or pregnancy can be established can be predicted with high accuracy.

The block diagram of the diagnostic imaging system of a fertilized egg which concerns on one Embodiment of this invention. The flowchart of the diagnostic imaging system which concerns on one Embodiment of this invention. The figure which showed an example of the age division which concerns on one Embodiment of this invention. The schematic diagram of an example of the normal image data of a fertilized egg. The schematic diagram of an example of the abnormal image data of a fertilized egg. The schematic diagram which showed the image data which amplified this in addition to the normal image data shown in FIG. 6 is a flowchart showing a process of determining the optimum number of learning data in an embodiment of the present invention. The figure which shows the relationship between the number of the data for learning of the group 1, and the correct diagnosis rate in one Embodiment of this invention. 6 is a flowchart illustrating a process for determining an optimal classifier in an embodiment of the present invention. The figure which shows the relationship between the number of the data for learning of the group 2, and the correct diagnosis rate in one Embodiment of this invention. The figure which shows the relationship between the number of the data for learning of the group 3, and the correct diagnosis rate in one Embodiment of this invention. The figure which shows the relationship between the number of the data for learning of the group 4, and the correct diagnosis rate in one Embodiment of this invention. The figure which shows the relationship between the number of the data for learning of the group 5, and the correct diagnosis rate in one Embodiment of this invention. The figure which shows the relationship between the optimal discrimination device and diagnostic object image data in one Embodiment of this invention. The figure which showed the comparison result of the correct diagnosis rate of the optimal discrimination device and comparison discrimination device which concerns on one Embodiment of this invention.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 shows a configuration diagram of a fertilized egg image diagnostic system (hereinafter simply referred to as “image diagnostic system”) according to an embodiment of the present invention. The image diagnosis system 1 is a system that outputs determination information regarding the feasibility of pregnancy based on image data of a fertilized egg. The diagnostic imaging system 1 delivers various data such as image data of a fertilized egg to a plurality of medical institutions 1 and 2 via the Internet 8. Various calculation processes are performed by the analysis means 6 (CPU: central processing unit). Various data are input and set by the input means 4 such as a keyboard. An output result is output by output means 5 such as a display or a printer. The image data from the medical institution received by the receiving unit 2 is stored in the storage unit 7 for each medical institution. The analysis result by the analysis unit 6 is transmitted to a plurality of medical institutions 1, 2 and the like by the transmission unit 3.

  The image data handled by the image diagnostic system 1 includes image data whose pregnancy prognosis is known by in vitro fertilization or microinsemination (hereinafter referred to as “pregnancy prognosis image data”) and image data to be diagnosed (hereinafter referred to as “pregnancy prognosis image data”) , “Diagnosis target image data.” The diagnosis target image data is data for which it is unknown whether pregnancy is established or not .. In Fig. 1, the receiving means 2 receives the diagnosis target image data 11 from a plurality of medical institutions 1, 2, etc. Are separately received via the Internet 8. Separately, the receiving means 2 receives the pregnancy prognosis finding image data 10. Both image data are in, for example, jpeg format, and both are stored in the storage means 7. .

  All the image data are subjected to image processing before being analyzed. The image processing is, for example, erasing of background artifacts and shadows of the storage container, and standardization processing of a digitized image. Therefore, the image data analyzed by the analysis means 6 is subjected to image processing in advance. The image processing only needs to be performed before analysis, may be performed by the same program, or may be performed by another program for image processing. Normally, in any case, the analysis unit 6 analyzes the image data after image processing. However, when performing image processing with the same program, the analysis unit 6 performs image processing on the image data before analysis. You may have the function to perform.

  The analysis means 6 derives judgment information regarding whether or not pregnancy is possible for the diagnostic object image data. As a premise for deriving such judgment information, an optimum classifier for image diagnosis, which will be described later, is created in the image diagnosis system 1 based on the pregnancy prognosis finding image data 10. That is, the information processing by the image diagnostic system 1 is roughly divided into information processing for creating an optimal classifier based on the pregnancy prognosis image data 10 and information processing for diagnosing diagnosis target image data using the optimal classifier. . In the present embodiment, an example in which these two information processes are executed by one program will be described, but each information process may be executed by another program.

  An outline of the information processing procedure by the diagnostic imaging system 1 will be described first with reference to FIG. FIG. 2 shows a flowchart of the diagnostic imaging system 1 according to an embodiment of the present invention. FIG. 7 described later shows a flowchart showing the process of determining the optimum number of learning data, and FIG. 9 shows a flowchart showing the process for determining the optimum classifier. The image diagnosis system 1 is used by causing a computer to execute an image diagnosis program. In the present embodiment, the present invention will be described as the diagnostic imaging system 1. The flow of information processing of the diagnostic imaging system 1 shown in FIGS. 2, 7, and 9 shows steps that the diagnostic imaging program causes the computer to execute. It is also a thing. Therefore, the contents of the information processing by the image diagnostic system 1 described in the present embodiment are the contents that the image diagnostic program causes the computer to execute and the process of the method for creating the image diagnostic discriminator.

  In FIG. 2, after initial setting (step 100), image data is read (step 101). As described above, the image data handled by the image diagnosis system 1 includes two types of image data 10 (see FIG. 1) for which the pregnancy prognosis is known and diagnosis target image data 11 (see FIG. 1) to be diagnosed. It is. Both image data may be read in succession, but if at least the pregnancy prognosis finding image data 10 is read, an optimum discriminator can be created (step 103). Thereafter, the diagnostic object image data 11 may be read. .

  Steps 102 and 103 are steps in the previous stage before diagnosing the diagnosis object image data 11, and in this embodiment, in order to improve the accuracy of diagnosis, the pregnancy prognosis finding image data 10 (see FIG. 1) is converted to age. The optimum classifier is determined for each age category (step 103). For this reason, the diagnostic object image data 11 includes patient age information. FIG. 3 shows an example of age classification. In the example of FIG. 3, the diagnosis target image data 11 is classified into five groups according to the age division shown in FIG.

  When the optimum classifier is determined for each age category, the image data is diagnosed by the optimum classifier corresponding to the age of the diagnosis target image data 11 (see FIG. 1) (step 104), and the analysis result is output (step) 105). For example, for image data 11 to be diagnosed belonging to group 1 (42 years or older) in FIG. 3, the image data is diagnosed by an optimum classifier corresponding to group 1.

  Both the analysis process for determining the optimum classifier and the analysis process for diagnosing image data by the optimum classifier utilize pattern recognition of supervised machine learning by artificial intelligence. An example of a program language for analysis processing is the Wolfram Language. As for machine learning pattern recognition, there are recognition methods such as logistic regression, naive Bayes, random forest, nearest neighbors, neural network, and deep learning. The Wolfram Language already has functions for handling these recognition methods, and you can select the best one from each recognition method.

  Next, an image handled by the diagnostic imaging system 1 will be described. In FIG. 1, pregnancy prognosis finding image data 10 is classified into normal image data and abnormal image data. Normal image data is image data that has been found to have resulted in the acquisition of newborns, and abnormal image data is image data that has been determined to have not resulted in the acquisition of newborns. The reason why the baby was not acquired may be clinically chromosomal abnormalities, uterine factors, etc., but there is no reason why the baby was not acquired when classifying abnormal image data.

  When both the normal image data and the abnormal image data are analyzed by the analysis means 6, the formats are unified. For example, the image resolution is converted to, for example, a square of 113 × 113 pixels at 72 pixels / inch. Reference numeral 12 in FIG. 4 shows a schematic diagram of an example of normal image data, and reference numeral 13 in FIG. 5 shows a schematic diagram of an example of abnormal image data. Since each image data is converted into a square, the same square image data can be amplified by rotation and inversion.

  FIG. 6 shows image data obtained by amplifying the normal image data shown in FIG. As shown in FIG. 6, three images are added to one image data by 90 ° rotation, 180 ° rotation, and 270 ° rotation, and are inverted, inverted 90 ° rotated, inverted 180 ° rotated, inverted 270 °. Four images are added by rotation, and when the original image data is included, eight image data are obtained. The same applies to the abnormal image data 13 shown in FIG. In this way, by amplifying the image data, one image data can be amplified to 8 times the number of image data, so that the characteristics of the image data can be expressed in many ways, and the classifier described later Accuracy can be increased.

  The unification of the format of the image data is not limited to the above, but may be changed as appropriate. For example, each image data may be converted into a circle, and the image data may be amplified by rotation and inversion. In the case of circular image data, by rotating the image data in small increments at a small angle, the number of amplified images can be increased, and the characteristics of the image data can be expressed in many ways, which is advantageous for improving the accuracy of the discriminator.

  The diagnostic imaging system 1 creates a discriminator using a plurality of learning data selected from pregnancy prognosis finding image data including normal image data and abnormal image data. In the present embodiment, the optimal number of learning data is determined in advance in order to prevent a decrease in diagnostic accuracy due to overlearning of machine learning. Hereinafter, the determination of the optimum number of learning data will be described with reference to FIG.

  FIG. 7 is a flowchart showing a process of determining the optimum number of learning data. For the sake of convenience, a numerical example will be described. However, this is an example and the present invention is not limited to this. In the present embodiment, the diagnostic imaging system 1 has received 5691 images of fertilized eggs (blastocysts) on the fifth day after the fertilization as the pregnancy prognosis finding image data 10 in FIG. These images were classified into five groups as shown in the example of FIG. 3 (step 103 in FIG. 2). As described above, the image data is amplified to an arbitrary number by amplifying the image data.

  Hereinafter, the optimum number of learning data is determined for each of the five groups. FIG. 7 shows a process of determining the optimum number of learning data for group 1 (42 years old or older). First, 669 pieces of image data of the age group of group 1 are divided into test data and learning data, and a certain number of pieces of learning data is used as verification data (step 110). This division was random, with 20% of all image data being test data, the remaining 80% being learning data, and 20% of learning data being verification data. N1 pieces are selected from the learning data, the n1 pieces are amplified to create n pieces of learning data (step 111), and a discriminator is formed using the n pieces of learning data including the verification data. (Step 112). In this step, the discriminator is learned while adjusting with the verification data, and the best discriminator is created from the n pieces of learning data. The ratio of the number of normal image data and abnormal image data in the n pieces of learning data is not limited and takes various values.

  The discriminator is a discriminant function program that classifies input data into several classes. Here, classes are classified into two categories: normal and abnormal. In the present embodiment, the classifier is an image diagnostic classifier, learning data including verification data is used to create the classifier, and test data is used to evaluate the diagnostic performance of the classifier (step 113). An image feature is extracted from learning data, and a discriminator is obtained by setting parameters that can distinguish between normal and abnormal as much as possible. When image data to be diagnosed is input to the discriminator, either normal or abnormal results are output. In creating the discriminator, various parameters (L2 regularization value, learning image size, resolution, etc.) are adjusted. Test data is diagnosed using each of these classifiers, and a classifier having a high correct diagnosis rate and a small standard deviation is set as a temporary classifier.

There are two types of diagnosis, normal or abnormal, and diagnosis results for the number of image data diagnosed are obtained. Since the verification data is image data that is known to be normal or abnormal, it is possible to determine whether the diagnosis result is correct. That is, when the diagnosis result is normal and the test data is normal, and when the diagnosis result is abnormal and the test data is also abnormal, a positive determination is made. On the other hand, when the diagnosis result is normal and the test data is abnormal, and when the diagnosis result is abnormal and the test data is normal, the determination is negative. In this embodiment, the correct diagnosis rate is defined by the following equation (1).
Formula (1) Correct diagnosis rate = number of correct image data / total number of image data

  After the provisional discriminator is created, the test data is diagnosed by this discriminator (step 113), and the correct diagnosis rate is calculated (step 114). Since the test data is also image data that is known to be normal or abnormal, it is possible to determine whether the diagnosis result is correct. The calculation of the correct diagnosis rate is as shown in Equation (1). After calculating the correct diagnosis rate, m pieces of learning data are increased (step 115), n pieces of learning data are selected again (step 111), and steps 112 to 1114 are repeated. That is, a new provisional classifier is repeatedly created, test data is diagnosed for each new classifier, and a correct diagnosis rate is calculated.

  By repeating Steps 111 to 115, a plurality of provisional discriminators are created for each number of learning data, and the calculation result of the correct diagnosis rate by each discriminator is obtained. That is, a correct diagnosis rate is obtained for each number of learning data. The analysis means 6 determines the optimum learning data number N from the viewpoint of a high correct diagnosis rate and a small standard deviation.

  FIG. 8 is a diagram showing the relationship between the number of learning data for group 1 (42 years and older) and the correct diagnosis rate. Taking the portion near 1750 learning data as an example, the lower end 16 shows a value obtained by subtracting the standard deviation value from the average value, the upper end 17 shows a value obtained by adding the standard deviation value to the average value, and a point 15 is Average values are shown. The further the distance between the lower end 16 and the upper end 17, the greater the variation and the greater the standard deviation. In the result of FIG. 8, when the number of pieces of learning data is near 12,000, the correct diagnosis rate is maintained at a high value and the variation is small (see part A), and the high correct check rate is stably maintained. It can be secured. The analysis means 6 determines 12144 as the optimum learning data number N using the average value and standard deviation of the correct diagnosis rate as a scale (step 116).

  After determining the optimum number N of learning data for groups 1 to 5, the analysis means 6 performs arithmetic processing for determining the optimum classifier. FIG. 9 is a flowchart illustrating a process for determining an optimal classifier. The optimum discriminator is determined for each of groups 1-5. Hereinafter, determination of the optimum classifier of group 1 will be described. First, 669 pieces of image data of the age group of group 1 are divided into test data and learning data, and a certain number of pieces of learning data is used as verification data (step 120). This division was random, with 20% of all image data being test data, the remaining 80% being learning data (535), and 20% of learning data being verification data.

  Next, 535 pieces of learning data are amplified to obtain 12144 pieces of image data, which is the optimum number of pieces of learning data for group 1 (step 121). The ratio of normal image data and abnormal image data of 12144 pieces of learning data is not limited and takes various values. A discriminator is created from these 12144 pieces of learning data (including verification data) (step 122). In this step, like the step 112 in FIG. 7, the classifier is learned while adjusting with the verification data, and the best classifier is created from the N pieces of learning data.

  Next, N2 pieces of test data to be diagnosed by the classifier are extracted (step 123). The test data is 135 pieces of image data corresponding to the remaining 20% not used for the learning data among all the data of the group 1. The breakdown of the 135 test data is the number of test data according to the ratio of normal image data to abnormal image data of all image data of group 1. The 135 test data are diagnosed by the discriminator created in step 122 (step 124).

  The diagnosis procedure is the same as step 113 in FIG. 7, and there are two types, normal or abnormal, and 135 diagnosis results are obtained for 135 pieces of test data, and the correct diagnosis rate defined by the above equation (1). The correct diagnosis rate is calculated for the diagnosis results of 135 pieces of test data (step 125).

  After calculating the correct diagnosis rate, change the parameters (L2 regularization value, learning data, etc.) to create a new discriminator, calculate the correct diagnosis rate again, and use the average value and standard deviation of the correct diagnosis rate as a scale. The classifier created in step 121 may be replaced with a classifier having a better result.

  When the calculation of the correct diagnosis rate in step 124 is completed, 12144 image data are newly created from the newly selected 535 image data, and steps 120 to 125 are repeated. can get. Thereafter, by repeating the same procedure, 135 correct diagnosis rates can be obtained for the newly selected learning data. The analysis means 6 employs the discriminator that yielded the highest correct diagnosis rate as the optimum discriminator (step 126). This step corresponds to step 103 in FIG.

  Through the above analysis process, the optimum classifier for the age group of group 1 (42 years old and over) is created. Hereinafter, the analysis processes shown in FIGS. 7 and 9 are sequentially repeated to create the optimum classifiers of the age categories of groups 2 to 5. That is, after analyzing the image data of each group and determining the optimum learning data number N for each group (step 116 in FIG. 7), the optimum discriminator is determined for each group (step 126 in FIG. 9). ).

  By repeating steps 111 to 116 in FIG. 7 for each group, the relationship between the number of pieces of learning data and the correct diagnosis rate is obtained for each group. 10 to 13 show the relationship between the number of learning data in groups 2 to 5 and the correct diagnosis rate. FIG. 10 shows the relationship between the number of learning data and the correct diagnosis rate in the age group of Group 2 (40 to 42 years old). Using this result, the optimum learning data number N is determined. The determination procedure is the same as in the case of group 1 (42 years old and over), and the optimum learning data number N is determined using the average value and standard deviation of the correct diagnosis rate as a scale. In the result of FIG. 10, when the number of pieces of learning data is about 8000, the correct diagnosis rate is maintained at a high value and the variation is small (see part B). The number of data N is determined to be 8085 (step 116).

  The determination of the optimal learning data number N is the same for the other groups 3 to 5 as follows. FIG. 11 shows the relationship between the number of learning data and the correct diagnosis rate in the age group of group 3 (from 38 to under 40). In the result of FIG. 11, the number of learning data is about 8000. In this case, the correct diagnosis rate maintains a high value and the variation is small (see part C), and the analysis means 6 determines 7848 as the optimum learning data number N.

  FIG. 12 shows the relationship between the number of learning data and the correct diagnosis rate in the age group of Group 4 (35 to under 38). In the result of FIG. 12, the number of learning data is around 17,000. At this time, the correct diagnosis rate is maintained at a high value and the variation is small (see part D), and the analyzing means 6 determines 17112 as the optimal number N of data for learning.

  FIG. 13 shows the relationship between the number of learning data and the correct diagnosis rate in the age division of group 5 (under 35 years old). The result of FIG. 13 shows that when the number of learning data is around 16000. The correct diagnosis rate maintained a high value and the variation was small (see part E), and the analysis means 6 determined 16665 as the optimum learning data number N.

  When the optimum learning data number N is determined for each group, the optimum discriminator is determined through the process of FIG. 9, and the five optimum discriminators corresponding to the five age categories of the groups 1 to 5 are determined. Will be decided. Thereafter, the analysis target image data is analyzed. That is, in FIG. 2, after determining the optimum classifier for each age category (step 103), diagnosis target image data is diagnosed by this optimum classifier (step 104).

  Diagnosis using the optimum classifier is performed using an optimum classifier corresponding to the age of the diagnosis target image data. FIG. 14 shows the relationship between the optimum discriminator and diagnostic object image data. The optimum discriminators 1 to 5 correspond to groups 1 to 5 (see FIG. 3). The optimum discriminator 1 diagnoses data over 42 years of diagnosis object image data, and similarly, the optimum discriminators 2 to 5 diagnose the diagnosis object image data of the age classification shown in FIG.

  In order to verify the effects of the optimum discriminators 1 to 5, a comparative analysis was performed. For this comparative analysis, a new comparative classifier was created using the pregnancy prognosis image data (5691 images of fertilized eggs (blastocysts) on day 5 after fertilization) used in the above analysis. did. The creation procedure of the comparison classifier is the same as the creation procedure of the optimum classifiers 1 to 5, but the creation of the optimum classifiers 1 to 5 uses image data classified by age, whereas the classification for comparison is performed. The creation of the vessel is different in that image data that is not classified by age is used.

  FIG. 15 shows a comparison result of the correct diagnosis rate between the optimum classifiers 1 to 5 and the comparison classifier. As described with reference to FIG. 9, the optimum discriminators 1 to 5 are discriminators that have obtained the highest correct diagnosis rate in the derivation process. Similarly, the comparative discriminator also has the highest positive discriminator in the derivation process. This is a classifier with a diagnosis rate. The correct diagnosis rate of the optimum discriminators 1 to 5 shown in FIG. 15 is obtained in the derivation process of the optimum discriminators 1 to 5, and the correct diagnosis rate of the comparison discriminator is obtained in the derivation process of the comparison discriminator. It is obtained. According to FIG. 15, in the optimum discriminators 1 to 4, the correct diagnosis rate exceeds the correct diagnosis rate of the comparison discriminator, and the effect of creating the optimum discriminator according to age could be confirmed. The correct diagnosis rate 0.639 of the optimum discriminator 5 was lower than the correct diagnosis rate 0.721 of the comparative discriminator, but the correct diagnosis rate 0.639 is more than 60% and is sufficiently high. It can be said that it is a rate. For this reason, it can be said that the diagnosis result by the age-specific optimal classifier according to the present embodiment predicts whether or not a normal child can be acquired with high accuracy.

  In the above-described embodiment, the pregnancy prognosis finding image data 10 is classified into normal image data and abnormal image data, and the normal image data is image data that has been found to have resulted in the acquisition of a newborn, and abnormal image data. Was image data that was found to have not resulted in the acquisition of live children. Instead of this, normal image data may be image data that has been found to have become pregnant, and abnormal image data may be image data that has been found not to have become pregnant. In this case, the procedure for creating the optimum classifier for each age is the same as in the above embodiment, and the diagnosis result by the created optimum classifier for each age is a prediction of whether or not pregnancy is possible with high accuracy. it can.

  As described above, the present invention uses a machine learning method, and it is possible to perform subjective judgment by simply executing a computer program on image data of a fertilized egg without using equipment such as photographing and measurement. Diagnosis is possible without doing so. The classifier that diagnoses the data is statistically created by machine learning based on the pregnancy prognosis image data whose pregnancy prognosis is known, and is also created by age, so the diagnosis results are normal. The possibility of acquisition or pregnancy failure is predictable with high accuracy.

  As mentioned above, although one Embodiment of this invention was described, the process for determination of an optimal discriminator may be changed suitably. For example, in the above-described embodiment, an example has been described in which the discriminator is created by learning while adjusting the verification data when creating the discriminator (step 112 in FIG. 7). Only the diagnosis (step 113 in FIG. 7) may be used. Moreover, although the example of the age group of 5 groups was shown in FIG. 3, it is not restricted to this.

  In the above embodiment, the example of receiving the pregnancy prognosis image data 10 via the Internet 8 has been described (see FIG. 1). However, the present invention is not limited to this, and the reception means 2 does not involve the Internet 8. You may receive directly. In addition, the pregnancy prognosis finding image data 10 may be received by the receiving means 2 once stored in the storage means outside the image diagnostic system 1, but the image data can be obtained by connecting the image diagnostic system 1 to a microscope. May be received in real time without being stored, and the result may be displayed on the output unit 5 of the diagnostic imaging system 1 in real time.

  Further, the ratio of the number of normal image data and abnormal image data in n1 pieces of learning data in FIG. 7 (step 111 in FIG. 7) and N1 pieces of learning data in FIG. 9 (step 121 in FIG. 9) is not limited. However, the ratio of the two image data may be determined as an optimal ratio that increases the correct diagnosis rate of the discriminator, and may be 1: 1, 2: 1, 1: 2, or the like. Specifically, before determining the final ratio, the diagnostic imaging system 1 is operated on a trial basis, the ratio is appropriately changed, and the optimal ratio is determined while checking the correct diagnosis rate of the discriminator. May be.

DESCRIPTION OF SYMBOLS 1 Image diagnosis system of fertilized egg 2 Receiving means 3 Transmitting means 4 Input means 5 Output means 6 Analyzing means 7 Storage means

Claims (6)

An image diagnostic system for a fertilized egg using a machine learning technique,
Storage means for storing image data of a fertilized egg whose pregnancy prognosis is known;
Receiving means for capturing the image data of the fertilized egg into the storage means;
Analyzing means for analyzing the image data;
Output means for outputting an analysis result by the analysis means,
The image data is image data composed of normal image data that has been found to have resulted in the acquisition of newborns, and abnormal image data that has been determined to have not resulted in the acquisition of newborns, or pregnancy establishment. It is image data composed of normal image data that has been found to have reached, and abnormal image data that has been found to have not resulted in the establishment of pregnancy, and includes patient age information,
The analysis means includes
The image data is divided into a plurality of groups according to age based on the age information of the patient,
The analysis target group selected from the plurality of groups is classified into test data and learning data,
Creating a plurality of classifiers from the learning data,
Diagnose the test data with each classifier, calculate the correct diagnosis rate of the test data,
Determining an optimum classifier from the plurality of classifiers based on the correct diagnosis rate;
The process from the classification to the determination of the optimum classifier is performed by selecting the other group as an analysis target group, and the process is repeated to determine the optimum classifier for each age. Diagnostic imaging system.   Before determining the optimum classifier, the analysis unit repeats the creation of the classifier and the calculation of the correct diagnosis rate by changing the number of learning data to be selected, and the learning data having different numbers Based on the calculated correct diagnosis rate, the optimum number of learning data is determined, and the creation of the discriminator and the calculation of the correct diagnosis rate are selected with the number of learning data as the optimal number. The optimal discriminator is determined from the plurality of discriminators obtained by the second repetition based on the plurality of correct diagnosis rates obtained by the second iteration by changing the learning data again. The diagnostic imaging system for fertilized eggs according to 1. An image diagnostic program for a fertilized egg for causing a computer to perform image diagnosis of a fertilized egg using a machine learning technique,
A step of taking in image data of a fertilized egg into a storage means;
An analysis step of analyzing the image data;
Causing the computer to execute an output step of outputting an analysis result by the analysis means;
The image data is image data composed of normal image data that has been found to have resulted in the acquisition of newborns, and abnormal image data that has been determined to have not resulted in the acquisition of newborns, or pregnancy establishment. It is image data composed of normal image data that has been found to have reached, and abnormal image data that has been found to have not resulted in the establishment of pregnancy, and includes patient age information,
In the analysis step,
The image data is divided into a plurality of groups according to age based on the age information of the patient,
The analysis target group selected from the plurality of groups is classified into test data and learning data,
Creating a plurality of classifiers from the learning data,
Diagnose the test data with each classifier, calculate the correct diagnosis rate of the test data,
Determining an optimum classifier from the plurality of classifiers based on the correct diagnosis rate;
The process from the classification to the determination of the optimum classifier is performed by selecting the other group as an analysis target group, and the process is repeated to determine the optimum classifier for each age. Diagnostic imaging program.   In the analysis step, the analysis means repeats the creation of the discriminator and the calculation of the correct diagnosis rate by changing the number of learning data to be selected before determining the optimal discriminator. Based on the correct diagnosis rate calculated for the learning data, the optimum number of the learning data is determined, and the number of the learning data is set as the optimal number, the creation of the classifier and the correct diagnosis rate The calculation is repeated again by changing the learning data to be selected, and the optimum classifier is selected from the plurality of classifiers obtained by the second iteration based on the plurality of correct diagnosis rates obtained by the second iteration. The diagnostic imaging program for a fertilized egg according to claim 3, wherein: A method for creating a classifier for image diagnosis of a fertilized egg using a machine learning technique,
A step of taking in image data of a fertilized egg into a storage means;
An analysis step of analyzing the image data;
An output step of outputting an analysis result by the analysis means,
The image data is image data composed of normal image data that has been found to have resulted in the acquisition of newborns, and abnormal image data that has been determined to have not resulted in the acquisition of newborns, or pregnancy establishment. It is image data composed of normal image data that has been found to have reached, and abnormal image data that has been found to have not resulted in the establishment of pregnancy, and includes patient age information,
In the analysis step,
The image data is divided into a plurality of groups according to age based on the age information of the patient,
The analysis target group selected from the plurality of groups is classified into test data and learning data,
Creating a plurality of classifiers from the learning data,
Diagnose the test data with each classifier, calculate the correct diagnosis rate of the test data,
Determining an optimum classifier from the plurality of classifiers based on the correct diagnosis rate;
The process from the classification to the determination of the optimum classifier is performed by selecting the other group as an analysis target group, and the process is repeated to determine the optimum classifier for each age. A method for creating a classifier for diagnostic imaging. In the analysis step, the analysis means repeats the creation of the discriminator and the calculation of the correct diagnosis rate by changing the number of learning data to be selected before determining the optimal discriminator. Based on the correct diagnosis rate calculated for the learning data, the optimum number of the learning data is determined, and the number of the learning data is set as the optimal number, the creation of the classifier and the correct diagnosis rate The calculation is repeated again by changing the learning data to be selected, and the optimum classifier is selected from the plurality of classifiers obtained by the second iteration based on the plurality of correct diagnosis rates obtained by the second iteration. The method for producing a discriminator for image diagnosis of a fertilized egg according to claim 5, wherein:

Images