System, program, and method for determining hypermutated cancer

Abstract

To provide a method, a program, and a method determining hypermutated cancer with higher accuracy than before.SOLUTION: Provided is a system for determining hypermutated cancer comprising, an input unit, a holding unit, a machine learning execution unit and a determining unit, the input unit configured to be capable of inputting a plurality of pieces of first image data, a plurality of pieces of second image data and a plurality of pieces of third image data, the holding unit configured to be capable of holding the first image data and the second image data, the machine learning execution unit configured to be capable of generating a determination model determining whether a cancer is hypermutated or not, using the first image data and the second image data held by the holding unit as training data, and the determining unit configured to be capable of determining whether the third image data represents an image of hypermutated cancer or not, by inputting the third image data into the determination model.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a system, program, and method for discriminating hypermutated cancer. [Background technology]

[0002] By extensively investigating gene mutations in cancer, it has become clear that cancers can be classified according to the pattern of gene mutation. One of the mutation patterns characteristic of such cancers is hypermutation (hypermutation or hypermutated). Hypermutated cancers are distinguished by a higher somatic mutation rate than other types. It is known that some cancers, such as gastric cancer, breast cancer, colon cancer, glioblastoma, and uterine cancer, exhibit the characteristics of hypermutated cancers. Hypermutated cancers often simultaneously have the property of microsatellite instability, which indicates a defect or incompleteness of the mismatch repair mechanism during DNA replication. This is thought to be due to the mutation of the mismatch repair enzymes MLH1, MLH3, MSH2, MSH3, MSH6, and PMS2 genes, and the suppression of expression of the MLH1 gene by methylation. It is also known that mutations in polymerase ε (POLE), a DNA replication enzyme, cause somatic mutations at a particularly high frequency, resulting in hypermutated cancers (Non-Patent Documents 1 and 2).

[0003] Meanwhile, the mechanism of cancer immune escape has been elucidated, and new cancer immunotherapy methods targeting this mechanism have been applied to clinical practice. One of the most notable is the PD-1 (Programmed cell Death-1) / PD-L1 (PD-1 Ligand1) pathway, also known as the immune checkpoint pathway. Blocking the immunosuppressive co-signal PD-1 / PD-L1 pathway releases the immune suppression of T cells, activating T cells and suppressing tumors expressing cancer-specific antigens. In addition, CTLA-4 is also expressed on activated T cells, and when it binds to the CD28 ligand of antigen-presenting cells, it suppresses T cell activation. Therefore, blocking this pathway can also release the immune suppression of T cells and cause tumor suppression. Anticancer drugs that apply this principle have been put to practical use (e.g., nivolumab, ipilimumab).

[0004] Furthermore, there are several other such immunosuppressive mechanisms, and it is expected that antitumor agents that block these immunosuppressive mechanisms will be developed and put to practical use in the future. Hypermutation cancers have many cancer-specific antigens that are targets of the immune system, so it has been shown that therapy that blocks immunosuppressive signal pathways is highly effective, and a method that can easily determine whether a cancer is hypermutation is desired (Non-Patent Document 3).

[0005] Conventionally, a method for testing hypermutation-type cancers has been known in which the number of mutations is counted by comprehensive genetic analysis, but this method has the problem of requiring a lot of time and effort. In addition, a method for testing for defects or incompleteness of the mismatch repair mechanism, which is one of the causes of hypermutation in cancer, by immunohistochemistry or microsatellite instability testing of related genes has also been known, but this method has the problem of not being able to detect all hypermutation-type cancers.

[0006] On the other hand, a pathological diagnosis support program as disclosed in Patent Document 1 is known. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Nat Rev Cancer.2014 December;14(12):786‐800 [Non-Patent Document 2] J Pathol 2013;230:148‐153 [Non-Patent Document 3] Science 03 Apr 2015 Vol. 348, Issue 6230, pp.124-128 [Patent documents]

[0008] [Patent Document 1] Patent Application No. 2004-346911 Summary of the Invention [Problem to be solved by the invention]

[0009] Patent Document 1 claims that it is possible to determine the presence or absence of a tumor and whether it is benign or malignant, but makes no mention of a method for distinguishing hypermutation cancers.

[0010] The present invention has been made in consideration of the above circumstances, and provides a method, program, and method for discriminating hypermutated cancers with higher accuracy than conventional methods. [Means for solving the problem]

[0011] According to the present invention, there is provided a system for discriminating hypermutated cancer, comprising an input unit, a storage unit, a machine learning execution unit, and a discrimination unit, wherein the input unit is configured to be capable of inputting a plurality of first image data, a plurality of second image data, and a plurality of third image data, the first image data being image data representing a stained pathological section of a hypermutated cancer, the second image data being image data representing a pathological section of a cancer that is not a hypermutated cancer and stained in the same manner as the pathological section of the cancer that is the basis of the first image data, and the third image data being image data representing a pathological section of a cancer that is to be newly discriminated as to whether or not it is a hypermutated cancer and stained in the same manner as the pathological section of the cancer that is the basis of the first image data, the storage unit is configured to be capable of holding the first image data and the second image data, the machine learning execution unit is configured to be capable of generating a discrimination model for discriminating whether or not it is a hypermutated cancer using the first image data and the second image data held by the storage unit as teacher data, and the discrimination unit is configured to input the third image data to the discrimination model and discriminate whether or not the third image data is a hypermutated cancer.

[0012] According to the present invention, a discrimination model for discriminating whether or not a cancer is a hypermutated cancer is generated using the first image data and the second image data as teacher data. Here, the first image data is image data representing a stained pathological section of a hypermutated cancer. The second image data is image data representing a pathological section of a cancer that is not a hypermutated cancer and stained in the same manner as the pathological section of the cancer that is the basis of the first image data. Then, the third image data is input to the discrimination model, and the third image data is configured to be able to discriminate whether or not the cancer is a hypermutated cancer. Here, the third image data is image data representing a pathological section of a cancer that is to be newly discriminated whether or not the cancer is a hypermutated cancer and stained in the same manner as the pathological section of the cancer that is the basis of the first image data. This allows for rapid and high-precision discrimination of whether or not a cancer is a hypermutated cancer, which was previously difficult without genetic analysis using a next-generation sequencer or the like, and makes it easy to select a drug that is effective for treatment.

[0013] Various embodiments of the present invention will be described below. The embodiments described below can be combined with each other. Preferably, the method for staining the pathological section is hematoxylin-eosin staining. Preferably, the input unit is configured to further input non-cancer image data, the non-cancer image data being image data that is not a pathological section of a cancer; the storage unit is configured to further store the non-cancer image data; the machine learning execution unit is configured to further generate a discrimination model that uses the non-cancer image data stored by the storage unit as teacher data and determines whether or not the image data is an image data of a pathological section of a cancer; and the discrimination unit is configured to further determine whether or not the third image data is image data of a cancer. Preferably, the image processing unit is configured to perform a Z-value conversion process for converting each RGB color for each pixel of at least one of the first image data, the second image data, and the non-cancer image data into a Z value in the CIE color system based on the color distribution of the first image data, the second image data, or the entire non-cancer image data. Preferably, the image processing unit is configured to be able to execute a division process of dividing at least one of the first image data, the second image data and the non-cancer image data input to the input unit. Preferably, the division process is configured to be capable of executing a division process for dividing image data of the same pathological section on at least one of the first image and the second image data. Preferably, the image processing unit executes the division process so that some areas in the divided images overlap each other. Preferably, the image processing unit is further configured to be capable of executing a division process for dividing third image data input to the input unit. Preferably, the discrimination unit discriminates whether the third image data is image data of a pathological section of cancer, and further discriminates whether the image data discriminated to be a pathological section of cancer is a highly mutated cancer. Preferably, the discrimination unit discriminates whether the cancer is a hypermutated cancer based on the ratio of image data determined to be the hypermutated cancer within image data determined to be image data of a pathological section of the cancer. According to another aspect, there is provided a program for making a computer function as an input unit, a storage unit, a machine learning execution unit, and a discrimination unit, wherein the input unit is configured to be able to input a plurality of first image data and a plurality of second image data, the first image data being image data representing a stained pathological section of a highly mutated cancer, and the second image data being image data representing a pathological section of a cancer that is not a highly mutated cancer and stained in the same manner as the pathological section of the cancer that is the source of the first image data, the storage unit is configured to be able to store the first image data and the second image data, and the machine learning execution unit is configured to use the first image data and the second image data stored by the storage unit as teacher data to generate a discrimination model that discriminates whether or not a cancer is a highly mutated cancer. According to another aspect, there is provided a method for distinguishing hypermutated cancer, the method being carried out using any one of the systems described above. According to another aspect, there is provided a method for distinguishing hypermutated cancer, which is executed using any one of the programs described above. Preferably, the method includes determining the efficacy of the immune checkpoint inhibitor. [Brief description of the drawings]

[0014] [Figure 1] 1 is a functional block diagram of a system 10 according to a first embodiment of the present invention. [Diagram 2] 1 is a flowchart showing a flow for generating a discrimination model for discriminating whether or not a cancer is a hypermutation cancer according to the first embodiment of the present invention. [Diagram 3] 13 is a conceptual diagram showing the flow of third image data when determining whether or not the third image data is a highly mutated cancer using a determination model. FIG. [Figure 4] 1 is a flowchart showing a flow of determining whether or not a cancer is a hypermutated cancer according to a first embodiment of the present invention. [Diagram 5] FIG. 5 is a conceptual diagram for explaining the analysis in S13 of FIG. [Figure 6] FIG. 13 is a conceptual diagram for explaining a weight w in a determination model. [Figure 7] FIG. 3 is a conceptual diagram for explaining the execution of machine learning in S5 of FIG. 2. [Figure 8] FIG. 11 is a functional block diagram of a system 20 according to a second embodiment. [Figure 9] 10A to 10C are diagrams illustrating the division process of an input image in the image processing unit 22. [Figure 10] 10A to 10C are diagrams illustrating the division process of an input image in the image processing unit 22. [Figure 11] 13 is a process flow of a determination process of third image data in this embodiment. [Figure 12] 10 is a diagram illustrating the discrimination process in a discriminator 25. FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described with reference to the drawings. Various characteristic features shown in the following embodiments can be combined with each other.

[0016] <1. First embodiment> 1.1. Identifying whether or not a cancer has a high frequency mutation A system 10 according to one embodiment of the present invention will be described below with reference to FIGS.

[0017] (1.1.1. System 10) As shown in FIG. 1, a system 10 includes an input unit 1, an image processing unit 2, a storage unit 3, a machine learning execution unit 4, and a discrimination unit 5.

[0018] The input unit 1 is configured to be able to input a plurality of first image data, a plurality of second image data, and a plurality of third image data. Here, the first image data is image data representing a stained pathological section of a highly mutated cancer. The second image data is image data representing a pathological section of a cancer that is not a highly mutated cancer, stained in the same manner as the pathological section of the cancer that is the basis of the first image data. Furthermore, the third image data is image data representing a pathological section of a cancer that is to be newly determined as being a highly mutated cancer, stained in the same manner as the pathological section of the cancer that is the basis of the first image data. Here, in this embodiment, the RGB values ​​of these image data can take values ​​of 0 to 255.

[0019] In this embodiment, 17 cases of colon cancer samples determined to be of the hypermutation type (high frequency mutation type) and non-hypermutation type (not of the high frequency mutation type) were obtained by analyzing the cancer genome DNA sequence. Here, these 17 cases were 17 cases that were determined to be of hypermutation as a result of cancer genome sequencing of 201 Japanese colon cancer patients (reference: Nagahashi et al. GenomeMed 2017). Then, the specimens were stained with hematoxylin and eosin to obtain the first image data and the second image data using digital pathology technology. Here, in this embodiment, the first image data and the second image data were stored as digital pathology image data conforming to the MIRAX format. Here, the above conditions are not limited to this, and a configuration may be adopted in which a predetermined number of cases of cancer samples other than colon cancer are obtained.

[0020] In this manner, in this embodiment, hematoxylin-eosin stained image data, which is often used in clinical cases, is used as the first image data and the second image data, making it possible to realize a highly versatile discrimination system.

[0021] However, other staining methods may be used depending on the conditions. Furthermore, the storage format of the image data is not limited to this.

[0022] The image processing unit 2 is configured to be able to execute a division process for dividing a plurality of first image data, second image data, and third image data input to the input unit 1. In this embodiment, the image processing unit 2 has a function for dividing the first image data, second image data, and third image data into predetermined tiles. As an example, the image processing unit 2 divides the first image data, second image data, and third image data into images of 300 pixels x 300 pixels. Although the division size is not particularly limited, it is preferable that the size is such that it is possible to identify whether the image data is a cancer tissue site or not. Then, in this embodiment, the division process divides each of the first image data and the second image data into 1000 or more pieces. Furthermore, in this embodiment, the image processing unit 2 is configured to be able to execute a division process for dividing image data of the same pathological section for at least one of the first image and the second image data. The division size and the number of divisions are not limited to these, and any condition may be adopted.

[0023] In this way, by dividing the image data input to the input unit 1, it is possible to increase the amount of training data used in the subsequent machine learning, thereby improving the accuracy of the machine learning.

[0024] In this embodiment, the image processing unit 2 is further configured to be capable of executing a conversion process for converting each RGB color for each pixel of the divided first image data and second image data into a Z value in the CIE color system based on the color distribution of the entire first image data and second image data. Specifically, since the Z value has a normal distribution centered on 0 and the RGB values ​​of the image data are values ​​from 0 to 255, it is desirable to keep the Z-value-converted value of each RGB color within a range of twice the standard deviation (σ). For this reason, the image processing unit 2 has a function of correcting values ​​of 2σ or more to 2σ and values ​​of -2σ or less to -2σ. The image processing unit 2 also has a function of adding 2 to these values ​​to convert all values ​​to values ​​of 0 or more, and then dividing by 4 to standardize the values ​​to values ​​of 0 to 1. Furthermore, the image processing unit 2 has a function of multiplying the values ​​by 255 to convert them into values ​​of normal color expression. In addition, the image processing unit 2 also performs a process of truncating the decimal point so that the values ​​become integer values. Note that the standardization method is not limited to this.

[0025] Here, if we define "x = int(((min(max(xz,-2),2)+2) / 4) × 255)", then "xz = z-valued RGB value" holds true.

[0026] In this way, by converting each RGB color of the first image data and the second image data into a Z value, it is possible to reduce the variation in color (shade of color) in the dyeing process and to suppress the influence of the degree of dyeing on the subsequent machine learning. As a result, it is possible to improve the accuracy of the machine learning.

[0027] The storage unit 3 is configured to be able to store the first image data and the second image data. Here, the storage unit 3 is configured with any memory, latch, HDD, SSD, or the like.

[0028] The machine learning execution unit 4 is configured to be capable of generating a discrimination model for discriminating whether or not a cancer is a highly mutated cancer, using the first image data and the second image data stored in the storage unit 3 as training data. Details of the discrimination model will be described later with reference to FIG.

[0029] The machine learning algorithm of the machine learning execution unit 4 is not particularly limited, and for example, a neural network or deep learning can be used. Also, for example, a CNN (Convolutional Neural Network) for image recognition called "Inception-v3" developed by Google can be used. Such a CNN can be executed using the "Keras" framework. In order to prevent overlearning, the machine learning itself can be executed using an "Early Stopping" method in which the accuracy of the model being trained is calculated for each epoch using images of a validation set, and the learning is stopped at the epoch in which the fluctuation of the accuracy index has settled. In this embodiment, the machine learning is repeatedly executed for 14 epochs in the learning with Z-value conversion.

[0030] The discrimination unit 5 is configured to input the third image data to the discrimination model and to be able to discriminate whether or not the third image data is a highly mutated cancer.

[0031] (1.1.2. Flowchart) Next, a flow for generating a discrimination model for discriminating whether or not a cancer is a hypermutation cancer according to one embodiment of the present invention will be described with reference to FIG.

[0032] First, in S1, the first image data and the second image data are input to the input unit 1.

[0033] Next, in S2, the image processing unit 2 executes a division process for dividing the first image data and the second image data. In this embodiment, each of the first image data and the second image data is divided into 1000 or more tiles. The number of divisions can be set appropriately, for example, 1000 to 3000, preferably 1000 to 2000, and more preferably 1000 to 1500. Specifically, for example, the number of divisions is 1000, 1100, 1200, 1300, 1400, 1500, 1600, 1700, 1800, 1900, 2000, 2100, 2200, 2300, 2400, 2500, 2600, 2700, 2800, 2900, or 3000, and may be within a range between any two of the numerical values ​​exemplified here.

[0034] At the same time, the image processing unit 2 performs Z-value conversion on the divided first image data and second image data.

[0035] Next, in S3, the Z-valued first image data and second image data are judged as to whether they are hypermutation type cancer tissue sites (hypermutation type) or non-hypermutation type cancer tissue sites (non-hypermutation type), and a label is attached to each image data. For example, the judgment may be made by a pathologist who specializes in cancer, or image data with labels already attached may be obtained from a server. Also, a mark may be attached to a portion of the image data before division that corresponds to the hypermutation type or non-hypermutation type, and if the tile image data after division corresponds to the marked portion, a label may be attached to the data after division.

[0036] Next, in S4, image data for 13 cases to be used for machine learning by the machine learning execution unit 4 is selected from the first image data and second image data for the 17 cases input to the input unit 1. Such selection may be made randomly or may be selected by a pathologist who specializes in cancer. Then, the first image data and second image data with the labels are stored in the storage unit 3. Such first image data and second image data become a "correct answer set" in the machine learning.

[0037] Next, in S5, the machine learning execution unit 4 executes machine learning to generate a discrimination model for discriminating whether or not the image data is a highly mutated cancer, using the first image data and the second image data stored in the storage unit 3 as training data. Specifically, using the first image data and the second image data for the 13 cases labeled in S4, machine learning is executed to discriminate whether or not the image data is a highly mutated cancer.

[0038] Next, in S6, it is determined whether the discrimination accuracy of the discrimination model is equal to or higher than a predetermined accuracy. If the discrimination accuracy of the discrimination model is not equal to or higher than the predetermined accuracy (NO), the process returns to S4 again, and from the first image data and second image data of 17 cases, image data of 13 cases with different combinations is selected, and the process in S5 is executed. On the other hand, if the discrimination accuracy of the discrimination model is equal to or higher than the predetermined accuracy (YES), the discrimination model is adopted, and the process proceeds to S7.

[0039] Finally, in S7, the discrimination unit 5 outputs the determination model determined in S6, and stores it in the holding unit 3 or a storage unit (not shown).

[0040] (1.1.3. Identification of whether or not a cancer has a high frequency mutation) Next, the flow of the third image data when determining whether or not the third image data is a highly mutated cancer using a determination model will be described with reference to FIGS.

[0041] 3, in this embodiment, the third image data input to the input unit 1 is output to the image processing unit 2, and the third image data on which the above-mentioned image processing (segmentation processing and Z-value processing) has been performed is output to the discrimination unit 5. Then, the discrimination unit 5 uses the discrimination model output in S7 of FIG. 2 to discriminate whether or not the third image data is a highly mutated cancer.

[0042] In this way, by performing division processing on the third image data as well, the size of the image data to be discriminated matches the size of the first and second image data, thereby improving the discrimination accuracy in the discrimination unit 5.

[0043] The flow chart at this time is as follows:

[0044] As shown in FIG. 4, first, the third image data is input to the input unit 1 in S11.

[0045] Next, in S12, the image processing unit 2 executes image processing (division processing and Z-value processing).

[0046] Next, in S13, the discrimination unit 5 uses the above-mentioned discrimination model to discriminate whether or not the third image data is a highly mutated cancer.

[0047] Finally, in S14, the discrimination result by the discrimination unit 5 is output. The manner in which the discrimination result is output is not particularly limited, and may be "highly mutated cancer", "not highly mutated cancer", "the probability of being highly mutated cancer is X%", etc.

[0048] (1.1.4. Discrimination by discriminant model) Next, the determination using the determination model in S13 in Fig. 4 will be described with reference to Fig. 5 and Fig. 6. In this embodiment, the machine learning algorithm is not particularly limited, and a neural network or deep learning can be used. In the following, for the sake of simplicity, an example using a neural network will be described.

[0049] As shown in FIG. 5, the neural network (hereinafter, referred to as NN in the drawings) is composed of multiple layers (first layer L1 to third layer L3) and multiple computation nodes N (N11 to N31). Here, Nij represents the j-th computation node N in the i-th layer. In this embodiment, the neural network is constructed with i=3 and j=5. Note that the values ​​of i and j are not limited to these, and can be, for example, i=1 to 100, j=an integer between 1 to 100, or an integer equal to or greater than 100.

[0050] A predetermined weight w is set for each computation node N. For example, when focusing on computation node N23 in the second layer, a weight w is set between computation node N23 and all computation nodes N11 to N15 in the first layer, which is the previous layer, as shown in Fig. 4. The weight w is set to a value between -1 and 1, for example.

[0051] The machine learning execution unit 4 inputs various parameters to the neural network. In this embodiment, the parameters input to the neural network are the Z value of the third image data, the distribution of the Z value of the third image data, the difference between the Z value of the third image data and the Z value of the first image data, the difference between the Z value of the third image data and the Z value of the second image data, and the difference between the Z value of the third image data and the distribution of the Z values ​​of the first image data and the second image data. Here, the Z values ​​of the first to third image data are Z values ​​in pixel units. The distribution of the Z values ​​of the first to third image data is the distribution of the Z values ​​in the image data (300 pixels x 300 pixels). The difference between the distribution of the Z values ​​of the third image data and the first image data and the second image data is the sum of the difference between the distribution of the Z values ​​of the third image data and the distribution of the Z values ​​for each corresponding pixel of the first image data and the second image data, or the difference between the Z values ​​for each corresponding pixel in the image data.

[0052] As described above, each parameter is normalized to a value between 0 and 1 before being input to the neural network. For example, when an input parameter is 0, 0 is input as an input signal. Also, when an input parameter is 1, 1 is input as an input signal.

[0053] Then, the discrimination unit 5 inputs an input signal defined by various parameters to the first layer L1. The input signal is output from the calculation nodes N11 to N15 of the first layer to the calculation nodes N21 to N25 of the second layer L2, respectively. At this time, the values ​​output from the calculation nodes N11 to N15 are multiplied by the weight w set for each calculation node N, and the resultant value is input to the calculation nodes N21 to N25. The calculation nodes N21 to N25 add up the input values, add the bias b shown in FIG. 6 to the sum, and input the sum to the activation function f(). Then, the output value of the activation function f() (the output value from the virtual calculation node N'23 in the example of FIG. 4) is propagated to the calculation node N31, which is the next node. At this time, the value obtained by multiplying the weight w set between the calculation nodes N21 to N25 and the calculation node N31 and the output value is input to the calculation node N31. The calculation node N31 adds up the input values, and outputs the sum as an output signal. At this time, the calculation node N31 may add up the input values, input the sum plus a bias to the activation function f(), and output the output value as an output signal. Here, in this embodiment, the value of the output signal is adjusted to be a value between 0 and 1. Then, the machine learning execution unit 4 outputs a value corresponding to the value of the output signal as the probability of determining whether or not it is a highly mutated cancer.

[0054] As described above, the system 10 of this embodiment uses the first image data and the second image data as teacher data, and executes machine learning by the machine learning execution unit 4 to generate a discrimination model (neural network and weight w) for discriminating whether or not the cancer is a hypermutation cancer. Then, the discrimination unit 5 discriminates whether or not the third image data is a hypermutation cancer using the discrimination model.

[0055] 1.2.Generating a discriminant model Next, the generation of the discrimination model in S5 to S6 in FIG. 2 will be described with reference to FIG.

[0056] As shown in FIG. 7, the machine learning execution unit 4 sets a weight w, for example, from -1 to 1, for each calculation node N constituting a neural network having the same configuration as the neural network shown in FIG. 5. At this time, in order to reduce the influence of the weight w, it is preferable that the absolute value of the weight w set initially is small. Then, five types of parameter sets are input to the neural network. In this embodiment, the Z value of the first image data, the Z value of the second image data, the distribution of the Z value of the first image data, the distribution of the Z value of the second image data, and the difference between the Z values ​​of the first image data and the second image data are used as parameters to be input to the neural network. Here, the Z value of the first image data and the Z value of the second image data are Z values ​​in pixel units. In addition, the distribution of the Z value of the first image data and the distribution of the Z value of the second image data are the distribution of the Z value within the image data (300 pixels x 300 pixels). In addition, the difference between the Z values ​​of the first image data and the second image data is the sum of the difference between the Z values ​​of the corresponding pixels of the first image data and the second image data, or the difference between the Z values ​​of the corresponding pixels within the image data.

[0057] Then, the output signal from the neural network is compared with the teacher data (determination by a specialist), and if the difference between the output signal and the teacher data (hereinafter referred to as the error) is equal to or greater than a predetermined threshold, the weight w is changed, and the five types of parameter sets are input again to the neural network. At this time, the weight w is changed by a known error propagation method or the like. By repeatedly executing such calculations (machine learning), the error between the output signal from the neural network and the previously given teacher data is minimized. At this time, the number of times of learning in the machine learning is not particularly limited, and can be, for example, 1000 to 20000 times. In addition, even if the error between the actual output signal and the previously given teacher data is not minimized, the machine learning may be terminated when the error becomes equal to or less than a predetermined threshold or at any timing of the developer.

[0058] Then, when the machine learning by the machine learning execution unit 4 is completed, the machine learning execution unit 4 sets the weight of each calculation node N at this time in the neural network. That is, in this embodiment, the weight w is stored in a storage unit such as a memory provided on the neural network. Then, the weight w set by the machine learning execution unit 4 is transmitted to a storage unit (not shown) provided in the system 10, and is set as the weight w of each calculation node N of the neural network of FIG. 5. In this embodiment, the weight w is stored in a storage unit such as a memory provided on the neural network of FIG. 5. Here, by making the configuration of the neural network of FIG. 7 the same as the configuration of the neural network of FIG. 5, it becomes possible to use the weight w set by the machine learning execution unit 4 as it is.

[0059] <2. Second embodiment> A second embodiment of the present invention will be described with reference to Figures 8 to 12. Note that the description of the same configurations and functions as those of the first embodiment will not be repeated.

[0060] As shown in Fig. 8, in the system 20 according to the second embodiment, the input unit 21 is configured to be able to input non-cancer image data in addition to the first image data and the second image data. Here, the non-cancer image data means image data other than the pathological slice of cancer. The image processing unit 22 performs a division process on the input image data. The details of the division process will be described later.

[0061] The storage unit 23 is configured to store the divided non-cancer image data in addition to the divided first image data and second image data. The machine learning execution unit 24 is configured to generate a discrimination model (hereinafter referred to as a first discrimination model) that discriminates whether or not an image of cancer is a cancer image and a discrimination model (hereinafter referred to as a second discrimination model) that discriminates whether or not an image of cancer is a highly mutated cancer, using the first image data, second image data, and non-cancer image data stored by the storage unit 3 as teacher data. The discrimination unit 25 is configured to input the third image data to the first and second discrimination models, and to discriminate whether or not the third image data is image data of cancer and whether or not it is image data of a highly mutated cancer.

[0062] 9 shows an example of image data P input to the input unit 21. The image data P has a tissue region T and a blank region BL (e.g., a preparation region). The tissue region T includes a region C1 of a cancer that is not a hypermutation type cancer, a region C2 of a hypermutation type cancer, and a tissue region NC that is not a cancer.

[0063] The image processing unit 22 performs a division process on the image data P input to the input unit 21. In the example shown in FIG. 9, the tissue region T is divided into 100 tiles, 10 vertically and 10 horizontally. That is, a tile D consisting of 100 tiles is formed so as to encompass the tissue region T. 00 ~D 99 Set.

[0064] In this example, a tile corresponding to the high-mutation cancer region C2 (e.g., tile D54) corresponds to the first image data, and a tile corresponding to the non-high-mutation cancer region C1 (e.g., tile D34) corresponds to the second image data. In addition, a tile corresponding only to the non-cancerous tissue region NC (e.g., tile D15), a tile corresponding only to the blank region BL (e.g., tile D49), and a tile including the non-cancerous tissue region NC and the blank region BL (e.g., tile D04) all correspond to non-cancerous image data.

[0065] In this manner, in the present embodiment, various images are input as non-cancer image data, such as tiles corresponding to non-cancerous tissue regions NC, tiles corresponding only to blank regions BL, tiles including non-cancerous tissue regions NC and blank regions BL, and machine learning is performed. Increasing the diversity of non-cancerous images in this manner improves the accuracy of determining whether the data to be examined is a cancer image.

[0066] In this embodiment, further division processing (hereinafter, referred to as second division processing) can be performed on the image data after the above division processing (hereinafter, referred to as first division processing). In FIG. 10, the tile Dnm after division by the first division processing is further divided into five tiles. Here, in the second division processing, the division processing is performed so that the tiles after division have some overlapping areas. That is, after the second division processing, the tiles Dnm1 and Dnm2 have some overlapping images. Also, the tiles Dnm2 and Dnm3 have some overlapping images.

[0067] In this way, by performing the division process so that some areas of the divided images overlap, it is possible to increase the number of images, thereby improving the learning efficiency in the subsequent machine learning.

[0068] Fig. 11 is a process flow of the discrimination process of the third image data in this embodiment. As shown in Fig. 11, in this embodiment, the discrimination unit 25 discriminates whether the third image data is a cancer image or not, and discriminates whether the third image data is a highly mutated cancer or not.

[0069] Specifically, in step S231 in step S23, the discrimination unit 25 discriminates whether the third image data is a cancer image or not. If the third image data is not a cancer image (No in step S231), in step S233, the discrimination unit 25 discriminates that the third image data is a non-cancerous image.

[0070] On the other hand, if it is a cancer image (Yes in step S231), the discrimination unit 25 discriminates whether or not the third image data is an image of a hypermutated cancer in step S232. If it is not a hypermutated cancer (No in step S232), the third image data is discriminated not to be an image of a hypermutated cancer in step S235. On the other hand, if it is a hypermutated cancer (Yes in step S232), the third image data is discriminated to be an image of a hypermutated cancer in step S234.

[0071] In this manner, in the present embodiment, the third image data is judged as to whether it is a cancer image and whether it is a highly mutated cancer. Therefore, it is not necessary for a pathologist or the like to diagnose in advance whether it is cancer image data, and the work efficiency in the judgment process can be improved.

[0072] Here, the discrimination unit 25 may discriminate whether or not the cancer is a hypermutation cancer based on the ratio of image data discriminated as hypermutation cancer within the image data discriminated as image data of cancer.

[0073] 12, in the third image data P2, an image E2 determined to be a highly mutated cancer is present in an image E1 determined to be image data of a cancer. In this case, the discrimination unit 25 determines that the region indicated by the image E1 is a highly mutated cancer when a ratio determined by (the number of tiles in E2) / (the number of tiles in E1) is greater than a predetermined threshold value.

[0074] In this way, it is possible to remove false positives that would lead to a localized identification of a highly mutated cancer as noise, thereby improving the accuracy of identification.

[0075] As described above, in the second embodiment, the input unit 21 is configured to further input non-cancer image data, and the machine learning execution unit 24 is configured to further generate a discrimination model that uses the non-cancer image data as training data and discriminates whether the image data is image data of a pathological section of cancer. The discrimination unit 25 is further configured to discriminate whether the third image data is image data of cancer. With this configuration, it is not necessary for a pathologist or the like to diagnose whether the third image data is cancer, and the work efficiency of the discrimination process is improved.

[0076] <3. Other embodiments> Although various embodiments have been described above, the present invention can also be carried out in the following aspects.

[0077] Computer, It functions as an input unit, a storage unit, a machine learning execution unit, and an analysis unit, The input unit is configured to be able to input a plurality of first image data and a plurality of second image data, the first image data is image data representing a stained pathology section of the hypermutated cancer; The second image data is image data representing a pathological section that is not a hypermutated cancer and that has been stained in the same manner as the pathological section of the cancer on which the first image data is based; The storage unit is configured to store first image data and second image data, The machine learning execution unit is configured to generate a discrimination model for discriminating whether or not the cancer is a highly mutated cancer by using the first image data and the second image data stored in the storage unit as teacher data. program.

[0078] A method for identifying hypermutated cancer, which is carried out using any one of the systems described above. The hypermutated cancers referred to here include any type of cancer, including, but not limited to, solid cancers such as brain tumors, head and neck cancer, breast cancer, lung cancer, esophageal cancer, gastric cancer, duodenal cancer, appendix cancer, colon cancer, rectal cancer, liver cancer, pancreatic cancer, gallbladder cancer, bile duct cancer, anal cancer, kidney cancer, ureter cancer, bladder cancer, prostate cancer, penile cancer, testicular cancer, uterine cancer, ovarian cancer, vulvar cancer, vaginal cancer, and skin cancer. For the purposes of the present invention, the hypermutated cancers are preferably colon cancer, lung cancer, gastric cancer, melanoma (malignant melanoma), head and neck cancer, and esophageal cancer.

[0079] A method for discriminating between hypermutated cancers, the method being carried out using the above-mentioned program.

[0080] The method for determining whether or not a cancer is hypermutated, comprising a step of determining the efficacy of an immune checkpoint inhibitor. The method for determining whether or not a cancer is hypermutated, may further comprise a step of indicating that administration of an immune checkpoint inhibitor is highly effective for a patient determined to have hypermutated cancer. Since hypermutated cancer has many cancer-specific antigens that are targets of the immune mechanism, it has been shown that therapy that blocks immunosuppressive signal pathways is highly effective. The method for determining whether or not a cancer is hypermutated is advantageous because it can be easily determined that the cancer is hypermutated. The term "immune checkpoint" as used herein is known in the art (Naidoo et al. British Journal of Cancer (2014) 111, 2214-2219), and examples of such immune checkpoints include CTLA4, PD1, and its ligand PDL-1. Other examples include TIM-3, KIR, LAG-3, VISTA, and BTLA. Immune checkpoint inhibitors inhibit their normal immune functions. For example, they inhibit the expression of immune checkpoint molecules by negatively regulating them, or by binding to the molecules and blocking normal receptor / ligand interactions. Immune checkpoints act to brake the immune system response to antigens, and their inhibitors reduce this immunosuppressive effect and enhance the immune response. Immune checkpoint inhibitors are known in the art, and preferred ones are anti-immune checkpoint antibodies such as anti-CTLA-4 antibodies (e.g., ipilimumab, tremelimumab), anti-PD-1 antibodies (e.g., nivolumab, pembrolizumab, pidilizumab, and RG7446 (Roche)), and anti-PDL-1 antibodies (e.g., BMS-936559 (Bristol-Myers Squibb), MPDL3280A (Genentech), MSB0010718C (EMD-Serono), and MEDI4736 (AstraZeneca)).

[0081] The storage unit 3 may be provided in an information processing device such as an external PC or server in the form of cloud computing. In this case, the external information processing device transmits data required for each calculation to the system 10.

[0082] The above-mentioned program may also be provided as a non-transitory computer-readable recording medium storing the program, or as an ASIC (application specific integrated circuit), FPGA (field-programmable gate array), or DRP (Dynamic ReConfigurable Processor) that implements the functions of the above-mentioned program. [Explanation of symbols]

[0083] 1,21: Input section 2,22: Image processing section 3,23: Holding part 4,24: Machine Learning Execution Department 5,25:Discrimination part 10,20: System

Claims

1. An input unit, a storage unit, a machine learning execution unit, and a discrimination unit are provided, the input unit is configured to be able to input a plurality of first image data, a plurality of second image data, and a plurality of third image data; the first image data is image data representing a stained pathology section of the hypermutated cancer; The second image data is image data representing a pathological section of a cancer that is not a hypermutated cancer and that has been stained in the same manner as the pathological section of the cancer on which the first image data is based; The third image data is a pathological section of a cancer to be newly determined as being a hypermutation cancer, and is image data representing a pathological section stained in the same manner as the pathological section of the cancer on which the first image data is based; The storage unit is configured to be able to store first image data and second image data, The machine learning execution unit is configured to generate a discrimination model for discriminating whether or not the cancer is a highly mutated cancer using the first image data and the second image data stored in the storage unit as teacher data, The discrimination unit is configured to input third image data to the discrimination model and to be able to discriminate whether the third image data is a highly mutated cancer or not. A system for identifying highly mutated cancers.

2. The staining method of the pathological section is hematoxylin-eosin staining. The system of claim 1 .

3. The input unit is configured to further input non-cancer image data, The non-cancer image data is image data that is not a pathological section of a cancer, The storage unit is configured to further store the non-cancer image data, The machine learning execution unit is configured to further generate a discrimination model that uses the non-cancer image data stored by the storage unit as teacher data and determines whether the image data is a pathological section of a cancer; The discrimination unit is configured to further discriminate whether the third image data is image data of a cancer. A system according to claim 1 or claim 2.

4. An image processing unit is provided, The image processing unit is configured to be capable of performing a Z-value conversion process for converting each RGB color of each pixel of at least one of the first image data, the second image data, and the non-cancer image data into a Z value in the CIE color system based on a color distribution of the first image data, the second image data, or the entire non-cancer image data. The system according to claims 1 to 3.

5. The image processing unit includes: A division process is configured to divide at least one of the first image data, the second image data, and the non-cancer image data input to the input unit. The system of claim 4.

6. The image processing unit includes: The system of claim 5 , wherein the segmentation process is performed such that some regions in the segmented image overlap.

7. The image processing unit includes: The system according to claim 5 or 6, further configured to be capable of executing a division process for dividing the third image data input to the input unit.

8. The discrimination unit is determining whether the third image data is image data of a pathological section of a cancer; The image data that has been determined to be a pathological section of cancer is further determined to be a highly mutated cancer or not. A system according to any one of claims 3 to 7.

9. The discrimination unit is determining whether the cancer is a hypermutation cancer based on a ratio of image data determined to be the hypermutation cancer in image data determined to be image data of a pathological section of the cancer; The system of claim 8.

10. Computer, The input unit, the storage unit, the machine learning execution unit, and the discrimination unit function as the input unit, The input unit is configured to be able to input a plurality of first image data and a plurality of second image data, the first image data is image data representing a stained pathology section of the hypermutated cancer; The second image data is image data representing a pathological section of a cancer that is not a hypermutated cancer and that has been stained in the same manner as the pathological section of the cancer on which the first image data is based; The storage unit is configured to be able to store first image data and second image data, The machine learning execution unit is configured to generate a discrimination model for discriminating whether or not the cancer is a highly mutated cancer by using the first image data and the second image data stored in the storage unit as teacher data. program.

11. Implemented using the system according to any one of claims 1 to 9, A method for identifying hypermutated cancers.

12. 11. A method for executing a program according to claim 10, A method for identifying hypermutated cancers.

13. determining the efficacy of an immune checkpoint inhibitor; The method according to claim 11 or 12.

Images