Classification device, learning device, classification method, learning method, and program

Abstract

In order to improve the accuracy of benignancy / malignancy classification of a specimen cell in pathological diagnosis, this classification device (1) comprises an acquisition unit (11) which acquires a pathological image, and a classification unit (12) which uses a classification model to classify whether the specimen cell is a benign cell or a malignant cell, the classification model having inputs of: a feature amount of a first weighted input image processed with first weighted information for emphasizing a first area of interest; and a feature amount of a second weighted input image processed with second weighted information for emphasizing a second area of interest that is different from the first area of interest.

Description

Technical Field

[0007] ,

[0001] The present invention relates to a classification device, a learning device, a classification method, a learning method, and a program.

Background Art

[0002] Techniques related to pathological diagnosis using a learned model for image recognition of a pathological image with the pathological image as an input have been disclosed.

[0003] Patent Document 1 describes a configuration that outputs a classification result based on a feature amount independent of the type of organ and a detection result of a special region based on a feature amount specialized for the organ to be inspected, based on an inspection target image obtained by imaging an organ.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] In the diagnosis of classifying whether the specimen cells included as a subject in a pathological image are benign cells or malignant cells, there is also a need for a technique that uses a learned model for image recognition of the pathological image and improves the accuracy of classifying the benignity and malignancy of specimen cells in pathological diagnosis.

[0006] [[ID= forty-three]]One aspect of the present invention has been made in view of the above problems, and an example of the object is to provide a technique for improving the accuracy of classifying the benignity and malignancy of specimen cells in pathological diagnosis.

Means for Solving the Problems

[0007] A classification device according to one aspect of the present invention includes an acquisition unit that acquires a pathological image including specimen cells as a subject, Using the pathological image as input, a first weighted input image is generated by processing the pathological image with the first weighted information using a first generative model trained to generate first weighted information for highlighting a first region of interest in the pathological image; using the pathological image as input, a second weighted input image is generated by processing the pathological image with the second weighted information using a second generative model trained to generate second weighted information for highlighting a second region of interest different from the first region of interest in the pathological image; the first weighted input image is input, and the first weighted input image The system includes a classification means that generates features of a first weighted input image using a first feature analysis model trained to generate features of a second weighted input image, receives the second weighted input image as input, generates features of the second weighted input image using a second feature analysis model trained to generate features of the second weighted input image, receives the features of the first weighted input image and the features of the second weighted input image as input, and classifies whether the sample cells are benign or malignant using a classification model trained to classify whether the sample cells are benign or malignant.

[0008] A learning device according to one aspect of the present invention includes an acquisition means for acquiring a pathological image containing sample cells as subjects, and classification information indicating whether the sample cells are benign or malignant; a first generation model that takes the pathological image as input and generates first weighting information for emphasizing a first region of interest in the pathological image, thereby generating a first weighted input image in which the pathological image has been processed by the first weighting information; a second generation model that takes the pathological image as input and generates second weighting information for emphasizing a second region of interest different from the first region of interest in the pathological image, thereby generating a second weighted input image in which the pathological image has been processed by the second weighting information; and the first weighted input image is input, and the features of the first weighted input image are... The system includes a learning means that generates features of the first weighted input image using a first feature analysis model for generating quantities, receives a second weighted input image as input, generates features of the second weighted input image using a second feature analysis model for generating features of the second weighted input image, receives the features of the first weighted input image and the features of the second weighted input image as input, classifies whether the sample cells are benign or malignant using a classification model for classifying whether the sample cells are benign or malignant, and updates the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information.

[0009] A classification method according to one aspect of the present invention involves a classification device acquiring a pathological image containing specimen cells as the subject, generating a first weighted input image in which the pathological image is processed by the first weighted information using a first generative model trained to generate first weighted information for highlighting a first region of interest in the pathological image, and generating a second weighted input image in which the pathological image is processed by the second weighted information using a second generative model trained to generate second weighted information for highlighting a second region of interest different from the first region of interest in the pathological image, and the first weighted... The process includes: receiving an input image and generating features of the first weighted input image using a first feature analysis model trained to generate features of the first weighted input image; receiving a second weighted input image and generating features of the second weighted input image using a second feature analysis model trained to generate features of the second weighted input image; and receiving the features of the first weighted input image and the features of the second weighted input image and classifying whether the sample cells are benign or malignant using a classification model trained to classify whether the sample cells are benign or malignant.

[0010] A learning method according to one aspect of the present invention involves a learning device acquiring a pathological image containing sample cells as the subject, and classification information indicating whether the sample cells are benign or malignant; using a first generative model that takes the pathological image as input and generates first weighting information to emphasize a first region of interest in the pathological image, it generates a first weighted input image in which the pathological image has been processed with the first weighting information; using a second generative model that takes the pathological image as input and generates second weighting information to emphasize a second region of interest different from the first region of interest in the pathological image, it generates a second weighted input image in which the pathological image has been processed with the second weighting information; and the first weighted input image is input, and the first weighted input image The process includes generating features of a first weighted input image using a first feature analysis model that generates features of a first weighted input image, receiving a second weighted input image as input, generating features of the second weighted input image using a second feature analysis model that generates features of the second weighted input image, receiving features of the first weighted input image and features of the second weighted input image as input, classifying whether the sample cells are benign or malignant using a classification model that classifies whether the sample cells are benign or malignant, and updating the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information.

[0011] A program relating to one aspect of the present invention is a program that causes a computer to function as a classification device, wherein the program includes an acquisition means for acquiring a pathological image containing specimen cells as the subject, and a first generation model trained to generate first weighting information for highlighting a first region of interest in the pathological image, using the pathological image as input to generate a first weighted input image in which the pathological image has been processed by the first weighting information, and a second generation model trained to generate second weighting information for highlighting a second region of interest different from the first region of interest in the pathological image, using the pathological image as input to generate a second weighted input image in which the pathological image has been processed by the second weighting information The system functions as a classification means that generates an image, takes the first weighted input image as input, generates the features of the first weighted input image using a first feature analysis model trained to generate features of the first weighted input image, takes the second weighted input image as input, generates the features of the second weighted input image using a second feature analysis model trained to generate features of the second weighted input image, and takes the features of the first weighted input image and the features of the second weighted input image as input, and uses a classification model trained to classify whether the sample cells are benign or malignant to classify whether the sample cells are benign or malignant.

[0012] A program relating to one aspect of the present invention is a program that causes a computer to function as a learning device, wherein the program uses an acquisition means to acquire a pathological image including specimen cells as subjects, and classification information indicating whether the specimen cells are benign or malignant cells, and a first generation model that takes the pathological image as input and generates first weighting information for emphasizing a first region of interest in the pathological image to generate a first weighted input image in which the pathological image has been processed by the first weighting information, and a second generation model that takes the pathological image as input and generates second weighting information for emphasizing a second region of interest different from the first region of interest in the pathological image to generate a second weighted input image in which the pathological image has been processed by the second weighting information, and the first weighted input image is input The system functions as a learning means that generates features of the first weighted input image using a first feature analysis model that generates features of the first weighted input image, receives a second weighted input image as input and generates features of the second weighted input image using a second feature analysis model that generates features of the second weighted input image, receives the features of the first weighted input image and the features of the second weighted input image as input and uses a classification model that classifies whether the sample cells are benign or malignant to classify whether the sample cells are benign or malignant, and updates the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information. [Effects of the Invention]

[0013] According to one aspect of the present invention, the accuracy of classifying specimen cells as benign or malignant in pathological diagnosis can be improved. [Brief explanation of the drawing]

[0014] [Figure 1] This is a block diagram showing the configuration of a classification device according to exemplary embodiment 1 of the present invention. [Figure 2] This is a flowchart showing the flow of the classification method according to Exemplary Embodiment 1 of the present invention. [Figure 3] This is a block diagram showing the configuration of the learning device according to Exemplary Embodiment 1 of the present invention. [Figure 4] This is a flowchart showing the flow of the learning method according to Exemplary Embodiment 1 of the present invention. [Figure 5] This is a block diagram showing the configuration of the classification device according to Exemplary Embodiment 2 of the present invention. [Figure 6] This is a schematic diagram showing the flow of the process executed by the classification unit according to Exemplary Embodiment 2 of the present invention. [Figure 7] This is a block diagram showing an example of the hardware configuration of the classification device and the learning device according to each exemplary embodiment of the present invention.

Mode for Carrying Out the Invention

[0015] 〔Exemplary Embodiment 1〕 The first exemplary embodiment of the present invention will be described in detail with reference to the drawings. This exemplary embodiment is a basic form of the exemplary embodiments described later.

[0016] (Overview of Classification Device 1) The classification device 1 according to this exemplary embodiment is a device that classifies whether a specimen cell is a benign cell or a malignant cell.

[0017] As an example, the classification device 1 first uses a learned generation model to generate weight information for emphasizing a target region in a pathological image with a pathological image including a specimen cell as a subject as an input, and performs a weighting process on the pathological image using the weight information to generate an image (hereinafter referred to as "input image processed by the weight information of the pathological image").

[0018] Next, the classification device 1 uses a learned feature analysis model to generate feature amounts of the weighted input image when the weighted input image is input, and generates feature amounts of the weighted input image.

[0019] Then, the classification device 1 receives the feature amounts of the weighted input images, and classifies whether the specimen cells are benign cells or malignant cells using a learned classification model for classifying whether the specimen cells are benign cells or malignant cells. As an example, the classification device 1 can be used for cytodiagnosis in intraoperative rapid diagnosis (ROSE: Rapid On-Site Evaluation).

[0020] Here, as an example of the weighting information, an attention map indicating a region of interest by setting weights for each region can be mentioned. Below, the case where the weighting information is an attention map will be described.

[0021] Also, the number of generation models and feature analysis models is not particularly limited, but below, as an example, the case of using two generation models (hereinafter, the first generation model and the second generation model) and feature analysis models (the first feature analysis model and the second feature analysis model) will be described.

[0022] Also, below, the attention ups generated by the first generation model and the second generation model are respectively referred to as the first attention map and the second attention map. Also, the input images processed by the first attention map and the second attention map are respectively referred to as the first weighted input image and the second weighted input image. Also, the regions of interest in the first weighted input image and the second weighted input image, that is, the regions emphasized in the attention map, are respectively referred to as the first region of interest and the second region of interest.

[0023] The specific configurations of the generation model, the feature analysis model, and the classification model do not limit this exemplary embodiment, but as an example, a deep neural network having a convolutional layer such as a CNN (Convolution Neural Network) can be used. As the convolutional layer, although not limited thereto, those having the following configuration may be used. • Batch normalization layer • Activation layer (e.g., ReLU (Rectified Linear Unit) layer) 2D convolutional layer • Pooling layer (Configuration of Classification Device 1) The configuration of the classification device 1 according to this exemplary embodiment will be described with reference to Figure 1. Figure 1 is a block diagram showing the configuration of the classification device 1 according to this exemplary embodiment.

[0024] As shown in Figure 1, the classification device 1 comprises an acquisition unit 11 and a classification unit 12. The acquisition unit 11 and the classification unit 12 are configured to implement an acquisition means and a classification means, respectively, in this exemplary embodiment.

[0025] The acquisition unit 11 acquires pathological images that include sample cells as subjects. The acquisition unit 11 supplies the acquired pathological images to the classification unit 12.

[0026] The classification unit 12 uses a first generative model to generate a first weighted input image in which the pathological image has been processed by a first attention map.

[0027] Furthermore, the classification unit 12 uses a second generative model to generate a second weighted input image in which the pathological image has been processed by a second attention map.

[0028] Furthermore, the classification unit 12 generates the first weighted input image features using the first feature analysis model.

[0029] Furthermore, the classification unit 12 generates a second weighted set of features for the input image using a second feature analysis model.

[0030] Furthermore, the classification unit 12 uses a classification model to classify whether the sample cells are benign or malignant.

[0031] As described above, the classification device 1 according to this exemplary embodiment includes an acquisition unit 11 that acquires a pathological image containing specimen cells as the subject, a first weighted input image in which the pathological image is processed by the first attention map using a first generative model trained to generate a first attention map for highlighting a first region of interest in the pathological image, and a second weighted input image in which the pathological image is processed by the second attention map using a second generative model trained to generate a second attention map for highlighting a second region of interest different from the first region of interest in the pathological image, and the first weighted The system employs a configuration comprising: a classification unit 12 that receives a deleted input image and generates features for a first weighted input image using a first feature analysis model trained to generate features for a first weighted input image; a classification unit 12 that receives features for a second weighted input image and generates features for a second weighted input image using a second feature analysis model trained to generate features for a second weighted input image; and a classification unit 12 that receives features for a first weighted input image and features for a second weighted input image and classifies whether the sample cells are benign or malignant using a classification model trained to classify whether the sample cells are benign or malignant.

[0032] Therefore, the classification device 1 according to this exemplary embodiment classifies whether the sample cells are benign or malignant by focusing on a first region of interest and a second region of interest, which are different regions in the pathological image. For example, in pathological diagnosis to classify whether the sample cells are benign or malignant, the classification device 1 according to this exemplary embodiment can differentiate the region of interest depending on the type of cells contained in the pathological image. Thus, the classification device 1 according to this exemplary embodiment has the effect of improving the accuracy of classifying sample cells as benign or malignant in pathological diagnosis.

[0033] (Process of classification method S1) The flow of the classification method S1 according to this exemplary embodiment will be explained with reference to Figure 2. Figure 2 is a flowchart showing the flow of the classification method S1 according to this exemplary embodiment.

[0034] (Step S11) In step S11, the acquisition unit 11 acquires a pathological image that includes the sample cells as the subject. The acquisition unit 11 supplies the acquired pathological image to the classification unit 12.

[0035] (Step S12) In step S12, the classification unit 12 generates a first weighted input image in which the pathological image has been processed by the first attention map, using the first generation model.

[0036] Furthermore, the classification unit 12 uses a second generative model to generate a second weighted input image in which the pathological image has been processed by a second attention map.

[0037] Furthermore, the classification unit 12 generates the first weighted input image features using the first feature analysis model.

[0038] Furthermore, the classification unit 12 generates a second weighted set of features for the input image using a second feature analysis model.

[0039] Furthermore, the classification unit 12 uses a classification model to classify whether the sample cells are benign or malignant.

[0040] As described above, in the classification method S1 according to this exemplary embodiment, a pathological image containing specimen cells as the subject is acquired, a first weighted input image is generated in which the pathological image is processed by the first attention map using a first generative model trained to generate a first attention map for highlighting a first region of interest in the pathological image, and a second weighted image is generated in which the pathological image is processed by the second attention map using a second generative model trained to generate a second attention map for highlighting a second region of interest different from the first region of interest in the pathological image, and the first weighted The system employs a configuration that includes the following steps: an erased input image is input, and a first feature analysis model trained to generate features of the first weighted input image is used to generate features of the first weighted input image; a second weighted input image is input, and a second feature analysis model trained to generate features of the second weighted input image is used to generate features of the second weighted input image; and the features of the first weighted input image and the second weighted input image are input, and a classification model trained to classify whether the sample cells are benign or malignant is used to classify whether the sample cells are benign or malignant. Therefore, according to the classification method S1 of this exemplary embodiment, the same effects as the classification device 1 described above can be obtained.

[0041] (Overview of Learning Device 2) The learning device 2 according to this exemplary embodiment is a device for updating the parameters of a first generative model, a second generative model, a first feature analysis model, a second feature analysis model, and a classification model.

[0042] As an example, the learning device 2 acquires pathological images containing sample cells as subjects, and classification information (correct labels) indicating whether the sample cells are benign or malignant. Based on a comparison of the classification results performed by the classification model with the classification information, it updates the parameters of each model.

[0043] The first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model are as described above.

[0044] (Configuration of learning device 2) The configuration of the learning device 2 according to this exemplary embodiment will be described with reference to Figure 3. Figure 3 is a block diagram showing the configuration of the learning device 2 according to this exemplary embodiment.

[0045] As shown in Figure 3, the learning device 2 comprises an acquisition unit 21 and a learning unit 22. The acquisition unit 21 and the learning unit 22 are configured to implement an acquisition means and a learning means, respectively, in this exemplary embodiment.

[0046] The acquisition unit 21 acquires pathological images that include the sample cells as subjects, and classification information indicating whether the sample cells are benign or malignant. The acquisition unit 11 supplies the acquired pathological images and classification information to the learning unit 22.

[0047] The learning unit 22 trains the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model by updating the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification results by the classification model with the classification information. The learning device 2 can be used, for example, to train a classification model used in cytology in rapid on-site evaluation (ROSE).

[0048] As described above, the learning device 2 according to this exemplary embodiment includes an acquisition unit 21 that acquires a pathological image containing specimen cells as the subject, and classification information indicating whether the specimen cells are benign or malignant cells; a first generation model that takes the pathological image as input and generates first weighting information to emphasize a first region of interest in the pathological image, thereby generating a first weighted input image in which the pathological image has been processed by the first weighting information; a second generation model that takes the pathological image as input and generates second weighting information to emphasize a second region of interest different from the first region of interest in the pathological image, thereby generating a second weighted input image in which the pathological image has been processed by the second weighting information; and the first weighted input image is input, and the first weighted input The system employs a configuration comprising: a first feature analysis model that generates features of a force image to generate features of a first weighted input image, a second weighted input image is input, a second feature analysis model that generates features of the second weighted input image to generate features of the second weighted input image, the features of the first weighted input image and the second weighted input image are input, a classification model that classifies whether the sample cells are benign or malignant is used to classify whether the sample cells are benign or malignant, and a learning unit 22 that updates the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information.

[0049] Therefore, the learning device 2 according to this exemplary embodiment can accurately train a first generation model for generating a first attention map, a second generation model for generating a second attention map, a first feature analysis model for generating a first weighted input image feature, a second feature analysis model for generating a second weighted input image feature, and a classification model for classifying whether a sample cell is benign or malignant. Accordingly, the learning device 2 according to this exemplary embodiment can improve the accuracy of classifying sample cells as benign or malignant in pathological diagnosis.

[0050] (Learning Method S2 Flow) The flow of the learning method S2 according to this exemplary embodiment will be explained with reference to Figure 4. 4 This is a flowchart showing the flow of the learning method S2 according to this exemplary embodiment.

[0051] (Step S21) In step S21, the acquisition unit 21 acquires a pathological image containing the sample cells as the subject, and classification information (correct label) indicating whether the sample cells are benign or malignant. The acquisition unit 11 supplies the acquired pathological image and classification information to the learning unit 22.

[0052] (Step S22) Learning Department 22 is , minutes Based on a comparison of the classification results from the type model with the classification information, the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model are updated.

[0053] As described above, in the learning method S2 according to this exemplary embodiment, a pathological image including specimen cells as the subject, and classification information indicating whether the specimen cells are benign or malignant cells are obtained, a first generative model that takes the pathological image as input and generates first weighting information to emphasize a first region of interest in the pathological image is used to generate a first weighted input image in which the pathological image has been processed with the first weighting information, a second generative model that takes the pathological image as input and generates second weighting information to emphasize a second region of interest different from the first region of interest in the pathological image is used to generate a second weighted input image in which the pathological image has been processed with the second weighting information, and the first weighted input image is input, and the first weighted The learning method employs a configuration that includes: generating features of a first weighted input image using a first feature analysis model that generates features of an input image; receiving a second weighted input image; generating features of a second weighted input image using a second feature analysis model that generates features of the second weighted input image; receiving features of the first weighted input image and features of the second weighted input image; classifying whether the sample cells are benign or malignant using a classification model that classifies whether the sample cells are benign or malignant; and updating the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information. Therefore, the learning method S2 according to this exemplary embodiment can be obtained to have the same effect as the learning device 2 described above.

[0054] [Exemplary Embodiment 2] A second exemplary embodiment of the present invention will be described in detail with reference to the drawings. Components having the same function as those described in Exemplary Embodiment 1 will be denoted by the same reference numerals, and their descriptions will be omitted as appropriate.

[0055] (Overview of Classification Device 1A) The classification device 1A according to this exemplary embodiment is a device that classifies sample cells to determine whether they are benign or malignant cells.

[0056] As an example, the classification device 1A first takes a pathological image containing specimen cells as input and uses a pre-trained generative model to generate weighting information to highlight regions of interest in the pathological image. The device then generates an image in which the pathological image has been weighted using this weighting information (hereinafter referred to as "input image processed with weighting information").

[0057] Next, the classification device 1A receives a weighted input image and generates weighted input image features using a feature analysis model that has been trained to generate features for the weighted input image.

[0058] The classification device 1A then receives weighted feature quantities from the input image and uses a pre-trained classification model to classify whether the sample cells are benign or malignant.

[0059] The sample cells included as subjects in pathological images are cells that have been stained beforehand using a staining method that stains at least the cytoplasm, such as Diff-Quik staining or Papanicolaou staining.

[0060] Furthermore, in this exemplary embodiment, we will also describe the case where the weighting information is an attention map.

[0061] Furthermore, in this exemplary embodiment, the number of generative models and feature analysis models is not particularly limited, but below, as an example, we will describe the case in which two generative models (hereinafter referred to as the first generative model and the second generative model) and feature analysis models (the first feature analysis model and the second feature analysis model) are used.

[0062] The first attention map, the second attention map, the first weighted input image, the second weighted input image, the first region of interest, the second region of interest, the generative model, the feature analysis model, and the classification model are as described in the embodiments above.

[0063] Here, the first region of interest in the first weighted input image and the second region of interest in the second weighted input image are not particularly limited as long as they are different regions from each other. For example, the sample cells are stained at least in the cytoplasm, and the first region of interest and the second region of interest are regions with different staining intensities from each other.

[0064] As an example of the first and second regions of interest, the sample cells may be stained at least in the cytoplasm, and in the weighted input image, regions where the staining intensity is equal to or greater than the first threshold may be set as the first region of interest, and regions where the staining intensity is equal to or greater than the second threshold but lower than the first threshold may be set as the second region of interest.

[0065] Other examples of the first and second regions of interest include the first region being an area with higher staining intensity than the second region, and the second region being an area containing stained cilia.

[0066] Further examples of the first and second regions of interest include the first region of interest being a region within the cell where the distance from the pixel with the highest staining intensity is shorter than a first predetermined length (in other words, a region close to the center of the cell or nucleolus), and the second region of interest being a region where the distance from the center of the cell or nucleolus is shorter than a second predetermined length which is longer than the first predetermined length, and longer than the first predetermined length.

[0067] Furthermore, the classification device 1A is a device that updates the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model. In other words, the classification device 1A also has the configuration of the learning device 2 described above.

[0068] Classification device 1A, for example, acquires pathological images containing sample cells as subjects, and classification information indicating whether the sample cells are benign or malignant. Based on a comparison of the classification results obtained by the classification model with the classification information, it updates the parameters of each model.

[0069] (Configuration of Classification Device 1A) As shown in Figure 5, the classification device 1A according to this exemplary embodiment includes a control unit 10, a storage unit 30, and an input / output unit 31.

[0070] The memory unit 30 stores data that the control unit 10, which will be described later, references. Examples of data stored in the memory unit 30 include pathology image PP and classification information CI.

[0071] The input / output unit 31 is an interface that acquires data from other connected devices or outputs data to other connected devices. The input / output unit 31 supplies data acquired from other devices to the control unit 10, and outputs data supplied from the control unit 10 to other devices.

[0072] Furthermore, the input / output unit 31 may be a communication module that communicates with other devices via a network. The specific configuration of the network is not limited to this exemplary embodiment, but as an example, a wireless LAN (Local Area Network), wired LAN, WAN (Wide Area Network), public telephone network, mobile data communication network, or a combination of these networks can be used.

[0073] (Control Unit 10) The control unit 10 controls each component of the classification device 1A. Furthermore, as shown in Figure 5, the control unit 10 also functions as an acquisition unit 11A, a classification unit 12, and a learning unit 22. In this exemplary embodiment, the acquisition unit 11A, the classification unit 12, and the learning unit 22 are configured to implement the acquisition means, classification means, and learning means, respectively.

[0074] The acquisition unit 11A acquires data supplied from the input / output unit 31. The acquisition unit 11A stores the acquired data in the storage unit 30. The acquisition unit 11A also includes a first acquisition unit 110 and a second acquisition unit 120, as shown in Figure 5.

[0075] The first acquisition unit 110 acquires a pathological image PP that includes specimen cells as the subject. In other words, the first acquisition unit 110 has the functions of the acquisition unit 11 in the exemplary embodiment described above.

[0076] The second acquisition unit 120 acquires pathological images that include the sample cells as subjects, and classification information indicating whether the sample cells are benign or malignant. In other words, the second acquisition unit 120 has the same functions as the acquisition unit 21 in the exemplary embodiment described above.

[0077] The classification unit 12 uses a first generative model to generate a first weighted input image in which the pathological image has been processed by a first attention map.

[0078] Furthermore, the classification unit 12 uses a second generative model to generate a second weighted input image in which the pathological image has been processed by a second attention map.

[0079] Furthermore, the classification unit 12 generates the first weighted input image features using the first feature analysis model.

[0080] Furthermore, the classification unit 12 generates a second weighted set of features for the input image using a second feature analysis model.

[0081] Furthermore, the classification unit 12 uses a classification model to classify whether the sample cells are benign or malignant.

[0082] The learning unit 22 updates the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification results from the classification model with the classification information.

[0083] (Overview of the processes performed by the classification unit 12) The processing performed by the classification unit 12 will be explained with reference to Figure 6. Figure 6 is a schematic diagram showing the flow of processing performed by the classification unit 12 according to this exemplary embodiment.

[0084] First, the classification unit 12 acquires a pathology image PP containing sample cells as the subject from the storage unit 30. The classification unit 12 then supplies the acquired pathology image PP to the first generation model GM1 and the second generation model GM2.

[0085] Next, the classification unit 12 generates a weighted input image processed by the attention map AM generated by the generative model GM.

[0086] For example, in the case of an attention map in which areas of interest are represented using weights from 0 to 1, the classification unit 12 takes the pathology image PP as input and generates a first weighted input image processed by the first attention map AM1 by multiplying the first attention map AM1 generated by the first generation model GM1 with the corresponding pixel values ​​in the pathology image PP.

[0087] Similarly, the classification unit 12 takes the pathology image PP as input and generates a second weighted input image processed by the second attention map AM2 by multiplying the second attention map AM2 generated by the second generative model GM2 by the corresponding pixel values ​​in the pathology image PP.

[0088] Next, the classification unit 12 generates a feature vector F3 by combining the feature vector F1 of the first weighted input image, which is generated by the first feature analysis model using the generated first weighted input image as input, and the feature vector F2 of the second weighted input image, which is generated by the second feature analysis model using the generated second weighted input image as input. As an example, feature vectors F1 and F2 are typically vectors of 1024 elements.

[0089] The classification unit 12 then inputs the feature quantity F3 into the classification model CM and, referring to the classification result CR from the classification model, classifies whether the sample cells are benign or malignant.

[0090] For example, if the classification result CR indicates that the sample cells included as subjects in the pathology image PP are benign cells, the classification unit 12 classifies the sample cells included as subjects in the pathology image PP as benign cells. On the other hand, if the classification result CR indicates that the sample cells included as subjects in the pathology image PP are malignant cells, the classification unit 12 classifies the sample cells included as subjects in the pathology image PP as malignant cells.

[0091] Here, the classification unit 12 may skip the processing performed by the first generative model GM1 and the second generative model GM2, which generate the attention map.

[0092] For example, in the case of an attention map in which the region of interest is represented using weights from 0 to 1, the attention maps generated by the first generation model GM1 and the second generation model GM2 may be set so that the weight is 1 for all regions of the pathological image PP.

[0093] The classification unit 12 generates a weighted input image by multiplying the corresponding pixel values ​​in the attention map AM and the pathology image PP. Therefore, if the attention map is set so that the weight is 1 in all regions of the pathology image PP, the classification unit 12 skips the processing by the first generation model GM1 and the second generation model GM2, and inputs the pathology image PP as a weighted input image to the feature analysis model.

[0094] (Processing performed by the learning unit 22) The processes executed by the learning unit 22 are described below.

[0095] As an example, the learning unit 22 is trained to generate a first attention map AM1 for highlighting a first area of ​​interest with high staining intensity, and a second attention map AM2 for highlighting a second area of ​​interest with lower staining intensity than the first area of ​​interest but with moderate staining intensity, so that it can accurately determine whether the sample cells are malignant or benign.

[0096] Furthermore, as another example, the learning unit 22 is trained to generate a second attention map AM2 for highlighting a second region containing stained cilia, and a first attention map AM1 for highlighting a first region that has a higher staining intensity than the second region, so that it can accurately determine whether the sample cells are malignant or benign.

[0097] In the following, as an example, we will describe a configuration in which a first attention map AM1 is generated by designating regions with a staining intensity equal to or greater than a first threshold as the first region of focus, and a second generation model GM2 generates a second attention map AM2 by designating regions with a staining intensity equal to or greater than a second threshold that is lower than the first threshold, and lower than the first threshold, as the second region of focus.

[0098] Learning Department 22 first, Set the initial values ​​for the weights. First generative model GM1 and second generative model GM 2 Configure it. As an example, the learning unit 22 is: As initial values, The first area of ​​interest, where the staining intensity is above the first threshold (for example, in a 256-level image, the area with a pixel value of 180 or higher), is weighted to 1, while the weight of all other areas is set to 0. Yo Set this to the first generative model GM1.

[0099] Furthermore, the learning section 22 is, As initial values, The second area of ​​focus is the region where the staining intensity is above the second threshold but lower than the first threshold, and also lower than the first threshold (for example, in a 256-level image, the region with pixel values ​​from 64 to 179). The weight of this region is set to 1, and the weight of all other regions is set to 0. Yo This is then set to the second generative model GM2. In this way, the learning unit 22 may set the initial values ​​of the first and second attention regions in the first attention map AM1 and the second attention map AM2 to be relatively higher than the other regions.

[0100] Here, the learning unit 22 may further divide the regions other than the first region of interest and the regions other than the second region of interest as described above, and set weights for each. For example, the learning unit 22 may set an initial value of 0.5 for the second generative model GM2, which is the weight of regions where the color intensity is equal to or greater than the first threshold (for example, in the case of a 256-level image, regions where the pixel value is 180 or greater).

[0101] Furthermore, the learning unit 22 may set different initial weight values ​​depending on the pixel values ​​within the first and second areas of interest. In other words, the learning unit 22 may set the initial values ​​of the first attention map AM1 and the second attention map AM2 such that the first and second areas of interest are regions with different staining intensities. For example, the learning unit 22 may set the initial values ​​for the first generative model GM1 to 1 for the region with pixel values ​​from 180 to 209, 0.9 for the region with pixel values ​​from 210 to 229, and 0.8 for the region with pixel values ​​from 230 to 255. Similarly, the learning unit 22 may also set different initial weight values ​​depending on the pixel values ​​in regions other than the first area of ​​interest and regions other than the second area of ​​interest.

[0102] Next, the learning unit 22 updates the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model using an optimization algorithm such as the gradient descent method, based on a comparison of the classification result CR and the classification information CI.

[0103] For example, if the classification result CR indicates that the sample cells are benign, and the classification information CI indicates that the sample cells are malignant, the learning unit 22 updates the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model so that when the pathology image PP is input to the classification unit 12, it outputs a classification result indicating that the sample cells are malignant.

[0104] By repeating the above process, the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model can be trained with high accuracy.

[0105] (Effect 1 of Classification Device 1A) As described above, the classification device 1A according to this exemplary embodiment employs a configuration comprising: an acquisition unit 11A that acquires a pathological image PP including specimen cells as the subject; and a classification unit 12 that classifies whether specimen cells are benign or malignant based on a first weighted input image feature quantity F1 generated using a first attention map AM1 for emphasizing a first area of ​​interest, and a second weighted input image feature quantity F2 generated using a second attention map AM2 for emphasizing a second area of ​​interest.

[0106] For example, if a certain cell can be classified as benign or malignant by focusing on a region with high staining intensity, and other cells different from that cell can be classified as benign or malignant by focusing on a region with relatively low staining intensity, then in the classification device 1A according to this exemplary embodiment, the first region of interest and the second region of interest can be set to regions with different staining intensities.

[0107] Therefore, the classification device 1A according to this exemplary embodiment can improve the accuracy of classifying specimen cells as benign or malignant in pathological diagnosis.

[0108] Another example is that a certain cell can be classified as malignant by focusing on the area around the nucleolus, and other cells different from that cell can be classified as malignant by focusing on the cilia. other If the cells can be classified as benign cells, in the classification device 1A according to this exemplary embodiment, the first area of ​​interest can be set to an area with a higher staining intensity than the second area of ​​interest, and the second area of ​​interest can be set to an area containing stained cilia.

[0109] Therefore, the classification device 1A according to this exemplary embodiment can improve the accuracy of classifying specimen cells as benign or malignant in pathological diagnosis.

[0110] (Effect 2 of Classification Device 1A) Furthermore, in the classification device 1A according to this exemplary embodiment, the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model are updated based on a comparison of the classification result CR and the classification information CI. Therefore, the learning device 2 according to this exemplary embodiment allows for accurate training of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model, thereby improving the accuracy of the classification of benign and malignant specimen cells in pathological diagnosis.

[0111] [Examples of implementation using software] Classifier 1, 1A Some or all of the functions of the learning device 2 may be implemented by hardware such as integrated circuits (IC chips), or by software.

[0112] In the latter case, Classification device 1, 1A The learning device 2 is implemented, for example, by a computer that executes program instructions, which are software that realizes each function. An example of such a computer (hereinafter referred to as computer C) is shown in Figure 7. Computer C comprises at least one processor C1 and at least one memory C2. Memory C2 contains the computer C that classifies the classifying device 1. 1A The program P for operating as learning device 2 is recorded. In computer C, processor C1 reads program P from memory C2 and executes it, thereby enabling classification device 1, 1A Then, each function of the learning device 2 is realized.

[0113] Processor C1 can include, for example, a CPU (Central Processing Unit), GPU (Graphic Processing Unit), DSP (Digital Signal Processor), MPU (Micro Processing Unit), FPU (Floating Point Number Processing Unit), PPU (Physics Processing Unit), microcontroller, or a combination thereof. Memory C2 can include, for example, flash memory, HDD (Hard Disk Drive), SSD (Solid State Drive), or a combination thereof.

[0114] Computer C may also be equipped with RAM (Random Access Memory) for loading program P at runtime and for temporarily storing various data. Furthermore, computer C may be equipped with communication interfaces for sending and receiving data with other devices. Additionally, computer C may be equipped with input / output interfaces for connecting input / output devices such as keyboards, mice, displays, and printers.

[0115] Furthermore, program P can be recorded on a non-temporary, tangible recording medium M that is readable by computer C. Such a recording medium M could be, for example, tape, disk, card, semiconductor memory, or programmable logic circuitry. Computer C can acquire program P via such a recording medium M. Program P can also be transmitted via a transmission medium. Such a transmission medium could be, for example, a communication network or broadcast waves. Computer C can also acquire program P via such a transmission medium.

[0116] [Additional Note 1] The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means disclosed in the embodiments described above are also included in the technical scope of the present invention.

[0117] [Additional Note 2] Some or all of the embodiments described above may also be described as follows. However, the present invention is not limited to the embodiments described below.

[0118] (Note 1) A means for acquiring pathological images that include sample cells as subjects, Using the pathological image as input, a first weighted input image is generated by processing the pathological image with the first weighted information using a first generative model trained to generate first weighted information for highlighting a first region of interest in the pathological image; using the pathological image as input, a second weighted input image is generated by processing the pathological image with the second weighted information using a second generative model trained to generate second weighted information for highlighting a second region of interest different from the first region of interest in the pathological image; and the first weighted input image is input, and the features of the first weighted input image are... A classification device comprising: a first feature analysis model trained to generate features of the first weighted input image, a second weighted input image is input, a second feature analysis model trained to generate features of the second weighted input image is input, and a classification means is input to classify whether the sample cells are benign or malignant cells using a classification model trained to classify whether the sample cells are benign or malignant cells.

[0119] (Note 2) The classification apparatus as described in Appendix 1, wherein at least the cytoplasm of the sample cells is stained, and the first region of interest and the second region of interest are regions with different staining intensities from each other.

[0120] (Note 3) The classification apparatus as described in Appendix 2, wherein the second region of interest is a region containing stained cilia, and the first region of interest is a region with a higher staining intensity than the second region of interest.

[0121] (Note 4) The system includes an acquisition means for acquiring a pathological image containing sample cells as the subject, and classification information indicating whether the sample cells are benign or malignant; a first generation model that takes the pathological image as input and generates first weighting information to emphasize a first region of interest in the pathological image, thereby generating a first weighted input image processed by the first weighting information; a second generation model that takes the pathological image as input and generates second weighting information to emphasize a second region of interest different from the first region of interest in the pathological image, thereby generating a second weighted input image processed by the second weighting information; and a first feature that takes the first weighted input image as input and generates feature quantities of the first weighted input image. A learning device comprising: a learning means that generates features of the first weighted input image using a feature analysis model; generates features of the second weighted input image using a second feature analysis model that takes the second weighted input image as input and generates features of the second weighted input image; classifies whether the sample cells are benign or malignant using a classification model that takes the features of the first weighted input image and the features of the second weighted input image as input and classifies whether the sample cells are benign or malignant using a classification model that classifies whether the sample cells are benign or malignant; and updates the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information.

[0122] (Note 5) The learning device according to Appendix 4, wherein the sample cells are stained at least in the cytoplasm, and the learning means sets initial values ​​for the first weighting information and the second weighting information such that the first region of interest and the second region of interest are regions with different staining intensities from each other.

[0123] (Note 6) The learning device according to Appendix 5, wherein the second region of interest is a region containing stained cilia, the first region of interest is a region with a higher staining intensity than the second region of interest, and the learning means sets the initial values ​​of the first region of interest and the second region of interest in the first weighting information and the second weighting information to be relatively higher than other regions.

[0124] (Note 7) The classification device acquires a pathological image containing sample cells as the subject, and uses a first generative model trained to generate first weighting information for highlighting a first region of interest in the pathological image as input to generate a first weighted input image in which the pathological image has been processed with the first weighting information, and uses a second generative model trained to generate second weighting information for highlighting a second region of interest different from the first region of interest in the pathological image as input to generate a second weighted input image in which the pathological image has been processed with the second weighting information, and the first weighted input image is input, A classification method comprising: generating features of the first weighted input image using a first feature analysis model trained to generate features of the first weighted input image; receiving the second weighted input image as input and generating features of the second weighted input image using a second feature analysis model trained to generate features of the second weighted input image; and receiving the features of the first weighted input image and the features of the second weighted input image as input and classifying whether the sample cells are benign or malignant using a classification model trained to classify whether the sample cells are benign or malignant.

[0125] (Note 8) The learning device acquires a pathological image containing sample cells as the subject, and classification information indicating whether the sample cells are benign or malignant. Using a first generative model that takes the pathological image as input and generates first weighting information to emphasize a first region of interest in the pathological image, it generates a first weighted input image in which the pathological image has been processed with the first weighting information. Using a second generative model that takes the pathological image as input and generates second weighting information to emphasize a second region of interest different from the first region of interest in the pathological image, it generates a second weighted input image in which the pathological image has been processed with the second weighting information. The first weighted input image is then input, and a second weighted input image is generated to generate feature quantities of the first weighted input image. A learning method comprising: generating features of the first weighted input image using a feature analysis model; generating features of the second weighted input image using a second feature analysis model that takes the second weighted input image as input and generates features of the second weighted input image; classifying whether the sample cells are benign or malignant using a classification model that takes the features of the first weighted input image and the features of the second weighted input image as input and classifies whether the sample cells are benign or malignant; and updating the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information.

[0126] (Note 9) A program that causes a computer to function as a classification device, wherein the program causes the computer to function as an acquisition means for acquiring a pathological image containing specimen cells as the subject, and using a first generative model trained to generate first weighting information for highlighting a first region of interest in the pathological image as input, the program generates a first weighted input image in which the pathological image is processed by the first weighting information, and using a second generative model trained to generate second weighting information for highlighting a second region of interest different from the first region of interest in the pathological image as input, the program generates a second weighted input image in which the pathological image is processed by the second weighting information, and the first A program that functions as a classification means that takes a weighted input image as input and generates features of the first weighted input image using a first feature analysis model trained to generate features of the first weighted input image; takes a second weighted input image as input and generates features of the second weighted input image using a second feature analysis model trained to generate features of the second weighted input image; and takes the features of the first weighted input image and the features of the second weighted input image as input and uses a classification model trained to classify whether the sample cells are benign or malignant as input.

[0127] (Note 10) A program that causes a computer to function as a learning device, wherein the program causes the computer to use acquisition means to acquire a pathological image including specimen cells as subjects, and classification information indicating whether the specimen cells are benign or malignant cells, and using a first generation model that takes the pathological image as input and generates first weighting information for emphasizing a first region of interest in the pathological image, the program generates a first weighted input image in which the pathological image has been processed by the first weighting information, and using a second generation model that takes the pathological image as input and generates second weighting information for emphasizing a second region of interest different from the first region of interest in the pathological image, the program generates a second weighted input image in which the pathological image has been processed by the second weighting information, and the first weighted input image is input and the first weighted A program that functions as a learning means that generates features of a first weighted input image using a first feature analysis model that generates features of a deleted input image, receives a second weighted input image as input, generates features of the second weighted input image using a second feature analysis model that generates features of the second weighted input image, receives the features of the first weighted input image and the features of the second weighted input image as input, classifies whether the sample cells are benign or malignant using a classification model that classifies whether the sample cells are benign or malignant, and updates the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information.

[0128] [Additional Note 3] Some or all of the embodiments described above can also be expressed as follows:

[0129] The system comprises at least one processor, the processor performing an acquisition process to acquire a pathological image containing sample cells as the subject; taking the pathological image as input, generating a first weighted input image in which the pathological image is processed with the first weighted information using a first generative model trained to generate first weighted information for highlighting a first region of interest in the pathological image; taking the pathological image as input, generating a second weighted input image in which the pathological image is processed with the second weighted information using a second generative model trained to generate second weighted information for highlighting a second region of interest different from the first region of interest in the pathological image; and the first weighted input A classification device that takes a force image as input and generates features of the first weighted input image using a first feature analysis model trained to generate features of the first weighted input image; takes a second weighted input image as input and generates features of the second weighted input image using a second feature analysis model trained to generate features of the second weighted input image; and takes the features of the first weighted input image and the features of the second weighted input image as input and performs a classification process to classify whether the sample cells are benign or malignant using a classification model trained to classify whether the sample cells are benign or malignant.

[0130] Furthermore, this classification device may also be equipped with memory, and this memory may store a program that causes the processor to perform the acquisition process and the classification process. This program may also be recorded on a computer-readable, non-temporary, tangible recording medium.

[0131] The system comprises at least one processor, the processor performing an acquisition process to acquire a pathological image including sample cells as the subject, and classification information indicating whether the sample cells are benign or malignant; a first generative model that takes the pathological image as input and generates first weighting information for highlighting a first region of interest in the pathological image, thereby generating a first weighted input image processed by the first weighting information; a second generative model that takes the pathological image as input and generates second weighting information for highlighting a second region of interest different from the first region of interest in the pathological image, thereby generating a second weighted input image processed by the second weighting information; and the first weighted input image is input, and the first weighted input image A learning device that performs a learning process which involves generating features of a first weighted input image using a first feature analysis model for generating features, receiving a second weighted input image as input, generating features of the second weighted input image using a second feature analysis model for generating features of the second weighted input image, receiving the features of the first weighted input image and the features of the second weighted input image as input, classifying whether the sample cells are benign or malignant using a classification model for classifying whether the sample cells are benign or malignant, and updating the parameters of the first generation model, the second generation model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification result with the classification information.

[0132] Furthermore, this learning device may also be equipped with memory, and this memory may store a program that causes the processor to execute the acquisition process and the learning process. This program may also be recorded on a computer-readable, non-temporary, tangible recording medium. [Explanation of Symbols]

[0133] 1, 1A classification device 2 Learning device 11, 11A, 21 Acquisition Department 110 First acquisition part 120 Second acquisition part 12 Classification section 22 Learning Department

Claims

1. A means for acquiring pathological images that include sample cells as subjects, Using the aforementioned pathological image as input, a first weighted input image is generated in which the pathological image has been processed with the first weighted information, using a first generative model that has been trained to generate first weighting information for highlighting a first region of interest in the pathological image. Using the aforementioned pathological image as input, a second weighted input image is generated in which the pathological image is processed with the second weighted information, using a second generative model that has been trained to generate second weighting information for highlighting a second region of interest different from the first region of interest in the pathological image. The first weighted input image is input, and the first weighted input image features are generated using a first feature analysis model that has been trained to generate the features of the first weighted input image. The second weighted input image is input, and the features of the second weighted input image are generated using a second feature analysis model that has been trained to generate features of the second weighted input image. A classification means that takes the features of the first weighted input image and the features of the second weighted input image as input and uses a classification model trained to classify whether the sample cells are benign or malignant to classify whether the sample cells are benign or malignant, Equipped with, The second region of interest is the region containing stained cilia, and the first region of interest is the region with a higher staining intensity than the second region of interest. Classification device.

2. A means for acquiring pathological images containing sample cells as subjects, and classification information indicating whether the sample cells are benign or malignant cells, Using a first generative model that takes the aforementioned pathological image as input and generates first weighting information for highlighting a first region of interest in the pathological image, a first weighted input image is generated in which the pathological image has been processed with the first weighting information. Using the aforementioned pathological image as input, a second generative model is used to generate second weighting information for emphasizing a second region of interest different from the first region of interest in the pathological image, thereby generating a second weighted input image in which the pathological image has been processed by the second weighting information. The first weighted input image is input, and the first feature analysis model that generates the feature quantities of the first weighted input image is used to generate the feature quantities of the first weighted input image. The second weighted input image is input, and the second feature analysis model that generates the feature quantities of the second weighted input image is used to generate the feature quantities of the second weighted input image. The features of the first weighted input image and the features of the second weighted input image are input, and a classification model is used to classify whether the sample cells are benign or malignant, A learning means that updates the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification results with the classification information, A learning device equipped with the following features.

3. The aforementioned sample cells were stained at least in the cytoplasm. The learning means sets initial values ​​for the first weighting information and the second weighting information such that the first area of ​​interest and the second area of ​​interest are areas with different staining intensities from each other. The learning device according to claim 2.

4. The second region of interest is the region containing stained cilia, and the first region of interest is the region with a higher staining intensity than the second region of interest. The learning means sets the initial values ​​of the first and second areas of interest in the first and second weighting information to be relatively higher than those of other areas. The learning device according to claim 3.

5. The classification device, To obtain pathological images that include the sample cells as the subject, Using the aforementioned pathological image as input, a first weighted input image is generated in which the pathological image has been processed with the first weighted information, using a first generative model that has been trained to generate first weighting information for highlighting a first region of interest in the pathological image. Using the aforementioned pathological image as input, a second weighted input image is generated in which the pathological image is processed with the second weighted information, using a second generative model that has been trained to generate second weighting information for highlighting a second region of interest different from the first region of interest in the pathological image. The first weighted input image is input, and the first weighted input image features are generated using a first feature analysis model that has been trained to generate the features of the first weighted input image. The second weighted input image is input, and the features of the second weighted input image are generated using a second feature analysis model that has been trained to generate features of the second weighted input image. The first weighted input image features and the second weighted input image features are input, and a classification model trained to classify whether the sample cells are benign or malignant is used to classify whether the sample cells are benign or malignant. Includes, The second region of interest is the region containing stained cilia, and the first region of interest is the region with a higher staining intensity than the second region of interest. Classification method.

6. The learning device, Obtaining pathological images containing the sample cells as subjects, and classification information indicating whether the sample cells are benign or malignant, Using a first generative model that takes the aforementioned pathological image as input and generates first weighting information for highlighting a first region of interest in the pathological image, a first weighted input image is generated in which the pathological image has been processed with the first weighting information. Using the aforementioned pathological image as input, a second generative model is used to generate second weighting information for emphasizing a second region of interest different from the first region of interest in the pathological image, thereby generating a second weighted input image in which the pathological image has been processed by the second weighting information. The first weighted input image is input, and the first feature analysis model that generates the feature quantities of the first weighted input image is used to generate the feature quantities of the first weighted input image. The second weighted input image is input, and the second feature analysis model that generates the feature quantities of the second weighted input image is used to generate the feature quantities of the second weighted input image. The features of the first weighted input image and the features of the second weighted input image are input, and a classification model is used to classify whether the sample cells are benign or malignant, Based on a comparison of the classification results with the classification information, the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model are updated. Learning methods that include this.

7. A program that makes a computer function as a classification device, The aforementioned program, the computer, A means for acquiring pathological images that include sample cells as subjects, Using the aforementioned pathological image as input, a first weighted input image is generated in which the pathological image has been processed with the first weighted information, using a first generative model that has been trained to generate first weighting information for highlighting a first region of interest in the pathological image. Using the aforementioned pathological image as input, a second weighted input image is generated in which the pathological image is processed with the second weighted information, using a second generative model that has been trained to generate second weighting information for highlighting a second region of interest different from the first region of interest in the pathological image. The first weighted input image is input, and the first weighted input image features are generated using a first feature analysis model that has been trained to generate the features of the first weighted input image. The second weighted input image is input, and the features of the second weighted input image are generated using a second feature analysis model that has been trained to generate features of the second weighted input image. A classification means that takes the features of the first weighted input image and the features of the second weighted input image as input and uses a classification model trained to classify whether the sample cells are benign or malignant to classify whether the sample cells are benign or malignant, To make it function as, The second region of interest is the region containing stained cilia, and the first region of interest is the region with a higher staining intensity than the second region of interest. program.

8. A program that makes a computer function as a learning device, The aforementioned program, the computer, A means for acquiring pathological images containing sample cells as subjects, and classification information indicating whether the sample cells are benign or malignant cells, Using a first generative model that takes the aforementioned pathological image as input and generates first weighting information for highlighting a first region of interest in the pathological image, a first weighted input image is generated in which the pathological image has been processed with the first weighting information. Using the aforementioned pathological image as input, a second generative model is used to generate second weighting information for emphasizing a second region of interest different from the first region of interest in the pathological image, thereby generating a second weighted input image in which the pathological image has been processed by the second weighting information. The first weighted input image is input, and the first feature analysis model that generates the feature quantities of the first weighted input image is used to generate the feature quantities of the first weighted input image. The second weighted input image is input, and the second feature analysis model that generates the feature quantities of the second weighted input image is used to generate the feature quantities of the second weighted input image. The features of the first weighted input image and the features of the second weighted input image are input, and a classification model is used to classify whether the sample cells are benign or malignant, A learning means that updates the parameters of the first generative model, the second generative model, the first feature analysis model, the second feature analysis model, and the classification model based on a comparison of the classification results with the classification information, A program that makes it function as such.

Images