JP7383939B2 - Learning device, learning method, cell discrimination device, cell discrimination method, cell discrimination learning program, and cell discrimination program - Google Patents

Description

The present invention relates to a learning device and a learning method for constructing a learned model, a cell discriminating device and a cell discriminating method that use the learned model, and a cell discriminating learning program and a cell discriminating program.

Conventionally, cell types have been determined by visually observing bright field images, phase contrast images, differential interference images, immunofluorescence staining images, etc. taken under various microscopes. In recent years, in addition to visual discrimination, a discrimination model has been constructed by machine learning such as deep learning, and discrimination has been performed using this model, as disclosed in Non-Patent Document 1, for example.

Ronald Wihal Oei et al., "Convolutional neural network for cell classification using microscope images of intracellular actin networks" PLOS ONE, 2019.

When building a discriminant model with higher accuracy, it is considered advantageous to have a larger amount of training data. On the other hand, for example, if the number of data differs for each classification in the learning data, the discriminant model that is constructed will be biased, so there is a risk that the discrimination accuracy will decrease. Therefore, learning is usually performed after generating training data whose quantity is matched to the classification with the least amount of data. Therefore, in machine learning, it is desirable to acquire a large amount of data for all classification items.

However, depending on the sample, there may be large deviations in the abundance ratio of cells, making it difficult to obtain a large amount of data. For example, when classifying cells in the blood, there are cell types for which it is difficult to obtain data, including circulating tumor cells (hereinafter referred to as CTCs). In addition to CTCs, blood contains cells such as red blood cells and white blood cells, as well as components other than cells. When attempting to classify these by machine learning, the number of data will be adjusted to the CTC with the smallest number of data. As a result, since machine learning is performed with a small amount of data, there is a problem that the accuracy of the discriminant model decreases.

Therefore, the present invention has been made in view of the above problems, and its purpose is to provide a cell discrimination device and a cell discrimination method capable of accurately discriminating cell types, and a learning method for constructing a trained model used therein. The objective is to provide devices and learning methods.

In order to solve the above-mentioned problems, a learning device according to the present invention is a learning device that constructs a trained model used for cell discrimination, and includes first training data including data indicating cells and data indicating non-cells. a first learning unit that constructs a first trained model for determining whether cells are cells or non-cells by machine learning using the method; and a plurality of types of data indicating mutually different types of cells; and the first learning for determining the type of cell of the data determined as a cell by the first trained model by machine learning using second training data that does not include data indicating non-cells. and a second learning section that constructs a second trained model different from the trained model.

In order to solve the above problems, a cell discrimination device according to the present invention is a cell discrimination device that discriminates the type of cell in input data, and uses a first trained model constructed by the above learning device. a first discrimination section that discriminates whether the input data indicates cells or non-cells; a second discriminator that discriminates which type of cell the input data indicates using the trained model of No. 2, and a second discriminator that discriminates which type of cell the input data indicates; is equipped with a discrimination result output section that outputs the discrimination result, and outputs the discrimination result of the second discrimination section when the discrimination result by the first discrimination section indicates a cell.

Moreover, in order to solve the above-mentioned problem, the learning method according to the present invention is a learning method for constructing a trained model used for cell discrimination, and includes a first model including data indicating cells and data indicating non-cells. A first learning step in which a first trained model is constructed to determine whether cells are cells or non-cells by machine learning using training data, and multiple types of data indicating different types of cells are combined. The above-mentioned first method for determining the cell type of data that is determined to be a cell by the first trained model by machine learning using the second training data that includes the data and does not include data indicating non-cells. and a second learning step of constructing a second trained model different from the trained model.

In addition, in order to solve the above problems, the cell discrimination method according to the present invention is a cell discrimination method that discriminates the type of cell in input data, and includes a data input step of inputting data, and the above-mentioned learning device. A first discrimination step in which the input data is determined to indicate cells or non-cells using the first trained model constructed by a second discrimination step of determining which type of cell the input data indicates using a second trained model constructed by the learning device; and a second discrimination step of determining which type of cell the input data indicates; a discrimination result output step of outputting the discrimination result when the discrimination result in the step indicates non-cells, and outputting the discrimination result in the second discrimination step when the discrimination result in the first discrimination step indicates cells; The configuration includes .

According to the present invention, it is possible to generate a discrimination model that can accurately discriminate the type of cells in an image.

FIG. 1 is a functional block diagram showing a schematic configuration of a cell discrimination device according to an embodiment of the present invention. FIG. 3 is a diagram showing an example of a first table. It is a figure which shows an example of a 2nd table. It is a flowchart showing the flow of cell discrimination processing according to one embodiment of the present invention. FIG. 3 is a diagram showing an example of an input image.

[Cell discrimination device]
Hereinafter, one embodiment of the present invention will be described in detail. FIG. 1 is a functional block diagram showing an example of a schematic configuration of a cell discrimination device 1 according to this embodiment. The cell discrimination device 1 includes an input section 10, a main control section 20, and a display section 30. The main control section 20 includes a discrimination section 40, a learning model 50, a learning section 60, and a discrimination result output section 70. The determining section 40 includes a first determining section 41 and a second determining section 42. Further, the learning section 60 functions as a learning device that constructs a trained model used for cell discrimination, and includes a first learning section 61, a second learning section 62, and a teacher data generation section 63. The learning model 50 includes a first trained model 51 and a second trained model 52.

The cell discrimination device 1 is a device that discriminates the type of cell in input data. In the following description, an example will be described in which image data (hereinafter, sometimes simply referred to as an image) is used as data.

(input section)
The input unit 10 receives input of an image to be determined. The input unit 10 receives the above-described image input by reading a data file stored in a storage medium or by receiving an image from another device via a wired or wireless network. The input unit 10 transmits the received image data to the main control unit 20.

(image data)
The image data is an image obtained by imaging cells in an analysis sample. Note that cell imaging is not limited to imaging individual cells one by one; it can also involve arranging a large number of cells on an array, imaging the entire array or a portion of it, and dividing the image after imaging. You can. Particularly, when performing comprehensive processing and later dividing to obtain individual cell images, the proportion of images taken of compartments or wells that do not contain cells increases. From the above, images can be broadly classified into images of cells and images of non-cells.

The image of the cell is an image of any type of cell among multiple types of cells that may exist in the analysis sample. As a non-limiting example, when a blood sample subjected to a certain treatment is used as an analysis sample, cells that may be present in the sample include leukocytes, cytokeratin-positive circulating tumor cells, and cytokeratin-negative circulating tumor cells. Examples include blood components such as cells. Therefore, the cell image may be an image of any of leukocytes, cytokeratin-positive circulating tumor cells, and cytokeratin-negative circulating tumor cells.

On the other hand, images other than cells are classified into images in which substances other than cells such as garbage are captured, and images in which neither cells nor substances other than cells are captured. Here, an image in which no substances other than cells are captured means that neither cells nor dirt are captured in the micropores or wells that capture cells, and only the culture medium and dispersion medium used to disperse the cells are included. An image of a micropore or well is intended.

The image is not limited to an image obtained by one specific imaging method, but may be a combination of images obtained by a plurality of imaging methods. For example, two or more observation images selected from bright field observation images, dark field observation images, phase contrast observation images, and fluorescence observation images can be used in combination. Further, as the fluorescence observation image, a combination of fluorescence observation images using various antibodies, nuclear staining images using DAPI, etc. can be used. Moreover, when using a plurality of images in combination, the plurality of images can be combined into one image. For example, the images may be combined into one image by performing an overlapping process, or they may be arranged and combined without overlapping to form a single image.

Further, each image may be subjected to pre-processing such as contrast adjustment and adjustment of another channel based on a particular channel. For example, processing may be performed to normalize the brightness values of other channels based on the brightness values of a specific channel (for example, a fluorescent observation image by DAPI as described later).

As described above, in this embodiment, an example in which image data is used as input data is described, but input data is not limited to image data. For example, numerical data digitized or parameterized based on an image and a specific index that can be extracted from the image may be used.

(learning model)
The first trained model 51 used in the first discrimination unit 41 receives the image input to the cell discrimination device 1 and determines whether the image indicates a cell or a non-cell. This is a trained model whose output is the probability value of . As will be described later, the first learned model 51 is constructed in the first learning unit 61 by machine learning using first teacher data.

In the output of the first trained model 51, images of non-cellular components and images in which neither cells nor non-cellular components are imaged may be output without distinguishing between them, or may be output as showing non-cells. It may be possible to distinguish between those showing cellular components and those in which neither cells nor non-cellular components are imaged and output them. That is, in the former case, for example, a learning model that outputs the probability values of "cell" and "non-cell" can be used, and in the latter case, for example, the probability value of "cell", "non-cell" and This can be a learning model that outputs an "empty" probability value.

On the other hand, the second trained model 52 used in the second discriminator 42 is a trained model different from the first trained model 51, and uses the image input to the cell discriminator 1 as input, Each is a trained model whose output is the probability value of that cell. As will be described later, the second learned model 52 is constructed in the second learning unit 62 by machine learning using second teacher data.

In the present embodiment, the first trained model 51 and the second trained model 52 are constructed using a convolutional neural network (CNN) having a plurality of convolutional layers using first training data and second training data, respectively. This is a model built by training a neural network. As the CNN, functions in known software such as Chainer (Preferred Networks) can be used. However, the neural network used for learning is not limited to CNN, and other known neural networks may be used.

(Teacher data generation section)
The teacher data generation unit 63 generates first teacher data and second teacher data from the image data input for teacher data generation. Specifically, the image data input for generating training data is associated in advance with information as to which of multiple types of cells or non-cells the image data is. There is. The teacher data generation unit 63 generates first teacher data and second teacher data by referring to the information associated with the image data. To associate the image data with the information, create a first table that shows the correspondence between the image data and the information, input the first table along with the image data, and refer to it. It is something. However, the association between the image data and the information is not limited to the case where a table is used. For example, the information may be included as metadata of the image data. An example of the first table is shown in FIG. In the table shown in FIG. 2, each piece of image data is associated with information such as "CK+", "CK-", "WBC", "DST", and "EMP", which will be described later.

Information as to whether the image data is of a plurality of types of cells or non-cells is determined in advance by humans. The images of cells included in the image data input for generating teacher data include images of types of cells that may exist in the sample that contained the cells to be determined. On the other hand, non-cellular images include images in which substances other than cells such as garbage are captured, and images in which neither cells nor substances other than cells are captured.

First, the generation of first teacher data in the teacher data generation section 63 will be explained.

If the information associated with the input image data is information indicating that it is non-cell, that is, information indicating the type of non-cell, the teacher data generation unit 63 labels it "non-cell". On the other hand, if the information associated with the input image data is information indicating which of a plurality of types of cells, that is, information indicating the type of cell, a label of "cell" is attached. Note that attaching a label means (i) attaching label information as metadata to image data, (ii) updating a data file indicating a label attached to each image data, or (iii) adding a label to image data. This can be achieved by saving it in a directory corresponding to the . In addition, whether the information associated with the input image data corresponds to information that should be labeled "non-cell" or "cell" is determined by information indicating the type of cell and information indicating the type of cell. This is determined by referring to a second table showing the correspondence between information indicating the type and the label to be attached. The second table may be created in advance and input together with the image data. Alternatively, the second table is not limited to the case where the information is input in advance, and the user can specify the correspondence relationship with the label for information such as "CK+", "CK-", "WBC", "DST", and "EMP". It may be created separately after inputting the image data by instructing. An example of the second table is shown in FIG. The table shown in FIG. 3 shows the correspondence between the information indicated by "CK+", "CK-", "WBC", "DST", and "EMP" and the labels to be attached.

As a result, the teacher data generation unit 63 generates data indicating cells and data indicating non-cells from the image data associated with information regarding which of multiple types of cells or non-cells. Generate first training data including: In addition, even if the input image data itself does not have information on whether it is a cell or a non-cell, a first trained model is constructed to determine whether the input image data is a cell or a non-cell. It is possible to create first training data for

Note that in this embodiment, image data associated with information indicating which of a plurality of types of cells is associated is selected and labeled as "cell." However, the input image data may be associated in advance with information indicating which of a plurality of types of cells it is, as well as information indicating that it is a cell. Accordingly, if the information associated with the input image data indicates that it is a cell, the image data may be considered to be labeled with "cell".

Next, generation of second teacher data in the teacher data generation section 63 will be explained.

The teacher data generation unit 63 extracts only data associated with information indicating which of a plurality of types of cells belongs to the input image data. Then, by attaching a label indicating the cell type indicated by the associated information to the input image data, second teacher data including a plurality of types of data indicating mutually different types of cells is generated. The teacher data generation unit 63 extracts only data associated with information regarding which of the plurality of types of cells it is, and generates second teacher data. Therefore, the second teacher data no longer includes data indicating non-cells. Further, even if input data including image data indicating non-cells used to generate the first teacher data is used, it is possible to generate second teacher data that does not include image data indicating non-cells.

In the teacher data generation unit 63 in this embodiment, when generating second teacher data, among the plurality of cell types, according to the number of data of the cell type that exists in the input data the least number, Data for each cell type is extracted. If there is a large variation in the number of data for each cell type, it is possible to build a trained model with high discrimination accuracy by matching the number of data for the cell type with the smallest number in the input data. . For example, if the number of data of the cell type with the smallest number in the input data is less than 90% of the number of data of the cell type with the largest number, then according to the number of data of the smallest cell type, It may be configured to extract data. However, it is not essential to match the number of data of each cell type included in the second teacher data.

When generating the first teacher data and the second teacher data, the teacher data generation unit 63 may perform data expansion using known techniques such as image rotation and inversion.

(1st study part and 2nd study part)
The first learning unit 61 constructs the first trained model 51 using the first teacher data generated by the teacher data generation unit 63 by a known machine learning method. On the other hand, the second learning unit 62 constructs the second trained model 52 using the second teacher data generated by the teacher data generation unit 63 by a known machine learning method. In this embodiment, as described above, the first learning unit 61 and the second learning unit 62 calculate the correspondence between input images and their information (whether it is a cell or what type of cell is it) using CNN. I am letting them learn.

In addition, in this embodiment, the cell discriminating device 1 including the learning section 60 including the first learning section 61, the second learning section 62, and the teacher data generating section 63 is described. However, it is not necessary to provide the learning section 60 in the cell discrimination device. That is, the first trained model 51 and the second trained model are generated by another learning device that exists independently of the cell discrimination device 1 and includes a first learning section 61, a second learning section 62, and a teacher data generation section 63. The model 52 may also be constructed. When the learning device exists independently of the cell discriminating device 1, the cell discriminating device 1 can load each learned model stored in a storage medium or communicate with other devices via a wired or wireless network. By receiving each learned model from the cell discriminator 1, each learned model becomes available for use in the cell discriminating device 1.

(First discrimination part)
The first discriminator 41 inputs the image to the first trained model 51, and determines whether the image input to the cell discriminator 1 represents a cell or not, based on the output result of the first trained model 51. Determine whether it indicates. Specifically, the one with the highest probability is adopted as the discrimination result. The first discrimination section 41 transmits the discrimination result to the discrimination result output section 70. Note that the first discrimination section 41 may transmit the discrimination result to the discrimination result output section 70 only when the discrimination result indicates non-cells.

(Second discrimination part)
The second discriminator 42 inputs the image to the second trained model 52 only when the first discriminator 41 obtains a discrimination result between the image and the image showing cells, and outputs the output result of the second trained model 52. From this, the type of cell in the image input to the cell discriminating device 1 is discriminated. Specifically, the one with the highest probability is adopted as the discrimination result. The second discrimination section 42 transmits the discrimination result to the discrimination result output section 70.

(Discrimination result output section)
The determination result output unit 70 outputs the determination result transmitted from the first determination unit 41 or the second determination unit 42 to the display unit 30. When the discrimination result sent from the first discrimination section 41 indicates a non-cell, the discrimination result output section 70 outputs the discrimination result sent from the first discrimination section 41. On the other hand, if the discrimination result sent from the first discrimination section 41 indicates a cell, or the discrimination result is not sent from the first discrimination section 41 and the discrimination result is sent from the second discrimination section 42. In this case, the determination result output unit 70 outputs the determination result sent from the second determination unit 42. Thereby, the determination result output unit 70 notifies the user of the final determination result via the display unit 30.

(Display)
The display unit 30 is a device that displays the final determination result output from the determination result output unit 70. In one embodiment, the display unit 30 is a display device that displays the final determination result as image data or character data. Note that the display unit is not limited to being provided in the cell discriminating device 1, and may be provided as an external device connectable to the cell discriminating device 1.

(Operation of cell discrimination device)
Next, an example of the flow when performing a discrimination process using the cell discrimination device 1 will be described with reference to FIG. 4. FIG. 4 is a flowchart illustrating an example of the flow of operation of the cell discrimination device 1.

First, the determination unit 40 acquires an image input by a user's input operation via the input unit 10 (step S11; data input step). Here, the images to be determined are processed by the same processing operations as the images included in the first training data and second training data used to construct the first trained model 51 and the second trained model. It was acquired. Next, the first discrimination unit 41 inputs the image to the first trained model 51 and obtains the output from the first trained model 51. Then, based on the output result, it is determined whether the image shows cells or non-cells (step S12; first determination step). If the determination result is that the image shows cells (yes in step S13), then the second determination unit 42 uses the image input to the first trained model 51 for the second trained model 52. The same image as is input and the output from the second trained model 52 is obtained. Then, from the output result, it is determined which type of cell the image represents (second determination step), and the determination result is transmitted to the determination result output unit 70 (step S14). On the other hand, if the first discrimination section 41 determines that the image shows non-cells (no in step S13), the discrimination result is transmitted to the discrimination result output section 70. When the discrimination result is sent from the second discrimination section 42, the discrimination result output section 70 outputs the discrimination result from the second discrimination section 42 to the display section 30, and determines whether the image from the first discrimination section 41 is non-identical. When the determination result that it is a cell is sent, the determination result from the first determination section 41 is output to the display section 30 (step S15; determination result output step).

As described above, in the discrimination process using the cell discrimination device 1, cells in an image are discriminated through two discrimination steps. This is compared to the case where discrimination is performed using a trained model that is constructed to discriminate whether the cell is a cell or a non-cell, and if it is a cell, what type of cell it is. Therefore, it is possible to accurately determine which type of cell the cell in the image is.

[Example of implementation using software]
The control blocks of the cell discriminating device 1 (main control unit 20, especially the discriminating unit 40, learning unit 60, and discrimination result output unit 70) are realized by logic circuits (hardware) formed on integrated circuits (IC chips), etc. or may be realized by software.

In the latter case, the cell discrimination device 1 includes a computer that executes instructions of a program that is software that implements each function. This computer includes, for example, at least one processor (control device) and at least one computer-readable recording medium storing the above program. In the computer, the processor reads the program from the recording medium and executes the program, thereby achieving the object of the present invention. As the processor, for example, a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) can be used. As the recording medium, in addition to "non-temporary tangible media" such as ROM (Read Only Memory), tapes, disks, cards, semiconductor memories, programmable logic circuits, etc. can be used. Further, the computer may further include a RAM (Random Access Memory) for expanding the above program. Furthermore, the program may be supplied to the computer via any transmission medium (communication network, broadcast waves, etc.) that can transmit the program. Note that one aspect of the present invention can also be realized in the form of a data signal embedded in a carrier wave, in which the program is embodied by electronic transmission.

〔summary〕
A learning device according to aspect 1 of the present invention is a learning device that constructs a trained model used for cell discrimination, and is a learning device that performs machine learning using first teacher data including data indicating cells and data indicating non-cells. a first learning unit that constructs a first trained model for determining whether cells are cells or non-cells; By machine learning using second training data that does not include data, a second trained model different from the first trained model is used to determine the cell type of data that has been determined as a cell by the first trained model. and a second learning section that constructs the second trained model.

In the learning device according to aspect 2 of the present invention, in aspect 1, the teacher data generation unit generates the first teacher data and the second teacher data by referring to information associated with the input data. It also has:

In the learning device according to Aspect 3 of the present invention, in Aspect 2, the teacher data generation unit generates information indicating the type of cell, information indicating the type of non-cell, and information indicating whether the cell is a cell or a non-cell. The above-mentioned first teacher data is generated by referring to the table showing the correspondence relationship with the above.

In the learning device according to aspect 4 of the present invention, in any one of aspects 1 to 3 above, the input data is image data.

A cell discriminating device according to aspect 5 of the present invention is a cell discriminating device that discriminates the type of cell in input data, and the cell discriminating device according to aspect 5 of the present invention is a cell discriminating device that discriminates the type of cell in input data, and includes a first learning constructed by the learning device according to any one of aspects 1 to 4 above. a first discriminator that discriminates whether the input data indicates a cell or a non-cell using a predetermined model, and the learning device A second discriminator determines which type of cell the input data indicates, using a second trained model constructed by If the discrimination result by the first discrimination section indicates a cell, the discrimination result output section outputs the discrimination result by the second discrimination section.

A cell discrimination device according to aspect 6 of the present invention includes the learning device described above.

A learning method according to aspect 7 of the present invention is a learning method for constructing a trained model used for cell discrimination, and includes machine learning using first training data including data indicating cells and data indicating non-cells. a first learning step of constructing a first trained model for determining whether the cell is a cell or a non-cell; By machine learning using second training data that does not include data, a second trained model different from the first trained model is used to determine the cell type of data that has been determined as a cell by the first trained model. and a second learning step of constructing a second trained model.

In the learning method according to aspect 8 of the present invention, in aspect 7, the data indicating cells or non-cells is image data obtained by imaging blood components, and as the type of cells, at least blood A composition containing circulating tumor cells.

A cell discrimination method according to aspect 9 of the present invention is a cell discrimination method that discriminates the type of cell in input data, and includes a data input step of inputting data, and a cell discrimination method according to any one of aspects 1 to 4 above. A first discrimination step of determining whether the input data indicates cells or non-cells using the first trained model constructed by the learning device; and a discrimination result of the first discrimination step described above. when indicating a cell, a second discrimination step of determining which type of cell the input data indicates using a second trained model constructed by the learning device; If the discrimination result in the first discrimination step indicates a non-cell, the discrimination result is output; if the discrimination result in the first discrimination step indicates a cell, the discrimination result in the second discrimination step is output. process.

The learning device and the cell discrimination device according to each aspect of the present invention may be realized by a computer, and in this case, by operating the computer as each part (software element) included in the learning device or the cell discrimination device. The scope of the present invention also includes a learning program or a cell discrimination program for realizing the above learning device or the above cell discrimination device on a computer, and a computer-readable recording medium on which the program is recorded.

The present invention is not limited to the embodiments described above, and various modifications can be made within the scope of the claims. That is, embodiments obtained by combining technical means appropriately modified within the scope of the claims are also included in the technical scope of the present invention.

One embodiment of the invention was implemented as follows. For the learning phase of machine learning and the performance evaluation of the discriminant model, we preprocessed the blood sample provided by Kyorin University, captured the extracted cells in each micropore of a micropore array, and arranged the cells on the array. , using images acquired using a microscope and a digital camera. Pretreatment is a process that mainly removes red blood cells and white blood cells and performs fluorescent immunostaining. Note that although many red blood cells and white blood cells can be removed by pretreatment, a certain number remain. The micropore array is a micropore array described in a non-patent document: Tosoh Research and Technical Report, Vol. It is a structure that can be aligned. Note that sample pretreatment and cell alignment into the micropore array were performed according to the method described in the non-patent document.

The microscope used was IX83 manufactured by Olympus Corporation, and the objective lens was set to 10 times. The digital camera used was ORCA-FLASH 4.0 manufactured by Hamamatsu Photonics Co., Ltd. Using this microscope and digital camera, a bright field image, a fluorescence image using a cytokeratin antibody, a fluorescence image using a CD45 antibody, and a nuclear staining image using DAPI were obtained. The acquired images were divided into micropores. The image per micropore was 62 x 62 pixels. For each micropore, the acquired images were arranged side by side and combined to form image data used for learning and discrimination. An example of image data is shown in FIG. Figure 5 shows white blood cells (WBC), cytokeratin-positive circulating tumor cells (CK+), cytokeratin-negative circulating tumor cells (CK-), noncellular components (DST), and empty micropores (EMP). An example of image data is shown. In addition, as shown in Figure 5, each image data includes a bright field image (the image shown in the "Bright field" column in the figure), a nuclear staining image by DAPI (the image shown in the "Fluorescence 1 DAPI" column in the figure). ), a fluorescence image by the cytokeratin antibody (image shown in the column "Fluorescence 2 CK" in the figure), and a fluorescence image by the CD45 antibody (image shown in the column "Fluorescence 3 CD45" in the figure). They are arranged in this order without any gaps and combined into a single image. Note that the correct label was determined by humans.

The number of data is 500 for WBC, 151 for CK+, 500 for CK-, 500 for DST, and 481 for EMP. From this data, 80% was randomly extracted as training data and 20% was used as verification data.

A program equipped with a learning section was executed on a computer to construct a two-stage discrimination model (a first trained model and a second trained model). Chainer from Preferred Networks, Inc. was used to implement the program. The processing described below uses functions and classes included in Chainer. In constructing the second trained model, the number of data was adjusted to cytokeratin-positive circulating tumor cells (CK+), which had the least amount of data.

A neural network consisting of convolutional layers and fully connected layers was used as the machine learning method, and Adaptive Moment Estimation (Adam) was used as the optimization algorithm. In the neural network, convolutional layers were arranged in the first layer, second layer, and third layer, and the number of channels 64, stride 3, and zero padding were specified as these hyperparameters. Then, ReLU was used as the activation function of the convolutional layer. A fully connected layer was placed in the next layer, and a softmax function was used as the activation function. Additionally, during learning, data was augmented by rotating or inverting images. Learning was performed for 150 epochs, and the mini-batch size was 64. Note that data labeled "CK+", "CK-", or "WBC" was designated to be subjected to the first stage learning with the label "Cell".

As described above, the first trained model that discriminates between the three classifications of "Cell", "DST", and "EPM", and the second trained model that discriminates between the three classifications of "CK+", "CK-", and "WBC" are created. A two-stage discriminant model was constructed.

As a result of verifying cell discrimination using a two-stage discrimination model, the correct answer rate for the validation data was 96.7%.

On the other hand, in the same data set, using data labeled "CK+", "CK-", "WBC", "DST", or "EMP", "CK+", "CK-", "WBC", We constructed a one-stage discrimination model that discriminates between the five categories of "DST" and "EMP" at once.

As a result of verifying cell discrimination using this one-stage discrimination model, the correct answer rate for the validation data was 93.7%.

INDUSTRIAL APPLICATION This invention can be utilized for the technique of classifying a cell.

1 Cell discrimination device 10 Input section 20 Main control section 30 Display section 40 Discrimination section 41 First discrimination section 42 Second discrimination section 50 Learning model 51 First learned model 52 Second learned model 60 Learning section (learning device)
61 First learning section 62 Second learning section 63 Teacher data generation section 70 Discrimination result output section

Claims (11)

A learning device that constructs a trained model used for cell discrimination,
First learning to construct a first trained model for determining whether it is a cell or a non-cell by machine learning using first training data including data indicating cells and data indicating non-cells. Department and
By machine learning using second training data that includes multiple types of data indicating different types of cells and not including data indicating non-cells, data that has been determined to be cells by the first trained model is determined. A learning device comprising: a second learning unit that constructs a second trained model different from the first trained model for determining cell types. 2. The computer according to claim 1, further comprising a teacher data generation unit that generates the first teacher data and the second teacher data by referring to information associated with the input data. learning device. The teacher data generation unit refers to a table showing the correspondence between information indicating the type of cell, information indicating the type of non-cell, and information indicating whether the cell is a cell or a non-cell. 3. The learning device according to claim 2, wherein the learning device generates one piece of teacher data. 4. The learning device according to claim 1, wherein the data included in the first teacher data and the second teacher data is image data. A cell discrimination device that discriminates the type of cell in input data,
A step of determining whether the input data indicates cells or non-cells using the first learned model constructed by the learning device according to any one of claims 1 to 4. 1 discrimination section;
When the discrimination result by the first discrimination unit indicates a cell, the second learned model constructed by the learning device is used to discriminate which type of cell the input data indicates. a second determining unit that
Discrimination that outputs the discrimination result when the discrimination result by the first discrimination section indicates non-cell, and outputs the discrimination result by the second discrimination section when the discrimination result by the first discrimination section indicates cell. A cell discrimination device comprising: a result output unit. The cell discrimination device according to claim 5, comprising the learning device. A learning method for constructing a trained model executed by at least one processor and used for cell discrimination, the method comprising:
First learning to construct a first trained model for determining whether it is a cell or a non-cell by machine learning using first training data including data indicating cells and data indicating non-cells. process and
Through machine learning using second training data that includes multiple types of data that indicate different types of cells and does not include data that indicates non-cells, cells in the data that are determined to be cells by the first trained model are A learning method comprising: a second learning step of constructing a second trained model different from the first trained model for determining the type of the first trained model. 8. The data indicating cells or non-cells is image data obtained by imaging blood components, and the types of cells include at least circulating tumor cells. How to learn. A cell discrimination method executed by at least one processor to discriminate the type of cell in input data, the method comprising:
a data input step of inputting data;
A first method for determining whether the input data indicates cells or non-cells using the first trained model constructed by the learning device according to any one of claims 1 to 4. Discrimination process;
When the discrimination result in the first discrimination step indicates a cell, the second trained model constructed by the learning device is used to discriminate which type of cell the input data indicates. a second discrimination step of
If the discrimination result in the first discrimination step indicates non-cells, the discrimination result is output, and if the discrimination result in the first discrimination step indicates cells, the discrimination result in the second discrimination step is output. A cell discrimination method comprising: a result output step. A learning program for causing a computer to function as the learning device according to any one of claims 1 to 4. A cell discrimination program for causing a computer to function as the cell discrimination apparatus according to claim 5 or 6.

Images