Discrimination device, discrimination method, discrimination program, and recording medium

Abstract

The present invention assesses, with excellent precision, whether an object being assessed is a blood cell.　An assessment device 10 assesses whether an object being assessed is a blood cell on the basis of an image of the object being assessed, said assessment device 10 comprising: a blood cell feature extraction unit 11 that acquires an image of a blood cell and extracts features of the blood cell from the acquired image; an object-being-assessed feature extraction unit 12 that acquires the image of the object being assessed and extracts features of the object being assessed from the acquired image; and an assessment unit 13 that assesses, through abnormality detection and on the basis of the extracted features of the blood cell and the extracted features of the object being assessed, whether the object being assessed is a blood cell.

Description

[Technical field]

[0001] The present invention relates to a discrimination device, a discrimination method, a discrimination program, and a recording medium for discriminating whether or not an object to be discriminated is a blood cell based on an image of the object. [Background technology]

[0002] Cells that have been released from primary tumor tissue or metastatic tumor tissue and infiltrated into the blood are called circulating tumor cells (CTCs). These CTCs are present in extremely small amounts in the peripheral blood of solid cancer patients and are thought to be involved in metastasis, and research into them has been actively conducted in recent years. On the other hand, since most of the nucleated cells in peripheral blood are white blood cells, it is important to distinguish between white blood cells and cancer cells. In clinical applications of CTCs, it has been reported that in breast cancer patients, if there are less than five CTCs in 7.5 mL of whole blood, the one-year mortality rate was 19%, and if there are five or more CTCs, the one-year mortality rate was 53%. Thus, identifying and testing CTCs is considered to have high clinical application value, such as being useful for predicting prognosis. It has also been reported that, depending on the type of cancer, it is difficult to distinguish between white blood cells and cancer cells based on their size. A method for identifying cells using morphological information of cells (image information that reflects not only the external shape but also the internal structure) has been reported (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2016 / 017533 Summary of the Invention [Problem to be solved by the invention]

[0004] The method shown in Patent Document 1 requires that images of all cell types to be identified are prepared in advance and trained. Due to the rarity of CTCs (said to be present at a ratio of one per one million white blood cells), it is difficult to prepare a large amount of CTCs themselves as training data (teaching data). Therefore, Patent Document 1 uses images of cultured cancer cells instead of images of CTCs. It is said that CTCs and cultured cancer cells are very similar, but this is thought to depend on the type of cancer. Therefore, when discrimination is performed using images of cultured cancer cells, there is a risk that discrimination cannot always be performed with high accuracy.

[0005] One embodiment of the present invention has been made in consideration of the above, and aims to provide a discrimination device, a discrimination method, a discrimination program, and a recording medium that can accurately determine whether or not a cell is a blood cell. [Means for solving the problem]

[0006] In order to achieve the above-mentioned object, a discrimination device according to one embodiment of the present invention is a discrimination device that discriminates whether an object to be discriminated is a blood cell or not based on an image of the object to be discriminated, and includes a blood cell feature extraction means that acquires an image of the blood cell and extracts features of the blood cell from the acquired image, a discrimination object feature extraction means that acquires an image of the object to be discriminated and extracts features of the object to be discriminated from the acquired image, and a discrimination means that discriminates whether the object to be discriminated is a blood cell or not by abnormality detection based on the features of the blood cells extracted by the blood cell feature extraction means and the features of the object to be discriminated extracted by the discrimination object feature extraction means.

[0007] In the discrimination device according to one embodiment of the present invention, whether or not the object to be discriminated is a blood cell is discriminated by abnormality detection based on the characteristics of blood cells and the characteristics of the object to be discriminated. As a result, for example, when discriminating whether the object to be discriminated is a white blood cell or a CTC, it is only necessary to extract the characteristics of the white blood cell, and discrimination can be performed without preparing the characteristics of the CTC itself or the characteristics of the cultured cancer cells. In other words, discrimination can be performed without preparing the characteristics of non-blood cells. Therefore, according to the discrimination device according to one embodiment of the present invention, for example, a decrease in the accuracy of discrimination when the characteristics of the cultured cancer cells are used can be prevented. In other words, according to the discrimination device according to one embodiment of the present invention, discrimination as to whether or not the object is a blood cell can be performed with high accuracy.

[0008] The discrimination means may discriminate whether the object to be discriminated is a blood cell or not by using a local outlier factor (LOF) method or a one-class support vector machine (SVM) as an abnormality detection. With this configuration, discrimination can be performed appropriately and reliably.

[0009] The blood cell characteristic extraction means may obtain an image of the optical thickness distribution of the blood cells, and the discrimination object characteristic extraction means may obtain an image of the optical thickness distribution of the discrimination object. With this configuration, characteristics for performing discrimination with high accuracy can be reliably extracted, and discrimination can be performed appropriately and reliably.

[0010] The blood cells may include white blood cells. According to this configuration, for example, it is possible to accurately estimate whether the object to be discriminated is a CTC.

[0011] The blood cell characteristic extraction means extracts characteristics of a plurality of blood cells, and the discrimination means performs clustering of the plurality of blood cells based on the characteristics of the plurality of blood cells, and Variation of features by cluster after Based on the result, blood cells to be used for determining whether the object to be determined is a blood cell or not are selected. do.According to this configuration, the characteristics of the blood cells used for discrimination can be made more appropriate, and as a result, discrimination can be performed with higher accuracy.

[0012] Incidentally, one embodiment of the present invention can be described not only as an invention of a discrimination device as described above, but also as an invention of a discrimination method, a discrimination program, and a recording medium as described below. These are essentially the same invention, except for being in different categories, and have the same functions and effects.

[0013] That is, a discrimination method according to one embodiment of the present invention is a discrimination method which is an operating method of a discrimination device which discriminates whether or not an object to be discriminated is a blood cell based on an image of the object to be discriminated, and includes a blood cell feature extraction step of acquiring an image of the blood cell and extracting features of the blood cell from the acquired image, a discrimination object feature extraction step of acquiring an image of the object to be discriminated and extracting features of the object to be discriminated from the acquired image, and a discrimination step of discriminating whether or not the object to be discriminated is a blood cell by abnormality detection based on the features of the blood cell extracted in the blood cell feature extraction step and the features of the object to be discriminated extracted in the discrimination object feature extraction step. In the blood cell feature extraction step, features of a plurality of blood cells are extracted, and in the discrimination step, clustering of the plurality of blood cells is performed based on the features of the plurality of blood cells, and blood cells to be used for discriminating whether or not the discrimination target is a blood cell are selected based on the variation in features for each cluster after clustering. .

[0014] A discrimination program according to one embodiment of the present invention is a discrimination program that causes a computer to operate as a discrimination device that discriminates whether or not an object to be discriminated is a blood cell based on an image of the object to be discriminated, and causes the computer to function as: a blood cell characteristic extraction means that acquires an image of a blood cell and extracts characteristics of the blood cell from the acquired image; a discrimination object characteristic extraction means that acquires an image of the object to be discriminated and extracts characteristics of the object to be discriminated from the acquired image; and a discrimination means that discriminates whether or not the object to be discriminated is a blood cell by abnormality detection based on the characteristics of the blood cell extracted by the blood cell characteristic extraction means and the characteristics of the object to be discriminated extracted by the discrimination object characteristic extraction means. The blood cell characteristic extraction means extracts characteristics of a plurality of blood cells, and the discrimination means performs clustering of the plurality of blood cells based on the characteristics of the plurality of blood cells, and selects blood cells to be used for discriminating whether the discrimination object is a blood cell or not based on the variation in characteristics for each cluster after the clustering. .

[0015] A recording medium according to one embodiment of the present invention is a computer-readable recording medium having the above-mentioned discrimination program recorded thereon. Effect of the Invention

[0016] According to one embodiment of the present invention, it is possible to accurately determine whether a cell is a blood cell or not. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 is a diagram showing a configuration of a discrimination device according to an embodiment of the present invention. [Diagram 2] FIG. 1 is a diagram showing an outline of determination in an embodiment of the present invention. [Diagram 3] 4 is a flowchart showing a discrimination method which is a process executed by the discrimination device according to the embodiment of the present invention. [Figure 4] 1 is a graph illustrating a receiver operating characteristic (ROC) curve according to an embodiment of the present invention. [Diagram 5] 1 is a graph showing the location of clustered white blood cell features. [Figure 6] 13 is a graph showing an ROC curve based on discrimination using clustered white blood cells. [Figure 7] FIG. 13 is a diagram showing a configuration of a discrimination program according to an embodiment of the present invention, together with a recording medium. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] Hereinafter, an embodiment of a discrimination device, a discrimination method, a discrimination program, and a recording medium according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are given the same reference numerals, and duplicated description will be omitted.

[0019] FIG. 1 shows a discrimination device 10 according to the present embodiment. The discrimination device 10 is a device that discriminates (identifies) whether or not an object to be discriminated is a blood cell based on an image of the object to be discriminated. For example, the discrimination device 10 discriminates whether or not the cells in the blood of a subject are white blood cells, using cells in the blood of the subject as the object to be discriminated. If the cells in the blood are discriminated to be not white blood cells, the cells in the blood are presumed to be CTCs. That is, the discrimination device 10 estimates whether or not the cells in the blood of the subject are CTCs. As described above, the estimation of whether or not CTCs are contained in blood has high clinical application value. Note that the discrimination device 10 is not limited to one that performs the above discrimination, and may be one that discriminates whether or not the object to be discriminated is a predetermined blood cell.

[0020] The discrimination device 10 is, for example, a conventional computer including hardware such as a CPU (Central Processing Unit), a memory, and a communication module. The discrimination device 10 may also be a computer system including multiple computers. The discrimination device 10 may also be configured using cloud computing. Each function of the discrimination device 10, which will be described later, is achieved by these components operating through a program or the like.

[0021] The imaging device 20 is a device that captures blood cells and a discrimination object to obtain an image used for discrimination in the discrimination device 10. For example, the image obtained by the imaging device 20 is a quantitative phase image. The quantitative phase image is an image of the optical thickness distribution of a cell. The optical thickness is the product of the physical length along the traveling direction of light and the refractive index. As the imaging device 20 that obtains the quantitative phase image, for example, an imaging flow cytometer (quantitative phase microscope) as shown in JP 2017-166848 A is used. Note that the image obtained by the imaging device 20 is not limited to the above, and may be any image that can extract the characteristics of blood cells such as white blood cells and discrimination objects. In addition, as the imaging device 20, an imaging device other than a quantitative phase microscope according to the type of image to be obtained may be used.

[0022] Specifically, the imaging device 20 captures the image of a sample obtained from blood. There are two types of images captured by the imaging device 20: an image used as a reference for discrimination, and an image to be discriminated, i.e., an image showing a discrimination target. The image used as the reference for discrimination is an image of white blood cells, and is an image of a sample that does not contain CTCs, i.e., contains only white blood cells. Usually, images of a plurality of white blood cells are used as the image used as the reference for discrimination. Therefore, the imaging device 20 captures images of a plurality of white blood cells. The image used as the reference for discrimination can be an image of a sample obtained from the subject's blood. Alternatively, when it can be assumed that the morphology of white blood cells does not vary between individuals, an image of a sample obtained from blood other than that of the subject may be used. The image to be discriminated is an image of a sample obtained from the subject's blood.

[0023] The discrimination device 10 and the imaging device 20 are connected to each other so that an image can be transmitted from the imaging device 20 to the discrimination device 10. The imaging device 20 transmits the acquired image to the discrimination device 10.

[0024] Next, a description will be given of the functions of the discrimination device 10 according to this embodiment. As shown in FIG. 1, the discrimination device 10 includes a blood cell characteristic extraction unit 11, a discrimination object characteristic extraction unit 12, and a discrimination unit 13.

[0025] The blood cell characteristic extraction unit 11 is a blood cell characteristic extraction means that acquires an image of a blood cell and extracts characteristics of the blood cell from the acquired image. The blood cell characteristic extraction unit 11 may acquire an image of an optical thickness distribution of the blood cell. Specifically, the blood cell characteristic extraction unit 11 extracts the characteristics of the blood cell as follows.

[0026] The blood cell feature extraction unit 11 acquires an image to be used as a reference for discrimination, i.e., an image of a white blood cell, transmitted from the imaging device 20. Note that the acquisition of the image does not necessarily have to be performed by receiving the image transmitted from the imaging device 20, and may be performed by an input operation by the user of the discrimination device 10, etc.

[0027] The blood cell feature extraction unit 11 extracts features of white blood cells from the acquired image. These features are features (morphological information) related to the morphology of white blood cells. The features may reflect at least one of the shape and the internal structure that can be confirmed from the outside. The features of white blood cells are, for example, spatial variations in optical thickness. The spatial variations in optical thickness are, for example, both or either of the gradient magnitude and gradient direction of a vector that represents the gradient of the optical thickness at a position in the image of the optical thickness distribution. Specifically, the spatial variations in optical thickness are HOG features that are histograms of the gradient (gradient direction, angle) of the optical thickness for each position of white blood cells in the image obtained by HOG (Histograms of Oriented Gradients). Furthermore, the features of white blood cells may include features that can be extracted from an image other than HOG features in addition to or instead of the HOG features. For example, features may be extracted from an image using a differential filter (Sobel filter).

[0028] The features of white blood cells are extracted as vectors of multiple dimensions. For example, in the case of HOG features, they are extracted as vectors of values ​​for each gradient (angle). For example, when 64 values ​​are extracted as HOG features, the vector becomes 64 dimensions. Features are extracted from an image by a conventional method. For example, when extracting HOG features, the method shown in Patent Document 1 may be used.

[0029] When the portion of the image in which the white blood cells to be extracted are included is a part of the acquired image, the blood cell feature extraction unit 11 may cut out the area (range) in which the white blood cells are included from the acquired image, i.e., the area to be extracted, and extract the features of the white blood cells from the cut-out image area. The above-mentioned area may be cut out by a conventional method. For example, the area may be cut out by object detection using machine learning. Alternatively, the area may be cut out according to a rule (rule-based) by setting a threshold value for the brightness value of the image.

[0030] The blood cell feature extraction unit 11 extracts features for each white blood cell from an image of multiple white blood cells. When multiple white blood cells are captured in one image, features may be extracted by cutting out an area from the image for each white blood cell. The blood cell feature extraction unit 11 outputs information indicating the extracted white blood cell features (for example, the above-mentioned vector) to the discrimination unit 13.

[0031] The discrimination object feature extraction unit 12 is a discrimination object feature extraction means that acquires an image of the discrimination object and extracts features of the discrimination object from the acquired image. The discrimination object feature extraction unit 12 may acquire an image of the optical thickness distribution of the discrimination object. The discrimination object feature extraction unit 12 acquires an image to be discriminated transmitted from the imaging device 20, i.e., an image showing the discrimination object. The acquisition of the image does not necessarily have to be performed by receiving one transmitted from the imaging device 20, as in the case of the acquisition of an image by the blood cell feature extraction unit 11, but may be performed by an input operation by the user of the discrimination device 10.

[0032] The discrimination object feature extraction unit 12 extracts features of the discrimination object in the same manner as the blood cell feature extraction unit 11 extracts features. There may be multiple discrimination objects from which features are extracted (i.e., discrimination objects). In that case, discrimination is performed for each discrimination object. The discrimination object feature extraction unit 12 outputs information indicating the extracted features of the discrimination object (for example, the above-mentioned vector) to the discrimination unit 13.

[0033] The discrimination unit 13 is a discrimination means that discriminates whether or not the discrimination object is a blood cell by abnormality detection based on the characteristics of the blood cells extracted by the blood cell characteristic extraction unit 11 and the characteristics of the discrimination object extracted by the discrimination object characteristic extraction unit 12. The discrimination unit 13 may discriminate whether or not the discrimination object is a blood cell by LOF or One class SVM as the abnormality detection.

[0034] Here, the discrimination in this embodiment will be compared with the discrimination shown in Patent Document 1. In Patent Document 1, a discriminator (classifier) ​​is created by machine learning such as SVM using features extracted from an image as training data to discriminate between white blood cells and CTCs, and discrimination is performed. The features of white blood cells and the features of cultured cancer cells are used as learning data. The features of cultured cancer cells are used as a substitute for the features of CTCs. This is because, as described above, it is difficult to prepare a large amount of CTCs as learning data.

[0035] Figure 2 illustrates the characteristics of white blood cells and the characteristics of cultured cancer cells. For simplicity, Figure 2 shows the characteristics as two-dimensional data. The classifier created by the method shown in Patent Document 1 corresponds to boundary B1 for making the distinction shown in Figure 2. Boundary B1 is generated so as to maximize the difference between the characteristics of white blood cells and the characteristics of cultured cancer cells.

[0036] The method shown in Patent Document 1 is based on the premise that the characteristics of CTCs and the characteristics of cultured cancer cells are very similar. However, depending on the type of cancer, it is considered that the characteristics of CTCs and the characteristics of cultured cancer cells are not necessarily very similar. In that case, as shown in FIG. 2, the position of the CTC characteristic in the feature space may be away from the position of the cultured cancer cell characteristic and may be closer to the white blood cell side than the boundary B1. In this case, the CTC is actually misclassified as a white blood cell.

[0037] In the present embodiment, discrimination is performed by anomaly detection based on the characteristics of white blood cells without using the characteristics corresponding to CTCs (in the above example, the characteristics of cultured cancer cells). That is, discrimination is performed by anomaly detection without using the characteristics of non-blood cells. For example, as shown in FIG. 2, a boundary B2 is used for discrimination. The boundary B2 is based on the distribution of the positions of the characteristics of white blood cells. When this is used, the position in the feature space of the CTC characteristics shown in FIG. 2 is not on the side closer to the white blood cells than the boundary B2, and misdiscrimination can be prevented. Note that the discrimination in the present embodiment may be performed by anomaly detection based on the characteristics of blood cells without using the characteristics of non-blood cells, and does not necessarily have to generate the boundary B2.

[0038] Specifically, the discrimination unit 13 performs discrimination as follows. The discrimination unit 13 inputs information indicating the characteristics of white blood cells from the blood cell characteristic extraction unit 11. The discrimination unit 13 inputs information indicating the information indicating the characteristics of the discrimination object from the discrimination object characteristic extraction unit 12. From the viewpoint of improving processing efficiency, the discrimination unit 13 may compress the dimension of a vector, which is information indicating the characteristics, and perform discrimination using the compressed vector. The dimensionality compression can be performed by a conventional method, for example, by principal component analysis (PCA).

[0039] The discrimination unit 13 performs discrimination by abnormality detection based on the characteristics of white blood cells and the characteristics of the discrimination object, determining that the discrimination object is a white blood cell as normal and that the discrimination object is not a white blood cell as abnormal. When discrimination is performed by LOF, the discrimination unit 13 calculates the LOF score for the judgment object by the following formula.

number

[0040] The LOF score for an object to be judged indicates how abnormal the object is, i.e., the degree of abnormality. The higher the LOF score for an object to be judged, the more abnormal the object to be judged is, i.e., it is not a white blood cell. LOF uses the ratio of average reach distances as an index, so calculations are performed taking into account not only the distribution distance of the white blood cell characteristics in the feature space, but also the variance in the distribution, and can score how abnormal the characteristics of the object to be judged are in the feature space.

[0041] The discrimination unit 13 compares the calculated LOF score with a preset threshold value. If the calculated LOF score exceeds the threshold value, the discrimination unit 13 determines that the object to be discriminated is not a white blood cell. In this case, the object to be discriminated is presumed to be a CTC as described above. On the other hand, if the calculated LOF score is less than the threshold value, the discrimination unit 13 determines that the object to be discriminated is a white blood cell.

[0042] The discrimination unit 13 outputs the discrimination result. For example, the discrimination unit 13 displays information indicating the discrimination result on a display device provided in the discrimination device 10. The display is referred to, for example, by a user of the discrimination device 10. Note that the output by the discrimination unit 13 may be performed in a manner other than the above. For example, the information indicating the discrimination result may be transmitted from the discrimination device 10 to another device for output.

[0043] In the above example, the case where the discrimination is performed by LOF is shown, but the discrimination may be performed by One class SVM. The above discrimination by One class SVM can be performed in the same manner as the conventional method by One class SVM. In addition, the discrimination may be performed by a method other than the above, as long as it uses the features of white blood cells as the judgment criterion (without using the features (images) of CTCs and cultured cancer cells as the judgment criterion) to discriminate whether the discrimination target is a blood cell or not by anomaly detection. Even if the image that is the subject of feature extraction by the blood cell feature extraction unit 11 contains images of non-white blood cells such as CTCs, there is no major problem in the above discrimination as long as the number of images of non-white blood cells is extremely small compared to the number of the entire images. In this embodiment, it is sufficient not to use features that are known in advance as the features of CTCs and cultured cancer cells.

[0044] Furthermore, when performing discrimination by the above-mentioned LOF, the score is calculated using the position in the feature space of each white blood cell feature, but discrimination may be performed by other methods. For example, a discriminator (trained model) for discrimination may be generated by machine learning from the white blood cell features extracted by the blood cell feature extraction unit 11 (without using the features (images) of CTCs and cultured cancer cells), and discrimination may be performed using the generated discriminator. The above is the function of the discrimination device 10 according to this embodiment.

[0045] Next, a discrimination method which is a process executed by the discrimination device 10 according to this embodiment (operation method of the discrimination device 10) will be described with reference to the flowchart of FIG. 3. In this process, the blood cell feature extraction unit 11 acquires an image of a white blood cell (S01, blood cell feature extraction step). Then, the blood cell feature extraction unit 11 extracts features of the white blood cell from the acquired image (S02, blood cell feature extraction step). In addition, the discrimination object feature extraction unit 12 acquires an image of the discrimination object (S03, discrimination object feature extraction step). Then, the discrimination object feature extraction unit 12 extracts features of the discrimination object from the acquired image (S04, discrimination object feature extraction step). It should be noted that the process (S01, S02) for extracting the characteristics of white blood cells by the blood cell characteristic extraction unit 11 and the process (S03, S04) for extracting the characteristics of the object to be discriminated by the discrimination object characteristic extraction unit 12 do not necessarily have to be performed in the above order, as long as the respective characteristics are extracted.

[0046] Next, the discrimination unit 13 calculates an LOF score for the object to be discriminated based on the characteristics of the white blood cells and the characteristics of the object to be discriminated (S05, discrimination step). Next, the discrimination unit 13 discriminates whether the object to be discriminated is a white blood cell or not based on the LOF score (S06, discrimination step). Note that in this process, an example is shown in which LOF is used for abnormality detection, but when another method of abnormality detection is used, a process according to the method is performed. Next, the discrimination unit 13 outputs information indicating the discrimination result (S07). The above is the process executed by the discrimination device 10 according to this embodiment.

[0047] As described above, in this embodiment, whether or not the object to be discriminated is a blood cell is discriminated by anomaly detection based on the characteristics of blood cells (e.g., white blood cells) and the characteristics of the object to be discriminated. As a result, for example, when discriminating whether the object to be discriminated is a white blood cell or a CTC, it is sufficient to extract the characteristics of the white blood cell, and discrimination can be performed without preparing the characteristics of the CTC itself or the characteristics (image) of the cultured cancer cells. That is, discrimination can be performed without preparing the characteristics of non-blood cells. Therefore, according to this embodiment, for example, a decrease in discrimination accuracy when the characteristics of the cultured cancer cells are used can be prevented. That is, according to this embodiment, discrimination whether or not the object is a blood cell can be performed with high accuracy. Furthermore, according to this embodiment, since it is not necessary to prepare the characteristics of the CTC itself or the characteristics (image) of the cultured cancer cells, discrimination can be easily performed.

[0048] Also, as in the above-described embodiment, the determination may be performed using LOF or One class SVM for anomaly detection. With this configuration, the determination can be performed appropriately and reliably. However, as described above, it is not necessary to use LOF or One class SVM, and the determination may be performed using other anomaly detection methods.

[0049] Also, as in the above-described embodiment, the image used for extracting the features may be an image of the optical thickness distribution. With this configuration, the features for performing the discrimination with high accuracy can be extracted reliably, and the discrimination can be performed appropriately and reliably. However, the image used for extracting the features does not necessarily have to be an image of the optical thickness distribution, and any image may be used as long as the features used for the discrimination can be extracted.

[0050] Furthermore, as in the above-described embodiment, the blood cells used to determine whether or not the object to be determined is a blood cell may include a white blood cell. According to this configuration, for example, it is possible to accurately estimate whether or not the object to be determined is a CTC. However, as described above, the blood cells may be any blood cells other than white blood cells. For example, the blood cells may be hematopoietic stem cells (G0 stage), hematopoietic stem cells, lymphoblasts, lymphocytes, precursor cells, monoblasts, monocytes, myeloblasts, promyelocytes, myelocytes, metamyelocytes, band cells, segmented cells, neutrophils, eosinophils, basophils, proerythroblasts, basophilic erythroblasts, polychromatic erythroblasts, normochromatic erythroblasts, erythrocytes, promegakaryocytes, megakaryocytes, etc.

[0051] To confirm the effectiveness of this embodiment, an evaluation was performed using a data set of 1000 images each of white blood cells and cultured cancer cells. FIG. 4 shows the ROC curve for this embodiment (Example, LOF in the figure). For comparison, the ROC curve for the SVM method described in Patent Document 1 (Comparative Example, SVM in the figure) is also shown. Although the shapes of the ROC curves are different, the true positive rate at a false positive rate of 50% is higher in this embodiment. Note that the above evaluation was performed using cultured cancer cells as an alternative, since evaluation using CTCs is actually difficult.

[0052] Next, a modified example of this embodiment will be described. In the above-described embodiment, the discrimination unit 13 uses the features of a plurality of white blood cells extracted by the blood cell feature extraction unit 11. In this modified example, only some of the features of a plurality of white blood cells are used for discrimination.

[0053] This modified example has the following configuration. Any points not described below are the same as those in the above-mentioned embodiment unless inconsistencies arise. As described above, since only the features of some of the white blood cells are used for discrimination, the blood cell feature extraction unit 11 extracts the features of multiple blood cells.

[0054] The discrimination unit 13 performs clustering of the plurality of blood cells based on the characteristics of the plurality of blood cells, and selects blood cells to be used for discriminating whether the discrimination target is a blood cell or not based on the clustering result. Specifically, the discrimination unit 13 selects (characteristics of) white blood cells to be used for discrimination as follows.

[0055] The discrimination unit 13 performs clustering based on the position of the feature of each white blood cell in the feature space. The clustering can be performed by a conventional method, for example, the K-means method. The number of clusters to be generated may be set in advance or may be determined (automatically) by the clustering process. The discrimination unit 13 selects white blood cells belonging to any of the clusters obtained by clustering as white blood cells to be used for discrimination (selection will be described in detail later). The discrimination unit 13 performs discrimination in the same manner as in the above-mentioned embodiment using the features of the selected white blood cells.

[0056] Figure 5(a) shows an example graph of white blood cells clustered and visualized by PCA. Each plotted point in the graph in Figure 5(a) indicates the position of a white blood cell feature obtained from an image of a single white blood cell. In the example shown in Figure 5(a), white blood cells are clustered into three clusters: cluster 0, cluster 1, and cluster 2. Figure 5(b) shows a graph in which points of white blood cell features are overlaid with points of cultured cancer cells (HCT).

[0057] FIG. 6(a) shows the ROC curve when discrimination is performed using all the white blood cells in the example shown in FIG. 5 (i.e., when discrimination is performed according to the above-mentioned embodiment). FIG. 6(b), (c), and (d) show the ROC curves when discrimination is performed using the white blood cells belonging to clusters 0, 1, and 2 shown in FIG. 5, respectively. Looking at the ROC curves shown in FIG. 6, cluster 0 has the best performance. The range in which the positions of the characteristics of white blood cells in clusters 1 and 2 are distributed and the range in which the positions of the characteristics of cultured cancer cells are distributed overlap more than in cluster 0. In addition, the positions of the characteristics of white blood cells in clusters 1 and 2 are more scattered than in cluster 0. From these, cluster 0 is considered to be a typical white blood cell group, since the positions of the characteristics of white blood cells are clustered to some extent. Clusters 1 and 2 are considered to be groups that are not typical white blood cells or cannot be distinguished from non-white blood cells, since the positions of the characteristics of white blood cells are not clustered.

[0058] Furthermore, when the images of white blood cells in each cluster were examined, the images of white blood cells in cluster 0 were typical images of white blood cells. In cluster 1, there were many images in which multiple cells were captured. In cluster 2, there were many images of cells with more complex internal structures than other cells (images of cells suspected to be cancer cells).

[0059] The discrimination unit 13 may perform discrimination for each cluster using test data such as the positions of characteristics of cultured cancer cells prepared in advance, and may select the cluster with the best performance (e.g., the AUC (Area Under the Curve) of the ROC curve) as the cluster to be used for discrimination. Alternatively, the discrimination unit 13 may calculate the variation in the positions of characteristics for each cluster, and select a cluster based on the variation. The variation is, for example, at least one of the variance and the average distance for each cluster. For example, the cluster with the smallest variation is selected. Alternatively, the cluster with the largest number of elements is selected.

[0060] By selecting blood cells such as white blood cells to be used for discrimination based on clusters as described above, the features of the blood cells to be used for discrimination can be made more appropriate. For example, only features extracted from images of typical white blood cells can be used for discrimination, and images of cells whose features overlap with those of cultured cancer cells can be excluded. In other words, a high-quality white blood cell dataset can be constructed. As a result, discrimination can be performed more accurately, that is, discrimination performance can be improved.

[0061] Next, a description will be given of a discrimination program for executing the above-mentioned series of processes by the discrimination device 10. As shown in Fig. 7, the discrimination program 40 is stored in a program storage area 31 formed in a computer-readable recording medium 30 that is inserted into a computer and accessed, or that is provided in the computer. The recording medium 30 may be a non-transitory recording medium.

[0062] The discrimination program 40 is configured to include a blood cell characteristic extraction module 41, a discrimination object characteristic extraction module 42, and a discrimination module 43. Functions realized by executing the blood cell characteristic extraction module 41, the discrimination object characteristic extraction module 42, and the discrimination module 43 are similar to the functions of the blood cell characteristic extraction unit 11, the discrimination object characteristic extraction unit 12, and the discrimination unit 13 of the discrimination device 10 described above, respectively.

[0063] The discrimination program 40 may be configured such that a part or the whole of it is transmitted via a transmission medium such as a communication line, and is received and recorded (including installed) by another device. Also, each module of the discrimination program 40 may be installed in one of a plurality of computers, not just one computer. In that case, the above-mentioned series of processes are performed by a computer system consisting of the plurality of computers. [Explanation of symbols]

[0064] 10...discrimination device, 11...blood cell characteristic extraction unit, 12...discrimination object characteristic extraction unit, 13...discrimination unit, 20...imaging device, 30...recording medium, 31...program storage area, 40...discrimination program, 41...blood cell characteristic extraction module, 42...discrimination object characteristic extraction module, 43...discrimination module.

Claims

1. A discrimination device for discriminating whether an object to be discriminated is a blood cell or not based on an image of the object, comprising: a blood cell characteristic extraction means for acquiring an image of a blood cell and extracting characteristics of the blood cell from the acquired image; A discrimination object feature extraction means for acquiring an image of the discrimination object and extracting features of the discrimination object from the acquired image; a discrimination means for discriminating whether or not the object to be discriminated is a blood cell by detecting an abnormality based on the characteristics of the blood cells extracted by the blood cell characteristic extraction means and the characteristics of the object to be discriminated extracted by the discrimination object characteristic extraction means; Equipped with The blood cell characteristic extraction means extracts characteristics of a plurality of blood cells, The discrimination means is a discrimination device that performs clustering of a plurality of blood cells based on the characteristics of the plurality of blood cells, and selects blood cells to be used for discriminating whether or not the discrimination target is a blood cell based on the variation in the characteristics of each cluster after clustering.

2. 2. The discrimination device according to claim 1, wherein the discrimination means discriminates whether the discrimination object is a blood cell or not by LOF or One-class SVM as the abnormality detection.

3. The blood cell characteristic extraction means acquires an image of an optical thickness distribution of blood cells, 3. The discrimination device according to claim 1, wherein the discrimination object feature extraction means acquires an image of an optical thickness distribution of the discrimination object.

4. 4. The discrimination device according to claim 1, wherein the blood cells include white blood cells.

5. A discrimination method which is an operating method of a discrimination device for discriminating whether or not an object to be discriminated is a blood cell based on an image of the object to be discriminated, comprising: a blood cell feature extraction step of acquiring an image of a blood cell and extracting features of the blood cell from the acquired image; A discrimination object feature extraction step of acquiring an image of the discrimination object and extracting features of the discrimination object from the acquired image; a discrimination step of discriminating whether or not the object to be discriminated is a blood cell by abnormality detection based on the characteristics of the blood cells extracted in the blood cell characteristic extraction step and the characteristics of the object to be discriminated extracted in the discrimination object characteristic extraction step; Including, In the blood cell characteristic extraction step, characteristics of a plurality of blood cells are extracted; In the discrimination step, a plurality of blood cells are clustered based on the characteristics of the plurality of blood cells, and blood cells to be used for discriminating whether or not the object to be discriminated is a blood cell are selected based on the variation in the characteristics of each cluster after clustering.

6. A discrimination program for causing a computer to operate as a discrimination device for discriminating whether or not an object to be discriminated is a blood cell based on an image of the object, comprising: The computer, a blood cell characteristic extraction means for acquiring an image of a blood cell and extracting characteristics of the blood cell from the acquired image; A discrimination object feature extraction means for acquiring an image of the discrimination object and extracting features of the discrimination object from the acquired image; a discrimination means for discriminating whether or not the object to be discriminated is a blood cell by detecting an abnormality based on the characteristics of the blood cells extracted by the blood cell characteristic extraction means and the characteristics of the object to be discriminated extracted by the discrimination object characteristic extraction means; Function as a The blood cell characteristic extraction means extracts characteristics of a plurality of blood cells, The discrimination means is a discrimination program that performs clustering of a plurality of blood cells based on the characteristics of the plurality of blood cells, and selects blood cells to be used for discriminating whether or not the object to be discriminated is a blood cell based on the variation in the characteristics of each cluster after clustering.

7. A computer-readable recording medium having the discrimination program according to claim 6 recorded thereon.

Images