Cancer cell identification method and cancer cell identification device

Abstract

Disclosed are a cancer cell identification method and cancer cell identification device with which it is possible to easily and precisely perform cancer cell identification for identifying and diagnosing lung cancer. Cells to be identified are captured using a CTC-chip, and the captured cells to be identified are fluorescently stained using a prescribed antibody. In-blood cell images of the stained cells to be identified are captured, and features of the in-blood cell images are clustered. An anomaly detector for determining whether the features are minority cell features or majority cell features is trained by deep learning via an Efficient-GAN technique using in-blood cell images sampled from healthy individuals and cancer patients. Majority cell features are determined on the basis of the graph morphologies of luminance histograms of the intensities of a plurality of morphological features and the occurrence frequencies thereof. Minority cell features are determined by using a minority cell detector to perform minority cell detection in the in-blood cell images of the cells to be identified, clustering the in-blood cell images of minority cells by means of a prescribed criterion, and using the probability of the appearance of each classified cluster.

Description

Cancer cell testing method and cancer cell testing device

[0001] The present invention relates to a cancer cell testing method and a cancer cell testing device that utilize machine learning to screen for cancer cells from cell images and perform cancer testing.

[0002] Conventionally, the most common method for detecting lung cancer involves first screening with X-rays, and then, if a positive result is detected, a definitive diagnosis is made using CT scans or biopsies. However, X-rays, the current mainstream screening method, have high specificity but low sensitivity, resulting in an issue of approximately 40% of cancer patients remaining undetected. Early diagnosis is an important factor in lung cancer treatment that directly affects survival rates, and therefore, more sensitive screening is needed to reduce the number of undetected cases.

[0003] In addition, cancer testing methods using liquid biopsies are used to perform various diagnoses using exosomes and circulating tumor cells (CTCs) extracted from bodily fluids such as blood and urine, and research is also being conducted into lung cancer diagnosis due to its minimally invasive characteristics. In lung cancer, CTCs and other cells are used as markers for predicting prognosis and screening, but the current situation is that predictions and genetic analysis are limited to a certain number of cells.

[0004] As another cancer testing method, a method for early diagnosis and post-treatment monitoring of breast cancer using multiple cancer gene biomarkers has been proposed, as shown in Patent Document 1. Specifically, the method disclosed includes the steps of (a) extracting mRNA from exosomes isolated from biological samples of normal individuals and subjects, (b) measuring the mRNA levels of adenine nucleotide translocase 2 and voltage-dependent anion-selective channel 1 genes using the extracted mRNA as a template, and (c) determining that breast cancer exists when the mRNA levels of ANT2 and VDAC1 genes are increased compared to normal individuals.

[0005] Meanwhile, in recent years, machine learning has become increasingly popular as a component technology of AI (Artificial Intelligence). Deep learning, in particular, has been widely applied to fields such as image processing, natural language processing, and speech recognition. Furthermore, there is a type of generative AI called generative adversarial networks (GAN). GANs, which are generative models, can generate non-existent data or transform the features of existing data by learning features from data. GANs are an unsupervised learning method that learns features without providing correct answer data.

[0006] One of the suitable applications of deep learning is anomaly detection in images, etc. Supervised learning in machine learning is difficult to apply when there are countless abnormal patterns, and in such cases, unsupervised learning, which does not use labels and uses only normal examples as training data, is effective. Various methods have been proposed that use GANs as an unsupervised learning method to improve anomaly detection accuracy compared to conventional machine learning methods.

[0007] Therefore, by combining AI-based machine learning with lung cancer testing, it is expected that highly accurate lung cancer testing will be possible using not only the number of CTCs but also characteristics from various other blood cells.

[0008] A digital pathology image processing system disclosed in Patent Literature 2 (PTL 2) has been proposed as a cancer testing method using AI. Specifically, a digital pathology image of a tissue sample is received and subdivided into multiple patches. The digital pathology image of the tissue sample is a whole-slide scan image of a tumor sample from a patient diagnosed with non-small cell lung cancer (NSCLC), and the digital pathology image or the entire slide image is a hematoxylin and eosin (H&E)-stained image. This method includes, for each patch, identifying image features detected within the patch and using a machine learning model to generate one or more labels corresponding to the identified image features within the patch. The machine learning model can be a deep learning neural network. In one embodiment, the image features include a tissue image, and the one or more labels applied to the patch include cancer regions of adenocarcinoma (ADC) and squamous cell carcinoma (SCC). This method also determines a heterogeneity metric for the tissue sample based on the generated labels, and the heterogeneity metric is used to evaluate the degree of heterogeneity of the identified image features and corresponding labels in the tissue sample.

[0009] Patent No. 7187081 Publication Special Publication No. 2023-534448

[0010] Hyunku Shin, Seunghyun Oh, Soonwoo Hong, Minsung Kang, DaehyeonKang, Yong-gu Ji, Byeong Hyeon Choi, Ka-Won Kang, Hyesun Jeong, Yong Park,Sunghoi Hong, Hyun Koo Kim, and Yeonho Choi “Early-Stage Lung Cancer Diagnosisby Deep Learning-Based Spectroscopic Analysis of Circulating Exosomes” Nano2020 14 (5), 5435-5444

[0011] The diagnostic method disclosed in Patent Document 1 involves measuring mRNA levels using gene profiling to diagnose breast cancer in order to reduce the number of undetected cases through more sensitive screening, but this requires specialized equipment, involves many steps in the testing and diagnosis, and is expensive to diagnose.

[0012] The method disclosed in the non-patent document diagnoses lung cancer from cells in the blood, but the target is circulating exosomes, and the method for extracting features from circulating exosomes is spectral analysis using surface-enhanced Raman scattering, which costs several tens of thousands of yen per test.

[0013] The AI-based cancer screening method disclosed in Patent Literature 2 enables more accurate screening by training a machine learning model (e.g., a deep learning neural network) to identify image features and generate labels corresponding to the identified image features shown in multiple patches from digital pathology images. However, the machine learning model is trained based on pre-labeled or pre-annotated tumor regions in a set of digital pathology images by human experts. For example, the tumor regions are manually selected by pathologists, physicians, clinical specialists, lung cancer diagnosis experts, etc. Therefore, there is a need for a system that enables easier and more accurate diagnosis.

[0014] This invention has been made in consideration of the above-mentioned background art, and aims to provide a cancer cell testing method and a cancer cell testing device that can perform cancer cell testing more easily and with higher accuracy in the testing and diagnosis of lung cancer, and can accurately screen cancer cells.

[0015] This invention uses deep abnormality detection technology to detect small numbers of cells contained in blood cell images in order to obtain information useful for testing. This cancer cell testing method uses machine learning to perform lung cancer testing and other tests using features obtained from the small number of cells and features obtained from other large numbers of cells, thereby obtaining information on cells that appear only in cancer patients and enabling highly accurate predictions.

[0016] (Claim 1) This invention relates to a cancer cell testing method for testing the possibility of cancer cells using blood cell images taken of collected blood, which comprises capturing a group of cells to be tested using a CTC-chip, fluorescently staining the captured group of cells to be tested with a predetermined antibody, and capturing blood cell images of the stained group of cells to be tested; thereafter, clustering is performed in machine learning processing on the features of the blood cell images, and the features are determined to be minority cell features and majority cell features; beforehand, training of an anomaly detector by Efficient-GAN is performed using blood cell images sampled from healthy subjects and cancer patients by deep learning; for the majority cell features in the blood cell images, a brightness histogram of brightness and its occurrence frequency is formed from a plurality of morphological features of the blood cell images, and multiple types of the majority cell features are obtained in the form of a graph, and the majority cell features are identified using a predetermined majority cell feature threshold; The cancer cell testing method includes: detecting minority cells in the blood cell image of the test subject cell group using a minority cell detector; clustering the blood cell image of the minority cells according to a predetermined criterion; determining the minority cell features using the occurrence probability of each classified cluster; removing white blood cell images and other noise images from the blood cell image; identifying the minority cell features for the test subject cell group using a predetermined minority cell feature threshold; and determining that the test subject cell group determined to be abnormal based on the majority cell feature threshold and the minority cell feature threshold is likely to be cancerous based on the identified majority cell feature and minority cell feature.

[0017] (Claim 2) The antibodies used for the fluorescent staining include three types: Hoechst, which is expressed in the DNA of cell nuclei; CD45, which is expressed in white blood cells; and CK, which is expressed in epithelial cells.

[0018] (Claim 3) The multiple cell features consist of multiple types of average brightness, height of the brightness histogram, width of the brightness histogram, median of the brightness histogram, skewness of the brightness histogram, and kurtosis of the brightness histogram for each of Hoechst, CD45, and CK in the blood cell image.

[0019] (Claim 4) The majority cell feature threshold is set based on the difference in the occurrence frequency or the brightness value of the brightness histogram for the majority cell feature between a healthy subject and a cancer patient.

[0020] (Claim 5) The removal of the white blood cell images appearing in the blood cell image using the small number of cells detector removes white blood cells appearing singly in the blood cell image, and further removes white blood cells appearing in multiple instances in the blood cell image.

[0021] (Claim 6) The small number of cell features are clustered, and the trained anomaly detector determines whether a cell is likely to be cancerous based on the differences in the small number of cell features between healthy individuals and cancer patients.

[0022] (Claim 7) The removal of the white blood cells involves, in a first step, determining a cell number threshold based on the outliers and the number of cells appearing in the image, binarizing the image and counting the number of detected cellular components, setting the cell number threshold so that cells that appear alone in the image and have a low outlier are considered normal, next, in order to remove images that contain multiple cells that were considered normal in the first step, determining an outlier threshold for detecting cells that are specifically emitting light by the fluorescent staining, using the area of ​​the cell at the center of the image as a criterion for determining the outlier threshold, determining the outliers and area of ​​each cell, and sorting the outliers in ascending order, using the average cell area, which is the cell's outlier value, as a second threshold to remove the white blood cell images, and clustering the test cell group that has an outlier value equal to or greater than the second threshold, and identifying cancer cells from the clusters obtained by this process using the occurrence probability compared to that of a normal person as a feature.

[0023] (Claim 8) This invention relates to a cancer cell testing device that uses blood cell images taken of collected blood to test for the possibility of cancer cells, and includes: a blood cell image capturing device that captures a group of cells to be tested using a CTC-chip, fluorescently stains the captured group of cells to be tested with a predetermined antibody, and captures blood cell images of the stained group of cells to be tested; and a feature extraction device that performs clustering on the features of the blood cell images, and obtains minority cell features and majority cell features as the features; and the feature extraction device has previously trained an anomaly detector using Efficient-GAN by deep learning using blood cell images sampled from healthy subjects and cancer patients, The feature extraction device comprises a processing device that forms a brightness histogram of brightness and its occurrence frequency among multiple morphological features of the blood cell image for the majority cell features in the blood cell image, acquires multiple types of the majority cell features in the form of a graph, identifies the majority cell features based on a predetermined majority cell feature threshold, detects minority cells in the blood cell image of the test subject cell group using a minority cell detector, clusters the blood cell image of the minority cells based on a predetermined criterion, obtains the minority cell features using the appearance probability of each classified cluster, removes white blood cell images and other noise images from the blood cell image, and identifies the minority cell features for the test subject cell group based on the predetermined minority cell feature threshold, and is a cancer cell testing device that can identify whether or not the test subject cell group determined to be abnormal based on the majority cell feature threshold and the minority cell feature threshold are cancerous cells based on the identified majority cell features and minority cell features.

[0024] The cancer cell testing method using the blood imaging testing system of the present invention enables more sensitive testing than the conventional X-ray testing method, and can test with only about 1 ml of blood, making lung cancer testing simple and inexpensive. The cancer cell testing method of the present invention, when used in combination with the conventional method, has been shown to reduce the number of undetected cases by 74.9%, enabling extremely high-precision testing.

[0025] 1 is a flowchart showing a cancer cell testing method according to an embodiment of the present invention; 2 is a diagram showing the order of capturing blood cell images in the cancer cell testing method according to an embodiment of the present invention; 3 is a diagram showing a fluorescent stained image and an antibody used in the cancer cell testing method according to an embodiment of the present invention; 4 is a diagram showing a phase-contrast image without staining used in the cancer cell testing method according to an embodiment of the present invention; 5 is a schematic diagram showing detection of lung cancer cells (lung cancer) and the like using AI from a fluorescent stained image and a phase-contrast image using a minority cell feature and a majority cell feature in the cancer cell testing method according to an embodiment of the present invention; 6 is a diagram showing histograms of fluorescence intensity and occurrence frequency of each fluorescence channel and six types of feature amounts which are morphological features obtained from three types of histograms in the cancer cell testing method according to an embodiment of the present invention; 7 is a diagram showing detection of 21 types of majority cell feature amounts which are morphological features obtained from histograms of each fluorescence channel, cell area, and number of CTCs in the cancer cell testing method according to an embodiment of the present invention; 8 is a schematic diagram showing detection of minority cell feature amounts in the cancer cell testing method according to an embodiment of the present invention; and 9 is a conceptual diagram showing detection of minority cells and removal of majority cells using an AI anomaly detector in the cancer cell testing method according to an embodiment of the present invention. 1 is a conceptual diagram showing a process of performing two-stage anomaly detection using an AI anomaly detector in a cancer cell testing method according to one embodiment of the present invention.

[0034] FIG. 1 is a conceptual diagram showing feature extraction using AI deep learning in a cancer cell testing method according to one embodiment of the present invention.

[0035] FIG. 1 shows a cell image (a) depicting a single white blood cell, which is a majority cell type removed in the first stage by the anomaly detector, and a cell image (b) depicting multiple white blood cells in a cancer cell testing method according to one embodiment of the present invention.

[0036] FIG. 1 is a graph and conceptual diagram showing the first-stage threshold processing by the AI ​​anomaly detector in a cancer cell testing method according to one embodiment of the present invention.

[0037] FIG. 1 is a graph showing an example of determining the threshold for the first-stage threshold processing by the AI ​​anomaly detector in a cancer cell testing method according to one embodiment of the present invention.

[0038] FIG. 1 is a graph showing an example of determining the threshold for the second-stage threshold processing by the AI ​​anomaly detector in a cancer cell testing method according to one embodiment of the present invention.

[0039] FIG. 2 is an image showing an example of clusters obtained by performing k-means clustering in the second-stage threshold processing by the AI ​​anomaly detector in a cancer cell testing method according to one embodiment of the present invention.1 is a graph showing the rate of reduction in undetected cancer patients when the cancer cell testing method of the present invention is used in combination with X-ray radiography in an embodiment.

[0026] A cancer cell testing method according to one embodiment of the present invention uses deep learning based on AI (artificial intelligence) to detect cancer cells, and consists of four steps: blood sampling (S1), capturing blood cell images (S2), extracting blood cell images (S3), and extracting features from the blood cell images (S4), as shown in the flowchart in Figure 1. In this embodiment, the detection of lung cancer cells will be described as an example.

[0027] In the blood collection (S1), the cancer cell testing method of this embodiment can be performed by using a small amount of blood collected during a health checkup or the like, and it is also possible to use surplus blood from various other blood collection tests.

[0028] In the blood cell imaging (S2), as shown in Figure 2, first, screening is performed using a CTC-chip to capture circulating tumor cells (CTCs) to select blood cells significant for lung cancer (S21). Next, fluorescent staining is performed using the three antibodies shown in Table 1 (S22), and then blood cell images are captured using a fluorescence microscope (S23).

[0029] The fluorescent staining using three types of antibodies involves three colors: Hoechst, which is expressed in the DNA of cell nuclei, CD45, which is expressed in leukocytes, and CK, which is expressed in epithelial cells, as shown in Figures 2 and 3. In this embodiment, staining is performed using these antibodies, and the resulting three-color stained image and an unstained phase-contrast image, as shown in Figure 4, are combined to create a total of four channels for use as blood cell images.

[0030] In the step of extracting blood cell images (S3), cells are detected and extracted using a region binarization method on the Hoechst channel, which is a fluorescent light that is specifically expressed in the DNA of cell nuclei. Cell extraction is performed by identifying and extracting cells based on the binarized image through computer processing.

[0031] Feature extraction (S4) for lung cancer testing from blood cell images is performed using two methods (S41, S42) as shown in Figures 1 and 5. Feature extraction (S4) is performed by computer processing using AI, which is a feature extraction device, after going through the processing steps described below.

[0032] The first method focuses on the morphological characteristics of histograms of feature values ​​for multiple cell images of a large number of cells in a cell population under examination (S41). This method utilizes 21 types of feature values, as shown in Table 2. These feature values ​​include 19 features based on cell brightness histograms and their morphology, one feature extracted from cell area, and one feature based on the number of circulating tumor cells (CTCs). Each feature value is calculated using the histograms of cell area and fluorescence brightness for each fluorescence channel and the number of CTCs, as shown in Figures 6 and 7. For the feature based on the number of CTCs, a CTC detector is not used. Instead, a CK brightness histogram is used, as shown in Figure 7. Cells with CK brightness values ​​in the rightmost 5% or more of the horizontal axis (brightness value) are detected as CTCs, and their number is treated as a feature value.

[0033] Among the multiple cell features, the feature utilizing the cell brightness histogram and its morphology utilizes the morphological features of the brightness histogram, which shows the brightness (horizontal axis) of multiple morphological features in blood cell images and the frequency of occurrence of each brightness value (vertical axis) when the overall brightness is set to 1, as shown in Figure 6. Specifically, the following six features are obtained from three histograms: CK brightness, Hoechstno brightness, and CD45 brightness: (1) brightness histogram height, (2) brightness histogram width (variance), (3) brightness histogram median (mode), (4) brightness histogram skewness, (5) brightness histogram kurtosis, and (6) brightness histogram mean brightness. Each histogram is expressed as a different histogram shape, as shown in Figure 6.

[0034] As described above, the AI ​​distinguishes between healthy individuals and cancer patients by setting a predetermined threshold for multiple cell features based on the morphological differences in the brightness histograms of each histogram.

[0035] The second method uses AI deep learning to detect anomalies, extracting and characterizing minority cells. First, using 20,000 randomly sampled images from healthy individuals and cancer patients, the AI ​​trains an anomaly detector using Efficient-GAN (Generative Adversarial Networks), as shown in Figures 8 to 11. As shown in Figures 9 and 11(a), Efficient-GAN detects anomalies A by calculating the root mean square error (RMSE) between the input image Iin of the Encoder (E) and the output image Iout of the Discriminator (D). To extract minority cells using AI, as shown in Figure 8, the AI ​​detects minority cells using an AI minority cell detector, then clusters the minority cells and creates minority cell features using the occurrence probability of the classified clusters. These clusters are divided into, for example, 160, and there are 160 types of minority cell features, corresponding to the number of clusters. In Efficient-GAN, as shown in Figures 9 and 11(a), the AI ​​uses an anomaly detection model trained to output an image identical to the input image. By machine-learning this anomaly detection model into the AI ​​anomaly detector, the error between the input image and the output image is reduced when a large number of normal cells, such as white blood cells, are input. However, the error increases when other small cells are input to the anomaly detector. Based on the anomaly value A calculated by the AI ​​anomaly detector, the AI ​​anomaly detector detects abnormal images using a predetermined threshold, removing the large number of cells and detecting the small number of cells. Detection of abnormal images using the anomaly value A is performed by calculating the root mean square error (RMSE) between the input image Iin of the AI ​​Encoder (E) and the output image Iout of the Discriminator (D), as shown in Figures 9 and 11(a) above.

[0036] Furthermore, AI-based detection of rare cells is performed in two stages using an AI rare cell detector, as shown in Figure 10. The first stage detects white blood cells that appear alone in the image, and the second stage detects white blood cells that appear multiple times in the image.

[0037] In the first stage of detection using the small number of cells detector, as shown in Figure 12, an abnormal value threshold is set based on the number of cells in the image. The number of cells is calculated by binarizing the image and counting the number of detected cells (connected components). In this process, a threshold is set so that cells that appear alone in the image and have a low abnormal value are considered normal, and white blood cells, which account for the majority of blood cells, are mainly removed. Images containing multiple white blood cells or images of cells that emit specific light due to fluorescent staining are then detected as abnormal cells and used for training the second stage of the anomaly detector.

[0038] In the first stage of AI-based detection of a small number of cells, as shown in Figure 13, a normal large number of cells has one cell and the error A M1 is the threshold T M1 This determines the image of a single cell and a white blood cell as a normal cell. M1 The threshold T M1 As shown in Figure 14, M1 The most frequent value when the histogram was drawn is multiplied by 3. Images of multiple cells or images with abnormal values ​​above a predetermined threshold are then regarded as abnormal cells and proceed to the second stage of processing.

[0039] In the second stage, in order to remove images that contain multiple cells in an image that was determined to be normal in the first stage, an abnormality threshold T is set to detect cells that emit specific light due to fluorescent staining. M２ Determine the abnormal value A M2 Threshold T for M２ As shown in Figure 15, when abnormal values ​​are sorted in ascending order, the area of ​​the central cell is set to a threshold T M２ This makes it possible to remove images in which multiple cells are scattered, and to detect other small numbers of cells.

[0040] The second-stage majority cell threshold T M２ is the anomaly value A calculated by the second-stage anomaly detector. M2As shown in FIG. 15, the vertical axis shows the area of ​​the cells, and the horizontal axis shows the abnormal value A of every 1000 blood cell images. M2 The graph is sorted in ascending order of the average values ​​of the cell area (horizontal axis in Figure 15). The first outlier A intersects with the average cell area (26.5) (horizontal line in Figure 15) in this graph. M2 is the threshold T M2 As a result, the threshold T M2 Cells that are:

[0041] As shown in Figure 11(b), the detected abnormal cell groups are clustered using k-means, and the occurrence probability of the resulting clusters is used as a feature. By setting feature values ​​for specific, rare cells, as shown in an example of a cluster (Figure 16), the characteristics of cells that appear extremely rarely in a specimen can be used for lung cancer screening. Furthermore, cancer patient screening can be performed by classifying the extracted feature values ​​using, for example, Random Forest.

[0042] The cancer cell testing method using the blood image testing system of this invention uses fluorescently stained cell images to identify multiple morphological features of the blood cell image by using a brightness histogram of the blood cell image's brightness and its occurrence frequency. For the minority cell features, the blood cell images are clustered according to a predetermined criterion, and the minority cell features are determined using the occurrence probability of each classified cluster. Images of white blood cells and other noise are removed from the blood cell image, and the minority cell features are identified for the test cell group using a predetermined minority cell feature threshold. The cancer cell testing method using the blood image testing system of this invention enables cancer testing with higher sensitivity than conventional X-ray testing, thanks to the simple steps described above. Specifically, testing is possible with only about 1 ml of blood, enabling simple and inexpensive cancer testing for lung cancer and other cancers with extremely high accuracy. Furthermore, the steps of extracting cell images and determining whether or not a cell is cancerous are automatically calculated and executed by a predetermined AI program that executes the above processes, enabling extremely rapid and accurate screening and diagnosis of cancer cells.

[0043] The multiple cell features used in the cancer cell testing method of this invention are selected from multiple types of brightness histogram height, brightness histogram width, brightness histogram median, brightness histogram skewness, brightness histogram kurtosis, and brightness histogram mean brightness for each of the antibodies Hoechst, CD45, and CK in a blood cell image. This enables more accurate understanding and identification of multiple cell features, enabling accurate judgment with simple processing steps.

[0044] The cancer cell testing method of this invention clusters minority cell features and uses AI to detect cancer cells, enabling simple and accurate diagnosis. Furthermore, a minority cell detector is used to detect minority cells in blood cell images of the cell group being tested, and images of normal white blood cells appearing in the blood cell images are removed. By removing white blood cells that appear singly in the blood cell images and also removing white blood cells that appear multiple times in the blood cell images, cancer cells can be accurately identified from minority cell features.

[0045] The cancer cell testing method of the present invention can be applied to cancer cells other than lung cancer cells in the above embodiment, and by using fluorescent stained images of cancer cells, it can be used for screening and cancer testing of various cancer cells.

[0046] Accuracy verification in this study was performed using images taken of 26 healthy individuals and 14 cancer patients. The leave-one-out method was used. The accuracy achieved was 85.7% sensitivity and 73.1% specificity. By combining this with low-sensitivity X-ray testing (59.4%), the undetected rate was reduced by 74.9%, as shown in Figure 17. In other words, while X-ray testing would have missed approximately 40% of cancer patients, the cancer cell testing method of the present invention reduced this 40% miss by 74.9%, thereby reducing the undetected rate to approximately 10%. Here, sensitivity, undetected rate, reduction in the number of undetected cases, and specificity are expressed by the following formulas. Sensitivity = (number of people correctly identified as "cancer patients") ÷ (number of "cancer patients" who underwent the test) × 100 [%] Undetected rate = 100 - sensitivity [%] Reduction rate of undetected cases = ((undetected rate with "X-ray test only") - (undetected rate with "X-ray test and blood cell imaging test combined")) ÷ (undetected rate with "X-ray test only") × 100 [%] Specificity = (number of people correctly identified as "healthy individuals") ÷ (number of "healthy individuals" who underwent the test) × 100 [%]

[0047] The cancer cell testing method using the blood image testing system of this invention can be performed using a small amount of blood, and can easily detect cells suspected of being lung cancer, etc., making it suitable for use in computer-based medical systems. Furthermore, since it can easily test for lung cancer cells, it can also be applied to regular health checkups.

Claims

1. A cancer cell testing method for testing the possibility of cancer cells using blood cell images taken of collected blood, comprising: capturing a group of cells to be tested using a CTC-chip; fluorescently staining the captured group of cells to be tested with a predetermined antibody; and capturing blood cell images of the stained group of cells to be tested; followed by clustering in machine learning processing of the features of the blood cell images, the features being determined as minority cell features and majority cell features; previously training an anomaly detector using Efficient-GAN by deep learning using blood cell images sampled from healthy individuals and cancer patients; forming a brightness histogram of brightness and its occurrence frequency among multiple morphological features of the blood cell image for the majority cell features in the blood cell images; obtaining multiple types of the majority cell features in the form of a graph; and identifying the majority cell features using a predetermined majority cell feature threshold; A cancer cell testing method comprising: detecting minority cells in the blood cell image of the test subject cell group using a minority cell detector; clustering the blood cell image of the minority cells according to a predetermined criterion; determining the minority cell features using the occurrence probability of each classified cluster; removing white blood cell images and other noise images from the blood cell image; identifying the minority cell features for the test subject cell group using a predetermined minority cell feature threshold; and determining that the test subject cell group determined to be abnormal based on the majority cell feature threshold and the minority cell feature threshold are likely to be cancer cells based on the identified majority cell feature and minority cell feature.

2. A cancer cell testing method according to claim 1, wherein the antibodies used for the fluorescent staining include three types: Hoechst, which is expressed in the DNA of cell nuclei; CD45, which is expressed in leukocytes; and CK, which is expressed in epithelial cells.

3. A cancer cell testing method as described in claim 2, wherein the multiple cell features comprise a plurality of types of average brightness, height of the brightness histogram, width of the brightness histogram, median of the brightness histogram, skewness of the brightness histogram, and kurtosis of the brightness histogram for each of Hoechst, CD45, and CK in the blood cell image.

4. A cancer cell testing method as described in claim 3, wherein the multiple cell feature threshold is set based on the difference in the occurrence frequency or the brightness value of the brightness histogram for the multiple cell feature between healthy subjects and cancer patients.

5. A cancer cell testing method as described in claim 1, wherein the removal of the white blood cell images appearing in the blood cell image using the small number of cells detector removes white blood cells appearing singly in the blood cell image, and further removes white blood cells appearing multiple times in the blood cell image.

6. A cancer cell testing method as described in claim 1, wherein the small number of cell features are clustered, and the trained anomaly detector determines whether a cell is likely to be cancerous based on the differences in the small number of cell features between healthy individuals and cancer patients.

7. The method for cancer cell testing according to claim 4, wherein the removal of the white blood cells comprises, in a first step, determining a cell number threshold based on outliers and the number of cells appearing in the image, binarizing the image and counting the number of cellular components detected, and setting the cell number threshold so that cells appearing alone in the image and with a low outlier are considered normal; next, in order to remove images in which multiple cells are shown in the image that were considered normal in the first step, determining an outlier threshold for detecting cells that are specifically luminescent by the fluorescent staining, using the area of the cells in the image as a criterion for determining the outlier threshold, determining the outlier and area of each cell, and sorting the outliers in ascending order, using the average cell area as the abnormal value of the cells as a second threshold to remove the images of the white blood cells; clustering the test cell group having an outlier value equal to or greater than the second threshold, and identifying cancer cells from the clusters obtained thereby using the occurrence probability compared to that of normal individuals as a feature.

8. A cancer cell testing device that uses blood cell images taken of collected blood to test for the possibility of cancer cells, comprising: a blood cell image capturing device that captures a group of cells to be tested using a CTC-chip, fluorescently stains the captured group of cells to be tested with a predetermined antibody, and captures blood cell images of the stained group of cells to be tested; and a feature extraction device that performs clustering on the features of the blood cell images, and obtains minority cell features and majority cell features as the features; the feature extraction device has previously trained an anomaly detector using Efficient-GAN by deep learning using blood cell images sampled from healthy subjects and cancer patients; the feature extraction device comprises a processing device that forms a brightness histogram of brightness and its occurrence frequency among a plurality of morphological features of the blood cell image for the majority cell features in the blood cell image, acquires the plurality of types of majority cell features in the form of a graph, identifies the majority cell features based on a predetermined majority cell feature threshold, detects minority cells in the blood cell image of the test subject cell group using a minority cell detector, clusters the blood cell image of the minority cells based on a predetermined criterion, obtains the minority cell features using the appearance probability of each classified cluster, removes white blood cell images and other noise images from the blood cell image, and identifies the minority cell features for the test subject cell group based on the predetermined minority cell feature threshold; and the cancer cell testing device makes it possible to identify whether or not the test subject cell group determined to be abnormal based on the majority cell feature threshold and the minority cell feature threshold are cancerous cells based on the identified majority cell features and minority cell features.

Images