Method and apparatus for cell identification using electromagnetic wave scattering spectra

Abstract

Provided is a method for using a sample containing cells collected from a patient to discriminate, with high sensitivity and high accuracy, normal cells and abnormal cells contained in the sample. A sample containing cells collected from a patient is irradiated with electromagnetic waves, an electromagnetic wave scattering spectrum from the cells in the sample is measured, and normal cells and abnormal cells are discriminated on the basis of the characteristics of the electromagnetic wave scattering spectrum. The discrimination uses a trained model in which the electromagnetic wave scattering spectra of normal cells and abnormal cells are trained. The sample may be a living cell, or may be a sample that is immobilized by an alcohol or formalin and dyed. The electromagnetic wave scattering spectrum may use either or both of a scattering spectrum obtained through Rayleigh scattering and a scattering spectrum obtained through Mie scattering.

Description

[Technical Field]

[0001] This invention relates to a method and apparatus for easily, sensitively, and accurately detecting abnormal cells (also called abnormal cells) that have degenerated from normal cells in a sample containing cells taken from a patient. In particular, it relates to a method and apparatus for distinguishing tumor cells that are difficult to distinguish with an optical microscope. [Background technology]

[0002] One method for examining malignant tumors is pathological examination, which involves observing tissue sections from suspected patients under a light microscope. In this diagnosis, clinicians diagnose suspected malignant tumors, and if suspected, pathologists receive pathological specimens such as tissue and cytological samples from the clinicians to diagnose the type and condition of the tumor, and a treatment plan is formulated based on that diagnosis. Thus, pathological diagnosis plays an extremely important role in medical practice. However, compared to other developed countries, Japan has a small number of certified pathologists, and a small number of pathologists diagnose a very large number of specimens. Furthermore, in pathological diagnosis using a light microscope, it is often difficult to distinguish between tumor cells and normal cells, and the diagnosis can sometimes depend on the pathologist's experience. In this situation, in recent years, AI (Artificial Intelligence) analysis based on machine learning of light microscope images of pathological specimens has attracted attention as a way to solve these problems.

[0003] Some tumor cells are indistinguishable from normal cells even to experienced pathologists when observed using only a light microscope. The differences between tumor cells and normal cells are not limited to the morphology of micrometer-sized cells and intracellular organelles observed with a light microscope, but also manifest in nanoscale (sub-micrometer) cytoskeleton, dysstructure within intracellular organelles, and protein denaturation. If this nanoscale information can be detected and analyzed using light scattering spectra in addition to light microscope images, the diagnostic accuracy of pathological specimens can be improved. Furthermore, if high-density information can be obtained faster than with light microscope observation, it is expected that high-speed and high-precision AI analysis will become possible.

[0004] Prior to analysis, it is essential that appropriate sample preparation and accurate testing methods are employed. In cytology, the collected sample is typically smeared onto a glass slide and immediately dehydrated and fixed with ethanol. If the sample is dry, it is rehydrated before being fixed with 95% ethanol. After fixation, staining is performed, such as Papanicolaou staining. In pathology, the collected sample is typically fixed with formalin, and a paraffin-embedded block is prepared. Sections are then prepared from this block, and staining such as HE (Hematoxylin Eosin) staining is performed. If a tumor is suspected, immunohistochemical staining methods corresponding to various tumor markers are applied, and highly accurate identification is performed by the pathologist. The stained sample is observed, and identification is made based on characteristics of tumor cells, such as enlarged nuclear size, multinucleation, and abnormalities in the intranuclear ultrastructure. However, there are many types of tumor cells that are extremely difficult to distinguish based on these differences. For example, in cytology for malignant mesothelioma, a rare type of cancer, distinguishing mesothelioma cells from reactive mesothelial cells and adenocarcinoma cells is often difficult even for specialists, and there is a need for more appropriate methods for differentiation.

[0005] The process of preparing pathological specimens (a series of steps including fixation and staining) is very time-consuming and laborious. Therefore, it is desirable to be able to identify tumors during surgery or immediately after collection, before fixation and staining. As a method for this, for example, Patent Document 1 discloses a method for detecting abnormal cells (tumor cells) by directly irradiating living cells with laser light without any staining treatment, and then spectrally analyzing the stimulated emission light emitted from the living cells. Specifically, this method attempts to detect abnormal cells by irradiating cells with a visible light laser and spectrally analyzing (spectral measurement) the stimulated emission light from the biological sample. In the present invention, light scattering spectra are used, but in the embodiments of this invention, the scattered light component of the laser light irradiated onto the living cells is cut off by a filter, and the light scattering targeted by the present invention is not included in the detection information. The spectrum of stimulated emission light has low accuracy in discriminating the state inside the cell, and some types of tumor cells cannot be detected.

[0006] Alternatively, there are methods for staining live cells during surgery or immediately after collection without fixation, such as the fluorescent staining using 5-aminolevulinic acid (5-ALA) described in Non-Patent Literature 1. 5-ALA is specifically taken up by tumor cells, metabolized (converted) to the fluorescent molecule protoporphyrin IX within the tumor cells, and accumulates there. This method of detecting the fluorescence of protoporphyrin IX to identify tumor cells is commonly used in the treatment of skin cancer in Europe and the United States, and in Japan, it is used as an intraoperative fluorescent diagnostic agent for brain tumors and bladder cancer to ensure complete excision of the tumor. However, this method, which identifies tumor cells solely based on the presence or absence of fluorescence from protoporphyrin IX, makes it difficult to distinguish tumor cells with higher accuracy than observing pathological specimens stained with various dyes, and it cannot distinguish between different types of tumor cells. Furthermore, staining methods that utilize the metabolism of live tumor cells by 5-ALA cannot be applied to staining fixed specimens, and therefore cannot be used as a simple, highly sensitive, and accurate method for detecting normal cells and abnormal cells (tumor cells) in pathological specimens containing cells collected from patients, which is the objective of this invention.

[0007] In addition, various methods have been investigated to distinguish between normal cells and abnormal cells (tumor cells) by directly irradiating living cells with electromagnetic waves such as light without staining treatment, and using the scattered electromagnetic wave spectrum. However, these methods have not been able to distinguish tumor cells with higher accuracy than observation of pathological specimens.

[0008] For example, examples of detecting tumor cells by analyzing scattered light, which is observed as backscattered light after being reflected from the surface of biological tissue when visible light is irradiated, have been reported in Non-Patent Documents 2 and 3. Non-Patent Document 2 shows that the enlarged nuclear size, a characteristic of tumor cells, can be detected by analyzing the angular dependence of the Mie scattered light intensity scattered backwards, which originates from the size of the nucleus of living cells. Similarly, Non-Patent Document 3 shows that the presence of enlarged nuclei contained in tumor cells can be detected by analyzing the wavelength dependence of the Mie scattered light intensity scattered backwards from living cells. As these examples show, changes in the scattering direction and spectral shape of Mie scattered and Rayleigh scattered light due to the size of the nucleus of a cell, which has a simple shape close to a sphere, can be theoretically predicted based on physics. However, it is physically impossible to uniquely predict changes in the scattering direction and spectral shape of Mie scattered and Rayleigh scattered light due to the more complex shape inside the cell, and therefore the type of tumor cell cannot be distinguished.

[0009] Patent Document 2 discloses a method for identifying cell types by using a flow cytometer to measure a fluid containing a mixture of various cell types, and analyzing the time-series signals of optical images or spectra obtained while varying the relative position of the illumination pattern and the object being observed within a microchannel using machine learning-based AI, which does not rely on unambiguous physical prediction. This method allows for the distinction between normal and abnormal cells in a large number of cells to be analyzed using machine learning-based AI. However, its applicability is limited to cases where time-series signals can be obtained from cells flowing at a constant speed, such as with a flow cytometer, and it is difficult to use it for immobilized cells attached to a substrate, such as pathological specimens. On the other hand, the present invention is characterized by not requiring time-series signals, and therefore, the invention described in Patent Document 2 cannot be used for the method of easily, sensitively, and accurately detecting normal cells and tumor cells contained in a stationary sample containing cells collected from a patient, which is the target of the present invention.

[0010] As described above, studies attempting to distinguish between normal and abnormal cells using various measurement methods targeting living cells are primarily aimed at identifying the extent of tumor tissue during surgery. However, since the above methods do not utilize more detailed discrimination methods using tumor markers expressed in various types of tumors, they have only been able to detect the presence of tumor cells or tissues containing tumor cells.

[0011] The present invention aims to provide a method for distinguishing between normal and abnormal cells in living cells, and to provide a method for distinguishing between normal and abnormal cells in fixed and stained pathological specimens prepared in accordance with existing pathological diagnostic processes, based on nanometer to subnanometer morphological characteristics such as the fine shape of cells, the shape of the nucleus, and the intranuclear ultrastructure, by measuring the light scattering spectrum from the pathological specimen and analyzing the light scattering spectrum using machine learning. In other words, the present invention aims to provide a method for easily, sensitively, and accurately detecting normal and abnormal cells in a sample containing cells collected from a patient. The sample may include pathological specimens collected from a patient, fixed, and stained according to existing pathological diagnostic procedures, as well as specimens in the pre-fixation and staining stages. The aim is to provide a means to make diagnosis, which previously relied on the experience and skill of pathologists, easier, more sensitive, and more accurate, without relying on the experience and skill of pathologists. In particular, the aim is to provide a method for distinguishing tumor cells that are difficult to distinguish with an optical microscope. [Prior art documents] [Patent Documents]

[0012] [Patent Document 1] Japanese Patent Publication No. 58-118948 [Patent Document 2] Patent No. 6959614 [Non-patent literature]

[0013] [Non-Patent Document 1] Shunichiro Ogura, et al., "Photodynamic Therapy Targeting Dormant Cancer Cells Using 5-Aminolevulinic Acid", Journal of the Japan Laser Medicine Society, Vol. 43, No. 4, pp. 238-248 (2023).

Non-Patent Document 2

Non-Patent Document 3

Summary of the Invention

Problems to be Solved by the Invention

[0014] An object is to provide a method and an apparatus for highly sensitively and highly accurately detecting the discrimination between normal cells and abnormal cells contained in a sample using a sample containing cells collected from a patient. In particular, an object is to provide a method and an apparatus for discriminating tumor cells that are difficult to discriminate with an optical microscope.

Means for Solving the Problems

[0015] To solve the above problems, the cell discrimination method of the present invention involves irradiating a sample containing cells taken from a patient (cell sample) with electromagnetic waves, obtaining an electromagnetic wave scattering spectrum from the sample, and distinguishing between normal and abnormal cells based on the information obtained from the spectrum. Light scattering is affected by the reflection and refraction of structures that are on the same scale as or smaller than the wavelength of light (100 nm to 1 μm). Light scattering by structures on the same scale as the wavelength (micrometer scale) is called Mie scattering, and light scattering by structures smaller than the wavelength (submicrometer scale) is called Rayleigh scattering. The present invention is characterized by utilizing the information (spectral shape) contained in the spectral spectra of Rayleigh scattering and Mie scattering to identify abnormal cells. Optical images also contain information due to Rayleigh scattering and Mie scattering, but optical images do not contain information beyond the diffraction limit. By analyzing the light scattering spectrum, which is a feature of the present invention, it is possible to obtain information on submicron-scale intracellular structures that cannot be observed with optical microscopes that are beyond the diffraction limit. What is novel about this invention is that the information is extracted from the light scattering spectrum using mathematical methods, rather than based on physical or chemical interpretation. When light is shone on a cell sample, the scattering direction and spectral shape of the light change due to the micrometer-scale morphology of the cells and cell nuclei, as well as the submicrometer-scale features such as the fine morphology of the cells, the shape of the nucleus, and the intranuclear ultrastructure. Furthermore, if the cells are stained, as in pathological specimens, and the structural features of abnormal cells and the types of proteins they contain are highlighted by dyes, the light scattering is strongly influenced by those dyes. It is thought that the light scattering spectrum from pathological specimens comprehensively reflects all of this information, and it is too complex to uniquely link the simple spectral shape features, such as the undulations observed in the light scattering spectrum, with the structure of the cells and cells within them, nor is it possible to theoretically analyze the spectrum based on physics. To solve this problem, the present invention presents a method for analyzing a light scattering spectrum by machine learning and predicting a pathological diagnosis result. In machine learning, without relying on physical logic, feature quantities of the light scattering spectrum are numerically extracted, and these feature quantities are statistically analyzed to train a prediction model (also referred to as a learning model). By inputting another light scattering spectrum into this prediction model, a method for obtaining a prediction of whether it is a normal cell or an abnormal cell becomes AI analysis. The numerically obtained feature quantities do not necessarily coincide with the features of the spectrum visually perceived by humans, and may be determined by minute spectral differences that cannot be distinguished visually, resulting in an interpretation that is essentially different from the conventional theoretical analysis based on physics. That is, in the present invention, by using machine learning, which is a numerical analysis not based on physical logic, in the causal relationship between the cell and the fine structure inside the cell that reflects the difference between normal cells and abnormal cells, the problem in the data interpretation of the light scattering spectrum that the causal relationship is too complex to be analyzed theoretically based on physics is solved.

Advantages of the Invention

[0016] According to the present invention, there is an effect that a method for highly sensitively and highly accurately detecting the discrimination between normal cells and abnormal cells can be provided. Conventionally, it is possible to facilitate the discrimination of tumor cells that are difficult to discriminate by the diagnosis of a pathologist through optical microscope observation, and it can support a diagnosis that does not depend on the experience and proficiency of the pathologist and the high accuracy of the diagnosis.

Brief Description of the Drawings

[0017] [Figure 1] Diagram showing the outline of the measurement system (a) Overall configuration diagram of the measurement system (b) Enlarged view of the illumination part for dark-field image observation (c) Enlarged view of the illumination part for bright-field image observation [Figure 2] Composition of the MDCK cell sample [Figure 3] Comparison of bright-field images and dark-field images of MDCK cells when cytochalasin D is not added (a) and when it is added (b) [Figure 4](a) Comparison of normalized light scattering spectrum (SESn710) from MDCK cells immersed in culture medium immediately after measurement and 60 minutes later. (b) Changes in the normalized light scattering spectrum (SESn710) from cells every 10 minutes from immediately after addition of Cytochalasin D to 60 minutes later. [Figure 5] Light scattering spectrum (SES) of cells contained in specimens A-F from normal patients. [Figure 6] Bright-field and dark-field images of cells from specimens A-F from normal patients: (a) Cells in which a peak was observed around 700 nm in the SES spectrum; (b) Cells in which the SES spectrum was flat. [Figure 7] Light scattering spectrum (SES) of cells in specimens G, H, and I from patients suspected of having mesothelioma. [Figure 8] Bright-field and dark-field images of cells contained in specimens G, H, and I from a patient suspected of having mesothelioma. [Figure 9] (a) Loading of the first and second principal components obtained by principal component analysis of the light scattering spectrum, (b) Scatter plot of the percentage ratio (score) of the first and second principal components included in the light scattering spectrum. [Figure 10] Conceptual diagrams of two-class classification using k-nearest neighbors (a), decision tree (b), and random forest (c). [Figure 11] This diagram illustrates how to allocate training and test data for samples A-I to verify the classification accuracy (correctness) of light scattering spectra using a machine learning model. [Figure 12] Block diagram of the cell discrimination device [Figure 13] Normalized light scattering spectrum (SESn710) for cells contained in samples A-I. [Figure 14] Contribution rates and loading results for each of the 1st principal component (PC1) through the 10th principal component. [Figure 15] Scatter plot of the nth and mth principal component scores for SESn710 in cells. [Figure 16] Diagram illustrating the degree of malignancy of tumor cells [Figure 17] Diagram illustrating support for pathological diagnosis [Figure 18] Normalized light scattering spectrum (SESn710) for two specific types of living cells. [Figure 19] Scatter plot of the nth and mth principal component scores for SESn710 in two specific cell types. [Figure 20] Diagram illustrating the method for measuring scattered light spectra. [Modes for carrying out the invention]

[0018] This invention involves obtaining a light scattering spectrum from scattered light generated by irradiating a sample containing cells taken from a patient with electromagnetic waves, and in order to distinguish between normal cells and abnormal cells based on the differences in the characteristics of their respective light scattering spectra, it numerically extracts the feature quantities contained in the light scattering spectrum, creates a learning model based on those feature quantities, and distinguishes between unknown cells.

[0019] Cells contain various organelles, including the cell nucleus, within the cytosol, which is separated from the outside by the cell membrane. It is known that characteristic changes occur in the size and ultrastructure of various organelles between normal and abnormal cells. This invention aims to capture the characteristics of abnormal cells, including various tumor cells, which play a particularly important role in pathological diagnosis, by capturing the characteristic quantities of the light scattering spectrum. However, it is not limited to these, and provides a method for detecting and analyzing various intracellular and extracellular ultrastructures using light scattering spectral measurement.

[0020] Cellular and intracellular structure degeneration can occur due to various causes, including aging, chemical and physical damage, and induction by various physiologically active substances, infections, and cell carcinogenesis. Therefore, early detection of cellular abnormalities is crucial by observing morphological changes in the nucleus and other organelles. Among cellular degenerations, cell shape and nuclear degeneration, in particular, play a significant role in various pathological diagnoses. For example, tumor cells may exhibit characteristics such as cell cluster formation, enlargement and multinucleation of the cell nucleus, enlargement of the nucleolus, and narrowing of the cytosol. Furthermore, tumor cells may exhibit abnormal polymerization of cytoskeletal proteins such as actin and keratin.

[0021] Furthermore, many malignant tumors that have become cancerous exhibit increased mobility, leading to invasion into the stroma beyond the basement membrane and metastasis to other organs due to degeneration of epithelial cells (epithelial-mesenchymal metastasis). It is known that during this process, the reorganization of actin filaments, which maintain the cytoskeleton, occurs, and the formation of filopodia on the cell membrane increases cell mobility. Therefore, cell degeneration due to tumors and the like can be detected as changes in shape from micrometers to submicrometers, such as changes in cell shape, the shape and size of the cell nucleus, intracellular organelles, and the cytoskeleton. In this invention, a cell discrimination method is used in which a sample containing cells collected from a patient is irradiated with electromagnetic waves, the electromagnetic wave scattering spectrum from the cells in the sample is obtained, and a learning model that has learned the characteristics of the electromagnetic wave scattering spectra of normal and abnormal cells is used to distinguish between normal and abnormal cells. The sample may be living cells, or it may be a sample fixed with alcohol or formalin and stained.

[0022] Research using light scattering to capture the characteristics of cells has been actively conducted for a long time. Light scattering phenomena are characterized by the particle size parameter (α; α = πd / λ), which is expressed as the ratio of the diameter (d) of the scattering body to the wavelength (λ) of the electromagnetic wave being observed, and the relative refractive index of the scattering body. When α << 1, Rayleigh scattering is dominant, and the scattered light intensity increases as the wavelength decreases. Near α = 1, Mie scattering becomes dominant, and the scattered light intensity takes a constant value independent of α. Furthermore, when α is 10 or more, diffraction phenomena occur due to the interference of scattered light, and the scattering angle and α begin to behave in a complex manner. In the present invention, the scattering spectrum can be either a scattering spectrum due to Rayleigh scattering or a scattering spectrum due to Mie scattering, or both. More preferably, the scattering spectrum can be either a forward scattering spectrum due to Rayleigh scattering or a forward scattering spectrum due to Mie scattering, or both. Furthermore, the difference in electromagnetic wave scattering spectra between normal cells and abnormal cells is due to an increase or decrease in scattering caused by changes in cell structure that are smaller than or equal in size to the wavelength of visible light.

[0023] In this invention, the wavelength of the electromagnetic waves used for observation is preferably 100 nm to 1 μm, and visible light in the wavelength range of 400 to 750 nm is particularly preferred. Ultraviolet to infrared light can also be used. Furthermore, when using this invention, it is also possible to use electromagnetic waves in the wavelength range from X-rays to ultraviolet light to obtain electromagnetic wave scattering spectra from even smaller regions of cells in a sample, and to distinguish between normal and abnormal cells using a learning model that has learned the characteristics of the electromagnetic wave scattering spectra of normal and abnormal cells.

[0024] The present invention is characterized by its ability to detect the structural characteristics of scattering bodies with high sensitivity and accuracy by observing scattered light from individual cells. In particular, since the scattering bodies include structures such as the nucleus and cytoskeleton that make up cells, the range of α includes 0.1 < α < 150. Since the size of cell nuclei ranges from a few micrometers to about 10 micrometers, when the wavelength of the electromagnetic wave being observed is around 500 nm, the value of α is in the range of 20 to 120. In this case, the scattered light intensity depends greatly on the scattering angle, and most of the scattering is observed as forward scattering in the same direction as the incident light. When using the method of the present invention, the observation of scattered light from cells is performed by detecting the scattered light scattered in an arbitrary direction with respect to the incident angle of the electromagnetic wave irradiating the cell, thereby obtaining the electromagnetic wave scattering spectrum from cells in the sample. It is possible to perform a cell discrimination method to distinguish between normal cells and abnormal cells using a learning model that has learned the characteristics of the electromagnetic wave scattering spectra of normal cells and abnormal cells.

[0025] When the cross-sectional area of ​​the scatterer is larger than the wavelength of the observed electromagnetic wave, the scattering efficiency changes in a complex manner with respect to α, and it is known that when there is a relative refractive index of cellular components in a medium such as water, it depends significantly on the size of the scatterer and the wavelength of the electromagnetic wave. Non-patent document 4 (M. Xu et. al., “Unified Mie and fractal scattering by cells and experimental study on application in optical characterization of cellular and subcellular structures”, Journal of biomedical optics, 13(2), 024015-024015 (2008)) shows the results of comparing light scattering spectra measured by irradiating cells with visible light and varying the scattering angle in the range of 1.1 to 165° with calculated values. Furthermore, scattered light from cells is represented as a superposition of scattered light components from the nucleus within the cell, scattered light components from the homogeneous cytoplasm, and scattered light components from minute organelles other than the nucleus, and the results of measuring the light scattering spectrum at various scattering angles are explained by comparing them with calculated values. According to this, in forward scattering with a scattering angle of 1.1°, the scattered light intensity increases with wavelength, while in backscattering, the phase is reversed, and the scattered light intensity tends to decrease with wavelength. The contributions from each component of the scattered light show that scattering from the cell body cytoplasm contributes as background, and in addition to the fractal "fluctuation" component from organelles, the contribution of the scattering component from the nucleus accounts for a large proportion. As described above, it has been shown that by irradiating cells with visible light, it is possible to accurately measure the light scattering spectrum from cells in the scattering angle range of 1.1 to 165°. In the method of the present invention, regardless of whether forward scattering or backscattering is observed, an electromagnetic wave scattering spectrum from cells can be obtained, and a cell discrimination method can be used to distinguish between normal and abnormal cells using a learning model that has learned the characteristics of the electromagnetic wave scattering spectra of normal and abnormal cells.

[0026] Non-Patent Literature 4 states that as the size of the nucleus increases, the dependence of the scattered light intensity on wavelength becomes more pronounced in forward scattering, increasing significantly with wavelength, while the opposite trend is observed in backscattering. Non-Patent Literature 4 also investigates the scattering behavior when cells are denatured by acetic acid, and it is found that while organelles denature and whiten due to acetic acid, the scattering intensity increases significantly, and the contribution of the "fluctuation" component increases as the wavelength decreases. Thus, in this invention, it is possible to capture and analyze scattered light from either forward or backscattering angles, thereby accurately identifying the differences between normal and abnormal cells that were unclear with various conventional methods.

[0027] In the embodiments of the present invention, the light scattering spectrum obtained by measurement is represented as two-dimensional data showing the relationship between wavelength and scattered light intensity. In this invention, since a large number of light scattering spectra obtained for individual cells are collected and analyzed, principal component analysis was used as a method for extracting features from the two-dimensional data. Principal components are extracted by diagonalizing the variance-covariance matrix of scattered light intensity at each wavelength. The first principal component is the eigenvector with the largest eigenvalue, and is the component with the maximum variance of scattering intensity. In addition to the first principal component, the second principal component, which is the eigenvector with the second largest eigenvalue, was used to analyze the characteristics of scattered light from cells and create a predictive model that classifies the differences in features between normal and abnormal cells. Specifically, the k-Nearest Neighbors algorithm (kNN), decision tree, and random forest (RF) were used as predictive models. For accuracy evaluation, the trained predictive model was used to predict whether a cell was normal or abnormal based on the light scattering spectral features of normal and abnormal cells that were not used in training. The validity and accuracy of the analysis method were then evaluated based on the accuracy of these predictions.

[0028] The present invention provides an AI analysis method for determining whether the cells in a sample are normal or abnormal by measuring the light scattering spectra of individual cells in a sample containing cells taken from a patient, analyzing the characteristic quantities of the obtained light scattering spectra using the method described above, and using a learning model constructed from previously acquired light scattering spectra. Furthermore, when performing such AI analysis, it is preferable to use a cell discrimination device 20, as shown in Figure 12, which includes a scattering spectrum acquisition unit 21 that irradiates an electromagnetic wave scattering spectrum measuring device 31 with electromagnetic waves and acquires a scattering spectrum 32, a learning model 22 that has been pre-trained on the scattering spectra of normal and abnormal cells, and a discrimination unit 23 that uses the learning model 22 to distinguish between normal and abnormal cells. The discrimination result from the discrimination unit 23 of the cell discrimination device 20 is output to a display device 33, such as a display that can numerically classify and display normal cells and abnormal cells (tumor cells). The cell discrimination device 20 is a computer, and the learning model 22, the scattering spectrum acquisition unit 21, and the discrimination unit 23 programs are stored in memory, and the processor executes the programs. Hereinafter, an example of an embodiment of the present invention will be described in detail with reference to the drawings. It should be noted that the scope of the present invention is not limited to the following embodiments or illustrated examples, and numerous modifications and variations are possible. [Examples]

[0029] In this example, experiments using cultured animal cells demonstrate that changes in submicrometer intracellular structures, which are difficult to observe with an optical microscope, are causally related to the light scattering spectrum, and that the light scattering spectrum contains information about the nuclear ultrastructure. Light scattering spectra of single cells were measured in the wavelength range of 420-720 nm using a dark-field inverted microscope combined with a spectrometer. In this example, the polymerization inhibitor Cytochalasin D was added to cultured animal cells to fragment the fibrous actin filaments that make up the cytoskeleton, and experiments were conducted to investigate the changes in the light scattering spectra of single cells during this process. Since the light scattering spectrum from cells includes the shape of the white light spectrum from the light source, the light scattering efficiency spectrum (SES), obtained by dividing the measured value of scattered light by the scattered light intensity from the light source, was used as the standard for the light scattering spectrum. SES shows the wavelength dependence of the scattered light intensity when light of the same intensity is applied at all wavelengths. The absolute value of SES varies greatly from cell to cell depending on the state of the cell, the light irradiation conditions to the cell, and the adjustment of the objective lens that collects the scattered light. To correct for this measurement variation, the entire spectrum was normalized by the SES value at 710 nm, resulting in a normalized light scattering efficiency spectrum (SES). n710 ) was used for analysis using machine learning.

[0030] As the cells used, a cell line derived from canine renal tubular epithelial cells (MDCK; RCB0995, RIKEN BioResource Center) was used. These cells were cultured on a cell culture dish using Dulbecco's modified Eagle medium (phenol red-free) supplemented with fetal bovine serum (FBS) and antibiotics. Just before the cells reached confluence (80% coverage), the medium was changed and the cell culture dish was placed on the sample stage. The center of the cell culture dish was observed using a dark-field inverted microscope (Olympus IX81 with a dark-field condenser lens). A ring-shaped spacer 11 (30 μm thick) was placed in the center of the cell culture dish 10, and medium 12 was added to an area with a diameter of 1.70 cm inside the ring, into which MDCK cells 14 were seeded and cultured. A coverslip 13 was placed on top of the dish to prevent the medium from drying out during observation.

[0031] Figure 1 shows an overview of the measurement system. Figure 2 shows the configuration of the cell culture dish used for the measurement. As shown in Figure 1(b), a halogen lamp was used as the illumination light source, and the illumination light was shone onto the cells on the cell culture dish through a dark-field condenser lens, and the scattered light from the cells was collected by the objective lens. Because the dark-field condenser lens is used to illuminate the cells with a larger numerical aperture than the objective lens, the light collected by the objective lens does not contain the component of light that has passed through the cells (transmitted light). On the other hand, in a normal optical microscope image (bright-field image), as shown in Figure 1(c), the light that has passed through the cells is collected by the objective lens. The scattered light collected by the objective lens is split into two paths. One is imaged onto a CMOS camera to obtain a dark-field image. The other is introduced into a spectrometer via an optical fiber, where the scattered light intensity spectrum measured by the spectrometer is divided by the illumination spectrum to obtain the SES. In the scattered image imaged on the optical fiber side, the core diameter of the optical fiber corresponds to an area with a diameter of approximately 30 μm on the cell culture dish, so the SES of about one cell can be selectively measured. Here, in the spectrometer, Raman scattering is detected in addition to Rayleigh scattering and Mie scattering. However, the intensity of Raman scattering is overwhelmingly lower than that of Mie scattered light and Rayleigh scattered light, below the detection sensitivity of the spectrometer, and is either negligible or undetectable in this measurement. That is, as shown in Figure 20(1), in the present invention, the sample (scatterer) is irradiated with light containing all wavelengths in the visible range, such as white light, as the excitation light 41. Since the scattered light, Mie scattered light and Rayleigh scattered light 43, has an overwhelmingly greater intensity than Raman scattered light and fluorescence 44, Raman scattered light and fluorescence 44 are ignored as scattering spectral components in the spectrometer 45. On the other hand, as shown in Figure 20(2), in an optical system for selectively detecting Raman scattered light and fluorescence 44, unless light with a strong intensity at a specific wavelength is used as the excitation light 46 and a powerful optical filter (excitation light cut filter) 47 is used to cut out Mie scattered light and Rayleigh scattered light, the main components of light scattering detected will be Mie scattering and Rayleigh scattering. Note that Mie scattered light and Rayleigh scattered light have the same wavelength as the light used to illuminate the cells (excitation light), but Raman scattered light and fluorescence have different wavelengths from the excitation light.

[0032] Figure 3 shows the bright-field and dark-field images of MDCK cells observed with a CMOS camera. Figure 3(a) shows MDCK cells without Cytochalasin D, and Figure 3(b) shows MDCK cells with Cytochalasin D added, resulting in fragmented cytoskeletons. Regarding the bright-field and dark-field images in Figures 3(a) and 3(b), no significant difference can be observed visually between the images immediately after (0 minutes) and 60 minutes after Cytochalasin D was added and placed under the microscope. The actual fragmentation of actin filaments by adding Cytochalasin D under the above conditions was confirmed separately by observing a sample stained with fluorescence (SPY555-actin, SPIROCHROME, SC202) using a confocal microscope. Actin filaments, which make up the cytoskeleton, do not absorb light in the visible light region and have a diameter of 10 nm or less, making them difficult to observe with an optical microscope. Therefore, it is reasonable that actin filament fragmentation cannot be observed in the bright-field images on the left side of Figure 3(a) and (b). In the dark-field images on the right side of Figure 3(a) and (b), actin filament fragmentation may be observed, but no clear difference can be found with the naked eye. However, by applying the machine learning analysis described later, it may be possible to quantify the structural changes and classify cells in which actin filaments have fragmented.

[0033] Figure 4(a) shows a comparison of SES from MDCK cells in a medium without Cytochalasin D supplementation at 0 minutes and 60 minutes after microscopy placement. n710 The results were almost identical. Even without the addition of Cytochalasin D, if cells are left for a long time, their structure changes, and SES n710 Although there were concerns that the structure might change, this experiment showed that under these experimental conditions, there were no structural changes in the cells observed in the light scattering spectrum for at least 60 minutes.

[0034] Next, when MDCK cells were treated with cytochalasin D at a concentration of 200 μM to inhibit actin filament polymerization, SES from the cells was measured every 10 minutes from immediately after treatment to 60 minutes later. n710 The changes are shown in Figure 4(b). From the results in Figure 4(b), a change in the light scattering spectrum was observed immediately after the addition of Cytochalasin D, and according to these results, the light scattering intensity tended to increase over time in the shorter wavelength range from around 500 nm. It is thought that the inhibition of actin filament polymerization leads to light scattering by actin filament fragments of various lengths contributing to Rayleigh scattering, and the increase in scattering intensity in the shorter wavelength range seen in Figure 4(b) can be explained by this contribution. Thus, this embodiment clearly demonstrates that the visible light scattering spectrum has a causal relationship with intracellular structural changes, and furthermore, that it is possible to measure differences in intracellular structure that cannot be clearly distinguished in bright-field or dark-field images. As shown in Figure 4, the visible light scattering spectrum of an unstained, transparent cell exhibits a flat spectral characteristic without any particularly clear features. However, this spectrum includes not only light scattering from actin filaments, which are part of the cytoskeleton, but also light scattering from the cell's fine structure, such as the cell nucleus and its intranuclear ultrastructure. It is thought that the spectrum appears flat overall as a result of the combined reflection of all these light scattering processes. However, it is possible that faint irregularities that appear as noise on the flat spectrum contain information that is causally related to these fine structures. However, it is virtually impossible to predict these irregularities uniquely using physics. Therefore, as a solution, this invention attempts to analyze information originating from intracellular ultrastructure from the light scattering spectrum by using machine learning-based AI, which does not rely on unique physical predictions. [Examples]

[0035] In this example, we demonstrate that machine learning can be used to distinguish light scattering spectra, which are extremely difficult to interpret from a physical standpoint, using pathological specimens of normal and tumor cells obtained from actual patients, fixed and stained according to existing pathological diagnostic procedures. For the study, pleural fluid was collected using a syringe from the upper edge of the ribs of normal specimens (A-F) and specimens containing cells suspected of being pleural mesothelioma (G, H, I). The cells contained in the pleural fluid were smeared onto glass slides, spray-fixed, and then fixed in 95% alcohol fixative. For staining, the cell nuclei were stained with hematoxylin according to the Papanicolaou standard staining method, followed by staining of the cytoplasm with OG-6 stain, then with EA-50 stain, cleared in a xylol bath, and finally mounted. The cells in the specimens were identified by a specialist pathologist, who distinguished between normal pleural mesothelial cells and other tumor cells. Light scattering spectra were measured using specimens in which individual cells were marked.

[0036] For normal cells, measurements were taken on 23 to 25 pre-marked normal cells contained in samples A to F. For tumor cells, measurements were taken on 13 to 27 pre-marked tumor cells contained in samples G to I. Here, tumor cells include both mesothelioma cells and adenocarcinoma cells. Furthermore, the possibility that reactive mesothelial cells are also included in the cells classified as normal cells and tumor cells cannot be ruled out.

[0037] One characteristic of mesothelioma cells is that they have a higher probability of having two nuclei compared to other adenocarcinoma cells and reactive mesothelial cells. Furthermore, both mesothelioma and adenocarcinoma cells tend to form cell clusters, which allows for differentiation from solitary reactive mesothelial cells. However, while mesothelioma cells fall between adenocarcinoma cells and reactive mesothelial cells in terms of nuclear size, nuclear size is distributed, making differentiation difficult without observing the distribution of nuclear size across a large number of cells, rather than simply observing individual cell nuclei. While the nuclei of adenocarcinoma cells are centrally located within the cell, those of adenocarcinoma cells are often ubiquitous at the edges of the cell; however, mesothelioma cells cannot be distinguished from reactive mesothelial cells based solely on nuclear location.

[0038] In this embodiment, the purpose is to verify that the light scattering spectrum from tumor cells differs from that of normal cells, and not to differentiate between adenocarcinoma cells, reactive mesothelial cells, and mesothelioma cells. In other words, whether cells in a specimen are mesothelioma cells or normal cells is determined from both pathological and clinical perspectives, and the present invention aims solely at detecting abnormal cells that are different from normal cells.

[0039] Figure 5 shows the light scattering efficiency (SES) spectra measured for cells in normal specimens A through F. The characteristics of the SES spectra differed for each cell type. The SES spectra from cells in specimens A and F were generally flat, while those from specimens B, C, and D were characterized by the presence of cells showing a scattering efficiency peak around 700 nm. In the case of specimen E, a peak around 700 nm was observed in the SES spectra of only some of the measured cells, while the rest showed a generally flat spectrum.

[0040] Figure 6 shows the bright-field (top) and dark-field (bottom) images of cells in samples A-F when SES was measured. (a) shows the bright-field (top) and dark-field (bottom) images of cells in which a scattering peak was observed around 700 nm, and (b) shows the bright-field (top) and dark-field (bottom) images of cells in which the SES was flat. For cells in sample B, it was confirmed that the nuclear size of the cells in which a scattering peak was observed was large, but for cells in sample C, no clear correlation was found between nuclear size and the scattering peak. From this, it was found that the appearance of a scattering peak may be influenced by factors other than nuclear size enlargement, such as differences in relative refractive index due to staining ease.

[0041] Next, Figure 7 shows the results of SES measurements for cells contained in specimens G to I, which are suspected to be mesothelioma. Figure 8 also shows light microscope images of cells in each specimen (bright-field images at the top, dark-field images at the bottom).

[0042] In Figure 6, all cells in specimen G showed flat SES, in specimen H a small number of cells showed SES with a peak around 700 nm, and in specimen I cells with a peak around 700 nm were the majority. Furthermore, in the light microscope images of the cells shown in Figure 8, there was a distribution in the size of the nuclei contained in each cell, and the position of the nuclei was sometimes central and sometimes ubiquitous. From these results, it was shown that cells showing a peak around 700 nm and cells with flat SES can be clearly distinguished, but these results alone did not necessarily show that normal cells and tumor cells can be distinguished. It is thought that the SES from this cytological specimen reflects the causal relationship between the reflection, absorption, and refraction of the dye bound to the cell, cell nucleus, cytoskeleton, and intracellular organelles by Papanicolaou staining. It is almost impossible to analyze all of these considerations in SES using physics-based theory, so in this invention, machine learning, which is a numerical analysis not based on physical logic, was used as a method for discrimination.

[0043] Normalized light scattering spectrum (SES) for cells contained in specimens A-I n710By performing principal component analysis on , the loadings of the first and second principal components shown in Fig. 9(a) were obtained. From these loadings, the percentage ratio (score) of the first and second principal components included in the SES of each cell was determined, and a plot with the score of the first principal component on the horizontal axis and the score of the second principal component on the vertical axis became the scatter plot shown in Fig. 9(b). For example, the SES where the scores of both the first and second principal components are in the positive region (the first quadrant in Fig. 9(b)) n710 can be expressed as the sum of the loadings of the first and second principal components, and the SES where the first and second principal components are in the positive and negative regions (the second quadrant in Fig. 9(b)) n710 means that it can be expressed by subtracting the loading of the second principal component from the loading of the first principal component. Also, the SES close to the origin n710 means that almost none of the first and second principal components are included. The black circles are the scores corresponding to the SES of normal cells included in the normal specimens A to F, and the white circles are the scores corresponding to the SES of tumor cells in the specimens G to I suspected of mesothelioma. n710 The scores of normal cells are distributed mainly in the third and fourth quadrants and also in the first and second quadrants, while the scores of tumor cells are concentrated in the first and second quadrants. This result shows that normal cells and tumor cells that cannot be clearly classified by the SES in Figs. 5 and 7 can be numerically classified by the second principal component score. n710 Here, the method for obtaining the scores of each principal component will be explained. The score of the first principal component of patient n (n = 1 to N) is shown in the following formula 1. As shown in the following formula 2, w1(λ) is the weighting (eigenvector) of the first principal component, and it is stipulated that the sum of the squares of w1(λ) is 1. As shown in the following formula 3, w1(λ) that maximizes the variance V1 of the score of the first principal component is determined. n710

Number

[0044]

Number

[0045]

Number

[0046]

Number

[0047]

number

[0048] Furthermore, the score of the second principal component for patient n (n=1 to N) is shown in Equation 4 below. Similar to the first principal component, w2(λ) is the weighting of the second principal component, and as shown in Equation 5 below, the sum of squares of w2(λ) must be 1. In addition, as shown in Equation 6 below, an orthogonality condition is imposed that w2(λ) is orthogonal to w1(λ), ​​and the w2(λ) that maximizes the variance of the score of the second principal component under this condition is determined.

[0049]

number

[0050]

number

[0051]

number

[0052] Then, the score of the nth principal component for patient n (n=1 to N) is obtained by adding orthogonality conditions until the variance of the score of the nth principal component is maximized. n (λ) is determined. Thus, spectral data is far simpler than image data, allowing for less arbitrary analysis.

[0053] As shown in Figure 9(b), the scatter plot was created, and the SES of unknown cells was used to determine whether they were normal cells or tumor cells. n710 By comparing these, the type of cell can be estimated. SES obtained from unknown cells. n710Next, the loading scores for the first and second principal components shown in Figure 9(a) were calculated, and the location where they were plotted in Figure 9(b) was used to estimate whether they were normal cells or tumor cells. For example, if the score of an unknown cell is in the third or fourth quadrant, where normal cells have a high score, the unknown cell is likely to be a normal cell. If it is in the first or second quadrant, where tumor cells have a high score, it is likely to be a tumor cell.

[0054] Methods for classifying cells into two classes, normal cells and tumor cells, include the k-nearest neighbors method, the decision tree method, and the random forest method. As shown in Figure 10(a), the k-nearest neighbors method classifies cells by taking a majority vote among the k nearest samples. As shown in Figure 10(b), the decision tree method classifies cells by creating partitions (decision trees) within a scatter plot and taking a majority vote within each partition. As shown in Figure 10(c), the random forest method classifies cells by taking a majority vote among multiple different decision trees and then taking a majority vote among those results.

[0055] As shown in Figure 11, SES cells of normal cells contained in one of the specimens A to F n710 Test data assuming unknown cells, SES of the remaining normal cells n710 The SES of tumor cells contained in one of the samples G-I, used as training data for a predictive model to create a scatter plot. n710 Test data assuming unknown cells, remaining tumor cells SES n710 The data was used as training data for creating scatter plots, and the test data of normal cells and tumor cells were classified into two classes using the method described above. As a result, the accuracy rates of the k-nearest neighbors method, decision tree method, and random forest method were 69.5%, 70.9%, and 72.4%, respectively. This analysis method was used to determine SES n710 This demonstrated that normal cells and tumor cells can be numerically classified. Thus, this embodiment clearly shows that the light scattering spectra of normal cells and tumor cells obtained from actual patients, which are extremely difficult to interpret from a physical standpoint, can be distinguished using machine learning.

[0056] In this analysis, only the first and second principal components are considered. However, by considering higher-order principal components as well, for example, the accuracy of classifying normal cells and tumor cells from minute shape characteristics of light scattering spectra that are invisible to the human eye can be improved, thereby increasing precision. Furthermore, by analyzing the vector information of the dark-field image shown in Figure 3 and the frequency information obtained from the 2D Fourier expansion in the same way as the light scattering spectrum, and incorporating them as features into the learning model along with the light scattering spectrum, even greater precision can be expected.

[0057] Furthermore, in pathological diagnosis, the important task is not to determine whether individual cells in a pathological specimen are normal or abnormal, but to diagnose whether abnormal cells are present in the specimen. For example, even if the discrimination accuracy (accuracy rate) of individual cells is 70%, if the discrimination of several hundred cells contained in a cytological specimen is performed and the results are statistically analyzed, the objective can be achieved at a high level if the presence of abnormal cells in the pathological specimen can be diagnosed with near 100% accuracy. In order to improve the diagnostic accuracy of pathological specimens, not only the discrimination accuracy of a single cell but also the measurement speed and measurement accuracy of the light scattering spectrum are important. Moreover, if the overwhelmingly large number of normal cells can be removed in advance during measurement, and if there is estimated information on how many abnormal cells are contained in the specimen, it is thought that the diagnostic accuracy can be improved dramatically. Thus, this embodiment demonstrates the possibility of achieving highly sensitive and accurate diagnosis of pathological specimens that surpasses conventional methods. [Examples]

[0058] In this example, as in Example 2, we used pathological specimens obtained from actual patients, which were fixed and stained according to existing pathological diagnostic procedures. However, unlike Example 2, we also considered higher-order principal components and present the results of classifying the light scattering spectrum using machine learning. Similar to Example 2, the normalized light scattering spectrum (SES) for cells contained in specimens A-I was obtained. n710)(See Fig. 13) Principal component analysis was performed, and the loadings of the first to tenth principal components were obtained. Fig. 14 shows the contribution rates and loading results of the first principal component (PC1), the second principal component (PC2), ···, and the tenth principal component. From these loadings, the SES of each cell n710 The percentage ratio (score) of the nth principal component (n is 1 to 10) and the mth principal component (m is 1 to 10) contained in the SES of each cell was determined, and a group of plots (scatter plots) with the score of the nth principal component on the horizontal axis and the score of the mth principal component on the vertical axis is shown in Fig. 15. Fig. 15 shows 100 scatter plots, including plots of the same principal component. Among these, since those with the vertical axis and the horizontal axis of each principal component swapped are substantially the same information, when checking 45 scatter plots (the upper right of the matrix in Fig. 15) where n < m for the nth principal component and the mth principal component, it was found that the difference between normal cells and tumor cells can be clearly distinguished by using the scatter plot of the scores of the second principal component (PC2) and the sixth principal component (PC6). Although a clear difference between normal cells and tumor cells is shown in the scatter plot of the scores of the second principal component and the sixth principal component, there is room for a clear difference in other components by changing the cell type or adjusting the analysis parameters.

[0059] As described above, it was found that the difference between normal cells and tumor cells can be clearly distinguished by using the scatter plot of the scores of the second principal component and the sixth principal component. On the other hand, it suggests the possibility that the scatter plot of the scores of the second principal component and the sixth principal component reflects the malignancy of tumor cells. That is, as shown in Fig. 16, the difference between tumor cells and normal cells is not a simple Yes / No, but there is also an intermediate state, and it may be possible to show the intermediate state as the malignancy of tumor cells. There are cells that were clearly diagnosed as mesothelioma by both a pathologist's judgment and a light scattering spectrum judgment (among these, cells with faded pigments were also discriminated), cells with scores near the origin and difficult to judge by the light scattering spectrum, and cells that were clearly diagnosed as reactive mesothelium by both a pathologist's judgment and a light scattering spectrum judgment. It can be shown as the malignancy of tumor cells, such that the malignancy increases as it goes towards the upper left and decreases as it goes towards the lower right.

[0060] Thus, once training data is established, pathologists can refer to the discrimination results from light scattering spectra when they are unsure of their judgment, thereby supporting pathological diagnosis using image data. Furthermore, as training data can be accumulated, the accuracy of discrimination using light scattering spectra can be further improved, and ultimately, fully automated diagnosis without relying on pathologists can be expected. For example, as shown in Figure 17(1), light scattering spectral data can be obtained from unknown cells (pathologists diagnose based on information about cell shape of 1 μm or larger from image data of unknown cells), and information about intracellular structures of 1 μm or smaller can be obtained from the light scattering spectra. Then, as shown in Figure 17(2), a score can be calculated from the light scattering spectral data, compared with the training data, and the malignancy of the cells can be calculated to support pathological diagnosis. Calculating scores from light scattering spectral data is simpler than diagnosis using image data, and tumor cells can be identified with high sensitivity and accuracy. Note that image data and light scattering spectral data contain different information, and complementary improvements in diagnostic accuracy are expected. By fusing the training data for images and the training data for light scattering spectra, improvements in the accuracy of determining cell malignancy are expected. [Examples]

[0061] This example shows the results of identifying unstained living cells in pathological specimens using light scattering spectra. Here, we show optical microscope images of two types of living cells, Jurkat T cell line and U937 monocyte cell line, and the normalized light scattering spectrum (SES) of individual cells. n710 ) was obtained. In the microscopic image shown in Figure 18(1), it is extremely difficult to distinguish between these two types of cells based on their morphological features. In the normalized light scattering spectrum shown in Figure 18(2), there appears to be a difference in shape between the upper spectra, but the difference is not clear in the lower spectra.

[0062] The normalized light scattering spectrum (SES) shown in Figure 18(2) n710 ) are subjected to principal component analysis to obtain loadings of the 1st to 10th principal components, and the SES of individual cells are analyzed. n710Scores were calculated for the nth principal component (n is 1 to 10) and the mth principal component (m is 1 to 10), and the scatter plot shown in Figure 19 was obtained. For almost all principal components, the scores differed between the two types of cells, and the two types of cells were clearly separated for all principal components up to the 10th (PC10). This example shows better results than Examples 2 and 3 described above, which used stained specimens, clearly demonstrating that the present invention can also be applied to living cells. Furthermore, the clear distinction observed in unstained living cells validates the hypothesis presented in Example 1. [Industrial applicability]

[0063] This invention makes it possible to easily distinguish tumor cells that are difficult to identify by conventional optical microscopy observations by pathologists, potentially enabling diagnosis that is independent of the pathologist's experience and skill level, and improving the accuracy of that diagnosis. Furthermore, by systematizing the measurement of light scattering spectra, it may be possible to acquire comprehensive characteristic information of cells with higher sensitivity and speed than optical microscopy images, potentially enabling unmanned and high-speed pathological diagnosis. [Explanation of symbols]

[0064] 1. Micro-scattering spectroscopy measurement system 2. Cell culture dish or specimen 3. Dark-field condenser lens 3' Brightfield Condenser Lens 4. Objective lens 5. Light source (halogen lamp) 6. Spectrometer 7 CMOS camera 10 cell culture dishes 11 Spacers 12 Culture medium 13 Cover glass 14 Cultured cells 20 Cell identification device 21 Scattering Spectrum Acquisition Unit 22 Learning Models 23. The part that distinguishes between normal and abnormal cells. 31 Electromagnetic wave scattering spectrum measuring device 32 Scattering Spectrum 33 Display device for discrimination results 41, 46 Excitation light 42 Samples (scattering material) 43 Mie scattered light and Rayleigh scattered light 44. Raman scattering light (or fluorescence) 45 Spectrometer 47. Optical filters (excitation light cut filters)

Claims

1. A cell discrimination method that involves irradiating a sample containing cells taken from a patient with electromagnetic waves containing all wavelengths in a specific wavelength range, obtaining an electromagnetic wave scattering spectrum due to Rayleigh scattering and Mie scattering that does not contain transmitted light and specular reflected light components from the sample, and using a learning model trained on the electromagnetic wave scattering spectra of normal and abnormal cells to distinguish between normal and abnormal cells from the electromagnetic wave scattering spectrum obtained from an unknown cell.

2. The cell discrimination method according to claim 1, characterized in that the electromagnetic wave scattering spectrum is measured with a dark-field optical system.

3. The aforementioned electromagnetic wave scattering spectrum is two-dimensional data representing the relationship between the wavelength of the electromagnetic wave and the intensity of the scattered electromagnetic wave. The cell discrimination method according to claim 1, wherein the learning model is trained on the two-dimensional data and classifies the differences between normal cells and abnormal cells.

4. The cell discrimination method according to claim 3, wherein the learning model extracts features from the two-dimensional data using principal component analysis, the principal components are extracted by diagonalizing the variance-covariance matrix of scattering intensity at each wavelength, and the differences in features between normal cells and abnormal cells are classified.

5. In the aforementioned principal component analysis, The first principal component is the component whose variance of scattering intensity is maximized at the eigenvector with the largest eigenvalue. The second principal component is the eigenvector with the second largest eigenvalue, which is orthogonal to the first principal component and has the greatest dispersion of scattering intensity. The cell discrimination method according to claim 4, wherein the nth principal component is an eigenvector with the nth largest eigenvalue that is orthogonal to the first (n-1) principal components and has the maximum dispersion of scattering intensity.

6. The cell identification method according to claim 1, wherein the sample is a sample that has been fixed with alcohol or formalin and stained.

7. The cell identification method according to claim 1, wherein the sample is a living cell.

8. The cell discrimination method according to claim 1, wherein the electromagnetic wave scattering spectrum is a forward scattering spectrum.

9. The cell discrimination method according to claim 1, wherein the electromagnetic wave to be irradiated has a wavelength of 100 nm to 1 μm.

10. The cell identification method according to claim 1, wherein the electromagnetic wave to be irradiated is visible light.

11. The cell discrimination method according to claim 1, wherein the difference in the electromagnetic wave scattering spectra between normal cells and abnormal cells is due to an increase or decrease in scattering caused by a change in cell structure that is smaller than or equal to the wavelength of visible light.

12. The cell identification method according to claim 1, wherein the abnormal cells are tumor cells.

13. The cell identification method according to claim 12, wherein the tumor cells are cancer cells.

14. The cell identification method according to claim 1, wherein the electromagnetic wave to be irradiated is ultraviolet light or X-rays.

15. The cell identification method according to claim 1, wherein the electromagnetic wave to be irradiated is infrared light.

16. A cell discrimination device comprising: an irradiation unit that irradiates a sample containing cells taken from a patient with electromagnetic waves including all wavelengths in a specific wavelength range; an electromagnetic wave scattering spectrum acquisition unit that acquires an electromagnetic wave scattering spectrum due to Rayleigh scattering and Mie scattering, which does not include transmitted light and specular reflected light components; a learning model that has been trained on the electromagnetic wave scattering spectra of normal and abnormal cells; and a discrimination unit that uses the learning model to distinguish between normal and abnormal cells from the electromagnetic wave scattering spectrum of a single unknown cell.

17. The cell discrimination device according to claim 16, characterized in that the electromagnetic wave scattering spectrum acquisition unit uses a dark-field optical system.

Images