Cell classification method and system based on AI

Abstract

The present disclosure provides a system and method of AI-based cell classification that cell classifies cells in a patient sample to determine whether cells indicative of lung cancer are present. In the system and method, cells that are less likely to indicate lung cancer are excluded from subsequent 3D imaging using 2D imaging, while cells that are likely to indicate lung cancer are 3D-imaged. The present disclosure also provides a method of training a 2D cell classifier for a system and method of AI-based cell classification.

Description

Technical Field

[0001] The present disclosure relates to an artificial intelligence (AI)-based 2D cell classification system and method for detecting cells indicative of cancer or cells with characteristics of normal bronchial epithelial cells (BECs) prior to 3D analysis, and a method of training a 2D cell classifier for use in the AI-based cell classification system and method. Background Art

[0002] Lung cancer is the deadliest cancer in the U.S., with more than 31 million patients at high risk for the disease. Early detection is the surest way to reduce lung cancer deaths, but many tests have poor sensitivity and specificity, leading to missed diagnoses that result in higher mortality rates, as well as increased costs and unnecessary suffering from invasive procedures. Summary of the invention

[0003] The present disclosure provides a cell classification method based on artificial intelligence (AI), which includes: a) generating a representative 2D image of cells from a patient sample including multiple cells through an optical tomography system; b) evaluating the representative 2D image using an abnormal cell 2D classifier to determine whether the cell has abnormal characteristics; c) evaluating the representative 2D image using a BEC 2D classifier to determine whether the cell has BEC-like characteristics; and d) if it is determined that the cell has abnormal characteristics or has BEC-like characteristics, generating a 3D image of the cell through the optical tomography system.

[0004] The method may further include one or more of the following additional features in any combination:

[0005] If it is determined that the cell does not have abnormal characteristics and does not have BEC-like characteristics, a 3D image of the cell is not generated;

[0006] e) repeating a) to d) for subset cells within a plurality of cells to generate patient sample data reflecting the total number of cells analyzed with abnormal characteristics, the total number of cells analyzed with BEC-like characteristics, or the total number of cells analyzed;

[0007] f) comparing the total number of cells analyzed by counting with BEC-like characteristics to a threshold value to determine whether the total number of cells analyzed by counting with BEC-like characteristics is sufficient for accurate determination because the total number is at or above the threshold value, or whether the total number of cells analyzed by counting with BEC-like characteristics is insufficient for accurate determination because the total number is below the threshold value;

[0008] Repeating a) to f) until the total number of cells analyzed by counting and analyzing having BEC-like characteristics is sufficient to detect the presence of cells indicative of lung cancer in the patient sample with a pre-selected accuracy;

[0009] The threshold is pre-selected by the number of counted cells having BEC-like characteristics that are statistically likely to detect the presence of cells indicative of lung cancer in the patient sample with a pre-selected accuracy;

[0010] The threshold was within the following ranges: 250 to 2500 counted cells with BEC-like characteristics, 400 to 1900 counted cells with BEC-like characteristics, 600 to 1800 counted cells with BEC-like characteristics, 800 to 1700 counted cells with BEC-like characteristics, 1000 to 1600 counted cells with BEC-like characteristics, 1200 to 1600 counted cells with BEC-like characteristics, or 1300 to 1500 counted cells with BEC-like characteristics;

[0011] The threshold is at least 250 counted cells with BEC-like characteristics, at least 400 counted cells with BEC-like characteristics, at least 500 counted cells with BEC-like characteristics, at least 600 counted cells with BEC-like characteristics, at least 700 counted cells with BEC-like characteristics, at least 800 counted cells with BEC-like characteristics, at least 900 counted cells with BEC-like characteristics, at least 1000 counted cells with BEC-like characteristics, at least 1100 counted cells with BEC-like characteristics, at least 1200 counted cells with BEC-like characteristics, at least 1300 counted cells with BEC-like characteristics, at least 1400 counted cells with BEC-like characteristics, at least 1500 counted cells with BEC-like characteristics, at least 2000 counted cells with BEC-like characteristics, or at least 2500 counted cells with BEC-like characteristics;

[0012] The method has a sensitivity value of at least 90% for pre-selected 2D abnormal cells;

[0013] The method has a specificity value of at least 65% for pre-selected 2D abnormal cells;

[0014] The method has a pre-selected 2D BEC sensitivity value of at least 85%;

[0015] Patient samples were obtained from sputum specimens;

[0016] The patient sample is obtained by separating and preserving a plurality of cells from a sputum sample, the patient sample comprising abnormal cells, BECs, squamous cells, monocytes, lymphocytes, polymorphonuclear leukocytes, other leukocytes, debris, cell fragments, cell clusters, and any combination thereof;

[0017] The abnormal cells are selected from the group consisting of cells exhibiting atypia, dysplastic cells, precancerous cells, pleomorphic dyskeratosis, type II pneumocytes, abnormal squamous cells, adenocarcinoma cells, bronchiolar lung carcinoma cells, abnormal neuroendocrine cells, small cell carcinoma cells, non-small cell carcinoma cells, tumor cells, neoplastic cells, bronchioalveolar carcinoma cells, and any combination thereof;

[0018] Prior to a), pre-treating the patient sample to stain a plurality of cells with a reagent that facilitates generation of a representative 2D image, evaluating the representative 2D image using an abnormal cell 2D classifier, or evaluating the representative 2D image using a BEC 2D classifier;

[0019] Prior to a), BECs or cells that may indicate cancer are enriched in patient samples;

[0020] Prior to a), i) embedding a patient sample in an optical medium and injecting the optical medium with the embedded sample into a capillary; and ii) loading the capillary into an optical tomography system such that the capillary is located between an illumination source and an objective lens of the optical tomography system;

[0021] wherein generating a 2D image of the cell by the optical tomography system comprises scanning the focal plane of the optical tomography system across the single cell in increments of 1 µm to generate a single-plane 2D image of the single cell at each increment, compiling a set of single-plane 2D images of the single cell, and filtering the set of single-plane 2D images of the single cell to generate a representative 2D image of the cell;

[0022] Among them, the representative 2D image of the cell is the image of the central part of the cell;

[0023] wherein evaluating the representative 2D image using the abnormal cell 2D classifier and evaluating the representative 2D image using the BEC 2D classifier both include determining values ​​of a plurality of cell characteristic measurements;

[0024] The cell feature measurements include object shape features, cell shape features, cytoplasm features, cell nucleolus features, chromatin distribution, nuclear size features, nuclear texture features, other morphological elements or any combination thereof.

[0025] The present disclosure also provides a method for training an AI-based cell classification system, which includes: a) operating an optical tomography system to generate a representative 2D image of a cell; b) using the representative 2D image of the cell and a known cell recognition method, assigning a known identifier to each of a plurality of cells, wherein the known identifier is i) abnormal or has an abnormal feature; ii) BEC or has a BEC-like feature; or iii) normal or non-BEC or does not have an abnormal feature or a BEC-like feature; c) calculating a plurality of cell feature measurements for each of the plurality of cells from the representative 2D image of the cell; d) performing regression between the cell feature measurements of each cell and the known identifier to assign an abnormal 2D cell classification score and a BEC 2D cell classification score for each cell; e) assigning a test identifier to each cell based on the abnormal 2D cell classification score and the BEC 2D classification score of the cell, namely i) has an abnormal feature; ii) has a BEC-like feature; or does not have an abnormal feature or a BEC-like feature; f) comparing each test identifier of each cell with the known identifier to calculate the accuracy of the abnormal cell 2D classifier and the BEC 2D classifier.

[0026] The method may further include one or more of the following additional features in any combination:

[0027] comparing the accuracies of the two 2D cell classifiers to a pre-selected accuracy requirement, and designating the 2D cell classifier as trained if the 2D cell classifier meets the pre-selected accuracy requirement;

[0028] comparing the accuracies of the two 2D cell classifiers with a pre-selected accuracy requirement, and if the 2D cell classifier does not meet the pre-selected accuracy requirement, adjusting the regression parameters of the affected 2D cell classifier and repeating steps d) to f);

[0029] The pre-selected accuracy requirement was at least 90% sensitivity value for 2D abnormal cells;

[0030] The preselected accuracy requirement was at least 65% 2D abnormal cell specificity value;

[0031] The pre-selected accuracy requirement is at least 85% of the 2D BEC sensitivity value;

[0032] The pre-selected accuracy requirement was a 2D BEC specificity value of at least 60%;

[0033] Regression includes adaptive boosted logistic regression, random forests, decision trees, or any combination of them;

[0034] Known methods of cell identification include cytological analysis and classification by pathologists;

[0035] generating a 3D image of the plurality of cells and using the 3D image in a known cell identification method;

[0036] The cell feature measurements include object shape features, cell shape features, cytoplasm features, cell nucleolus features, chromatin distribution, nuclear size features, nuclear texture features, other morphological elements or any combination thereof.

[0037] Additionally, the training method may be used to train an AI as described above or otherwise herein.

[0038] The present disclosure also provides a cell classification system, comprising: an optical tomography system, operable to: generate a representative 2D image of a cell from a patient sample; and generate a 3D image of the cell if the cell has abnormal features or BEC-like features; and a processor, operable to: compare the representative 2D image of the cell with an abnormal cell 2D classifier to determine whether the cell has abnormal features; and compare the representative 2D image of the cell with a BEC 2D classifier to determine whether the cell has BEC-like features.

[0039] In some embodiments, the system can be used in the cell sorting methods or training methods, or both, described above or otherwise herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] This patent or application file contains at least one drawing printed in color. Copies of this patent or patent application publication in color will be provided by the Office upon request and payment of the necessary fee.

[0041] The present disclosure may be better understood by reference to the following detailed description taken in conjunction with the accompanying drawings, which are provided by way of example only and in which like elements are designated by letters (e.g., 40a, 40b, 40c), and in which:

[0042] Figure 1 is a schematic diagram of an optical tomography system that may be used in the present disclosure;

[0043] Figure 2 is a schematic diagram of an optical tomography system that may be used in the present disclosure for obtaining multiple 2D images of cells;

[0044] Figure 3 is a flow chart of the lung cancer detection method disclosed herein;

[0045] Figure 4 This disclosure is used for Figure 3 Flowchart of AI-based cell classification method for lung cancer detection;

[0046] Figure 5 is a flow chart of a 2D classifier training method disclosed herein;

[0047] Figure 6 is an example 2D portion of a 3D image of a cell that can be obtained using the optical tomography system of the present disclosure;

[0048] Figure 7 is a cumulative probability density curve (CDF) graph, for example, an abnormal cell 2D classifier result according to the present disclosure, wherein a green dashed line represents a threshold value at which the classifier operates, a red line shows a trend of cells classified as having abnormal features (“abnormal”), a green line shows a trend of cells classified as having BEC-like features (“epithelial”), a light blue line shows a trend of miscellaneous cells (“rest”), and a dark blue line shows a trend of cells having BEC-like features and miscellaneous cells (“combined”);

[0049] Figure 8 is a CDF graph, for example, a BEC 2D classifier result according to the present disclosure, wherein the green dashed line represents the threshold at which the classifier operates, the green line shows the trend of cells classified as having BEC-like features, and the light blue line shows the trend of miscellaneous cells (normal). DETAILED DESCRIPTION

[0050] The following description and the accompanying drawings set forth certain specific details to provide a thorough understanding of the various embodiments disclosed. However, those skilled in the relevant art will recognize that the disclosed embodiments can be practiced in various combinations without one or more of these specific details, or with other methods, components, devices, materials, etc. In other cases, well-known structures or components related to the environment of the present disclosure are not shown or described to avoid unnecessarily obscuring the description of the embodiments. In addition, the various embodiments may be methods, systems, or devices.

[0051] The present disclosure relates to AI-based cell classification methods and systems that can be used to detect cells indicative of cancer in multiple cells in lung-related patient samples, such as sputum. Cells indicative of cancer may be cancer cells, but they may also include non-cancerous cells with abnormal characteristics. In some embodiments, a sample may be designated as containing cells indicative of cancer (and therefore positive for lung cancer) based solely on the detection of non-cancerous cells with abnormal characteristics. When a sample is designated as containing cells indicative of cancer, the patient may be referred for further testing.

[0052] The systems and methods use optical tomography, such as current optical tomography systems, to detect cells and generate 3D images of the cells. However, unlike existing optical tomography systems and methods, the systems and methods of the present disclosure first generate representative 2D images of the cells before generating 3D images of the cells, and use an AI-based 2D cell classifier to analyze the representative 2D images to determine whether a 3D image should be generated. Patient samples contain a large number of different cells that are unlikely to indicate lung cancer and are therefore useless in detecting lung cancer. 3D images of such cells are not required, and generating such images is a waste of resources. Avoiding unnecessary 3D imaging of cells that are useless for detecting lung cancer allows samples to be processed more efficiently.

[0053] The present disclosure also provides a method of training an AI of a 2D cell classifier for use in an AI-based cell classification system or method using cells with known identifiers.

[0054] In some embodiments, the lung cancer may be non-small cell lung cancer (NSCLC), particularly squamous cell carcinoma and adenocarcinoma. In some embodiments, the lung cancer may be small cell lung cancer (SCLC).

[0055] The term "cancer" refers to excessive cell proliferation resulting in uncontrolled growth, lack of differentiation, local tissue invasion or metastasis.

[0056] In some embodiments, cells with abnormal characteristics can include cells exhibiting atypia, dysplastic cells, precancerous cells, pleomorphic dyskeratosis, type II pneumocytes, abnormal squamous cells, adenocarcinoma cells, bronchioalveolar carcinoma cells, abnormal neuroendocrine cells, small cell carcinoma cells, non-small cell carcinoma cells, tumor cells, neoplastic cells, bronchioalveolar carcinoma cells, and any combination thereof.

[0057] In some embodiments, cells having BEC-like characteristics may include cells that normally line the airway lumens of a patient's lungs.

[0058] In some embodiments, the optical tomography system may be a CELL-CT ® system (VisionGate, Inc., Washington, USA).

[0059] AI-based cell classification system

[0060] Now refer to Figure 1 and Figure 2The AI-based cell classification system of the present disclosure may include an optical tomography system 100, or the AI-based cell classification method of the present disclosure may be performed using the optical tomography system 100, which may be used to generate a 3D image and a 2D image of a cell 10 (e.g., cell 10a or cell 10b). Although the operation of the optical tomography system is described as being used to acquire an image of one cell 10, in an optical medium volume in the optical path of a high-power microscope, images of multiple cells 10 within the same optical medium volume may be acquired.

[0061] The optical tomography system 100 may include a cell imaging system 110 including an illumination source 120 optically coupled to an objective lens 130 such that the illumination passes through the microcapillary 30 and any intervening cells 10 before reaching the objective lens 130. The illumination then passes through the objective lens 130 to a beam splitter 140, which causes a portion of the illumination to be deflected to a mirror 150 and reflected back to the beam splitter 140 and then transmitted to a high-speed camera 160, and another portion of the illumination is transmitted directly through the beam splitter 140 to the high-speed camera 160 to generate a pseudo-projection image 40 of the cell 10 in the optical medium 20 contained in the microcapillary 30. During 3D imaging, at least one pseudo-projection image 40 of the cell 10 is generated by vibrating the mirror 150 in a direction 60 (typically using an actuator not shown, such as a piezoelectric motor) to scan the volume occupied by the cell 10, thereby sweeping the focal plane 50 across the cell 10, and then integrating the images to create a pseudo-projection image from a single perspective. Additional pseudo-projection images are obtained by rotating the microcapillary 30. Each pseudo-projection image is a single image representing a sample volume that is larger than the depth of field of the objective lens 130. The high-speed camera 160 generates a plurality of pseudo-projection images 40 for each cell 10, the plurality of pseudo-projection images 40 corresponding to a plurality of axial microcapillary rotation positions, an example of which is shown in FIG. Figure 1 40a, 40b and 40c are shown in FIG. 40a, 40b and 40c. In some embodiments, when the microcapillary 30 is rotated 360°, 500 pseudo-projection images are generated.

[0062] In some embodiments, the optical tomography system 100 is communicatively coupled to a processor 170, which is operable to receive the plurality of pseudo-projection images 40 from the high-speed camera 160 and generate a 3D image (not shown) of the cell 10 using the pseudo-projection images 40. The image before or after processing by the processor 170 may be stored in a communicatively coupled memory 180. The processor 170 may send data about the cell 10, or data related to or derived from a plurality of cells 10 contained in a patient sample, to a communicatively coupled output 190. The data sent to the output 190 may include a 2D or 3D image of one or more cells 10, or a summary of cells in a patient sample analyzed by the optical tomography system 100, such as a graph or table, an index, such as an abnormality index, reflecting the likelihood that the patient has or is at high risk of having lung cancer, or even simply indicating that the sample is positive for cells indicative of lung cancer.

[0063] In embodiments disclosed herein, the optical tomography system 100 may also be operable to generate a representative 2D image of the cell 10. In such embodiments, the objective lens 130 has a focal plane 50 that moves to generate a plurality of 2D images (not shown) as the objective lens 130 sweeps over the microcapillary 30 and any cell 10 in a back-and-forth direction 60. This method of generating a 2D image by moving the objective lens 130 is different from the method of generating a 3D image using a pseudo-projection image generated by vibrating the mirror 150. The processor 170 may be operable to receive the plurality of 2D images and determine a representative 2D image of the cell 10. In some embodiments, the representative 2D image of the cell 10 is an image of a central portion of the cell, such as the center. The processor 170 may then be further operable to perform AI-based cell classification on the representative 2D image using the 2D cell classifier described herein to determine whether the cell 10 has abnormal features or BEC-like features, or does not have abnormal features or BEC-like features. Only when it is determined that the cell 10 has abnormal features or has BEC-like features, the processor 170 instructs the cell imaging system 110 to generate the pseudo-projection image 40 of the cell 10 .

[0064] A large number of cells10 in a patient sputum sample or other sample are unlikely to be indicative of lung cancer, so generating a 3D image of these cells is not useful for detecting lung cancer. Many of these cells do not have the morphology of normal or abnormal cells of the bronchial epithelium, so they can be excluded using only a 2D image. Prior to AI-based cell classification, it was difficult to exclude such cells from patient samples, so in previous cell classification methods, these cells were simply imaged along with cells that were likely to be indicative of lung cancer. The ability to first assess whether a cell is likely to be indicative of lung cancer using a 2D image before capturing a pseudo-projection image40 allows AI-based cell classification to focus on cells10 that provide meaningful information, without wasting processing and analysis time on cells that are not useful for detecting lung cancer.

[0065] In some embodiments, cell imaging system 110 includes an illumination source 120 , an objective lens 130 , a beam splitter 140 , a mirror 150 , and a high-speed camera 160 .

[0066] In some embodiments, the optical tomography system 100 also includes a processor 170 , any communicatively coupled memory 180 , and a communicatively coupled output 190 .

[0067] In some embodiments, the optical tomography system 100 also includes a microcapillary 30, an optical medium 20, or one or more cells 10, but in other embodiments, the optical tomography system 100 does not include one or more of these potential components, although they may be provided for operation of the system.

[0068] In certain embodiments, the AI-based cell classification system 100 is operable to generate images in at least two different modes. One mode includes a cell search mode, during which a 2D image of a cell 10 is generated and processed to generate a representative 2D image, which is then used to determine whether the cell 10 has abnormal features or BEC-like features. In this mode, a 2D image is generated by moving the objective lens 130 along a direction 60 to sweep the focal plane 50 across the cell. The second mode includes a projection image capture mode, during which, if the cell 10 is determined to have abnormal features or BEC-like features during the cell search mode, a pseudo-projection image 40 is generated and used to generate a 3D image of the cell 10. In this mode, a pseudo-projection image 40 is created by vibrating the mirror 150 in the direction 60, thereby generating a 3D image.

[0069] Lung cancer detection methods

[0070] Figure 3 and Figure 4A lung cancer detection method 200 is described, which can be performed using an optical tomography system, such as the optical tomography system 100. Elements of the optical tomography system 100 are referenced as examples in the description of the lung cancer detection method 200. Similar components of different optical tomography systems can also be used in conjunction with the lung cancer detection method 200.

[0071] The lung cancer detection method 200 includes step 210, in which a lung cell sample is collected from a patient. For example, the lung cell sample can be sputum, but other sample types can also be used, such as samples obtained by bronchoalveolar lavage (BAL), nasal swabs, or biopsies. In some embodiments, particularly for detecting other cancers or abnormal cells, the sample can include urine or blood.

[0072] The sample can be subjected to the AI-based cell classification method 300 at any time during which the cells are not degraded to the extent that only limited cell content remains. The duration of such time can depend on storage conditions, such as whether the sample is refrigerated or how or whether the sample is processed. In some embodiments, the sample can be subjected to the AI-based cell classification method 300 within 30 minutes, one hour, two hours, 6 hours, 12 hours, one day, two days, one week, or two weeks after collection.

[0073] Particularly in the case of sputum samples, patient samples can include many cell types that are unlikely to be indicative of lung cancer, such as leukocytes, including polymorphonuclear leukocytes, monocytes, and lymphocytes, cell clusters, oral squamous cells, and squamous intermediate cells (e.g. Figure 6 Samples also often contain large amounts of non-cellular or cellular debris.

[0074] In some embodiments, the sample comprises abnormal cells, BECs, squamous cells, monocytes, lymphocytes, polymorphonuclear leukocytes, other leukocytes, debris, cell fragments, cell clusters, and any combination thereof.

[0075] In step 220, the sample is processed for analysis. Processing may optionally include staining to make any one of the plurality of cells 10 or any feature, such as a nucleus of any one of the plurality of cells 10, more easily detectable using the optical tomography system 100, or otherwise staining or treating the cells with a reagent that facilitates generating a representative 2D image, evaluating the representative 2D image using an abnormal cell 2D classifier, evaluating the representative 2D image using a BEC 2D classifier, generating a 3D image of the cells, or analyzing a 3D image of the cells. In a particular embodiment, the cells may be stained with hematoxylin.

[0076] Processing may optionally include enriching cells of interest in the sample, either alone or in combination with staining. For example, in some embodiments, the sample is enriched for BECs.

[0077] In one embodiment, samples can be enriched for BECs by staining for cytoskeletal proteins and utilizing fluorescence activated cell sorting (FACS) based enrichment.

[0078] In one embodiment, BEC enrichment can include treating the sample with at least one antibody, typically multiple antibodies, having fluorescent conjugates that can be used for FACS-based enrichment. Specifically, the antibodies can bind to BECs, or they can bind to contaminating inflammatory cells, such as neutrophils and macrophages. The antibodies that bind to contaminating inflammatory cells can include anti-CD45 antibodies. In some embodiments, the sample is treated with a combination of antibodies that bind to BECs and antibodies that bind to contaminating inflammatory cells, wherein the antibodies have different fluorescent conjugates.

[0079] In one embodiment, cells can be stained with 4',6-diamidino-2-phenylindole (DAPI) alone or in combination with antibodies for FACS-based cell enrichment purposes.

[0080] Cells can be enriched by FACS, where gating is used to exclude DAPI-positive material (which is often doublets or debris), high side scatter objects, objects bound by anti-inflammatory cell antibodies, or any combination thereof. Cells can also be enriched by FACS, where gating is used to select cells bound by anti-BEC antibodies. In some embodiments, exclusive gating and inclusive gating can be used and can be implemented simultaneously or sequentially.

[0081] After any optional staining or enrichment, sample processing 220 includes placing the cells 10 contained in the sample in an optical medium 20 .

[0082] The optical medium 20 can be any medium that can be reasonably expected to keep the cells 10 intact for the expected duration before and during AI-based cell classification 300. The optical medium 20 can also have a viscosity that allows the optical medium 20 to move through the microcapillary 30. The optical medium 20 can also not interfere with the image generation of the optical tomography system 100. Specifically, the optical medium can have an optical index that matches the optical index of the other components of the optical tomography system 100 and the microcapillary 30 through which the light passes during image acquisition. Typically, the optical tomography system 100 components and the microcapillary 30 through which the light passes also have matching indices. For optimal optical tomography operation, any changes in the movement of the light should be due to encountering the object to be imaged, rather than changes in the optical index of other components or objects in the light path.

[0083] In step 230, an optical medium 20 containing a plurality of cells 10 from a patient sample is injected into a microcapillary 30. In some embodiments, the outer diameter of the microcapillary 30 can be 500 μm or less, for example, between 30 μm and 500 μm. In some embodiments, the inner diameter of the microcapillary 30 can be 400 μm or less, for example, between 30 μm and 400 μm, such as 50 μm. In one embodiment, the entire portion of the patient sample to be analyzed is placed in one microcapillary 30. In another embodiment, portions of the patient sample to be analyzed are placed in a plurality of microcapillaries 30, which can be sequentially subjected to the AI-based cell classification method 300. In yet another embodiment, the sample can be pumped from a sample container through the microcapillary 30.

[0084] In step 240 , the microcapillary 30 is loaded into the optical tomography system 100 such that the microcapillary 30 is located between the illumination source 120 and the objective lens 130 .

[0085] In step 250, the optical medium 20 and any cells 10 contained therein are advanced (prior to initial step 240) or through the microcapillary 30 by applying pressure at one end of the microcapillary such that different volumes of the optical medium 20 carrying different portions of the patient sample are located in the optical path of the optical tomography system 100, between the illumination source 120 and the objective lens 130. In some embodiments, a plunger (not shown) is used to advance the optical medium 20. For example, a plunger can be applied to a container (not shown) of the optical medium 20 and the patient sample connected to the microcapillary 30 to force additional optical medium 20 from the container into the microcapillary 30.

[0086] Next, the method proceeds to AI-based cell classification 300, which includes a cell search mode 310 and a projection image capture mode 320, both of which are Figure 4 Detailed description is given in further detail in .

[0087] In the cell search mode 310, the optical tomography system 100 generates 2D images of the cell 10, which are further processed to generate a representative 2D image, which is used to determine whether the cell 10 has abnormal characteristics or BEC-like characteristics. In some embodiments, the algorithm selects a central portion of the cell 10, such as an image of the center, as the representative 2D image. In some embodiments, in the cell search mode 310, the optical tomography system 100 sweeps the focal plane 50 across the cell 10 along the direction 60 at intervals of 1 μm to capture a series of 2D images. Multiple cells 10 can be identified in the same volume of optical medium 20 in the optical path of the optical tomography system 100.

[0088] In the projection image capture mode 320 , each cell 10 identified as having abnormal features or having BEC-like features is further imaged to generate a plurality of pseudo-projection images 40 , which are used to generate a 3D image of the cell 10 .

[0089] In step 260, it is determined whether a preselected number of cells having BEC-like characteristics have been imaged in 3D. If the preselected number has been reached, the process proceeds to step 270. Otherwise, the process returns to step 250 and the sample is pushed into the microcapillary.

[0090] In step 270, an image of the cell selected for 3D imaging is provided to the user, for example, using output 190. The results may include results specific to a single cell 10, such as a 3D image of the cell 10. The AI-based cell classification results may also include data based on an analysis of all or part of a patient sample or multiple cells 10, such as an abnormality index and / or cancer threshold, such as Figure 4 As further explained in . For example, the AI-based cell classification results may include a list or other indication of individual cells 10 that were found to have abnormal features or BEC-like features or were otherwise marked according to pre-selected criteria so that the user can further review the 3D image. The AI-based cell classification results may include graphical, statistical or numerical information, such as the number of cells detected (usually the total number of cells analyzed), the number of cells for which the 3D image was generated, the number of counted cells with abnormal features, the number of counted cells with BEC-like features, or the relative proportion of any of these cell groups. The AI-based cell classification results may also include information about the accuracy of the prediction. The AI-based cell classification results may include an index reflecting the likelihood that a patient has or is at high risk of developing lung cancer, such as an abnormality index, or even simply indicating that the sample is positive for cells indicative of lung cancer.

[0091] Sample identification data can also be provided together with the AI-based cell classification results.

[0092] Specific reference Figure 4 , the AI-based cell classification method 300 may include a cell search mode 310 and a projection image capture mode 320 .

[0093] The cell search mode 310 may include step 400 in which the optical tomography system 100 generates a plurality of single focal plane 2D images of a cross-sectional area of ​​the microcapillary 30 by moving the objective lens 130 in a direction 60 so that the focal plane 50 is swept back and forth across the microcapillary (and any cells in any focal plane 50), with each image being located at a different focal plane 50. In one embodiment, the single focal plane 2D images may be taken in the direction 60 in increments of 1 μm.

[0094] In step 410, a representative 2D image of the cell 10 is generated. To this end, the processor 170 compiles and filters at least a portion of the plurality of single focal plane 2D images to determine whether the 2D image contains features associated with the cell 10 or other solid object in the optical medium 20, such as being darker than the optical medium 20. If the plurality of single focal plane 2D images contain features associated with the cell 10 or other solid object, the processor 170 analyzes at least a portion of the plurality of single focal plane 2D images associated with the cell 10 (e.g., all or a portion of the 2D image contains a dark area) to determine which should be designated as the representative 2D image based on pre-selected criteria. In one embodiment, the processor determines which 2D image represents a central portion, such as the center, of the cell 10. The image representing the central portion of the cell 10 is designated as the representative 2D image of the cell 10.

[0095] If a representative 2D image is generated, the process proceeds to step 420. In some embodiments, rather than proceeding directly to step 250, the method determines whether another cell 10 is present in the volume of optical medium and proceeds to step 410 for the other cell (and optionally to step 400 if further 2D images are desired). This operation may be repeated for all, a set number, or a set proportion of the hypothetical cells 10 in the volume of optical medium 20 in the optical path of the optical tomography system 100.

[0096] In step 420, the processor 170 evaluates the representative 2D image using an abnormal cell 2D classifier to determine whether the cell 10 has abnormal features. The processor can use an algorithm trained or implemented by AI in step 420. The abnormal cell 2D classifier can be trained according to the training method 500 herein. Specifically, the abnormal cell 2D classifier can identify multiple features of the cell 10 and perform cell feature measurements, and then compare at least a subset of these cell feature measurements with the training values ​​of cell feature measurements related to cells with abnormal features and optionally cells without abnormal features. Based on the comparison, the cell can be identified as having abnormal features or not having abnormal features. In order to facilitate the identification and measurement of multiple cell features, the abnormal cell 2D classifier can segment the representative 2D image into a cell nucleus part and a non-cell nucleus part. Cell feature measurements can include object shape features, cell shape features, cytoplasm features, cell nucleolus features, chromatin distribution, nuclear size features, nuclear texture features, other morphological elements or any combination thereof.

[0097] The abnormal cell 2D classifier may have a pre-selected abnormal cell 2D accuracy, typically a pre-selected abnormal cell 2D sensitivity or a pre-selected abnormal cell 2D specificity.

[0098] If it is determined that the cell 10 has abnormal features in 2D, the method can proceed directly to the projection image capture mode 320, as shown. However, in some embodiments, even if it is determined that the cell 10 has abnormal features in 2D, the method can still proceed to step 430 (not shown) to determine whether the cell 10 has BEC-like features in 2D.

[0099] In the illustrated embodiment, if it is determined that the cell 10 does not have abnormal characteristics, the method proceeds to step 430 .

[0100] Next, in step 430, the processor 170 evaluates the representative 2D image using a BEC 2D classifier to determine whether the cell 10 has a BEC-like feature. The processor may use an algorithm trained or implemented by AI in step 430. The BEC 2D classifier may be trained according to the training method 500 herein. Specifically, the BEC 2D classifier may identify multiple features of the cell 10 and perform cell feature measurements, and then compare at least one set of identified and measured cell feature measurements with training values ​​of cell feature measurements associated with cells having BEC-like features and optionally cells not having BEC-like features. Based on the comparison, the cell may be identified as having a BEC-like feature or not having a BEC-like feature. Cell features and cell feature measurements may be selected from the same group for the abnormal cell 2D classifier described above. In some embodiments, cell feature identification and measurement may be performed separately for use in an abnormal cell 2D classifier and a BEC 2D classifier. Based on the comparison, the cell may be designated as having a BEC-like feature or not having a BEC-like feature.

[0101] A BEC 2D classifier can have a pre-selected 2D BEC accuracy, a typical 2D BEC sensitivity or a 2D BEC specificity.

[0102] The abnormal cell 2D classifier and the BEC 2D classifier, when operated together, can have a pre-selected abnormal cell referral rate or a pre-selected BEC referral rate.

[0103] If it is determined that the cell 10 has BEC-like features, the method proceeds directly to the projected image capture mode 320 as shown. If it is determined that the cell 10 does not have BEC-like features, no 3D image is captured and the method proceeds to step 250. In some embodiments (not shown), rather than proceeding directly to step 250, the method determines whether another cell 10 is present in the volume of the optical medium and continues to perform steps 410-430 for the other cell (and optionally step 400 if further 2D images are desired). This can be repeated for all, a set number, or a set proportion of the hypothetical cells 10 in the volume of the optical medium 20 in the optical path of the optical tomography system 100.

[0104] In the projection image capture mode 320, in step 440, the optical tomography system 100 generates a plurality of pseudo projection images 40 by vibrating the mirror 150. These images may be taken in a 360 degree rotation around the cell 10. However, the pseudo projections 40 may be taken in a rotation around the cell 10 as small as 180 degrees.

[0105] Next, in step 450, the processor 170 generates a 3D image of the cell 10 using at least a portion of the plurality of pseudo projection images 40. Typically, all pseudo projection images 40 are used to generate the 3D image of the cell 10. However, if it is determined that the pseudo projection images 40 are of poor quality or may contain errors, the 3D imaging of the cell may be stopped. In addition, the 3D image is typically generated using pseudo projection images 40 covering a 360-degree rotation around the cell 10, but pseudo projection images covering a rotation as small as 180 degrees around the cell 10 may be used.

[0106] In step 460, processor 170 uses a 3D classifier to determine whether the cell has abnormal characteristics or BEC-like characteristics. This information is used to determine whether the cell is likely to indicate lung cancer. The 3D classifier may also have a pre-selected accuracy value.

[0107] The process then returns to step 260 to determine whether the preselected number of BECs have been 3D imaged. If the preselected number has been reached, the process continues to step 270, which in some embodiments may include steps 470 and 480.

[0108] In step 470, the data associated with the cells are stored as patient sample data. The patient sample data can be used to determine an abnormality index. The patient sample data can also reflect the total number of cells counted and analyzed with abnormal characteristics, the total number of cells counted and analyzed with BEC-like characteristics, or the total number of cells counted and analyzed, and if additional 2D images are required, steps 410-470 and optional step 400 can be repeated for additional cells until a pre-selected threshold for any or all of these totals is reached.

[0109] In step 480, the processor compares the total number of cells counted and analyzed with abnormal characteristics to the total number of cells counted and analyzed with BEC-like characteristics or the total number of cells counted and analyzed to generate an abnormality index. For example, the abnormality index can be related to the ratio of any of these totals.

[0110] In step 490, the processor compares the abnormality index to a cancer threshold value set corresponding to patient samples designated as positive for lung cancer cells.

[0111] The overall method 300 includes 2D cell classification and 3D cell classification, which can have a pre-selected measurement accuracy, such as a pre-selected measurement sensitivity or a pre-selected measurement specificity. The cell classification method 300 described herein can be much faster than a similar method that does not use a 2D cell classifier but generates a 3D image of each identified cell or assumed cell. In some embodiments, the method 300 can reduce the cell classification time, such as for steps 400-460, by at least 25%, at least 35%, at least 40%, at least 45%, at least 50%, or in the range of 25% to 60%, 25% to 50%, 25% to 45%, 25% to 40%, 25% to 35%, 35% to 60%, 35% to 50%, 35% to 45%, 35% to 40%, 40% to 60%, 40% to 50%, 40% to 45%, 45% to 60% or 45% to 50%.

[0112] Although in the embodiments described in the cell classification method 300, 3D images are not generated for cells that do not have abnormal features or BEC-like features determined using 2D classification, in other embodiments, 3D images can be generated for at least one cell or multiple cells that do not have abnormal features or BEC-like features. For example, such images can be used to verify the performance or accuracy of one or both of the AI-based cell classification system or method and / or the component 2D cell classifier. Although generating such 3D images increases the total cell classification time, as long as 3D images are not generated for all cells, a reduction in time should still be achieved compared to methods that perform 3D imaging of all or nearly all cells. In some embodiments, the total number of cells that do not have abnormal features or BEC-like features for which 3D images are generated can be limited to 25% or less, 20% or less, 15% or less, 10% or less, 5% or less, 1% or less of the total number of counted and analyzed cells with abnormal features, the total number of counted and analyzed cells with BEC-like features, or the total number of counted and analyzed cells, or between 0.1% and 25%, 0.1% and 20%, 0. 1% to 15%, 0.1% to 10%, 0.1% to 5%, 0.1% to 1%, 1% to 25%, 1% to 20%, 1% to 15%, 1% to 10%, 1% to 5%, 5% to 25%, 5% to 20%, 5% to 15%, 5% to 10%, 10% to 25%, 10% to 20%, 10% to 15%, 15% to 15%, 15% to 20%, or 20% to 25%. Cells selected for 3D imaging that do not have abnormal or BEC-like characteristics can be randomly selected or selected based on pre-selected criteria, such as criteria that result in the generation of 3D images of at least one of the most common cell types in the sample.

[0113] Compared to similar methods that do not use abnormal cell 2D classifiers and BEC 2D classifiers (collectively referred to as 2D cell classifiers) but generate 3D images of each identified cell or hypothetical cell, the AI-based cell classification method 300 described herein can also significantly reduce the number of cells for which a 3D image is generated. In some embodiments, the method 300 can reduce the number of cells for which a 3D image is generated by at least 30%, at least 40%, at least 50%, at least 55%, at least 60%, or in the range of 30% to 70%, 30% to 60%, 30% to 55%, 30% to 50%, 30% to 40%, 40% to 70%, 40% to 60%, 40% to 55%, 40% to 50%, 50% to 70%, 50% to 60%, 50% to 55%, 55% to 70%, 55% to 60%, or 60% to 70%.

[0114] The abnormal cell 2D classifier, the BEC 2D classifier and the overall AI-based cell classification method can have a pre-selected accuracy.

[0115] For any assay, the “sensitivity” of the assay is defined as the percentage of assay items (e.g., cells or the entire sample) that are actually positive for a characteristic but are correctly identified as positive by the assay.

[0116] For any assay, the “specificity” of the assay is defined as the percentage of assay items (e.g., cells or the entire sample) that are actually negative for a property but are correctly identified as positive by the assay.

[0117] For the 2D cell classifiers used in the cell classification method 300, sensitivity is often a more important accuracy metric than specificity because only cells that are positive for abnormal features or BEC-like features will be imaged in 3D. Therefore, poor sensitivity of either 2D cell classifier may result in cells that may indicate cancer not being imaged in 3D and not being included in the lung cancer detection results, which may result in false negatives. Conversely, if the specificity is poor, more cells than necessary will be imaged in 3D, which will unnecessarily slow down the AI-based cell classification speed, but will not result in any cells that may indicate cancer being missed and false negatives.

[0118] In the context of an abnormal cell 2D classifier, the 2D abnormal cell sensitivity is the percentage of cells that are determined to be abnormal or will be determined to be abnormal using a known identification method, such as cytology (e.g., reviewed by a pathologist) and are identified as having abnormal characteristics by the abnormal cell 2D classifier. The 2D abnormal cell specificity is the percentage of cells that are determined to be normal or will be determined to be normal using a known identification method, such as cytology, and are identified as not having abnormal characteristics by the abnormal cell 2D classifier.

[0119] In some embodiments, the preselected 2D abnormal cell sensitivity can be at least 90%, at least 95%, at least 98%, or in the range of 90% to 100%, 90% to 99.9%, 90% to 99%, 90% to 98%, 90% to 95%, 95% to 100%, 95% to 99.9%, 95% to 98%, 98% to 100%, 98% to 99.9% or 98% to 99%.

[0120] In other embodiments, the pre-selected 2D abnormal cell specificity can be at least 65%, at least 70%, at least 75%, or in the range of 65% to 100%, 65% to 99%, 65% to 90%, 65% to 80%, 65% to 75%, 65% to 70%, 70% to 100%, 70% to 99%, 70% to 90%, 70% to 80%, 70% to 75%, 75% to 100%, 75% to 99%, 75% to 90% or 75% to 80%.

[0121] In the context of a BEC 2D classifier, the 2D BEC sensitivity is the percentage of cells that are or will be determined to be BECs using a known identification method, such as cytology, and are identified by the BEC 2D classifier as having BEC-like characteristics. The 2D BEC specificity is the percentage of cells that are or will be determined to be not BECs using a known identification method, such as cytology, and are identified by the BEC 2D classifier as not having BEC-like characteristics.

[0122] In some embodiments, the preselected 2D BEC sensitivity may be at least 85%, at least 90%, at least 95%, or in the range of 85% to 100%, 85% to 99%, 85% to 95%, 85% to 90%, 90% to 100%, 90% to 99%, 90% to 95%, 95% to 100%, or 95% to 99%.

[0123] In some embodiments, the preselected 2D BEC specificity may be at least 60%, at least 65%, at least 70%, or in the range of 60% to 100%, 60% to 90%, 60% to 80%, 60% to 70%, 65% to 100%, 65% to 90%, 65% to 80%, 65% to 70%, 70% to 100%, 70% to 90%, or 70% to 80%.

[0124] In the context of AI-based 2D cell classification using two 2D cell classifiers, an abnormal cell referral rate can also be calculated. The abnormal cell referral rate refers to the percentage of cells that are determined to be abnormal or will be determined to be abnormal using a known identification method, such as cytology, and are identified as having abnormal characteristics or having BEC-like characteristics and thus require 3D imaging. The abnormal cell referral rate may be higher than the 2D abnormal cell sensitivity because cells that are abnormal but not identified as having abnormal characteristics may still be identified as having BEC-like characteristics and therefore require 3D imaging.

[0125] In some embodiments, the preselected abnormal cell referral rate can be at least 70%, at least 80%, at least 85%, at least 90%, at least 95%, at least 98%, at least 99%, or between 70% and 99.9%, 80% and 99.9%, 85% and 99.9%, 90% and 99.9%, 95% and 99.9%, 98% and 99.9%, 99% and 99.9%, 70% and 99%, 8 ... % to 99%, 90% to 99%, 95% to 99%, 98% to 99%, 70% to 98%, 80% to 98%, 85% to 98%, 90% to 98%, 95% to 98%, 70% to 95%, 80% to 95%, 85% to 95%, 90% to 95%, 70% to 90%, 80% to 90%, 85% to 90%, 70% to 85%, 80% to 85%, or 70% to 80%.

[0126] In the context of AI-based 2D cell classification using the Abnormal Cell 2D Classifier alone or in combination with the BEC 2D Classifier, the Abnormal Cell True Positive Referral Rate can also be calculated. The Abnormal Cell True Positive Referral Rate is the percentage of cells that are or would be determined to be abnormal using a known identification method, such as cytology, and that are identified as having abnormal characteristics by the Abnormal Cell 2D Classifier. In practice, the Abnormal Cell True Positive Referral Rate may not be as important as the Abnormal Cell Referral Rate because accurate results can be obtained regardless of the abnormal cells being imaged in 3D. In some embodiments, the abnormal cell true positive referral rate can be at least 60%, at least 70%, at least 80%, at least 90%, or in the range of 60% to 99.9%, 70% to 99.9%, 80% to 99.9%, 90% to 99.9%, 60% to 99%, 70% to 99%, 80% to 99%, 90% to 99%, 60% to 90%, 70% to 90%, 80% to 90%, 60% to 80%, 70% to 80%, or 60% to 70%.

[0127] In the context of AI-based 2D cell classification using two 2D cell classifiers, the BEC referral rate can also be calculated and considered when determining accuracy. The BEC referral rate is the percentage of cells that are or will be determined to be BECs using a known identification method, such as cytology, and are identified as having abnormal characteristics or BEC-like characteristics and are therefore imaged in 3D. The BEC referral rate may be higher than the BEC 2D classifier sensitivity because BECs that are not identified as having BEC-like characteristics may still be identified as having abnormal characteristics and therefore imaged in 3D. A given number of BECs are typically imaged to ensure the accuracy of such lung cancer detection, so the actual number of BECs imaged in 3D will affect the time required to perform lung cancer testing on a given patient sample, or whether the sample is considered sufficient to obtain an accurate result.

[0128] In some embodiments, the preselected BEC referral rate may be at least 70%, at least 75%, at least 80%, at least 85%, at least 90%, at least 91%, at least 95%, or between 70% and 99.9%, 75% and 99.9%, 80% and 99.9%, 85% and 99.9%, 90% and 99.9%, 91% and 99.9%, 95% and 99.9%, 70% and 99%, 75% and 99%, 80% and 99%, 85% and 99.9%, 9%, 90% to 99%, 91% to 99%, 95% to 99%, 70% to 95%, 75% to 95%, 80% to 95%, 85% to 95%, 90% to 95%, 70% to 91%, 75% to 91%, 80% to 91%, 85% to 91%, 90% to 91%, 70% to 90%, 75% to 90%, 80% to 90%, 85% to 90%, 70% to 85%, 75% to 85%, or 80% to 85%.

[0129] In the context of AI-based 2D cell classification using the BEC 2D classifier alone or in combination with the abnormal cell 2D classifier, a BEC true positive referral rate can also be calculated. The BEC true positive referral rate is the percentage of cells that are or would be determined to be BEC using a known identification method, such as cytology, and that are identified as having BEC-like features by the BEC 2D classifier. In practice, the BEC true positive referral rate may not be as important as the BEC referral rate, as accurate results can be obtained regardless of why 3D imaging of BECs is performed. In some embodiments, the BEC true positive referral rate can be at least 50%, at least 60%, at least 70%, at least 80%, at least 90%, or in the range of 50% to 99.9%, 60% to 99.9%, 70% to 99.9%, 80% to 99.9%, 90% to 99.9%, 50% to 99%, 60% to 99%, 70% to 99%, 80% to 99%, 90% to 99%, 50% to 90%, 60% to 90%, 70% to 90%, 80% to 90%, 50% to 80%, 60% to 80%, 70% to 80%, 50% to 70%, 60% to 70%, or 50% to 60%.

[0130] The AI-based cell classification method 300 may also have a pre-selected 2D rejection rate. The 2D rejection rate is the proportion of the total number of cells counted and analyzed that are identified as not having abnormal or BEC-like features and therefore not subjected to 3D imaging. In some embodiments, the 2D rejection rate can be 70% or less, 60% or less, 55% or less, 50% or less, or in the range of 0% to 70%, 1% to 70%, 10% to 70%, 25% to 70%, 40% to 70%, 50% to 70%, 55% to 70%, 60% to 70%, 0% to 60%, 1% to 60%, 10% to 60%, 25% to 60%, 40% to 60%, 50% to 60%, 55% to 60%, 0% to 55%, 1% to 55%, 10% to 55%, 25% to 55%, 40% to 55%, 50% to 55%, 0% to 50%, 1% to 50%, 10% to 50%, 25% to 50%, or 40% to 50%.

[0131] Accuracy values ​​can also be calculated for 3D classifiers. Specifically, the AI-based cell classification method can have a 3D abnormal cell sensitivity, which is the percentage of cells that are determined to be abnormal or will be determined to be abnormal using a known identification method, such as cytology (e.g., examination by a pathologist) and are identified as abnormal (or positive, indicating lung cancer) by the 3D classifier. When calculating the 3D abnormal cell sensitivity, the number of false negative cells that the 2D classifier does not reference 3D imaging must be subtracted from the total number of known abnormal cells because the 3D classifier has no chance to classify these cells as abnormal. A pre-selected minimum 3D abnormal cell sensitivity helps avoid false negative test results. Similar calculations can be performed for other accuracy parameters of the 3D cell classifier.

[0132] The lung cancer detection method 200 also has a pre-selected lung cancer accuracy parameter for the entire lung cancer test. Lung cancer sensitivity is the percentage of patients who actually have lung cancer and are determined to have lung cancer by the lung cancer detection method (positive test results). The AI-based cell classification method 300 also has a pre-selected measurement accuracy parameter. In some embodiments, it may be the same as the lung cancer accuracy parameter.

[0133] In general, the sensitivity of any lung cancer detection method using 3D imaging of the type used in method 300 depends on the number of BECs imaged in 3D. The number of BECs imaged in 3D from a patient sample can be determined from the total number of BECs counted and can be set as a threshold so that a preselected assay accuracy, e.g., a preselected assay sensitivity, is statistically likely to have been achieved. The method 300 can continue analyzing the cells 10 until a preselected threshold reflecting the total number of BECs counted is reached, and the preselected threshold can be at least 250, at least 400, at least 500, at least 600, at least 700, at least 800, at least 900, at least 1000, at least 1100, at least 1200, at least 1300, at least 1400, at least 1500, at least 2000, at least 2500, or in the range of 250 to 2500, 250 to 2000, 400 to 1900, 600 to 1800, 800 to 1700, 1000 to 1600, 1200 to 1600, or 1300 to 1500.

[0134] Method 300 may reach a preselected sensitivity more quickly than similar methods that do not use abnormal cell 2D classifiers and BEC 2D classifiers but instead generate 3D images of each cell identified or hypothetical cell because time is not spent capturing 3D images of cells that are unlikely to be indicative of lung cancer.

[0135] The abnormal cell sensitivity is the percentage of cells that are identified or will be identified as abnormal using a known identification method and identified as abnormal by the AI-based cell classification method 300. The abnormal cell sensitivity can be approximated by multiplying the 2D abnormal cell sensitivity by the 3D abnormal cell sensitivity. The abnormal cell specificity is the percentage of cells that are identified or will be identified as normal using a known identification method and identified as normal by the AI-based cell classification method 300. The abnormal cell sensitivity can be approximated by multiplying the 2D abnormal cell specificity by the 3D abnormal cell specificity.

[0136] Meyer, MG, et al. (2015), CELL-CT ® A three-dimensional cell imaging technology platform enables the detection of lung cancer using the non-invasive LuCED sputum test. Cancer Cytopathology, 123: 512-523 (doi.org / 10.1002 / cncy.21576); Wilbur, DC, et al. (2015). Automated three-dimensional morphometric analysis of sputum samples for lung cancer detection: performance characteristics support use in lung cancer screening. Cancer Cytopathology, 123: 548-556 (doi.org / 10.1002 / cncy.21565); US6519355, US6522775, US6591003, US6636623, US6697508, U S7197355, US7494809, US7569789, US7738945, US7811825, US7835561, US7867778, US7787112, US7907765 , US7933010, US8090183, US8155420, US8947510, US9594072, US10753857, US11069054 and US20200018704 are all incorporated herein by reference in their entirety, particularly relating to components of the optical tomography system, basic operation (including potential cell staining and enrichment), image formation (including formation of pseudo-projection images) and 3D classifiers, and the lung cancer detection methods and systems described herein.

[0137] Training Methods

[0138] The present disclosure also includes Figure 5 Method 500, shown in a flowchart of FIG. , is used to train a 2D cell classifier that can be used in method 200, method 300, or method 310. The 2D classifier training is controlled by a binary ground truth; for the abnormal cell 2D classifier, a cell either has abnormal features or does not have abnormal features, and for the BEC 2D classifier, a cell either has BEC-like features or does not have BEC-like features.

[0139] Elements of the optical tomography system 100 are referenced as examples in the description of the training method 500. Similar components of different optical tomography systems may also be used in conjunction with the training method 500. However, the training values ​​generated using the training method 500, as implemented in the abnormal cell 2D classifier and the BEC 2D classifier, are most useful when used with the same optical tomography system used in the training method 500, or with an optical tomography system having the same or highly similar components and having performance specifications that are unlikely to produce any difference in results beyond a pre-selected accuracy.

[0140] In step 510, optical tomography system 100 is used to generate a set of abnormal cell representative 2D images of a plurality of cells identified as abnormal by a known identification method, such as cytology, a set of BEC representative 2D images of a plurality of cells identified as BEC by a known identification method, and a set of normal / non-BEC representative 2D images of a plurality of lung sample cells identified as normal and not BEC by a known identification method, for example, according to steps 400 and 410 of method 310. In some embodiments, for training purposes, a cell is identified as abnormal by a known identification method if a cytopathologist confirms that the cell belongs to one of the following diagnostic categories: 1) atypia, dysplasia, precancerous cell, malignant cell, or 2) polymorphic dyskeratosis, type II pneumocyte. In some embodiments, for training purposes, a cell is identified as a BEC by a known identification method if two cytopathologists independently confirm that the cell is a BEC. In some embodiments, if a cell is randomly obtained from a patient sample and is identified as a normal cell by a cytopathologist because it does not meet the abnormal cell criteria, the cell is identified as a normal cell, and if any cytopathologist identifies the cell as a non-BEC, the cell is identified as a non-BEC.

[0141] In some embodiments, cells used for training purposes may be obtained from the same type of patient sample (e.g., sputum) or prepared in the same manner (e.g., placed in the same optical medium or stained with the same dye) as used in conjunction with the AI-based cell classification methods and systems that will operate with the training values.

[0142] Although method 500 illustrates simultaneous training of an abnormal cell 2D classifier and a BEC 2D classifier for simplicity, the 2D cell classifier may alternatively be trained separately using a separate set of training cells, as described in the examples.

[0143] In step 520 , the processor calculates a plurality of cell characteristic measurements for each of the plurality of cells using the representative 2D image of the cells generated in step 510 .

[0144] In step 530, the processor performs a regression between the cell feature measurements and the known identifiers (i) abnormal, ii) BEC, or iii) normal and non-BEC, or in an approach where a separate 2D cell classifier is trained for iii) normal or non-BEC (whichever is the opposite of the other known identifier used), to assign an abnormal 2D cell classification score and a BEC 2D cell classification score to each cell.

[0145] In step 540 , the processor assigns a test identifier of i) having abnormal characteristics, ii) having BEC-like characteristics, or iii) not having abnormal or BEC-like characteristics based on the abnormal 2D cell classification score and the BEC 2D classification score for each cell.

[0146] In step 550, each test identifier is compared to the respective known identifier for each cell to calculate the accuracy of the two 2D cell classifiers.

[0147] In step 560, the accuracy of the 2D cell classifier can be compared with a pre-selected accuracy requirement, and if the accuracy requirement is met, the 2D cell classifier is designated as trained, or, if the accuracy requirement is not met, the regression parameters of the 2D cell classifier that does not meet the accuracy requirement are adjusted in step 570, and steps 530 to 560 are repeated for at least the adjusted 2D cell classifier.

[0148] In some embodiments, the processor adapts the two-dimensional cell classifier algorithm using adaptive boosted logistic regression, random forest, decision tree, or any combination thereof.Adaptive boosted logistic regression uses logistic regression in an iterative cycle to improve the overall classification value.

[0149] In some embodiments, the preselected 2D abnormal cell sensitivity value can be at least 90%, at least 95%, or in the range of 90% to 100%, 90% to 99.9%, 90% to 99%, 90% to 98%, 95% to 100%, 95% to 99.9%, 95% to 99%, or 95% to 98%, or any other 2D abnormal cell sensitivity value described herein for an abnormal cell 2D classifier.

[0150] In some embodiments, the preselected 2D abnormal cell specificity value can be at least 65%, at least 70%, at least 75%, or in the range of 65% to 100%, 65% to 99%, 65% to 90%, 65% to 80%, 65% to 75%, 70% to 100%, 70% to 99%, 70% to 90%, 70% to 80%, 70% to 75%, 75% to 100%, 75% to 90%, or 75% to 80%, or any other 2D abnormal cell specificity value described herein for an abnormal cell 2D classifier.

[0151] In some embodiments, the preselected 2D BEC sensitivity value can be at least 85%, at least 90%, at least 95%, or in the range of 85% to 100%, 85% to 99.9%, 85% to 99%, 95% to 90%, 90% to 100%, 90% to 99.9%, 90% to 95%, 95% to 100%, 95% to 99.9% or 95% to 99%, or any other 2D BEC sensitivity value described herein for a BEC 2D classifier.

[0152] In some embodiments, the pre-selected 2D BEC specificity value can be at least 60%, at least 65%, at least 70%, or in the range of 60% to 100%, 60% to 99%, 60% to 90%, 60% to 80%, 60% to 75%, 60% to 70%, 60% to 65%, 65% to 100%, 65% to 99%, 65% to 90%, 65% to 80%, 65% to 75%, 65% to 70%, 70% to 100%, 70% to 99%, 70% to 90%, 70% to 80%, or 70% to 75%, or any other 2D BEC specificity value described herein for a BEC 2D classifier.

[0153] Throughout the specification, claims, and drawings, the following terms have the meanings explicitly associated herein unless the context clearly dictates otherwise. The term "herein" refers to the specification, claims, and drawings associated with the present application. The phrases "in one embodiment," "in another embodiment," "in various embodiments," "in some embodiments," "in other embodiments," and other variations thereof refer to one or more features, structures, functions, limitations, or characteristics of the present disclosure, and are not limited to the same or different embodiments, unless the context clearly dictates otherwise.

[0154] In this description, unless otherwise specified, any concentration range, percentage range, ratio range or integer range should be understood to include any value or sub-range within the range. It should also be noted that, unless otherwise specified by the content, the term "or" is generally used in a sense that includes "or" (i.e., representing one, two or any combination of alternatives). In addition, as used in this specification and the appended claims, unless otherwise specified by the content, the singular forms "one", "an" and "the" include plural references. The terms "including" and "having" and their variants are used synonymously and should be interpreted as non-restrictive. The term "combination thereof" used herein refers to all possible combinations of the items listed before the term. For example, "A, B, C or a combination thereof" is intended to refer to any of the following: A, B, C, AB, AC, BC or ABC. Similarly, the term "combination thereof" used herein refers to all possible combinations of the items listed before the term. For example, "A, B, C and a combination thereof" is intended to refer to all of the following: A, B, C, AB, AC, BC and ABC.

[0155] Example

[0156] Example 1

[0157] Training of 2D cell classifier

[0158] Using CELL-CT ® The optical tomography system implements a lung cancer detection method that is designed to detect lung cancer in its pre-invasive stages, when treatment is easier and more likely to be successful.

[0159] Use the following method:

[0160] 1. Patient samples for testing are fixed, stained, and BECs are enriched.

[0161] 2. The patient sample is then suspended in an oil-based optical medium. The cells in the optical medium are then inserted into a glass microcapillary tube with an inner diameter of approximately 60 µm. Pressure is applied to the medium to move the cells into the optical path of a high-power microscope in an optical tomography system. (See Figure 1 ).

[0162] 3. Once a cell enters the optical path of the high-power microscope, the cell search mode will be activated. In the cell search mode, CELL-CT ® The focal plane was continuously swept across the lumen of the microcapillary in steps of 1 µm to identify the linear, radial, and angular positions of cells. Detected dark objects above a certain size were identified as potential 3D captures. The 2D image set was filtered to obtain the central image of the cell. This 2D image was then evaluated using a 2D cell classifier to determine whether the cell had abnormal or BEC-like features, as described below.

[0163] 4. Each cell was also analyzed in projection image capture mode, during which the capillary was rotated so that the optical tomography system generated 500 high-resolution images of the same diagonal cross section of the capillary, which were taken over 360 degrees of capillary rotation. These images are pseudo-projection images, which are simulations of the projection images created by integrating the light from the objective lens as the focal plane sweeps across the nucleus of any cell included in the image. Therefore, the pseudo-projection images represent the entire nuclear contents from a single image taken from a single perspective.

[0164] 5. Process the pseudo-projection images to correct for residual noise and motion artifacts.

[0165] 6. Processing the corrected pseudo-projection image using filtered back projection to produce a 3D tomographic representation of the cell, also referred to as a 3D image. A 3D tomographic representation is created for a plurality of cells, including a plurality of cells identified in a volume of optical medium within an optical path of a high power microscope before the optical medium is propelled into a different volume by pressure. Figure 6 Examples of such 3D images showing squamous intermediate cells and dysplastic cells.

[0166] 7. A cytopathologist evaluates the 3D images to determine if the cells are abnormal, BECs, normal, or not BECs.

[0167] The abnormal 2D cell classification score specifies the likelihood that the examined subject has abnormal characteristics. The abnormal 2D cell classification scores for a large number of imaged cells are used to prepare a composite abnormal 2D cell classification score, which determines whether a patient is likely to have lung cancer.

[0168] The BEC 2D cell classification score specifies the likelihood that the object under investigation has BEC-like characteristics. The BEC classification scores for a large number of imaged cells are used to prepare a composite BEC 2D cell classification score, which helps determine how many BECs have been analyzed.

[0169] The abnormal cell 2D classifier and BEC 2D classifier for lung cancer detection method were trained and tested. Using the cell images obtained as described above, the cytopathologist assigned a known identifier to each cell using a conventional known cell recognition method and gave a test identifier in the training method.

[0170] The development of the abnormal cell 2D classifier followed a supervised learning process controlled by a binary known identifier (abnormal / normal or with abnormal features / not with abnormal features). A cell was assigned the known identifier “abnormal” or “with abnormal features” if the cytopathologist confirmed that the cell belonged to one of the following diagnostic categories:

[0171] - Atypical, dysplastic, precancerous, malignant cells

[0172] - Polymorphous parakeratosis, type II pneumocytes.

[0173] If a cell was obtained randomly and designated as normal by a cytopathologist (due to the lack of the above abnormal criteria), the cell was designated as having no abnormal features or “normal”.

[0174] BEC 2D classifier development followed a supervised learning process controlled by a binary known identifier (BEC / non-BEC, or with BEC-like features / without BEC-like features). When two cytographers independently agreed that it was a BEC, the cell was assigned the known identifier “BEC” or “with BEC-like features”.

[0175] Cells classified as non-BECs by any cytopathologist are assigned the known identifier “non-BEC” or “not having BEC-like features”.

[0176] For training purposes, the abnormal cell 2D classifier used only the identifier “not having abnormal features” and the BEC 2D classifier used only the identifier “not having BEC-like features”, however even during training, cells may be designated as not having abnormal features or BEC-like features.

[0177] For both known and test identifiers, having an abnormal feature was assigned a value of 1. For both known and test identifiers, not having an abnormal feature was assigned a value of 0. A training set of 16,046 cells was imaged and provided with known identifiers by a cytopathologist and test identifiers by the abnormal cell 2D classifier. To provide test identifiers, cells and nuclei were segmented from the voxel blocks containing the cells, and the optical tomography system then automatically measured features of each cell, which were then used to assign a test identifier. The test identifier of each cell was then compared to the known identifier to determine if the test identifier was correct. This aggregated dataset formed the training set to create the abnormal cell 2D classifier. Adaptive boosted logistic regression (Ada-boost) was used to improve the abnormal cell 2D classifier. Ada-boost uses logistic regression in an iterative cycle to improve the overall classification value. Cross-validation was used to limit the complexity of the model to ensure that it had the ability to generalize.

[0178] The trained abnormal cell 2D classifier is used to classify cells that are not part of the training set. Figure 7 Results showing known identifiers and corresponding test identifiers for the trained abnormal cell 2D classifier.

[0179] The 2D abnormal cell sensitivity is 95% and the 2D abnormal cell specificity is 74.7%. The sensitivity needs to be greater than 90% and the specificity needs to be greater than 65% for the abnormal cell 2D classifier to be considered trained, so the trained abnormal cell 2D classifier exceeds the pre-selected accuracy value.

[0180] Table 1 - Abnormal cell 2D classifier training set

[0181] Category 0 (Test) Category 1 (Testing) total Category 0 (known) 11565 3922 15487 Category 1 (known) 27 532 559 total 11592 4454 16046

[0182] Figure 7 The cumulative probability density curve of the abnormal cell 2D classifier trained using the results of Table 1 is shown. For cells with known abnormal feature identifiers, 93.3% (1-0.067) obtained the correct test identifier (red line) and then underwent 3D imaging. 96.2% of cells had known BEC-like feature identifiers (green line) and were then identified by the BEC 2D classifier, and 87.5% of cells had known non-BEC-like feature identifiers (light blue line) and were then identified by the BEC 2D classifier. Therefore, 12.5% ​​of miscellaneous cells (cells without abnormal or BEC-like features) were ultimately imaged in 3D.

[0183] For both known and test identifiers, having BEC-like features was assigned a value of 1. For both known and test identifiers, not having BEC-like features was assigned a value of 0. A training set of 12,572 cells was imaged and provided with known identifiers by a cytopathologist and test identifiers by the BEC 2D classifier. To provide test identifiers, cells and nuclei were segmented from the voxel blocks containing the cells, and features of each cell were automatically measured by the optical tomography system and then used to assign test identifiers. The test identifier of each cell was then compared to the known identifier to determine if the test identifier was correct. This aggregated dataset formed the training set for creating the BEC 2D classifier. Adaptive boosted logistic regression (Ada-boost) was used to improve the abnormal cell 2D classifier. Ada-boost utilizes logistic regression in an iterative cycle to improve the overall classification value. Cross-validation was utilized to limit the complexity of the model to ensure that it had the ability to generalize.

[0184] The trained BEC 2D classifier was used to classify cells that were not part of the training set. Figure 8 Results showing known identifiers and corresponding test identifiers for the trained BEC cell classifier.

[0185] The 2D BEC sensitivity is 88% and the 2D BEC specificity is 66%. The BEC 2D classifier needs to have a sensitivity greater than 85% and a specificity greater than 60% to be considered trained, so the trained BEC 2D classifier exceeds the pre-selected accuracy values.

[0186] Table 2 - BEC 2D classifier training set

[0187] Category 0 (Test) Category 1 (Testing) total Category 0 (known) 5960 3061 9021 Category 1 (known) 426 3125 3551 total 6386 6186 12572

[0188] Figure 8 The cumulative probability density curve of the BEC 2D classifier trained using the results of Table 2 is shown. For cells with known BEC-like feature identifiers, 86.6% (1-0.134) obtained the correct test identifier (green line) and then 3D imaging was performed. For cells with known non-BEC-like feature identifiers, 81% obtained the correct test identifier (light blue line).

[0189] Figure 7 and Figure 8 The cumulative results are as follows:

[0190] 93.3% of cells with abnormal features (based on known identifiers) were imaged in 3D;

[0191] 5.8% (1-0.962) + 0.866*0.962*100% = 89% of cells with BEC-like features (based on known identifiers) were imaged in 3D; and

[0192] 0.875*0.81*100%=71% of cells without aberrant or BEC-like features were imaged only in 2D.

[0193] Example 2

[0194] Evaluation of 2D cell classifiers

[0195] The trained 2D cell classifier was tested three times independently after development to ensure robustness. In each test, the lung cancer detection method was performed using the trained 2D classifier and the trained 3D classifier, but, as during training in Example 1, 3D images of all objects for which 2D images were obtained were acquired in order to evaluate accuracy.

[0196] Algorithm Verification Research

[0197] The trained 2D cell classifier was embedded in executable code that also implemented the rest of the standard lung cancer test and was tested using cells from five adenocarcinoma and two squamous cell carcinoma cell lines. Various accuracy metrics were evaluated. The abnormal cell referral rate was 98.7%. The BEC 2D classifier demonstrated a BEC 2D sensitivity of 90% and a BEC 2D specificity of 44%.

[0198] Pilot Clinical Accuracy Study (CAS)

[0199] A pilot CAS was performed using a set of samples similar to those used to train the 2D cell classifier to evaluate the referral rate of the 2D cell classifier. 34 samples representing 267,659 cells were tested. The calculated abnormal cell referral rate was 95%. The BEC referral rate was 91%. The abnormal cell referral rate and BEC referral rate indicate that the training of the 2D algorithm generalizes well to cell populations that were not part of the training process. Case-level accuracy (whether a patient is positive or negative for cells indicative of lung cancer) was not altered by using the 2D cell classifier and excluding some cells from 3D imaging.

[0200] CAS-A

[0201] The CAS-A study was conducted to further evaluate the performance of the 2D cell classifier. As with the pilot CAS, 38 samples were tested, representing 757,627 cells. Because there were too few abnormal cells in CAS-A, the abnormal cell referral rate was calculated to be 90% by merging the data with the pilot CAS. The BEC referral rate was calculated to be 87%.

[0202] The 2D rejection rate was not specified but was estimated to be 53% during 2D cell classifier training. The 2D rejection rate measured in the CAS-A study was 60%, and the median 2D rejection rate per case was 52%.

[0203] Summary of Clinical Results

[0204] Overall, the abnormal cell referral rate and BEC referral rate are within specifications, while the 2D rejection rate is higher than the values ​​measured in training. This suggests that the training of the 2D cell classifier generalizes well to cell populations that were not part of the training process.

[0205] Example 3

[0206] Test duration

[0207] When performing a lung cancer detection method according to the present disclosure that excludes some cells from 3D imaging using a 2D cell classifier and comparing the optical tomography image acquisition portion of the AI-based cell classification method with the optical tomography image acquisition portion of a previous cell classification method in which all cells were imaged in 3D, it was observed that while maintaining the pre-selected accuracy, the number of cells in the patient's 3D image was reduced by 52% and the method duration was reduced by 45%.

[0208] The various embodiments described above can be combined to provide further embodiments. All U.S. patents, U.S. patent application publications, U.S. patent applications, foreign patents, foreign patent applications, and non-patent publications mentioned in this specification and / or listed in the application data sheet, including U.S. patent application serial number 63 / 394,550 filed on August 2, 2022, are incorporated herein by reference in their entirety. If necessary, aspects of the embodiments can be modified to adopt the concepts of various patents, applications, and publications to provide further embodiments.

[0209] In light of the above detailed description, these and other changes may be made to the embodiments. In general, in the claims, the terms used should not be interpreted as limiting the claims to the specific embodiments disclosed in the specification and claims, but should be interpreted to include all possible embodiments and the full scope of equivalents to which these claims are entitled. Therefore, the claims are not limited by the present disclosure.

Claims

1. A cell classification method based on artificial intelligence (AI), comprising: a) A representative 2D image of cells is generated from a patient sample including multiple cells using an optical tomography system. b) evaluating the representative 2D image using an abnormal cell 2D classifier to determine whether the cell has abnormal characteristics; c) evaluating the representative 2D image using a BEC 2D classifier to determine whether the cell has BEC-like features; as well as d) if it is determined that the cell has the abnormal feature or the BEC-like feature, generating a 3D image of the cell by the optical tomography system.

2. The method according to claim 1, wherein: If it is determined that the cell does not have the abnormal characteristics and does not have the BEC-like characteristics, then a 3D image of the cell is not generated.

3. The method according to claim 1 or claim 2, further comprising: e) repeating a) to d) for subset cells within the plurality of cells to generate patient sample data reflecting the total number of cells counted and analyzed having the abnormal characteristics, the total number of cells counted and analyzed having the BEC-like characteristics, or the total number of cells counted and analyzed.

4. The method according to claim 3, further comprising: f) comparing the total number of cells counted and analyzed having the BEC-like characteristics with a threshold value to determine whether the total number of cells counted and analyzed having the BEC-like characteristics is sufficient for accurate determination because the total number of cells counted and analyzed is at or above the threshold value, or whether the total number of cells counted and analyzed having the BEC-like characteristics is insufficient for accurate determination because the total number of cells counted and analyzed is below the threshold value.

5. The method according to claim 4, further comprising: a) through f) are repeated until the total number of cells counted and analyzed having the BEC-like characteristics is sufficient to detect the presence of cells indicative of lung cancer in the patient sample with a pre-selected accuracy.

6. The method according to claim 5, wherein: The threshold is pre-selected from the number of counted cells having the BEC-like characteristics that are statistically likely to detect the presence of cells indicative of lung cancer in the patient sample with the pre-selected accuracy.

7. The method according to any one of claims 4 to 6, wherein: The threshold is within the range of 250 to 2500 counted cells having the BEC-like characteristics, 400 to 1900 counted cells having the BEC-like characteristics, 600 to 1800 counted cells having the BEC-like characteristics, 800 to 1700 counted cells having the BEC-like characteristics, 1000 to 1600 counted cells having the BEC-like characteristics, 1200 to 1600 counted cells having the BEC-like characteristics, or 1300 to 1500 counted cells having the BEC-like characteristics.

8. The method according to any one of claims 4 to 7, wherein: The threshold value is at least 250 counted cells having the BEC-like characteristics, at least 400 counted cells having the BEC-like characteristics, at least 500 counted cells having the BEC-like characteristics, at least 600 counted cells having the BEC-like characteristics, at least 700 counted cells having the BEC-like characteristics, at least 800 counted cells having the BEC-like characteristics, at least 900 counted cells having the BEC-like characteristics, at least 1000 counted cells having the BEC-like characteristics, at least 1100 counted cells having the BEC-like characteristics, at least 1200 counted cells having the BEC-like characteristics, at least 1300 counted cells having the BEC-like characteristics, at least 1400 counted cells having the BEC-like characteristics, at least 1500 counted cells having the BEC-like characteristics, at least 2000 counted cells having the BEC-like characteristics, or at least 2500 counted cells having the BEC-like characteristics.

9. The method according to any one of claims 1 to 8, wherein: The method has a pre-selected 2D abnormal cell sensitivity value of at least 90%.

10. The method according to any one of claims 1 to 9, wherein: The method has a specificity value of at least 65% for pre-selected 2D abnormal cells.

11. The method according to any one of claims 1 to 10, wherein: The method has a pre-selected 2D BEC sensitivity value of at least 85%.

12. The method according to any one of claims 1 to 11, wherein: The method has a pre-selected 2D BEC specificity value of at least 60%.

13. The method according to any one of claims 1 to 12, wherein: The patient samples were obtained from sputum samples.

14. The method according to claim 11, wherein: The patient sample is obtained by isolating and preserving a plurality of cells from a sputum sample.

15. The method according to any one of claims 1 to 14, wherein: The patient sample includes abnormal cells, BECs, squamous cells, monocytes, lymphocytes, polymorphonuclear leukocytes, other leukocytes, debris, cell fragments, cell clusters, and any combination thereof.

16. The method according to claim 15, wherein: The abnormal cells are selected from the group consisting of cells exhibiting atypia, dysplastic cells, precancerous cells, pleomorphic dyskeratosis, type II pneumocytes, abnormal squamous cells, adenocarcinoma cells, bronchiolar lung carcinoma cells, abnormal neuroendocrine cells, small cell carcinoma cells, non-small cell carcinoma cells, tumor cells, neoplastic cells, bronchioalveolar carcinoma cells, and any combination thereof.

17. The method according to any one of claims 1 to 16, further comprising: Prior to a), the patient sample is pre-processed to stain the plurality of cells with a reagent that facilitates generating the representative 2D image, evaluating the representative 2D image using the abnormal cell 2D classifier, or evaluating the representative 2D image using the BEC 2D classifier.

18. The method according to any one of claims 1 to 17, further comprising: Prior to a), BECs or cells that may indicate cancer are enriched in the patient sample.

19. The method according to any one of claims 1 to 18, further comprising: Before a), i) embedding the patient sample into an optical medium and injecting the optical medium with the embedded sample into a capillary; as well as ii) loading the capillary into the optical tomography system such that the capillary is located between an illumination source and an objective lens of the optical tomography system.

20. The method according to any one of claims 1 to 19, wherein: Generating a 2D image of the cell by the optical tomography system includes scanning the focal plane of the optical tomography system over a single cell in increments of 1 µm to generate a single-plane 2D image of the single cell at each increment, compiling a group of single-plane 2D images of the single cell, and filtering the group of single-plane 2D images of the single cell to generate a representative 2D image of the cell.

21. The method according to claim 20, wherein: The representative 2D image of the cell is an image of the central portion of the cell.

22. The method according to any one of claims 1 to 21, wherein: Both evaluating the representative 2D image using the abnormal cell 2D classifier and evaluating the representative 2D image using the BEC 2D classifier include determining values ​​for a plurality of cell characteristic measurements.

23. The method according to claim 22, wherein: The cell feature measurements include object shape features, cell shape features, cytoplasm features, cell nucleolus features, chromatin distribution, nuclear size features, nuclear texture features, other morphological elements or any combination thereof.

24. A method for training an AI-based cell classification system, comprising: a) Operate the optical tomography system to generate representative 2D images of cells. b) assigning a known identifier to each cell in a plurality of cells using the representative 2D image of the cell and a known cell identification method, wherein the known identifier is i) abnormal or has abnormal characteristics, ii) a BEC or has BEC-like characteristics, or iii) normal or non-BEC or does not have abnormal characteristics or BEC-like characteristics; c) calculating a plurality of cell characteristic measurements for each of the plurality of cells based on the representative 2D image of the cell; d) performing a regression between the cell characteristic measurements and the known identifiers for each cell to assign an abnormal 2D cell classification score and a BEC 2D cell classification score to each cell; e) assigning, based on the abnormal 2D cell classification score and the BEC 2D classification score of the cell, a test identifier to each cell i) as having an abnormal characteristic; ii) has BEC-like characteristics; or have no abnormal characteristics or BEC-like features; f) comparing each test identifier of each cell with the known identifier to calculate the accuracy of the abnormal cell 2D classifier and the BEC 2D classifier.

25. The method according to claim 24, further comprising: The accuracies of the two 2D cell classifiers are compared to a pre-selected accuracy requirement, and if a 2D cell classifier meets the pre-selected accuracy requirement, the 2D cell classifier is designated as trained.

26. The method of claim 24 or claim 25, further comprising: The accuracies of the two 2D cell classifiers are compared to a pre-selected accuracy requirement, and if a 2D cell classifier does not meet the pre-selected accuracy requirement, the regression parameters of the affected 2D cell classifier are adjusted and steps d) to f) are repeated.

27. The method according to any one of claims 24 to 26, wherein: The pre-selected accuracy requirement is a 2D abnormal cell sensitivity value of at least 90%.

28. The method according to any one of claims 24 to 27, wherein: The pre-selected accuracy requirement is a 2D abnormal cell specificity value of at least 65%.

29. The method according to any one of claims 24 to 28, wherein: The pre-selected accuracy requirement is at least 85% of the 2D BEC sensitivity value.

30. The method according to any one of claims 24 to 29, wherein: The pre-selected accuracy requirement is a 2D BEC specificity value of at least 60%.

31. The method according to any one of claims 24 to 30, wherein: The regression includes adaptive boosted logistic regression, random forest, decision tree, or any combination thereof.

32. The method according to any one of claims 24 to 31, wherein: Such known cell identification methods include cytological analysis and classification by pathologists.

33. The method according to any one of claims 24 to 32, further comprising: A 3D image of a plurality of cells is generated and used in the known cell identification method.

34. The method according to any one of claims 24 to 33, wherein: The cell feature measurements include object shape features, cell shape features, cytoplasm features, cell nucleolus features, chromatin distribution, nuclear size features, nuclear texture features, other morphological elements or any combination thereof.

35. A cell classification system comprising: An optical tomography system capable of operating to: Generate representative 2D images of cells from patient samples; as well as If the cell has abnormal characteristics or BEC-like characteristics, generating a 3D image of the cell; and A processor operable to: comparing the representative 2D image of the cell with an abnormal cell 2D classifier to determine whether the cell has abnormal characteristics; as well as The representative 2D image of the cell is compared to a BEC 2D classifier to determine whether the cell has BEC-like features.

Images