System and method for unsupervised cell recognition using cell unwrapping

Abstract

Disclosed herein is a system and method for identifying abnormal cells, which can be used to replace or supplement traditional methods of visual identification of abnormal cells. The method described here uses a "cell unwrapping" technique wherein the cell is transformed into a polar coordinate graph and a normalized rectangular matrix and analyzed using machine learning models and / or statistical analysis methods to determine the likelihood that a given cell is abnormal.

Description

Attorney Docket: 8350.2023-236WOSYSTEM AND METHOD FOR UNSUPERVISED CELL RECOGNITION USING CELL UNWRAPPINGRelated Applications

[0001] This application claims the benefit of U.S. Provisional Patent Application No.63 / 657,303, filed June 7, 2024, the contents of which are herebyincorporated herein in their entirety.Background

[0002] The identification of abnormal cells is a cornerstone of medical diagnosis, providing critical insights into various disease processes, from inflammation to cancer. Pathologists, as specialists in disease diagnosis, meticulously examine tissue and cell samples to discern deviations from normal cellular appearance and behavior.

[0003] Abnormal cells are fundamentally characterized by their deviation from the typical appearance and behavior of healthy cells within a given tissue or organ. Unlike normal cells, which adhere to a predictable cycle of growth, division, and programmed death, abnormal cells may exhibit uncontrolled multiplication and fail to undergo natural cell death. The term "atypical" is frequently employed in pathology to describe cells that appear unusual or abnormal under microscopic examination, specifically noting alterations inAttorney Docket: 8350.2023-236WOtheir shape, size, or internal structure, particularly within the cytoplasm or nucleus.

[0004] One of the most striking features of abnormal cells is their significant variability in size and shape, in stark contrast to the typically uniform appearance of normal cells. For example, the cell nucleus, which houses the cell's genetic material (DNA), frequently undergoes profound changes in abnormal cells. It is often noticeably larger than a normal nucleus, and its size and shape can vary considerably from cell to cell. Abnormal nuclei may also appear "decentralized" within the cell. A critical diagnostic feature is an elevated ratio of the nucleus's size to the cytoplasm's volume, known as the "Nucleus-to-Cytoplasm (N:C) Ratio". This means the nucleus occupies a disproportionately large part of the cell. A high N:C ratio is a strong predictor of malignancy, often proving more indicative than just an increased nuclear size alone.

[0005] The identification of abnormal cells carries profound implications for patient management and prognosis. For instance, a diagnosis of dysplasia (i.e., wherein cells begin to look abnormal under the microscope), while indicating abnormality, does not automatically equate to cancer, but it necessitates careful monitoring or intervention to prevent potential progression.

[0006] Initial visual assessment of cells is foundational to the entire diagnostic process. If a significant abnormality is overlooked during the macroscopic assessment, it may not be sampled for microscopic review, potentiallyAttorney Docket: 8350.2023-236WOleading to a missed diagnosis, regardless of the pathologist's microscopic expertise. This highlights a fundamental workflow deficiency which is addressed by this invention.Summary of the Invention

[0007] The present invention automates the process of identifying abnormal cellsand can be used to replace or supplement traditional methods of visual identification of abnormal cells. The method described here uses a "cell unwrapping" technique wherein the cell is transformed into a polar coordinate graph and then to a normalized rectangular matrix and analyzed using machine learning models and / or statistical analysis methods to determine a probability that a given cell is abnormal.

[0008] In some embodiments, data and parameters from the rectangular matrix are fed to a correlation filter transformer network, for example, a neural network (NN), multilayer perceptron (MLP), a support vector machine (SVM) or a convolutional neural network (CNN).

[0009] In some embodiments, data and parameters can be further processedthrough a one-dimensional Fast Fourier Transform (1D-FFT) to present the magnitude of the signal through power spectrum analysis. In some embodiments, the power spectrum can have special properties like shiftAttorney Docket: 8350.2023-236WOinvariance which can be used to rotate the correlation filter transformer network.

[0010] In some embodiments, the center reference point can be computed by traversing the perimeter boundary of the radius and computing the centroid of the nucleus.Brief Description of the Drawings

[0011] By way of example, specific exemplary embodiments of the disclosed system and method will now be described, with reference to the accompanying drawings, in which:

[0012] FIG. 1 is a flowchart of the process of identifying abnormal cells inaccordance with the present invention.

[0013] FIG. 2 is an illustration of the cell unwrapping process wherein polar coordinates are transformed to rectangular coordinates in the form of a matrix.

[0014] FIG. 3 is a graph showing the distance of the boundaries of the cell wall andthe nucleus wall from the reference point.

[0015] FIG. 4. is a system / dataflow diagram of a system implementing thedescribed methods.Detailed DescriptionAttorney Docket: 8350.2023-236WO

[0016] A flowchart of the process is shown in FIG. 1. The method starts at 102 with the localization and semantic segmenting of an image of the cell. Cell images may be obtained by any known means, for example, using light microscopes and digital imaging equipment. In alternate embodiments, the method may be performed in images of cells that are stored in a database or on nonvolatile media.

[0017] The cell is then segmented and unwrapped at 104. The center of radius (i.e., the reference point) is computed by traversing the perimeter boundary of the radius and computing the centroid of the nucleus. A semantic segmentation for the detection of the nucleus and cytoplasm is then performed wherein the segmented region of interest is transformed into a polar coordinate graph and then transformed to a normalized rectangular matrix. The polar graph is defined by a polar curve established with the coordinates (r, 0) by measuring the radius (r) from the center reference point, located inside of the nucleus, to the boundaries of the segmentedareas (i.e., the nucleus and the cytoplasm), and establishing a correlation line between the radius value and the angle 6. In one embodiment, the value of pixels along each line 202, formed as the angle 6 changes, is recorded as a column in matrix 204. In other embodiments, other data may be recorded as the columns in the matrix. The angle 6 may be varied in any desired size of discrete step as it moves around the polar coordinate system from 0°to 360°.Attorney Docket: 8350.2023-236WO

[0018] At 106, various parameters of the unwrapped data may be calculated. For example, the ratio= Pn / Pc, representing the size of the nucleus of the cell to the overall size of the cell (a strong predictor of malignancy) and other cell parameters from cytoplasm and nuclei may be calculated.

[0019] FIG. 3 is a graph of the distance of the boundaries of the cell wall (Rc) and the nucleus wall (Rn) from the reference point, obtained as a result of the cell unwrapping, which can be used to calculate various characteristics indicative of an abnormal cell, for example, if the ratio of the area of the nucleus to the area of the overall cell is too large, if the overall shape of the cell is abnormal, if the nucleus is decentralized within the cell, or if the nucleus is misshaped.

[0020] At 108, the data and parameters may be further processed, in one embodiment, using a 1D-FFT, to present the magnitude of the signal through a power spectrum analysis. The power spectrum has special properties like shift invariance, which can be used to rotate the neural network orcorrelation filter transformer network.

[0021] At 110, the normalized data, any calculated parameters and / or the power spectrum are input to the transformer network, which extracts features from the data that are used to make the normal / abnormal determination. In various embodiments, any combination of the parameters, calculated at 106, the power spectrum features, calculated at 108, or the raw data from the cell unwrapping at 104 may be used as input to the transformer network at 110.Attorney Docket: 8350.2023-236WO

[0022] The extracted features at 110 can then be used to make a prediction, at 112, of the probability that the cell presents as abnormal. In one embodiment, the prediction may be made using a classifier network.

[0023] The feature extraction at 110, in various embodiments, can include the use of deep learning methods, CNNs, transformers, SVMs, correlation filters, statistical pattern recognition methods or signal processing features. In one exemplary embodiment, the feature extraction consists of (1) computing the Fourier transform at 108, (2) performing feature extraction (110) and classification (112) on this domain, and computing the power spectrum features or higher order spectra, and then using deep learning methods, CNNs, Transformers, SVMs, Correlation filters, NNs, statistical pattern recognition methods or other SP to do feature extraction at (110) and / or classification (112) to determine normal or abnormal cells. In another exemplary embodiment, the ratio of perimeters of cytoplasm and nucleus can be calculated at 106 and can be used to give a weighted indicator of abnormality.

[0024] In various embodiments of the invention, the system (See FIG. 4) upon which the processes and methods are implemented may consist of software implementing the processes and executing on a general-purpose computing device or processor. The cell images may be imaged in real time or the methods may operate on stored images of cells (402). The software may include a cell unwrapping module 404 for unwrapping the cell and creatingAttorney Docket: 8350.2023-236WOthe normalized data, a processing module 406 for calculating various parameters of the normalized data, including for example, the ratio of the area of the nucleus to the area of the cell and / or the power spectrum of the data, a feature extraction module 408 for extracting features from any combination of the data, as previously described, and a classifier 410, that classifies the calls as normal or abnormal, based on the extracted features.The system may further include an imaging component used to obtain the images of the cells.

[0025] As would be realized by one of skill in the art, many variations on implementations discussed herein which fall within the scope of theinvention are possible. Specifically, many variations of the parameters, the components used, and their arrangement could be used to obtain similar results. The invention is not meant to be limited to the particular exemplary embodiments disclosed herein. Moreover, it is to be understood that the features of the various embodiments described herein were not mutually exclusive and can exist in various combinations and permutations, even if such combinations or permutations were not made express herein, without departing from the spirit and scope of the invention. Accordingly, the method and apparatus disclosed herein are not to be taken as limitations on the invention but as an illustration thereof.

Claims

Attorney Docket: 8350.2023-236WOClaims:

1. A method for detecting abnormal cells comprising:obtaining an image of the cell;unwrapping the image of the cell into a matrix of normalized data; extracting features from the normalized data; andclassifying the cell as normal or abnormal based on the extracted features.

2. The method of claim 1 wherein the normalized data comprises values of pixels in the cell image.

3. The method of claim 2 wherein the pixel values are collected along a line as the line is rotated around a reference point in discrete steps and further wherein pixel values along the line at each discrete step are recorded in a column in a matrix.

4. The method of claim 3 wherein the reference point is located within the nucleus of the cell.

5. The method of claim 4 wherein the reference point is located at a centroid of the nucleus.

6. The method of claim 1 further comprising:Attorney Docket: 8350.2023-236WOcalculating one or more parameters based on the normalized data; andextracting features from the one or more parameters.

7. The method of claim 1 comprising:calculating a power spectrum of the normalized data; andextracting features from the power spectrum.

8. The method of claim 6 wherein one of the one or more parameters is a ratio of area of the nucleus of the cell to area of the whole cell.

9. The method of claim 1 wherein the feature extractor includes one or more of deep learning methods, CNNs, transformers, SVMs, correlation filters, statistical pattern recognition methods and / or signal processing features.

10. The method of claim 9 wherein the input to the feature extractor is any combination of the raw normalized data, parameters calculated from the raw normalized data and a Fourier transform of the raw normalized data.

11. The system of claim 1 wherein classifying the cell comprises:Attorney Docket: 8350.2023-236WOproviding a probability that the cell is abnormal.

12. A system comprising:a processor; andsoftware that, when executed by the processors, performs themethod of claim 1.

13. The system of claim 12 wherein the software comprises:a cell unwrapping module;a processing module for calculating parameters of the normalized data;a feature extraction module; anda classification module.

Images