A method and system for mitochondria-based single cell feature extraction and analysis

Abstract

The application relates to a kind of mitochondria-based single cell feature extraction and analysis method and system, comprising: obtaining the multiple modal images such as bright field image, nucleus fluorescent image and mitochondria fluorescent image of single cell;Image preprocessing is carried out to the three modal images obtained;For different structures such as mitochondria, morphological and texture features are extracted, and feature analysis is carried out;Further, through the fusion of mitochondria and machine learning technology, the automatic classification of cells is realized.The application is used for the classification of human cervical epithelial cells (H8) and cervical cancer cells (HeLa), and the machine learning analysis of morphological features and texture features shows the potential of mitochondria in the classification of cervical cells.The application has strong applicability, can be combined with machine learning and other analysis methods, and can be applied to various biological cells, has universality, and is easy to popularize.

Description

Technical Field

[0001] This invention relates to a method and system for single-cell feature extraction and analysis based on mitochondria, which can be used for the classification of human cervical epithelial cells (H8) and cervical cancer cells (HeLa). The analysis of clinical morphological and texture features shows the potential of intracellular mitochondria in cervical cell classification. Background Technology

[0002] Cells are the basic building blocks of organisms; all organisms are composed of cells. Mitochondria are organelles found in most cells and are the structures that produce energy within the cell. Besides providing energy, mitochondria participate in processes such as cell differentiation and apoptosis, and have the ability to regulate cell growth and the cell cycle. Generally, the number of mitochondria in a cell depends on its metabolic level; cells with more vigorous metabolic activity have more mitochondria. Under pathological conditions, the size and number of mitochondria change accordingly. Scientists have also discovered mitochondrial switches that promote cancer metastasis; when the function of mitochondria in tumor cells changes, it promotes cell migration, ultimately leading to successful tumor metastasis. Therefore, research on the morphological and textural characteristics of mitochondria is beneficial for the diagnosis and treatment of certain diseases and cancers. Currently, when clinicians use cytological experience to identify cells, they mainly focus on abnormal morphological changes and the size of the cell nucleus, while research on the morphological and textural characteristics of mitochondria is very limited.

[0003] Cervical cancer is a prevalent malignant tumor among women. Fortunately, cervical cancer develops slowly, and early screening offers a significant opportunity to reduce treatment costs and mortality. Common clinical screening methods for cervical cancer include human papillomavirus (HPV) testing, liquid-based cytology testing, and colposcopy, with biopsy pathology serving as the gold standard for diagnosis. These clinical examinations typically require close involvement of pathologists. However, the lack of experienced pathologists is a major obstacle to large-scale screening globally, especially in underdeveloped regions.

[0004] Currently, the feature extraction and analysis of mitochondria in single cells involves two main approaches. One is to use targeted probes to connect mitochondrial genes or proteins to analyze the relationship between mitochondria and diseases such as cancer. While this method is highly accurate, the process is complex and expensive. On the other hand, fluorescence imaging analysis of mitochondria in single cells mainly focuses on the quantitative analysis of their intensity, neglecting the quantitative analysis of their morphological characteristics. However, since morphological characteristics carry information about the distribution and quantity of mitochondria that are inseparable from diseases such as cancer, it is necessary to extract and quantify mitochondrial morphological characteristics. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for extracting and analyzing single-cell features based on mitochondria, and also provides a system for extracting and analyzing single-cell features based on mitochondria.

[0006] This invention verifies the potential of mitochondria as a biomarker for cell carcinogenesis and the potential of mitochondrial images for automatic cell classification using machine learning. The morphological and textural features of mitochondria demonstrate their ability to effectively perform feature analysis and automatic classification of single cells, and can be applied to the classification of human cervical epithelial cells and cervical cancer cells.

[0007] Terminology Explanation:

[0008] 1. Bright field image: Bright field is an imaging mode of modern microscopes. If only the transmitted beam is allowed to pass through the objective aperture to form an image, then it is a bright field image.

[0009] 2. Cell nuclear fluorescence images: After cell samples are stained and fixed with DAPI, the cell nuclei are specifically represented as blue cell fluorescence images under halogen lamp illumination, obtained through an excitation filter (excitation wavelength 325-369nm) and an emission filter (emission wavelength 430-490nm).

[0010] 3. Mitochondrial fluorescence images: After cell samples are stained and fixed with MitoTracker Red CMXRos, mitochondria are specifically represented as red cellular fluorescence images under halogen lamp illumination, obtained through an excitation filter (excitation wavelength 542-582nm) and an emission filter (emission wavelength 604-644nm).

[0011] 4. Median filtering for noise reduction: Median filtering is a non-linear smoothing technique that sets the gray value of each pixel to the median of the gray values ​​of all pixels within a certain neighborhood window of that pixel. This can remove noise and smooth the image.

[0012] 5. Canny operator edge extraction: The Canny operator is an edge detection operator and is currently one of the most theoretically complete edge detection algorithms. It can be calculated using toolkits in common image processing software such as MATLAB and OpenCV.

[0013] 6. The expansion operation expands the pixels of a binary image outwards, connecting previously independent neighboring regions and creating a more complete closed region to represent cell edges.

[0014] 7. Closure operation: Performs closure calculation on binary images to merge narrow gaps and slender bends, filling gaps in the contour.

[0015] 8. Hole filling: Fills in regions and "holes" in a binary image.

[0016] 9. Outline drawing: Use fine lines to outline the edges of the image, resulting in a closed line, which is the edge of the cell image.

[0017] 10. k-Nearest Neighbors (kNN) is a training dataset that, given a new input instance, finds the k nearest neighbors to the input instance in the training dataset. If the majority of these k nearest neighbors belong to a certain class, the input instance is classified into that class.

[0018] 11. Support Vector Machine (SVM) is a type of generalized linear classifier that performs binary classification of data using supervised learning. Its decision boundary is the hyperplane with the maximum margin calculated from the learning samples. SVM can perform non-linear classification using kernel methods and is one of the common kernel learning methods.

[0019] 12. Random Forest (RF) refers to a classifier that uses multiple trees to train and predict samples.

[0020] 13. Logistic Regression (LR) is a generalized linear regression analysis model that is commonly used in data mining, automatic disease diagnosis, economic forecasting, and other fields.

[0021] 14. CLGM operator, gray-level co-occurrence matrix operator.

[0022] 15. LBP operator, Local Binary Mode Operator.

[0023] 16. The light excitation system consists of a halogen lamp (U-HGLGPS, Olympus, Japan) whose light source is filtered by a neutral density filter and then propagates into a fluorescent filter cube.

[0024] 17. Sample system: The cell suspension is placed in a sandwich chip, which consists of two 1.1 mm thick glass slides as the bottom and top layers and two 170 μm covering layers as spacers, forming a static unit chamber, which is mounted on the stage of an inverted microscope (IX73, Olympus, Japan).

[0025] 18. A three-modal asynchronous imaging system, by switching filter banks, can asynchronously acquire bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of static cells on a chip through a 40 objective lens (LUCPLFLN40X, Olympus, Japan) onto a CMOS sensor (800D, Canon, Japan).

[0026] The technical solution of this invention is as follows:

[0027] A method for single-cell feature extraction and analysis based on mitochondria, comprising:

[0028] Acquire images of three modalities of a single cell; including: bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell;

[0029] Image preprocessing is performed on the acquired images of the three modalities;

[0030] For different data characteristics, features of three modalities were extracted and feature analysis was performed to obtain the ratio of nucleus to cytoplasm area and the ratio of mitochondrial to cytoplasm area, and the histogram distribution of the ratio of nucleus to cytoplasm and the ratio of mitochondrial to cytoplasm and their nuclear density estimation curves were obtained.

[0031] Automatic classification is achieved through machine learning.

[0032] According to a preferred embodiment of the present invention, image preprocessing includes:

[0033] The bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of single cells were cropped and normalized to remove large background blocks unrelated to cells.

[0034] Edge detection and contour extraction were performed on bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of individual cells, respectively. The contour extraction steps included median filtering for noise reduction, Canny edge extraction, dilation operation, closure operation, hole filling, and contour delineation.

[0035] According to a preferred embodiment of the present invention, extracting features from three modalities refers to: extracting morphological features; including:

[0036] For the preprocessed bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, seven morphological features are extracted from each image, including: area, perimeter, maximum axial length, minimum axial length, aspect ratio, rectangularity, and roundness. Thus, a total of 21 morphological features are obtained from the three-modal image of each cell, which are used for the quantitative display of cell morphological information.

[0037] By calculating the area ratio, the ratio of the cell nucleus area to the cytoplasm area and the ratio of the mitochondrial area to the cytoplasm area can be obtained.

[0038] Feature analysis refers to the comparative analysis of the obtained morphological feature values ​​to obtain the histogram distribution of the ratio of cell nucleus to cytoplasm and the ratio of mitochondria to cytoplasm, as well as the nuclear density estimation curve.

[0039] According to a preferred embodiment of the present invention, automatic classification includes: using multiple machine learning classifiers to classify cells based on quantified morphological features, namely 21 morphological features obtained, and the ratio of nucleus to cytoplasm area and the ratio of mitochondrial to cytoplasm area.

[0040] Further preferred options include a variety of machine learning classifiers, including k-nearest neighbor (kNN), support vector machine (SVM), random forest (RF), and logistic regression (LR).

[0041] According to a preferred embodiment of the present invention, extracting features from three modalities further includes:

[0042] Texture features were extracted using the gray-level co-occurrence matrix (CLGM) operator: For the preprocessed bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, six CLGM features in four directions were extracted for each image: contrast, dissimilarity, homogeneity, energy, autocorrelation, and second moment of angle. In this way, 24 CLGM features were obtained for each image, and a total of 72 multimodal CLGM features were obtained from the three types of images: bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell.

[0043] Meanwhile, texture features were extracted using the Local Binary Pattern (LBP) operator: for the preprocessed bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, LBP images with a neighborhood radius of 1 and a number of neighborhood points of 8 were obtained, and the histogram of gray values ​​of the LBP images was calculated. 59 LBP features were obtained for each image, and a total of 177 multimodal LBP features were obtained from the three types of images: bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell.

[0044] According to a preferred embodiment of the present invention, automatic classification further includes: combining the extracted texture features, including 72 multimodal CLGM features and 177 multimodal LBP features, with the extracted morphological features to achieve automatic cell classification;

[0045] Specifically, this involves concatenating 72 multimodal CLGM features and 177 multimodal LBP features with extracted morphological features to form a new feature array, which is then input into a machine learning classifier for training. The training is performed using the open-source Python Scikit-learn image processing package. The machine learning classifier outputs the image category results, which are compared with the true categories to statistically determine the classification accuracy, sensitivity, specificity, and precision.

[0046] A mitochondrial-based single-cell feature extraction and analysis system, comprising:

[0047] The static cytometry unit is configured to acquire images of three modalities of a single cell, including bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of a single cell.

[0048] The image preprocessing unit is configured to perform image preprocessing on the acquired images of the three modalities;

[0049] The image feature extraction unit is configured to extract features from three modalities.

[0050] The image feature analysis unit is configured to: perform feature analysis on the features of the three extracted modalities, obtain the ratio of the area of ​​the nucleus to the cytoplasm and the ratio of the area of ​​the mitochondria to the cytoplasm, and obtain the histogram distribution of the ratio of the nucleus to the cytoplasm and the ratio of the mitochondria to the cytoplasm and their nuclear density estimation curves.

[0051] The classification unit is configured to divide the collected sample results into a training set and a test set, and to automatically classify the samples by finding the optimal parameters and optimizing the classification function based on the input feature values.

[0052] According to a preferred embodiment of the present invention, the image feature extraction unit includes an image morphological feature extraction unit and an image texture feature extraction unit;

[0053] The image morphological feature extraction unit is configured to extract morphological features.

[0054] The image texture feature extraction unit is configured to extract texture features using the CLGM operator and the LBP operator.

[0055] According to a preferred embodiment of the present invention, the static cytology unit includes a light excitation system, a sample system, and a three-modal asynchronous imaging system for acquiring bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of a single cell.

[0056] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a mitochondrial-based single-cell feature extraction and analysis method.

[0057] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a mitochondrial-based single-cell feature extraction and analysis method.

[0058] The beneficial effects of this invention are as follows:

[0059] 1. This invention combines multiple features from bright-field images of single cells and mitochondrial fluorescence images, providing mitochondria as a powerful internal biomarker for single-cell classification;

[0060] 2. This invention extracts the clinical morphological and textural features of images of three single-cell modalities, which can be used as a clinical reference in the analysis, thus more effectively ensuring the reliability of experimental conclusions;

[0061] 3. This invention uses histogram distribution and its nuclear density estimation curve to intuitively compare the clinical morphological characteristics of cell nuclei and mitochondria in a single modality;

[0062] 4. This invention can achieve rapid and automatic classification of human cervical cancer cells and normal cells, eliminating the tedious process of manual slide reading;

[0063] 5. This invention is applicable to combination with various cell analysis methods, such as cell nuclear imaging analysis, two-dimensional light scattering technology of cells, etc., to achieve more effective and comprehensive analysis and automatic classification of cells;

[0064] 6. The exploration of the potential of mitochondrial cell markers in this invention is applicable to image analysis and classification of other biological cells and has broad applicability. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the single-cell feature extraction and analysis method based on mitochondria of the present invention.

[0066] Figure 2 Schematic diagrams of bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of H8 cells and HeLa cells provided for examples of the present invention;

[0067] Figure 3 Data preprocessing flowcharts for three modalities provided in this invention example;

[0068] Figure 4 Histograms and nuclear density estimation curves of two clinical morphological features of two cell lines provided for examples of the present invention;

[0069] Figure 5 A schematic diagram illustrating the machine learning classification results of unimodal morphological and textural features provided as an example of the present invention. Detailed Implementation

[0070] The present invention will be further described below with reference to the accompanying drawings and examples.

[0071] Example 1

[0072] A method for single-cell feature extraction and analysis based on mitochondria, comprising:

[0073] Acquire images of three modalities of a single cell; including: bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell;

[0074] Image preprocessing is performed on the acquired images of the three modalities;

[0075] For different data characteristics, features of three modalities are extracted and feature analysis is performed to obtain the ratio of nucleus to cytoplasm area and the ratio of mitochondrial to cytoplasm area, and the histogram distribution of the ratio of nucleus to cytoplasm and the ratio of mitochondrial to cytoplasm and their nuclear density estimation curves are obtained.

[0076] Automatic classification is achieved through machine learning.

[0077] Example 2

[0078] According to the single-cell feature extraction and analysis method based on mitochondria described in Example 1, such as... Figure 1 As shown, the difference lies in:

[0079] Image preprocessing includes cropping and normalizing the bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, respectively, and removing large background blocks unrelated to the cell.

[0080] Edge detection and contour extraction were performed on bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of individual cells, respectively. The contour extraction steps included median filtering for noise reduction, Canny edge extraction, dilation operation, closure operation, hole filling, and contour delineation.

[0081] Extracting features from three modalities refers to extracting morphological features, including:

[0082] For preprocessed bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of single cells, seven morphological features were extracted from each image, including: area, perimeter, maximum axial length, minimum axial length, aspect ratio, rectangularity, and roundness. In image preprocessing, based on the contours obtained after image edge detection, the number of pixels within the closed contours was counted, and the dimensions were then calculated using the image scale. The seven required morphological features were calculated and statistically analyzed according to the feature descriptions in Table 1.

[0083] Table 1

[0084]

[0085] In this way, a total of 21 morphological features are obtained from the trimodal image of each cell, which are used for the quantitative display of cell morphological information.

[0086] By calculating the area ratio, the ratio of the cell nucleus area to the cytoplasm area and the ratio of the mitochondrial area to the cytoplasm area can be obtained;

[0087] Feature analysis refers to the comparative analysis of obtained morphological feature values ​​to obtain histogram distributions of the ratio of cell nucleus to cytoplasm and the ratio of mitochondria to cytoplasm, as well as their nuclear density estimation curves. A histogram distribution is a statistical reporting graph used to display the distribution of quantitative data. The horizontal axis represents the data value of a continuous variable that needs to be statistically analyzed, and the vertical axis represents the cumulative frequency. The nuclear density estimation curve is similar to a histogram; its vertical axis represents the probability of the data value occurring, visually displaying the continuous probability density of the data. The higher the "peak" of the curve, the denser the data at that point.

[0088] Automatic classification includes: using multiple machine learning classifiers to classify cells based on quantified morphological features, namely 21 morphological features, the ratio of nucleus to cytoplasm area, and the ratio of mitochondrial to cytoplasm area. Each cell's trimodal image can simultaneously obtain 21 basic morphological features and 2 area ratio features for quantitative display of cell morphological information, and can automatically classify cells using these quantified morphological features.

[0089] Various machine learning classifiers include k-nearest neighbor (kNN), support vector machine (SVM), random forest (RF), and logistic regression (LR).

[0090] This embodiment uses single-cell feature extraction and analysis methods to analyze the clinical morphological features of three modalities of images of single human cervical epithelial cells (H8) and cervical cancer cells (HeLa). In this embodiment, the obtained images of the three modalities are preprocessed, and statistical calculations are performed on the preprocessed images. Seven clinical morphological features can be obtained from each image. Additionally, the mean and variance of the morphological features—the ratio of nucleus to cytoplasm area and the ratio of mitochondrial to cytoplasm area—are statistically obtained by calculating the proportion of pixels in the contour region. The specific operation steps are as follows:

[0091] (1) Culture of H8 and HeLa cells: The cultured cervical cells were centrifuged and resuspended in Dulbecco modified Eagle medium. The suspended cells were stained with MitoTracker Red CMXRos at 37°C in the dark for 30 minutes at a working concentration of 100 nM. Subsequently, the cells were washed with phosphate-buffered saline and fixed with 4% paraformaldehyde at 37°C for 10 minutes. Then, the cells were washed with PBS and fixed with DAPI at room temperature for 5 minutes. After washing the cells, a cell suspension was prepared.

[0092] (2) Obtain bright-field images, nuclei, and mitochondrial fluorescence images of the processed H8 and HeLa cells, such as... Figure 2 As shown. Figure 2 (a) is a schematic diagram of bright-field image, nuclear fluorescence image and mitochondrial fluorescence image of H8 cells; Figure 2 (b) shows a schematic diagram of bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of HeLa cells; a total of 2005 sets of trimodal data for the two types of cervical cells were obtained, including 990 sets of H8 cells (990 images per modality) and 1015 sets of HeLa cells (1015 images per modality).

[0093] (3) Figure 3 As shown, the obtained trimodal images are subjected to median filtering for noise reduction, Canny operator edge extraction, expansion, closure, and hole filling to obtain the cell contour edges;

[0094] (4) Statistical calculations were performed on the preprocessed trimodal images, and seven clinical morphological features were obtained for each image: area, perimeter, maximum axial length, minimum axial length, aspect ratio, rectangularity, and roundness.

[0095] (5) Calculate the ratio of nucleus to cytoplasm area and the ratio of mitochondrial to cytoplasm area under all H8 and HeLa cell trimodal images, and further explore the histogram distribution and nuclear density estimation curve to demonstrate the distribution characteristics of the ratio of nucleus to cytoplasm and the ratio of mitochondrial to cytoplasm, such as... Figure 4 As shown. In Figure 4 In (a), the ratio of cell nucleus to cytoplasm was not significantly different between the two cell types, and the density peaks basically overlapped at around 0.47. Figure 4 In (b), the mitochondrial-to-cytoplasmic density ratio peak for H8 cells was approximately 0.76, while for HeLa cells it was approximately 0.78. Statistical results of the mitochondrial-to-cytoplasmic ratio indicate that, compared to the traditional clinical method of classifying malignant tumor cells based on the cell nucleus, using mitochondria to classify malignant tumor cells has greater potential clinical application value.

[0096] (6) Seven morphological features from bright-field and nuclear fluorescence images of H8 and HeLa cells, along with 15 features including the ratio of nucleus to cytoplasm area, were input into four machine learning methods. A machine learning classifier was constructed using 5-fold cross-validation. 2005 data points from the two cell types were randomly divided into five datasets, each with 401 samples. Four datasets were used as the training set, and the remaining dataset as the test set. After five rounds of cross-validation, the average of the five prediction results was used as the final accuracy. Accuracy was defined as the percentage of correctly classified cells out of all cells; sensitivity was defined as the ratio of correctly classified HeLa cells to the total number of HeLa cells; specificity was defined as the ratio of correctly classified H8 cells to the total number of H8 cells; and precision was defined as the ratio of correctly predicted HeLa cells to the total number of predicted HeLa cells. The AUC parameter was used to evaluate the classifier's performance; a value closer to 1 indicated better classifier performance. Table 2 shows the four classification results of human cervical epithelial cells and cervical cancer cells based on cell morphology features.

[0097] Table 2

[0098] Classification methods accuracy Sensitivity Specificity Accuracy KNN 68.48％±2.90％ 72.81％±9.76％ 64.04％±13.32％ 68.37％±5.58％ SVM 69.13％±3.84％ 64.93％±14.21％ 73.43％±14.13％ 72.83％±6.21％ RF 69.63％±2.88％ 74.38％±12.33％ 64.75％±16.95％ 69.85％±6.99％ LR 71.07％±2.54％ 74.78％±12.94％ 67.27％±15.06％ 71.26％±5.90％

[0099] The classification accuracy of the four machine learning algorithms is around 70%, among which the LR method has the best classification accuracy of 71.07%, sensitivity of 74.78%, and specificity of 67.27%.

[0100] Example 3

[0101] The difference between the single-cell feature extraction and analysis method based on mitochondria described in Example 2 and the method described in Example 2 is as follows:

[0102] Extracting features from three modalities also includes:

[0103] Texture features were extracted using the CLGM operator: For preprocessed bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of single cells, six CLGM features in four directions were extracted for each image: contrast, dissimilarity, homogeneity, energy, autocorrelation, and angular second moment. Table 3 describes the gray-level co-occurrence matrix features.

[0104] Table 3

[0105]

[0106]

[0107] In this way, each image yields 24 CLGM features, and the three types of images—bright-field image of a single cell, nuclear fluorescence image, and mitochondrial fluorescence image—acquire a total of 72 multimodal CLGM features.

[0108] Simultaneously, texture features were extracted using the LBP operator: For preprocessed bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of single cells, the open-source Python package Scikit-Image was used to obtain LBP images with a neighborhood radius of 1 and 8 neighboring points. Histograms of the grayscale values ​​of the LBP images were then calculated, yielding 59 LBP features for each image. A total of 177 multimodal LBP features were obtained from the three types of images (bright-field, nuclear, and mitochondrial fluorescence images of single cells). This approach aims to obtain more information reflecting the texture differences in single-cell images.

[0109] Automatic classification also includes combining the extracted texture features, including 72 multimodal CLGM features and 177 multimodal LBP features, with the extracted morphological features to achieve automatic cell classification;

[0110] Specifically, this involves concatenating 72 multimodal CLGM features and 177 multimodal LBP features with extracted morphological features to form a new feature array, which is then input into a machine learning classifier for training. The training is performed using the open-source Python Scikit-learn image processing package. The machine learning classifier outputs the image category results, which are compared with the true categories to statistically determine the classification accuracy, sensitivity, specificity, and precision.

[0111] Four machine learning classifiers—k-nearest neighbor (kNN), support vector machine (SVM), random forest (RF), and logistic regression (LR)—are used to automatically classify the experimental results. The preprocessed experimental results are grouped for training and prediction, and classification results and evaluation values ​​are provided.

[0112] This embodiment uses single-cell feature extraction and analysis methods to identify and automatically classify H8 and HeLa cells. The obtained trimodal images are preprocessed, and statistical calculations are performed on the preprocessed images, yielding seven clinical morphological features per image. Texture features are extracted from bright-field and two-color fluorescence images using the gray-level co-occurrence matrix and local binary mode operator, resulting in 24 features using the CLGM operator and 59 features using the LBP operator. The clinical morphological and texture features are then input into four machine learning algorithms as feature parameters for automatic classification of the two cell types. The specific operation steps are as follows:

[0113] (1) Seven morphological features were selected from Example 1: area, perimeter, maximum axial length, minimum axial length, aspect ratio, rectangularity, and roundness;

[0114] (2) The CLGM and LBP operators were selected to extract texture features from bright-field images and two types of fluorescence images to obtain more information to reflect the texture differences in single-cell images. In the three modalities, each image yielded 24 features using the CLGM operator and 59 features using the LBP operator.

[0115] (3) The clinical morphological and texture features of bright-field images, nuclear fluorescence images and mitochondrial fluorescence images were used as classification features for four types of machine learning. The classification results of human cervical epithelial cells and cervical cancer cells based on the four machine learning methods are shown in Table 4.

[0116] Table 4

[0117]

[0118]

[0119] Figure 5 In Figure (a), the ROC curves for LR classification results based on the morphological and texture features of H8 and HeLa cells are shown. The classification results for mitochondrial fluorescence images, represented by the solid line, are the best, with an accuracy of 73.92% and an AUC of 0.84. In this embodiment, the t-distributed random neighborhood embedding dimensionality reduction method is used to visualize the input features of different modalities. Figure 5Images (b), (c), and (d) show the t-distribution random neighborhood results for three single-modality features: bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image, respectively. All three single-modality features exhibit some clustering, but the clustering is not significant, and the results are not significantly different. Furthermore, the mitochondrial fluorescence image shows the best classification performance among the three modalities, further demonstrating the potential of mitochondria for cancer cell classification.

[0120] Example 4

[0121] A mitochondrial-based single-cell feature extraction and analysis system, comprising:

[0122] The static cytology unit is configured to: acquire images of three modalities of a single cell; the image preprocessing unit is configured to: perform image preprocessing on the acquired images of the three modalities; the image feature extraction unit is configured to: extract features from the three modalities; the image feature analysis unit is configured to: perform feature analysis on the extracted features of the three modalities; and the classification unit is configured to: divide the collected sample results into a training set and a test set, and, based on the input feature values, optimize the classification function by finding the optimal parameters to achieve automatic classification of the samples.

[0123] The image feature extraction unit includes an image morphological feature extraction unit and an image texture feature extraction unit;

[0124] The image morphological feature extraction unit is configured to extract morphological features; the image texture feature extraction unit is configured to extract texture features using the CLGM operator and the LBP operator.

[0125] The static cytology unit includes a light-excitation system, a sample system, and a three-modal asynchronous imaging system for acquiring bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of individual cells.

[0126] Example 5

[0127] A computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any of the mitochondrial-based single-cell feature extraction and analysis methods in Examples 1-3.

[0128] Example 6

[0129] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the mitochondrial-based single-cell feature extraction and analysis methods of Examples 1-3.

[0130] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for single-cell feature extraction and analysis based on mitochondria, characterized in that, include: Acquire images of three modalities of a single cell, including: bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell; perform image preprocessing on the acquired images of the three modalities; For different data characteristics, features of three modalities are extracted and feature analysis is performed to obtain the ratio of nucleus to cytoplasm area and the ratio of mitochondrial to cytoplasm area. Histogram distributions of the ratio of nucleus to cytoplasm and the ratio of mitochondrial to cytoplasm and their nuclear density estimation curves are obtained. Automatic classification is achieved through machine learning. Extracting features from three modalities refers to extracting morphological features, including: For the preprocessed bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, seven morphological features were extracted from each image, including: area, perimeter, maximum axial length, minimum axial length, aspect ratio, rectangularity, and roundness; thus, a total of 21 morphological features were obtained from the three-modal image of each cell. By calculating the area ratio, the ratio of the cell nucleus area to the cytoplasm area and the ratio of the mitochondrial area to the cytoplasm area can be obtained; Feature analysis refers to the comparative analysis of the obtained morphological feature values ​​to obtain the histogram distribution of the ratio of cell nucleus to cytoplasm and the ratio of mitochondria to cytoplasm, as well as the nuclear density estimation curve.

2. The method for single-cell feature extraction and analysis based on mitochondria according to claim 1, characterized in that, Image preprocessing includes cropping and normalizing the bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, respectively, and removing large background blocks unrelated to the cell. Edge detection and contour extraction were performed on bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of individual cells, respectively. The contour extraction steps included median filtering for noise reduction, Canny edge extraction, dilation operation, closure operation, hole filling, and contour delineation.

3. The method for single-cell feature extraction and analysis based on mitochondria according to claim 1, characterized in that, Automatic classification includes: using multiple machine learning classifiers to classify cells based on quantified morphological features, namely 21 morphological features, the ratio of nucleus to cytoplasm area, and the ratio of mitochondrial to cytoplasm area; Various machine learning classifiers include k-nearest neighbor (kNN), support vector machine (SVM), random forest (RF), and logistic regression (LR).

4. The method for single-cell feature extraction and analysis based on mitochondria according to claim 1, characterized in that, Extracting features from three modalities refers to extracting texture features, including: Texture features were extracted using the CLGM operator: For the preprocessed bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, six CLGM features in four directions were extracted for each image: contrast, dissimilarity, homogeneity, energy, autocorrelation, and second moment of angle. In this way, 24 CLGM features were obtained for each image, and a total of 72 multimodal CLGM features were obtained from the three types of images: bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell. Meanwhile, texture features were extracted using the LBP operator: for the preprocessed bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, LBP images with a neighborhood radius of 1 and a number of neighborhood points of 8 were obtained, and the histogram of gray values ​​of the LBP images was calculated. Each image yielded 59 LBP features, and a total of 177 multimodal LBP features were obtained from the three types of images: bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell.

5. The method for single-cell feature extraction and analysis based on mitochondria according to claim 4, characterized in that, Automatic classification also includes combining the extracted texture features, including 72 multimodal CLGM features and 177 multimodal LBP features, with the extracted morphological features to achieve automatic cell classification; Specifically, this involves concatenating 72 multimodal CLGM features and 177 multimodal LBP features with extracted morphological features to form a new feature array, which is then input into a machine learning classifier for training. The training is performed using the open-source Python Scikit-learn image processing package. The machine learning classifier outputs the image category results, which are compared with the true categories to statistically determine the classification accuracy, sensitivity, specificity, and precision.

6. A single-cell feature extraction and analysis system based on mitochondria, characterized in that, include: The static cytology unit is configured to acquire images of three modalities of a single cell. The image preprocessing unit is configured to perform image preprocessing on the acquired images of the three modalities; The image feature extraction unit is configured to extract features from three modalities. The image feature analysis unit is configured to perform feature analysis on the extracted features of the three modalities. The classification unit is configured to divide the collected sample results into a training set and a test set, and to automatically classify the samples by finding the optimal parameters and optimizing the classification function based on the input feature values. Extracting features from three modalities refers to extracting morphological features, including: For the preprocessed bright-field image, nuclear fluorescence image, and mitochondrial fluorescence image of a single cell, seven morphological features were extracted from each image, including: area, perimeter, maximum axial length, minimum axial length, aspect ratio, rectangularity, and roundness; thus, a total of 21 morphological features were obtained from the three-modal image of each cell. By calculating the area ratio, the ratio of the cell nucleus area to the cytoplasm area and the ratio of the mitochondrial area to the cytoplasm area can be obtained; Feature analysis refers to the comparative analysis of the obtained morphological feature values ​​to obtain the histogram distribution of the ratio of cell nucleus to cytoplasm and the ratio of mitochondria to cytoplasm, as well as the nuclear density estimation curve.

7. A single-cell feature extraction and analysis system based on mitochondria according to claim 6, characterized in that, The image feature extraction unit includes an image morphological feature extraction unit and an image texture feature extraction unit; The image morphological feature extraction unit is configured to extract morphological features. The image texture feature extraction unit is configured to extract texture features using the CLGM operator and the LBP operator; The static cytology unit includes a light excitation system, a sample system, and a three-modal asynchronous imaging system for acquiring bright-field images, nuclear fluorescence images, and mitochondrial fluorescence images of individual cells.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the mitochondrial-based single-cell feature extraction and analysis method according to any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the mitochondrial-based single-cell feature extraction and analysis method according to any one of claims 1-5.

Images