Amnion epithelial stem cell screening method and device based on microscopic image feature extraction

Abstract

The application discloses an amnion epithelial stem cell screening method and device based on microscopic image feature extraction, relates to the technical field of microscopic image analysis and cell screening, and is used for solving the problem that the population topology and internal activity are difficult to be quantified in the lossless screening of amnion epithelial stem cells. Firstly, phase contrast microscopic images are subjected to enhancement processing, and nucleolus centers and global gradient energy maps are extracted; subsequently, Voronoi districts are constructed, and cytoplasm texture entropy and gray attenuation gradient are coupled to generate a cytoplasm compactness factor; meanwhile, a cell gap topology consistency index is calculated based on the mapping of the Voronoi boundary and the gradient map; finally, the compactness factor is weighted and corrected by using the topology index to obtain stemness maintenance confidence, and target coordinates are output. Through the dual verification of geometric topology and physical gradient, the application realizes the joint constraint screening of the features inside and outside the cells.

Description

Technical Field

[0001] This invention relates to the field of microscopic image analysis and cell screening technology, specifically to a method and apparatus for screening amniotic epithelial stem cells based on microscopic image feature extraction. Background Technology

[0002] Human amniotic epithelial stem cells (AECs) are derived from placental amniotic tissue. They possess low immunogenicity, are free from ethical controversy, and have the potential to differentiate into cells from all three germ layers. They demonstrate significant clinical application value in regenerative medicine fields such as neurodegenerative diseases, liver disease repair, and the treatment of autoimmune diseases. To meet the high standards required for clinical treatment and stem cell bank storage, obtaining highly pure, highly active AECs that have not undergone epithelial-mesenchymal transition is a crucial prerequisite for ensuring subsequent therapeutic efficacy. Therefore, establishing an efficient screening system is of great significance for promoting their large-scale application.

[0003] Current methods for screening and identifying amniotic epithelial stem cells primarily rely on flow cytometry to detect surface antigen markers or immunofluorescence staining. These methods typically require the use of exogenous antibody reagents to label cells, and the invasive procedures can easily alter the cells' original physiological state or even cause cell damage, making it difficult to directly use the screened cells for subsequent clinical reinfusion therapy. On the other hand, while manual morphological observation based on traditional optical microscopy achieves non-destructive testing, this process heavily depends on the operator's subjective experience, and conventional morphological parameters such as area and roundness are insufficient to quantify the non-linear relationship between the fine internal texture features of cells and their stemness state. In particular, the unique cobblestone-like population growth pattern of amniotic epithelial stem cells makes it easy to overlook the intercellular connectivity when analyzing individual cells, resulting in low accuracy in identifying cells that have undergone early epithelial-mesenchymal transition or have failed contact inhibition, making it difficult to meet the quality control requirements for high-quality cell preparation. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and apparatus for screening amniotic epithelial stem cells based on microscopic image feature extraction, thus solving the problems mentioned above in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a screening method for amniotic epithelial stem cells based on microscopic image feature extraction, comprising the following steps: S1. Acquiring original grayscale images of amniotic epithelial stem cells using an inverted phase-contrast microscope; performing contrast-limited adaptive histogram equalization and bilateral filtering on the original grayscale images to generate texture-enhanced images; extracting cell nuclear contours and calculating nuclear centroid coordinates by maximum inter-class variance thresholding on the texture-enhanced images; and calculating the global gradient energy map of the texture-enhanced images using the Sobel operator; S2. Generating Voronoi zeta polygon meshes using the nuclear centroid coordinates as seed points to define single-cell jurisdictions; calculating the local binary pattern texture entropy of the cytoplasm after removing the nuclear region within the single-cell jurisdiction; and calculating the entropy from the nuclear centroid coordinates along the polar radius direction to the Voronoi zeta polygon. S3. The gray-level decay gradient of the polygon boundary is used to calculate the cytoplasmic compactness factor by weighting the local binary mode texture entropy and the gray-level decay gradient; S4. The coordinates of the common boundary line of the Voronoi Thésen polygon mesh are extracted and mapped to the global gradient energy map. The ratio of the cumulative pixel gradient value along the common boundary line path to the path length is calculated to obtain the edge intensity response value. The ratio of the edge intensity response value to the image background noise gradient is defined as the intercellular gap topological consistency index; S5. The intercellular gap topological consistency index is compared with a preset topological integrity threshold. If it is higher than the threshold, the cytoplasmic compactness factor is weighted and corrected using the intercellular gap topological consistency index to obtain the stemness maintenance confidence. When the stemness maintenance confidence is in the preset high stemness range, the corresponding region is determined to be the target cell and the screening coordinate position is output.

[0006] Furthermore, the specific process of generating a texture-enhanced image by acquiring the original grayscale image of amniotic epithelial stem cells using an inverted phase contrast microscope and performing contrast-limited adaptive histogram equalization and bilateral filtering on the original grayscale image is as follows: The original grayscale image of amniotic epithelial stem cells is divided into non-overlapping rectangular context blocks. The local grayscale histogram of each rectangular context block is calculated, and the number of pixels in the local grayscale histogram that exceed the preset clipping threshold is evenly distributed to other grayscale levels in the local grayscale histogram. The adjacent rectangular context blocks after distribution are reconstructed and stitched together using a bilinear interpolation algorithm. A bilateral filter kernel containing a spatial distance weight factor and a pixel grayscale difference weight factor is constructed, and a sliding convolution operation is performed on the reconstructed and stitched image using the bilateral filter kernel to output the texture-enhanced image.

[0007] Further, the specific process of extracting cell kernel contours and calculating kernel centroid coordinates by performing maximum inter-class variance thresholding on the texture-enhanced image, and calculating the global gradient energy map of the texture-enhanced image using the Sobel operator is as follows: Traverse all gray levels of the texture-enhanced image, calculate the inter-class variance between foreground and background pixels, select the gray level corresponding to the maximum inter-class variance as the optimal segmentation threshold, and generate a binary mask using the optimal segmentation threshold; perform morphological closing operation on the binary mask, performing dilation followed by erosion, and extract the outer boundaries of the connected components after the operation as the cell kernel contours; calculate the zero-order spatial moment and first-order spatial moment of each connected component, and calculate the kernel centroid coordinates using the ratio of the first-order spatial moment to the zero-order spatial moment; convolve the texture-enhanced image using horizontal and vertical Sobel convolution kernels respectively, calculate the square root of the sum of the squares of the horizontal and vertical gradient components, and generate the global gradient energy map.

[0008] Furthermore, a Voronoi Thésen polygon mesh is generated using the nuclear centroid coordinates as seed points to define the single-cell jurisdiction. The specific process for calculating the cytoplasmic local binary pattern texture entropy after removing the nucleus region within the single-cell jurisdiction is as follows: The Euclidean distance from each pixel in the texture-enhanced image to all nuclear centroid coordinates is calculated. Pixels are then classified into the category represented by the nuclear centroid coordinate with the smallest Euclidean distance, generating a Voronoi Thésen polygon mesh enclosed by polygonal boundaries. The texture-enhanced image is masked using the Voronoi Thésen polygon mesh, setting the grayscale value of pixels within the corresponding nucleus contour to zero within each single-cell jurisdiction, retaining the cytoplasmic region outside the nucleus as the texture analysis object. Pixels within the cytoplasmic region are traversed, and the grayscale values ​​of the central pixel and neighboring pixels are compared to generate binary codes. These binary codes are converted to decimal numbers, and their histogram distribution is statistically analyzed. The information entropy of the histogram distribution is calculated using the Shannon entropy formula, yielding the cytoplasmic local binary pattern texture entropy.

[0009] Further, the specific process of calculating the gray-level attenuation gradient from the nuclear centroid coordinates along the polar radius direction to the boundary of the Voronoi polygon, and weighting the local binary pattern texture entropy and the gray-level attenuation gradient to obtain the cytoplasmic compactness factor is as follows: A local polar coordinate system is established with the nuclear centroid coordinates as the pole. Rays are emitted in all directions at set angular intervals until they intersect with the boundary of the Voronoi Thiessen polygon mesh, obtaining the pixel gray-level sequence along the ray path. The linear regression slope of the pixel gray-level sequence relative to the ray length is fitted using the least squares method. The mean of the absolute values ​​of the linear regression slopes of all angular rays is calculated to obtain the gray-level attenuation gradient. The gray-level attenuation gradient is subjected to max-min normalization. The normalized gray-level attenuation gradient and the cytoplasmic local binary pattern texture entropy are multiplied by their respective preset weight coefficients and then summed to output the cytoplasmic compactness factor.

[0010] Further, the process of extracting the coordinates of the common boundary line of the Voronoi Thésen polygon mesh and mapping them to the global gradient energy map, and calculating the ratio of the cumulative pixel gradient value along the common boundary line path to the path length to obtain the edge intensity response value is as follows: Traverse the adjacency matrix of the Voronoi Thésen polygon mesh, identify the perpendicular bisector connecting two adjacent centroid coordinates, and extract the coordinate set of all pixels on the perpendicular bisector as the common boundary line coordinates; use the common boundary line coordinates for indexing in the global gradient energy map to extract the corresponding gradient magnitude sequence on the common boundary line coordinate path; sum the gradient magnitude sequences to obtain the total gradient energy of the path, count the number of pixels in the common boundary line coordinate set as the path length, and calculate the ratio of the total gradient energy of the path to the path length to obtain the edge intensity response value.

[0011] Furthermore, the specific process of defining the ratio of edge intensity response value to image background noise gradient as the intercellular gap topological consistency index is as follows: The gradient magnitude of all pixels in the global gradient energy map is statistically analyzed, and the average gradient level of the entire image is calculated using the arithmetic mean method. This average gradient level is defined as the image background noise gradient. A division operation is performed to calculate the ratio of edge intensity response value to image background noise gradient, yielding the original signal-to-noise ratio. The original signal-to-noise ratio is then logarithmically transformed using the natural logarithm function, and the processed value is output as the intercellular gap topological consistency index.

[0012] Furthermore, the intercellular topological consistency index is compared with a preset topological integrity threshold. If it is higher than the threshold, the intercellular topological consistency index is used to weight and correct the cytoplasmic compactness factor to obtain the stemness maintenance confidence. The specific process is as follows: a segmented judgment logic is constructed. When the intercellular topological consistency index is lower than the preset topological integrity threshold, the stemness maintenance confidence is set to zero. When the intercellular topological consistency index is higher than the preset topological integrity threshold, a normalization mapping is performed on the intercellular topological consistency index to convert it into a gain coefficient between 0 and 1. An exponential enhancement function is constructed using the gain coefficient as the exponent. The cytoplasmic compactness factor is multiplied using the exponential enhancement function, and the result is output as the stemness maintenance confidence.

[0013] Furthermore, when the stemness maintenance confidence level is within the preset high stemness range, the specific process of determining the corresponding region as the target cell and outputting the screening coordinate position is as follows: Set a high stemness range consisting of a lower threshold and an upper threshold, and determine whether the stemness maintenance confidence level falls within this range; if the determination result is yes, retrieve the nuclear centroid coordinates corresponding to the region, and map the nuclear centroid coordinates from the image pixel coordinate system back to the physical space coordinate system of the microscope stage; encapsulate the coordinate data in the physical space coordinate system into a screening instruction, and generate a screening coordinate position list containing the target cell number and physical location information.

[0014] The amniotic epithelial stem cell screening device based on microscopic image feature extraction includes the following modules: an image acquisition and processing module, used to acquire raw grayscale images of amniotic epithelial stem cells using an inverted phase-contrast microscope, perform contrast-limited adaptive histogram equalization and bilateral filtering on the raw grayscale images to generate texture-enhanced images, extract cell nuclear contours and calculate nuclear centroid coordinates by maximum inter-class variance thresholding of the texture-enhanced images, and calculate the global gradient energy map of the texture-enhanced images using the Sobel operator; and a cytoplasmic feature analysis module, used to generate a Voronoi Thésen polygon mesh with the nuclear centroid coordinates as seed points to define single-cell regions, calculate the local binary pattern texture entropy of the cytoplasm after removing the cell nuclear region within the single-cell region, and calculate the grayscale attenuation gradient from the nuclear centroid coordinates along the polar radius direction to the Voronoi polygon boundary. The system employs several modules: a cytoplasmic compactness factor is obtained by weighting the local binary pattern texture entropy and grayscale decay gradient; an edge topology analysis module extracts the coordinates of the common boundary line of the Voronoi Thésen polygon mesh and maps them to the global gradient energy map; it calculates the ratio of the cumulative pixel gradient value along the common boundary line path to the path length to obtain the edge intensity response value; and the ratio of the edge intensity response value to the image background noise gradient is defined as the intercellular gap topology consistency index; a comprehensive screening decision module compares the intercellular gap topology consistency index with a preset topology integrity threshold. If the index is higher than the threshold, the cytoplasmic compactness factor is weighted and corrected using the intercellular gap topology consistency index to obtain the stemness maintenance confidence. When the stemness maintenance confidence is within a preset high stemness range, the corresponding region is determined to be the target cell, and the screening coordinates are output.

[0015] The present invention has the following beneficial effects:

[0016] (1) The amniotic epithelial stem cell screening method based on microscopic image feature extraction effectively suppresses phase halo noise, a common feature in phase-contrast microscopy, while enhancing the texture details of microscopic images through contrast-adaptive histogram equalization and bilateral filtering, providing a high-quality data foundation for subsequent feature extraction. Furthermore, this invention utilizes nuclear centroid coordinates to generate Voronoi Thésen polygon meshes to define single-cell regions, creatively solving the technical challenge of accurately segmenting cell boundaries of epithelial cells growing in sheets using traditional edge detection algorithms. Based on this, by calculating the local binary pattern texture entropy of the cytoplasm and the gray-level decay gradient along the polar radius, and weighting and coupling these two factors into a cytoplasmic compactness factor, a quantitative characterization of the internal microstructure of single cells is achieved. This multi-dimensional feature extraction method can keenly capture the subtle features of the cytoplasm transitioning from dense and uniform to loose and heterogeneous, thereby accurately assessing the stemness maintenance status of individual cells without the use of chemical reagents.

[0017] (2) An amniotic epithelial stem cell screening device based on microscopic image feature extraction was developed. A cell gap topological consistency index was constructed. By mapping the theoretical common boundary of the Voronoi Thiessen polygon to the global gradient energy map, the ratio of the actual pixel gradient energy to the background noise on the theoretical path was calculated, thereby realizing the physical verification of the cell population topological structure. This step can effectively distinguish between tightly packed high-quality epithelial cells and deteriorated cells with enlarged gaps and loose connections. Finally, by using the cell gap topological consistency index as a weighting coefficient to correct the cytoplasmic compactness factor, a hierarchical screening logic based on topological constraints was constructed. This ensures that only regions with both good population connectivity structure and high-quality internal texture features are identified as target cells, significantly reducing the false detection rate caused by cell aging or early differentiation, and improving the reliability and clinical usability of the screening results.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] Figure 1 This is a flowchart of the amniotic epithelial stem cell screening method based on microscopic image feature extraction according to the present invention.

[0020] Figure 2 This is a flowchart of the amniotic epithelial stem cell screening device based on microscopic image feature extraction according to the present invention. Detailed Implementation

[0021] This application provides a method and apparatus for screening amniotic epithelial stem cells based on microscopic image feature extraction. This method and apparatus address the technical problems of existing screening technologies that rely on invasive labeling, leading to impaired cell viability, and the difficulty of accurately identifying epithelial cell population topological anomalies and early differentiation states through traditional morphological analysis.

[0022] The overall concept of the solution in this application embodiment is as follows:

[0023] First, images were acquired using an inverted phase-contrast microscope and subjected to noise reduction and enhancement processing. The cell nucleus centroid was extracted, and a global gradient energy map was calculated. Next, Voronoi-Thieson polygons were constructed using the nuclear centroids to simulate the theoretical contact boundaries between cells. Within this framework, the cytoplasmic compactness factor, which reflects the complexity of the cell's internal structure, was extracted. Subsequently, the authenticity of the Voronoi theoretical boundary was verified using the global gradient energy map, and the intercellular topological consistency index, which reflects the tightness of intercellular connections, was calculated. Finally, the cytoplasmic compactness factor was weighted and corrected using the topological consistency index. Target cells with high stemness were screened out through the joint constraints of internal and external features, and their coordinates were output.

[0024] Please see Figure 1This invention provides a technical solution: a screening method for amniotic epithelial stem cells based on microscopic image feature extraction, comprising the following steps: S1. Acquiring original grayscale images of amniotic epithelial stem cells using an inverted phase-contrast microscope; performing contrast-limited adaptive histogram equalization and bilateral filtering on the original grayscale images to generate texture-enhanced images; extracting cell nuclear contours and calculating nuclear centroid coordinates by performing maximum inter-class variance threshold segmentation on the texture-enhanced images; and calculating the global gradient energy map of the texture-enhanced images using the Sobel operator; S2. Generating Voronoi Thésen polygon meshes using the nuclear centroid coordinates as seed points to define single-cell jurisdictions; calculating the local binary pattern texture entropy of the cytoplasm after removing the cell nuclear region within the single-cell jurisdiction; and calculating the entropy from the nuclear centroid coordinates along the polar radius direction to the edge of the Voronoi polygon. S3. The gray-level decay gradient of the boundary is used to calculate the cytoplasmic compactness factor by weighting the local binary pattern texture entropy and the gray-level decay gradient; S4. The coordinates of the common boundary line of the Voronoi Thiessen polygon mesh are extracted and mapped to the global gradient energy map. The ratio of the cumulative pixel gradient value along the path of the common boundary line to the path length is calculated to obtain the edge intensity response value. The ratio of the edge intensity response value to the image background noise gradient is defined as the intercellular gap topological consistency index; S5. The intercellular gap topological consistency index is compared with the preset topological integrity threshold. If it is higher than the threshold, the intercellular gap topological consistency index is used to weight and correct the cytoplasmic compactness factor to obtain the stemness maintenance confidence. When the stemness maintenance confidence is in the preset high stemness range, the corresponding region is determined to be the target cell and the screening coordinate position is output.

[0025] In this implementation scheme, the main function of step S1 is to complete high-quality image preprocessing and basic feature localization. This step first uses an inverted phase-contrast microscope to acquire images of living cells. Addressing the halo phenomenon unique to phase-contrast imaging, contrast-limited adaptive histogram equalization and bilateral filtering are employed. Contrast-limited adaptive histogram equalization is an image enhancement technique used to enhance local contrast while limiting noise amplification, while bilateral filtering is a nonlinear filter that smooths the image based on pixel spatial proximity and gray-level similarity. The combination of these two techniques aims to eliminate phase halo interference and preserve high-frequency details at cell edges. Subsequently, the optimal gray-level threshold is automatically established using maximum inter-class variance thresholding to separate the cell nucleus from the background and calculate its geometric center (nuclear centroid coordinates), providing a benchmark for subsequent localization. Simultaneously, the Sobel operator, a discrete differential operator, is used to calculate the approximate gradient of image gray levels, generating a global gradient energy map reflecting the intensity distribution of image edges, providing data support for subsequent morphological verification. The core of step S2 lies in establishing a single-cell morphological model and quantifying its internal texture features. Because amniotic epithelial stem cells grow in sheets, making boundary segmentation difficult using traditional methods, this step introduces the Voronoi lattice, a seed-point-based spatial partitioning algorithm that divides the plane into regions, minimizing the distance from each point within a region to its seed point. This is used to simulate and define the theoretical territory of a single cell. Within this territory, the uniformity and complexity of the cytoplasmic texture are measured by calculating the local binary pattern texture entropy, where the local binary pattern is an operator describing local texture features of an image. Simultaneously, the gray-level decay gradient from the nucleus centroid outwards along the radial direction is calculated to characterize the trend of cytoplasmic density changes. The cytoplasmic compactness factor, obtained by weighted coupling of these two parameters, effectively quantifies the density of the cell's internal structure, thereby determining whether the cell maintains the morphological characteristics unique to undifferentiated stem cells. Step S3 aims to verify the authenticity of the topological model constructed in the previous step using physical signals, thereby assessing the connectivity of the cell population. This step extracts the common boundary lines of the Voronoi lattice, i.e., the theoretical contact boundaries between adjacent cells, and maps them back to the global gradient energy map generated in S1. By calculating the ratio of the cumulative pixel gradient value to the path length on these boundary paths, the edge intensity response value is obtained. This value reflects whether there is a real cell membrane edge signal at the theoretical boundary location. This response value is then compared with the image background noise gradient. The defined intercellular gap topological consistency index is essentially a signal-to-noise ratio parameter used to quantify the tightness of physical contact between cells. A high index indicates that the Voronoi theoretical boundary coincides with the actual high-gradient edge, confirming that the cells exhibit a typical cobblestone-like tight arrangement, which is an important marker of high-quality epithelial stem cells. Step S4 is the final decision-making and screening execution stage based on multidimensional features.This step constructs a hierarchical screening logic. First, the intercellular topological consistency index is compared with a preset topological integrity threshold to initially screen out poorly performing cells that are loosely arranged or have undergone epithelial-mesenchymal transition. If the cells pass the initial screening, the index is used as a weight to weight and correct the cytoplasmic compactness factor obtained in S2, generating a stemness maintenance confidence score. This confidence score is a final score that comprehensively reflects the quality of the cell's external connectivity structure and internal texture. Only when this score falls into a preset high stemness range is the system determined to be a target cell, and the pixel coordinates are converted into physical coordinates for output. This process ensures that the screened cells possess both a tight population topology and dense internal stemness characteristics, achieving high-precision automated screening.

[0026] Specifically, the process of generating a texture-enhanced image by acquiring the original grayscale image of amniotic epithelial stem cells using an inverted phase contrast microscope and performing contrast-limited adaptive histogram equalization and bilateral filtering on the original grayscale image is as follows: The original grayscale image of amniotic epithelial stem cells is divided into non-overlapping rectangular context blocks. The local grayscale histogram of each rectangular context block is calculated, and the number of pixels in the local grayscale histogram that exceed a preset clipping threshold is evenly distributed to other grayscale levels in the local grayscale histogram. The adjacent rectangular context blocks after distribution are reconstructed and stitched together using a bilinear interpolation algorithm. A bilateral filter kernel containing a spatial distance weight factor and a pixel grayscale difference weight factor is constructed, and a sliding convolution operation is performed on the reconstructed and stitched image using the bilateral filter kernel to output the texture-enhanced image.

[0027] This implementation first addresses the common issues of uneven illumination and background noise during inverted phase-contrast microscopy imaging. The original grayscale images of amniotic epithelial stem cells often exhibit local brightness deviations due to differences in light source distribution or cell thickness; directly processing the entire image can easily lead to loss of detail. Therefore, by dividing the image into non-overlapping rectangular context blocks—essentially cutting a large image into several small local regions—a local grayscale histogram can be independently calculated for the brightness characteristics of each tiny region. The local grayscale histogram statistically analyzes the frequency of each grayscale level within that small region. To avoid amplifying random noise in the background while enhancing contrast, a preset clipping threshold is set to truncate the high-frequency portion of the histogram, evenly distributing the number of pixels in the truncated portion to the other low-frequency grayscale levels in the histogram. This operation limits excessive enhancement of local contrast. After equalizing each block, to eliminate obvious stitching gaps between blocks caused by different processing parameters, a bilinear interpolation algorithm is used. This algorithm uses the grayscale transformation function of the center points of four adjacent blocks to perform a weighted calculation on the current pixel, achieving a smooth transition. Subsequently, to preserve key edge information such as cell membranes while removing noise, a bilateral filter is introduced. The bilateral filter kernel is a non-linear combined weighted kernel, its core being the simultaneous consideration of the geometric proximity of pixels in spatial location and the similarity of pixel grayscale values. This causes the filter to behave as a Gaussian blur in flat regions to remove noise, while in edge regions, due to the large grayscale difference, the weights are reduced, thus preserving edges. The specific calculation process of the bilateral filter is shown in the following formula: ;in, : Represents the grayscale value at pixel p in the texture-enhanced image output after filtering; : Represents the original grayscale value at neighboring pixel q in the reconstructed and stitched intermediate image; This represents the area covered by the sliding convolution window centered at pixel p; This represents the spatial distance weighting factor, which follows a Gaussian distribution and is used to reduce the influence of distant pixels. This represents the Euclidean distance between pixel p and its neighboring pixel q; The pixel grayscale difference weighting factor follows a Gaussian distribution and is used to preserve edge information; : Represents the absolute value of the grayscale intensity difference between pixel p and its neighboring pixel q; This represents the normalization coefficient, used to ensure that the sum of the filter kernel weights is 1; This represents the standard deviation in the spatial domain. This parameter can be determined based on the image resolution, for example, by taking 2% to 5% of the number of pixels along the image diagonal. The value represents the standard deviation of the range, which can be determined based on the variance of the image background noise. Through the above calculations, the final output texture-enhanced image significantly enhances the texture details of the cell surface while suppressing the phase halo noise unique to phase contrast microscopy, providing a high-quality data foundation for subsequent feature extraction.

[0028] Specifically, the process of extracting cell kernel contours and calculating kernel centroid coordinates by performing maximum inter-class variance thresholding on the texture-enhanced image, and calculating the global gradient energy map of the texture-enhanced image using the Sobel operator is as follows: Traverse all gray levels of the texture-enhanced image, calculate the inter-class variance between foreground and background pixels, select the gray level corresponding to the maximum inter-class variance as the optimal segmentation threshold, and generate a binary mask using the optimal segmentation threshold; perform morphological closing operations on the binary mask, performing dilation followed by erosion, and extract the outer boundaries of the connected components after the operation as the cell kernel contours; calculate the zero-order spatial moment and first-order spatial moment of each connected component, and calculate the kernel centroid coordinates using the ratio of the first-order spatial moment to the zero-order spatial moment; convolve the texture-enhanced image using horizontal and vertical Sobel convolution kernels respectively, calculate the square root of the sum of the squares of the horizontal and vertical gradient components, and generate the global gradient energy map.

[0029] In this implementation scheme, to accurately locate cells and extract their morphological features, the cell nuclei first need to be separated from the background. The Otsu algorithm, also known as the maximum inter-class variance thresholding method, is employed. Its principle is to find an optimal grayscale threshold that maximizes the statistical variance between the foreground cell nuclei and the background, thus ensuring optimal separation. The specific formula for calculating the inter-class variance by iterating through all grayscale levels of the image is as follows: ;in, : Indicates the inter-class variance when gray level k is selected as the segmentation threshold; : Indicates the proportion of background pixels with gray values ​​less than the threshold k to the total number of pixels in the image; : Indicates the proportion of foreground cell nucleus pixels with gray values ​​greater than or equal to the threshold k to the total number of pixels in the image; : Represents the average grayscale value of the background pixels; : Represents the average grayscale value of the foreground cell kernel pixels. Selecting... The maximum value of k is used as the optimal segmentation threshold, and a binary mask is generated accordingly. For any internal holes or edge burrs that may exist within the cell nucleus in the binary mask, a morphological closing operation is performed. Morphological closing involves first dilating the image to expand the highlighted areas to connect breaks, then performing an erosion operation to restore the original size of the object, thereby filling the holes within the cell nucleus and smoothing the edges, obtaining a closed and smooth outer boundary of the connected domain as the cell nucleus outline. Next, to determine the seed points for generating the Voronoi diagram, the geometric center of each cell nucleus, i.e., the centroid coordinates, needs to be calculated. Using the principle of image moments, the centroid is determined by calculating the zeroth-order and first-order spatial moments of the connected domain, as shown in the following formula: ;in, : Represents the coordinates of the centroid of the nucleus in the horizontal direction of the image; : Represents the coordinates of the centroid in the vertical direction of the image; : Represents the zeroth-order spatial moment, which physically means the area of ​​the connected region, i.e., the total number of pixels contained in the cell kernel; : Represents the first-order spatial moment in the horizontal direction, reflecting the cumulative distribution of pixels along the horizontal axis; : Represents the first-order spatial moment in the vertical direction, reflecting the cumulative distribution of pixels along the vertical axis. Finally, to obtain global features that reflect changes in image edge intensity, the Sobel operator is used to perform convolution operations on the texture-enhanced image. The Sobel operator contains convolution kernels in both horizontal and vertical directions, and the gradient is obtained by calculating approximate gray-level differences between pixels in these two directions. The formula for calculating the global gradient energy map is as follows: ;in, : Represents the gradient energy value at coordinates (x, y) in the global gradient energy map; : Represents the gray value of the texture-enhanced image at coordinates (x, y); : Represents a horizontal Sobel convolution kernel, used for detecting vertical edges; The vertical Sobel convolution kernel is used to detect horizontal edges; * represents a 2D convolution operator. The global gradient energy map generated by this step highlights regions in the image with drastic grayscale changes, such as cell membrane boundaries and organelle edges, providing a physical basis for subsequent verification of the authenticity of Voronoi polygon boundaries.

[0030] Specifically, the process of generating a Voronoi Thésen polygon mesh using the nuclear centroid coordinates as seed points to define the single-cell jurisdiction, and calculating the cytoplasmic local binary pattern texture entropy after removing the nucleus region within the single-cell jurisdiction is as follows: Calculate the Euclidean distance from each pixel in the texture-enhanced image to all nuclear centroid coordinates, classify the pixels into the category represented by the nuclear centroid coordinate with the smallest Euclidean distance, and generate a Voronoi Thésen polygon mesh enclosed by polygonal boundaries; perform a masking operation on the texture-enhanced image using the Voronoi Thésen polygon mesh, setting the grayscale value of pixels within the corresponding cell nucleus contour to zero within each single-cell jurisdiction, retaining the cytoplasmic region outside the cell nucleus as the texture analysis object; traverse the pixels within the cytoplasmic region, compare the grayscale values ​​of the central pixel with those of neighboring pixels to generate binary codes, convert the binary codes to decimal numbers and statistically analyze the histogram distribution, calculate the information entropy of the histogram distribution using the Shannon entropy formula, and obtain the cytoplasmic local binary pattern texture entropy.

[0031] In this implementation scheme, to address the problem of blurred cell boundaries caused by the sheet-like growth of amniotic epithelial stem cells, making segmentation difficult using traditional edge detection algorithms, a Voronoi diagram partitioning technique from computational geometry is employed. The core idea of ​​this technique is to partition the space based on the "nearest neighbor principle." First, the Euclidean distance from each pixel in the texture-enhanced image to the coordinates of all identified centroids in the image is calculated, and the pixel is assigned to the category represented by the nearest centroid. In this way, the generated Voronoi Thiessen polygon mesh can simulate the theoretical growth territory of cells under contact inhibition, i.e., the single-cell territory, thus achieving approximate segmentation under conditions without obvious cell membrane boundaries. After determining the single-cell territory, to focus on analyzing the texture features of the cytoplasm (because changes in the cytoplasm best reflect the differentiation state of stem cells), it is necessary to exclude interference from the cell nucleus. Through a masking operation, the gray values ​​of pixels within the cell nucleus outline area of ​​the single-cell territory are forcibly set to zero, retaining only the annular or irregular regions outside the nucleus as the objects of cytoplasmic texture analysis. Next, to quantify the roughness and complexity of cytoplasmic texture, a Local Binary Pattern (LBP) algorithm is introduced. This algorithm encodes texture features by comparing the grayscale values ​​of the center pixel with its neighboring pixels. Specifically, it iterates through the cytoplasmic region, taking the current pixel as the center, and compares the grayscale values ​​of its eight surrounding pixels with the center pixel. If the neighboring pixel value is greater than the center pixel value, it is marked as 1; otherwise, it is marked as 0. The generated 8-bit binary code is the LBP value of that point. Finally, to statistically describe the degree of disorder in the distribution of these LBP texture features, the Shannon entropy formula from information theory is used to calculate the cytoplasmic local binary pattern texture entropy. The higher the entropy value, the more disordered the cytoplasmic texture, usually indicating that the cell has aged or undergone atypical differentiation. The calculation formula is as follows: ;in, : Represents the calculated local binary pattern texture entropy of the cytoplasm; : Represents the total number of categories of local binary pattern feature values. For the 8-neighborhood LBP algorithm, this value is 256. : Represents the feature value of the m-th LBP mode; : Represents the probability that the m-th LBP mode feature value appears in the cytoplasm region of the current single cell jurisdiction, that is, the ratio of the number of pixels that appear in this mode to the total number of pixels in the cytoplasm region; This represents an extremely small positive number to prevent zero values ​​from occurring in logarithmic operations, typically on the order of 10 to the power of -6. Through the above steps, the microscopic texture changes in the cytoplasm, which are difficult to discern with the naked eye, are transformed into quantifiable numerical indicators, providing crucial evidence for subsequent screening.

[0032] Specifically, the process of calculating the gray-level attenuation gradient from the nuclear centroid coordinates along the polar radius direction to the boundary of the Voronoi polygon, and weighting the local binary pattern texture entropy and the gray-level attenuation gradient to obtain the cytoplasmic compactness factor is as follows: A local polar coordinate system is established with the nuclear centroid coordinates as the pole. Rays are emitted in all directions at set angular intervals until they intersect with the boundary of the Voronoi Thiessen polygon mesh, obtaining the pixel gray-level sequence along the ray path. The linear regression slope of the pixel gray-level sequence relative to the ray length is fitted using the least squares method, and the mean of the absolute values ​​of the linear regression slopes of all angular rays is calculated to obtain the gray-level attenuation gradient. The gray-level attenuation gradient is subjected to max-min normalization. The normalized gray-level attenuation gradient and the cytoplasmic local binary pattern texture entropy are multiplied by their respective preset weight coefficients and then summed to output the cytoplasmic compactness factor.

[0033] In this implementation scheme, to capture the density variation pattern of amniotic epithelial stem cells from the nucleus center to the cell edge (high stem cell density typically has a high nucleocytoplasmic ratio and dense cytoplasm, resulting in rapid grayscale decay; while flattened or differentiated cells exhibit gradual grayscale decay), a grayscale decay gradient model based on a polar coordinate system was constructed. First, a local polar coordinate system was established with the nuclear centroid coordinates of each cell as the pole, and virtual rays were emitted outwards at set angular intervals (e.g., every 10 degrees). The process stopped when the ray intersected the boundary of a Voronoi polygon, and the grayscale values ​​of all pixels along the ray path were extracted, forming a discrete sequence reflecting the radial density distribution. To accurately describe the descent trend of this sequence and avoid single-point noise interference, a least squares method was used to perform linear regression fitting on the pixel grayscale distribution of each ray, and its slope was calculated. The formula for calculating the linear regression slope of a single ray is as follows: ;in, : indicates that at the launch angle is The slope of the linear regression of grayscale relative to distance on the ray path at that time; : Indicates the total number of pixels contained in this ray; : Represents the Euclidean distance from the nth pixel on the ray to the centroid of the nucleus; : Represents the average distance of all pixels on this ray from the centroid; : Represents the grayscale value of the nth pixel on the ray; This represents the average grayscale value of all pixels along the ray. After obtaining the slopes at all angles, the arithmetic mean of the absolute values ​​of these slopes is taken as the grayscale attenuation gradient of the cell. The larger this value, the more compact and three-dimensional the cytoplasm. Since the texture entropy obtained in the previous step and the attenuation gradient obtained in this step have different physical dimensions and numerical ranges, direct addition will lead to deviation. Therefore, the max-min normalization method is used to map both to the interval between 0 and 1. Finally, the cytoplasmic compactness factor is generated by weighted summation. This factor is a comprehensive evaluation index. The calculation formula is as follows: ;in, : Represents the final output cytoplasmic compaction factor; : Represents the local binary pattern texture entropy of the cytoplasm of the current cell calculated in the preceding steps; and : These represent the maximum and minimum values ​​of the texture entropy of all cells in the current field of view or training set, respectively; : Represents the mean of the absolute values ​​of the linear regression slopes of all angle rays of the current cell, i.e., the gray-level decay gradient; and : These represent the maximum and minimum values ​​of the grayscale decay gradient for all cells in the current field of view or training set, respectively; : Represents the preset weighting coefficient of texture entropy, which can be determined by measuring the contribution of sample data using principal component analysis; : Represents the preset weighting coefficient of the grayscale decay gradient, and and The sum is 1.

[0034] Specifically, the process of extracting the coordinates of the common boundary line of the Voronoi Thésen polygon mesh and mapping them to the global gradient energy map, and calculating the ratio of the cumulative pixel gradient value along the common boundary line path to the path length to obtain the edge intensity response value is as follows: Traverse the adjacency matrix of the Voronoi Thésen polygon mesh, identify the perpendicular bisector connecting two adjacent centroid coordinates, and extract the coordinate set of all pixels on the perpendicular bisector as the common boundary line coordinates; use the common boundary line coordinates for indexing in the global gradient energy map to extract the corresponding gradient magnitude sequence along the common boundary line coordinate path; sum the gradient magnitude sequences to obtain the total gradient energy of the path, count the number of pixels in the common boundary line coordinate set as the path length, and calculate the ratio of the total gradient energy to the path length to obtain the edge intensity response value.

[0035] In this implementation scheme, to verify whether the Voronoi mesh generated in the previous steps truly reflects the physical contact state between cells, it is first necessary to extract the verification path from the geometric structure. By traversing the adjacency matrix of the Voronoi Thiessen polygon mesh, a data structure that records which polygons each polygon is adjacent to, it is possible to quickly identify the centroid of each pair of adjacent cell nuclei. Geometrically, the perpendicular bisectors of two adjacent seed points constitute the Voronoi boundary between them. The coordinate set of all discrete pixels on this perpendicular bisector segment is extracted, i.e., the coordinates of the common boundary line. These coordinates represent the theoretical cell membrane position where two cells are in contact. Next, these coordinates are used for indexing and addressing in the global gradient energy map generated in step S1, i.e., the corresponding gradient magnitude is read according to the coordinate position. Since the bright pixels in the global gradient energy map represent strong edge signals in the image, if the gradient magnitude corresponding to the theoretical boundary line is high, it indicates that a real cell membrane does exist at that position. Finally, to eliminate the cumulative error caused by different boundary line lengths, the average gradient energy is calculated to obtain the edge intensity response value. This value quantifies the average signal intensity along the theoretical boundary path. The calculation formula is as follows: ;in, This represents the edge strength response value on the common boundary line connecting the u-th cell and the v-th cell; : Represents the total number of pixels contained in the coordinate set of the common boundary line, i.e., the path length; : Represents the index variable of a pixel on the path; and This represents the x and y coordinates of the Kth pixel on the common boundary line; : Represents the coordinates in the global gradient energy map The gradient magnitude at a given location. This calculation step transforms the geometric topological path into a specific physical signal intensity value. If this value is large, it means that the Voronoi boundary highly overlaps with the high gradient texture in the image, indicating tight intercellular connections.

[0036] Specifically, the process of defining the ratio of edge intensity response value to image background noise gradient as the cell gap topological consistency index is as follows: The gradient magnitudes of all pixels in the global gradient energy map are statistically analyzed, and the average gradient level of the entire image is calculated using the arithmetic mean method. This average gradient level is defined as the image background noise gradient. A division operation is performed to calculate the ratio of edge intensity response value to image background noise gradient, yielding the original signal-to-noise ratio. The original signal-to-noise ratio is then logarithmically transformed using the natural logarithm function, and the processed value is output as the cell gap topological consistency index.

[0037] In this implementation scheme, the simple edge intensity response value is easily affected by the brightness of the microscope light source or the complexity of the imaging background, and the absolute value is difficult to use directly for cross-image standardization. Therefore, establishing a relative benchmark is crucial. First, the gradient magnitude of all pixels in the global gradient energy map is statistically analyzed, and the average gradient level of the entire image is calculated using the arithmetic mean method. This value represents the basic signal noise floor under the current microscopic imaging environment, i.e., the image background noise gradient. This parameter can adaptively reflect the overall sharpness and texture richness of the image. The formula for calculating the image background noise gradient is as follows: ;in, : Represents the gradient of background noise in an image; : Indicates the vertical height in pixels of the image; : Indicates the horizontal width of the image in pixels; : Represents the gradient magnitude of the pixel in the h-th row and w-th column of the global gradient energy map. After obtaining the signal (edge ​​intensity) and noise (background gradient), a division operation is performed to calculate their ratio, which is essentially a signal-to-noise ratio (SNR) metric. To compress the dynamic range of the data and improve sensitivity to subtle differences, the original SNR is transformed using the natural logarithm function, outputting the intercellular gap topological consistency index. The formula for calculating the intercellular gap topological consistency index is as follows: ;in, : Represents the calculated intercellular gap topological consistency index. The higher the index, the more significantly the gradient signal at the theoretical boundary is higher than the background noise, and the tighter the cell connections. : Represents the edge strength response value calculated in the preceding steps; : Represents the background noise gradient of the current image; : Represents the numerical stability constant, usually taken as 1, used to ensure that the argument of the logarithmic function is always greater than 1, thereby ensuring that the output exponent is non-negative; : Represents the natural logarithm operation with base e. This step achieves adaptive normalization of features by introducing background noise as the denominator, enabling the final topological consistency index to robustly characterize the credibility of the coincidence between the Voronoi theoretical boundary and the actual cell membrane boundary, effectively distinguishing between real intercellular spaces and background artifacts.

[0038] Specifically, the intercellular topological consistency index is compared with a preset topological integrity threshold. If it is higher than the threshold, the intercellular topological consistency index is used to weight and correct the cytoplasmic compactness factor to obtain the stemness maintenance confidence. The specific process is as follows: a segmented judgment logic is constructed. When the intercellular topological consistency index is lower than the preset topological integrity threshold, the stemness maintenance confidence is set to zero. When the intercellular topological consistency index is higher than the preset topological integrity threshold, a normalization mapping is performed on the intercellular topological consistency index to convert it into a gain coefficient between 0 and 1. An exponential enhancement function is constructed using the gain coefficient as the exponent. The cytoplasmic compactness factor is multiplied using the exponential enhancement function, and the result is output as the stemness maintenance confidence.

[0039] In this implementation scheme, to ensure that the final selected amniotic epithelial stem cells possess both excellent internal activity and a favorable external growth environment, a hierarchical screening model based on topological constraints is constructed. First, a segmented decision-making logic is introduced, a decision mechanism similar to logic gating in circuits, designed to quickly eliminate cells whose internal texture is acceptable but whose population structure has collapsed. Specifically, the intercellular topological consistency index obtained in the previous steps is compared with a preset topological integrity threshold. This threshold is determined by analyzing the edge gradient characteristics of a large number of known cell samples that have undergone epithelial-mesenchymal transition, using the receiver operating characteristic (ROC) curve to determine the optimal cutoff point. When the cell's topological consistency index is below this threshold, it indicates that the tight junctions around the cell have loosened, classifying it as a high-risk differentiation cell. In this case, the stemness maintenance confidence is forcibly set to zero, resulting in a veto. Conversely, when it is above the threshold, it indicates that the cell is in a favorable epithelial cobblestone growth environment. In this case, topological quality is used to positively stimulate cytoplasmic quality. First, the topological index is converted into a gain coefficient through normalization mapping, and then an exponential enhancement function is constructed to nonlinearly correct the cytoplasmic compactness factor. The purpose of this step is to assign a non-linear score reward to high-quality cells with extremely clear edges and extremely tight connections, thereby differentiating high-quality cells from ordinary qualified cells. The formula for calculating the stemness maintenance confidence score is as follows: ;in, This represents the final calculated confidence level of stemness maintenance; the higher the value, the greater the cell's potential to serve as a seed cell. : Represents the intercellular space topological consistency index calculated in the preceding steps; : Indicates the preset topological integrity threshold, which is used to distinguish the critical point between epithelial-like morphology and mesenchymal-like morphology; : Represents the cytoplasmic compactness factor calculated in the preceding steps, reflecting the quality of the cell's internal texture; : Represents the theoretical maximum value of the topological consistency index observed in the current batch or training dataset, used to normalize the denominator; : Represents the exponential enhancement adjustment coefficient, used to control the sensitivity of weighted correction, and typically ranges from 0.5 to 2.0; This represents an exponential function with the natural constant e as its base. Its monotonically increasing nature and gradually increasing rate of change enable nonlinear signal amplification. Through this formula, the system achieves deep coupling between internal and external features, ensuring that the final score comprehensively reflects the biological state of amniotic epithelial stem cells.

[0040] Specifically, when the stemness maintenance confidence level is within a preset high stemness range, the process of determining the corresponding region as the target cell and outputting the screening coordinate position is as follows: Set a high stemness range consisting of a lower threshold and an upper threshold, and determine whether the stemness maintenance confidence level falls within this range; if the determination result is yes, retrieve the nuclear centroid coordinates corresponding to the region, and map the nuclear centroid coordinates from the image pixel coordinate system back to the physical space coordinate system of the microscope stage; encapsulate the coordinate data in the physical space coordinate system into a screening instruction, and generate a screening coordinate position list containing the target cell number and physical location information.

[0041] In this implementation scheme, after obtaining the stemness maintenance confidence score for each cell, the final decision output stage begins. To avoid false positives (such as abnormally high scores due to image artifacts) and false negatives (cells with barely acceptable quality), a high stemness interval consisting of a lower threshold and an upper threshold is set. Only when the confidence score falls within this strict interval is the corresponding region identified as a target cell. This dual-threshold screening strategy effectively ensures the high uniformity of the output samples. Once a target cell is identified, the system needs to convert the image pixel positions identified by the algorithm into physical spatial positions that can be recognized by automated equipment (such as microfluidic pipette tips or robotic arms). This process is called coordinate mapping. Since there is a scaling, rotation, and translation relationship between the microscope imaging field of view and the stage mechanical coordinate system, a pre-calibrated transformation matrix is ​​needed to map the nucleus centroid coordinates from the image pixel coordinate system back to the physical spatial coordinate system of the microscope stage. The specific calculation formula for coordinate mapping is as follows: ;in, : Represents the absolute lateral coordinates of the target cell in the physical coordinate system of the stage after mapping, usually in micrometers; : Represents the absolute longitudinal coordinates of the target cell in the physical coordinate system of the stage after mapping, usually in micrometers; and This represents the column and row coordinates of the target cell nucleus centroid extracted in the previous step in the image pixel coordinate system. and This represents the pixel equivalent of a microscope imaging system in the horizontal and vertical directions, that is, the actual physical length represented by each pixel, in micrometers per pixel; : Indicates the rotation angle of the camera sensor plane relative to the stage's moving axis, determined by a calibration slide; and : Indicates the physical offset of the origin at the top left corner of the current microscope field of view in the global coordinate system of the stage; and The colons represent the cosine and sine functions, used to handle rotational components. Finally, the calculated physical coordinate data is encapsulated into filtering instructions, generating a list of filtering coordinate positions containing the target cell number and physical location information. This list is directly transmitted as a control signal to the subsequent automated sorting device, achieving seamless integration from image recognition to physical grasping, and solving the technical gap problem where traditional microscopic observation cannot directly guide automated equipment operation.

[0042] Please see Figure 2 An amniotic epithelial stem cell screening device based on microscopic image feature extraction includes the following modules: an image acquisition and processing module, used to acquire raw grayscale images of amniotic epithelial stem cells using an inverted phase-contrast microscope, perform contrast-limited adaptive histogram equalization and bilateral filtering on the raw grayscale images to generate texture-enhanced images, perform maximum inter-class variance threshold segmentation on the texture-enhanced images to extract cell nuclear contours and calculate nuclear centroid coordinates, and calculate the global gradient energy map of the texture-enhanced images using the Sobel operator; and a cytoplasmic feature analysis module, used to generate a Voronoi Thésen polygon mesh with nuclear centroid coordinates as seed points to define single-cell jurisdictions, calculate the local binary pattern texture entropy of the cytoplasm after removing the cell nuclear region within the single-cell jurisdiction, and calculate the grayscale attenuation gradient from the nuclear centroid coordinates along the polar radius direction to the Voronoi polygon boundary. The system employs several modules: a local binary pattern texture entropy and grayscale decay gradient weighted calculation to obtain the cytoplasmic compactness factor; an edge topology analysis module to extract the coordinates of the common boundary line of the Voronoi Thésen polygon mesh and map them to the global gradient energy map; an edge intensity response value to be obtained by calculating the ratio of the cumulative pixel gradient value along the common boundary line path to the path length; and a cell gap topology consistency index to be defined as the ratio of the edge intensity response value to the image background noise gradient. A comprehensive screening decision module compares the cell gap topology consistency index with a preset topology integrity threshold. If the index is higher than the threshold, the cytoplasmic compactness factor is weighted and corrected using the cell gap topology consistency index to obtain the stemness maintenance confidence. When the stemness maintenance confidence is within a preset high stemness range, the corresponding region is identified as the target cell, and the screening coordinates are output.

[0043] In this implementation scheme, the image acquisition and processing module serves as the data sensing and preprocessing front-end of the entire device, and its main function is to construct a high-quality feature analysis foundation environment. This module first converts the optical analog signals acquired by the inverted phase-contrast microscope into digital grayscale signals, and uses image enhancement algorithms to eliminate physical noise interference from the imaging system, ensuring the visibility of fine textures. Subsequently, this module undertakes the dual tasks of "localization" and "mapping": on the one hand, it uses segmentation algorithms to locate the cell nucleus, a biomarker, providing seed points for subsequent single-cell segmentation; on the other hand, it generates a global energy map through gradient calculation, providing a physical reference base map for subsequent topological verification, thus completing the initial transformation from the original image to a feature map. The cytoplasmic feature analysis module, as one of the core computing units of the device, focuses on the quantitative assessment of the biological state inside single cells. Given the difficulty in accurately segmenting the boundaries of amniotic epithelial stem cells, this module utilizes the nuclear centroid coordinates provided by the preceding module to construct a virtual Voronoi region to simulate the cell growth space, solving the problem of traditional segmentation algorithm failure. Building upon this foundation, the module delves into the texture details and density variations within its jurisdiction. By coupling local texture entropy with radial gray-level decay gradients, it transforms microscopic cytoplasmic structural features, indistinguishable to the naked eye, into measurable numerical factors, thereby achieving a precise initial assessment of the "intrinsic stemness quality" of individual cells. The edge topology analysis module, serving as the device's topology verification unit, verifies cell growth status at the population structure level. This module utilizes the common boundary lines of the Voronoi grid as detection paths, retrieving corresponding physical edge signals from the global gradient energy map. By calculating the ratio of signal intensity to background noise along the theoretical path, the module objectively determines whether cells have formed tight epithelial-like connections, effectively identifying "false positive" cells that, while possessing acceptable internal texture, exhibit loose population structures or have undergone epithelial-mesenchymal transition. This achieves objective verification of the cell's "external growth environment." The comprehensive screening decision module is the control center of the entire device, responsible for executing hierarchical screening logic and outputting execution instructions. This module first uses an index provided by the edge topology analysis module to perform a "one-vote veto" threshold filtering, eliminating regions with abnormal population structures. Then, it uses topological quality to weight and correct the intrinsic quality score provided by the cytoplasmic feature analysis module, ensuring that the final selected targets possess both a dense internal structure and tight external connections. Finally, the module maps the calculation results from the image space back to the physical space, generating coordinate instructions that can be recognized by automated equipment, completing the final closed loop from "image recognition" to "physical sorting."

[0044] In summary, this application has at least the following effects:

[0045] A method and device for screening amniotic epithelial stem cells based on microscopic image feature extraction were developed. This method achieves label-free, non-destructive screening of live cells, effectively avoiding the potential impact of exogenous reagents on stem cell biological activity and clinical safety during traditional immunofluorescence or flow cytometry. Technically, the combined processing of bilateral filtering and adaptive histogram equalization overcomes the interference of phase halo on texture details in phase-contrast microscopy. Innovatively, Voronoi polygon meshes combined with cytoplasmic texture entropy and gray-level attenuation gradients are used to overcome the bottleneck of accurately segmenting and quantifying the internal microstructure of epithelial cells growing in sheets using traditional edge detection algorithms, achieving precise measurement of the "intrinsic stem quality" of single cells. Furthermore, by constructing a cell intercellular topological consistency index and using dual verification of geometric theoretical boundaries and physical gradient signals, the contact inhibition state of cell populations and the risk of early epithelial-mesenchymal transition (EMT) are accurately identified, ensuring that the screened cells possess both dense internal structure and compact population topological characteristics. The resulting physical space coordinate instructions achieve a seamless connection from algorithm recognition to automated sorting, significantly improving the purity, efficiency, and standardization of clinical-grade amniotic epithelial stem cell preparation.

[0046] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0047] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0048] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0049] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0050] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0051] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for screening amniotic epithelial stem cells based on microscopic image feature extraction, characterized in that, Includes the following steps: S1. Acquire the original grayscale image of amniotic epithelial stem cells using an inverted phase contrast microscope. Perform contrast-limited adaptive histogram equalization and bilateral filtering on the original grayscale image to generate a texture-enhanced image. Perform maximum inter-class variance threshold segmentation on the texture-enhanced image to extract the cell nuclear contour and calculate the nuclear centroid coordinates. Calculate the global gradient energy map of the texture-enhanced image using the Sobel operator. S2. Generate a Voronoi Thésen polygon mesh using the nuclear centroid coordinates as seed points to define the single-cell territory. After removing the nucleus region within the single-cell territory, calculate the local binary pattern texture entropy of the cytoplasm and calculate the gray-level decay gradient from the nuclear centroid coordinates along the polar radius direction to the Voronoi polygon boundary. The local binary pattern texture entropy and the gray-level decay gradient are weighted and calculated to obtain the cytoplasmic compactness factor. S3. Extract the coordinates of the common boundary line of the Voronoi Thésen polygon mesh and map them to the global gradient energy map. Calculate the ratio of the cumulative pixel gradient value along the path of the common boundary line to the path length to obtain the edge intensity response value. Define the ratio of the edge intensity response value to the image background noise gradient as the intercellular gap topological consistency index. S4. Compare the intercellular gap topological consistency index with the preset topological integrity threshold. If it is higher than the threshold, use the intercellular gap topological consistency index to weight and correct the cytoplasmic compactness factor to obtain the stemness maintenance confidence. When the stemness maintenance confidence is in the preset high stemness range, determine the corresponding region as the target cell and output the screening coordinate position.

2. The method for screening amniotic epithelial stem cells based on microscopic image feature extraction according to claim 1, characterized in that: The specific process of generating texture-enhanced images by acquiring raw grayscale images of amniotic epithelial stem cells using an inverted phase-contrast microscope and performing contrast-limited adaptive histogram equalization and bilateral filtering on the raw grayscale images is as follows: The original grayscale image of amniotic epithelial stem cells is divided into non-overlapping rectangular context blocks. The local grayscale histogram of each rectangular context block is calculated, and the number of pixels in the local grayscale histogram that exceed the preset clipping limit value is evenly distributed to other grayscale levels of the local grayscale histogram. The assigned adjacent rectangular context tiles are reconstructed and stitched together using a bilinear interpolation algorithm; A bilateral filter kernel containing spatial distance weighting factor and pixel gray-level difference weighting factor is constructed. The sliding convolution operation is performed on the reconstructed and stitched image through the bilateral filter kernel to output a texture-enhanced image.

3. The method for screening amniotic epithelial stem cells based on microscopic image feature extraction according to claim 2, characterized in that: The specific process of extracting cell nucleus contours and calculating nucleus centroid coordinates by performing maximum inter-class variance thresholding on the texture-enhanced image, and then calculating the global gradient energy map of the texture-enhanced image using the Sobel operator is as follows: Traverse all gray levels of the texture enhancement image, calculate the inter-class variance between foreground and background pixels, select the gray level corresponding to the maximum inter-class variance as the optimal segmentation threshold, and generate a binarization mask through the optimal segmentation threshold. Perform morphological closing operations on the binarized mask, perform dilation followed by erosion, and extract the outer boundary of the connected domain after the operation as the cell nucleus outline; Calculate the zeroth and first spatial moments of each connected domain, and calculate the core centroid coordinates by the ratio of the first spatial moment to the zeroth spatial moment. The texture enhancement image is convolved with Sobel convolution kernels in the horizontal and vertical directions respectively. The square root of the sum of the squares of the horizontal and vertical gradient components is calculated to generate a global gradient energy map.

4. The method for screening amniotic epithelial stem cells based on microscopic image feature extraction according to claim 1, characterized in that: The Voronoi Thiessen polygon mesh is generated using the nuclear centroid coordinates as seed points to define the single-cell territory. The specific process of calculating the local binary pattern texture entropy of the cytoplasm after removing the nucleus region within the single-cell territory is as follows: Calculate the Euclidean distance from each pixel in the texture-enhanced image to all kernel centroid coordinates, classify the pixels into the category represented by the kernel centroid coordinate with the smallest Euclidean distance, and generate a Voronoi Thésen polygon mesh bounded by polygon boundaries. The texture enhancement image is masked using Voronoi polygon mesh. Within each single cell region, the pixel grayscale value of the corresponding cell nucleus contour is set to zero, while the cytoplasm region outside the cell nucleus is retained as the texture analysis object. The pixels within the cytoplasm are traversed, and the gray values ​​of the central pixel and its neighboring pixels are compared to generate a binary code. The binary code is then converted to a decimal number, and the histogram distribution is statistically analyzed. The information entropy of the histogram distribution is calculated using the Shannon entropy formula, thus obtaining the local binary pattern texture entropy of the cytoplasm.

5. The method for screening amniotic epithelial stem cells based on microscopic image feature extraction according to claim 4, characterized in that: The specific process for calculating the gray-level attenuation gradient from the nuclear centroid coordinates along the polar radius direction to the boundary of the Voronoi polygon, and weighting the local binary pattern texture entropy with the gray-level attenuation gradient to obtain the cytoplasmic compactness factor is as follows: A local polar coordinate system is established with the core centroid coordinates as the pole. Rays are emitted in all directions at set angular intervals until they intersect with the boundary of the Voronoi Thésen polygon mesh, and the pixel grayscale sequence on the ray path is obtained. The grayscale attenuation gradient is obtained by fitting the linear regression slope of the pixel grayscale sequence with respect to the ray length using the least squares method and calculating the mean of the absolute values ​​of the linear regression slopes of all angle rays. The grayscale decay gradient is subjected to maximum and minimum normalization. The normalized grayscale decay gradient and the local binary pattern texture entropy of the cytoplasm are multiplied by the corresponding preset weight coefficients and then summed to output the cytoplasmic compactness factor.

6. The method for screening amniotic epithelial stem cells based on microscopic image feature extraction according to claim 1, characterized in that: The specific process of extracting the coordinates of the common boundary line of the Voronoi Thiessen polygon mesh and mapping them to the global gradient energy map, and calculating the ratio of the cumulative pixel gradient value along the path of the common boundary line to the path length to obtain the edge intensity response value is as follows: Traverse the adjacency matrix of the Voronoi Thésen polygon mesh, identify the perpendicular bisectors connecting the coordinates of two adjacent centroids, and extract the coordinate set of all pixels on the perpendicular bisector as the coordinates of the common boundary line. By indexing and addressing the global gradient energy map using the coordinates of the common boundary line, the gradient magnitude sequence corresponding to the coordinate path of the common boundary line is extracted. The total energy of the path gradient is obtained by summing the gradient magnitude sequences. The number of pixels in the coordinate set of the common boundary line is used as the path length. The ratio of the total energy of the path gradient to the path length is calculated to obtain the edge intensity response value.

7. The method for screening amniotic epithelial stem cells based on microscopic image feature extraction according to claim 6, characterized in that: The specific process of defining the ratio of edge intensity response value to image background noise gradient as the intercellular gap topological consistency index is as follows: The gradient magnitude of all pixels in the global gradient energy map is statistically analyzed, and the average gradient level of the entire image is calculated by the arithmetic mean method. The average gradient level of the entire image is defined as the image background noise gradient. The ratio of edge intensity response value to image background noise gradient is calculated by performing a division operation to obtain the original signal-to-noise ratio. The original signal-to-noise ratio is then logarithmically transformed using the natural logarithm function, and the processed value is output as the intercellular gap topological consistency index.

8. The method for screening amniotic epithelial stem cells based on microscopic image feature extraction according to claim 1, characterized in that: The process of comparing the intercellular gap topological consistency index with a preset topological integrity threshold, and if it is higher than the threshold, then using the intercellular gap topological consistency index to weight and correct the cytoplasmic compactness factor to obtain the stemness maintenance confidence is as follows: A segmented judgment logic is constructed, and when the intercellular topological consistency index is lower than the preset topological integrity threshold, the stemness maintenance confidence is set to zero. When the intercellular gap topological consistency index is higher than the preset topological integrity threshold, the intercellular gap topological consistency index is normalized and mapped to a gain coefficient between 0 and 1. An exponential enhancement function is constructed using the gain coefficient as the exponent. The cytoplasmic compactness factor is then multiplied using the exponential enhancement function, and the result is output as the stemness maintenance confidence level.

9. The method for screening amniotic epithelial stem cells based on microscopic image feature extraction according to claim 8, characterized in that: When the stemness maintenance confidence level is within the preset high stemness range, the specific process of determining the corresponding region as the target cell and outputting the screening coordinates is as follows: Define a high dryness interval consisting of a lower threshold and an upper threshold, and determine whether the dryness maintenance confidence falls within this interval. If the judgment result is yes, then retrieve the corresponding nuclear centroid coordinates of the region and map the nuclear centroid coordinates from the image pixel coordinate system back to the physical space coordinate system of the microscope stage. The coordinate data in the physical space coordinate system is encapsulated into filtering instructions to generate a list of filtered coordinate positions containing the target cell number and physical location information.

10. An amniotic epithelial stem cell screening device based on microscopic image feature extraction, applied to the amniotic epithelial stem cell screening method based on microscopic image feature extraction as described in any one of claims 1-9, characterized in that, Includes the following modules: The image acquisition and processing module is used to acquire the original grayscale image of amniotic epithelial stem cells through an inverted phase contrast microscope, perform contrast-limited adaptive histogram equalization and bilateral filtering on the original grayscale image to generate a texture-enhanced image, perform maximum inter-class variance threshold segmentation on the texture-enhanced image to extract the cell nuclear contour and calculate the nuclear centroid coordinates, and calculate the global gradient energy map of the texture-enhanced image through the Sobel operator. The cytoplasmic feature analysis module is used to generate a Voronoi Thésen polygon mesh with the nuclear centroid coordinates as the seed point to define the single-cell territory. After removing the nucleus region within the single-cell territory, the local binary pattern texture entropy of the cytoplasm is calculated, and the gray-level decay gradient from the nuclear centroid coordinates along the polar radius direction to the Voronoi polygon boundary is calculated. The local binary pattern texture entropy and the gray-level decay gradient are weighted and calculated to obtain the cytoplasmic compactness factor. The edge topology analysis module is used to extract the coordinates of the common boundary line of the Voronoi Thésen polygon mesh and map them to the global gradient energy map. It calculates the ratio of the cumulative pixel gradient value along the common boundary line path to the path length to obtain the edge intensity response value. The ratio of the edge intensity response value to the image background noise gradient is defined as the intercellular gap topology consistency index. The comprehensive screening decision module compares the intercellular topological consistency index with a preset topological integrity threshold. If the index is higher than the threshold, the intercellular topological consistency index is used to weight and correct the cytoplasmic compactness factor to obtain the stemness maintenance confidence. When the stemness maintenance confidence is in the preset high stemness range, the corresponding region is determined to be the target cell and the screening coordinate position is output.

Images