Unsupervised cell counting and positioning method and system based on multi-modal cell image

Abstract

The invention relates to the technical field of cell image analysis, and discloses an unsupervised cell counting and positioning method based on a multi-modal cell image, which comprises the following steps: acquiring the multi-modal cell image; analyzing the image; separating the different modal cell images into a single cell map and a cell cluster map; the average value and the standard deviation of the perimeters of all the single cells on the single cell map are calculated, cell morphological category judgment is carried out on the cells in the single cell map, and the category judgment comprises two categories: cells with uniform cell sizes and cells with non-uniform cell sizes. According to the method, the cell images are divided into two types based on the characteristics of the cell images of different modalities, and the corresponding cell cluster segmentation algorithm is selected according to the types of the cell images to carry out further cell cluster segmentation on the cell images, so that the counting of the multi-modal cell images is more accurate, better generalization performance is achieved, the burden of marking work can be effectively reduced, and the marking efficiency is improved. Meanwhile, relatively high cell counting precision is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of cell image analysis, and in particular to an unsupervised cell counting and positioning method and system based on multimodal cell images. Background Art

[0002] Cells are the basic units that make up life. In recent years, with the continuous advancement and development of microscopic imaging technology, a large amount of cell image data has been generated. Due to the diversity of cell types and cell staining patterns and the differences between cell morphologies, the types of cell images imaged are very rich. For a large amount of diverse cell image data, it becomes very difficult to obtain information of research value, such as the number of cells and the location of cells.

[0003] Cell counting plays an extremely important role in many fields such as medical image analysis. The current research difficulties are: cell images have low grayscale, uneven brightness distribution, and the unique complex structural characteristics of cell images themselves. The characteristics of fluorescence images are high variability in contrast, noise and local changes, which makes cell counting and cell positioning very difficult. The most primitive cell counting generally uses manual counting methods, but traditional manual counting methods are inefficient, time-consuming, and low in accuracy, and are easily affected by the counter's own experience. Obviously, the manual cell counting method is no longer suitable for cell image data at this stage, so an automated processing method is urgently needed to solve this problem. With the rapid development of computer imaging technology, it is an inevitable development trend to automatically complete cell counting of cell images with the help of image processing algorithms. However, due to the different cell types, microscopic imaging equipment, processing methods, imaging modes, and staining methods that produce various styles of cell images, it is full of huge challenges to develop a universal algorithm that can be applied to all styles of cell images. Therefore, the development of a universal unsupervised cell image processing system is of great significance and very difficult.

[0004] In recent years, with the rapid development of deep learning technology, it has achieved great success in the field of medical image processing. As one of the basic tasks in medical image analysis, cell image processing tasks have produced a large number of reliable deep learning-based methods. For example, the UNet-based model CellPose has additional structures or outputs, which can effectively segment dense cell images. However, most of the current cell counting and segmentation models are for single-modality image data models and lack universality. However, in practical applications, the image data set to be segmented is usually multimodal image data. Due to the large differences in texture, features, and cell shape and size of different modal images, it is difficult to obtain satisfactory segmentation results using existing methods for cell segmentation and counting. The mainstream image segmentation method uses deep neural networks, which have high robustness and good accuracy, but this task is still challenging. Since there are many imaging methods for cell images, the images will have different modal information, resulting in great differences in image features. And because cells are of various types and shapes, it is difficult for neural network models to accurately perform pixel-level segmentation tasks. In addition, since there may be hundreds or thousands of cells in an image, it usually takes a lot of effort for experts to label the cell images. Manual labeling is a labor-intensive, inefficient and time-consuming task with subjective biases. In addition, it relies on the professional knowledge of the labeler, which greatly increases the research cost. Therefore, there is an urgent need to provide a technical solution to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to solve the problems existing in the prior art and to propose an unsupervised cell counting and positioning method and system based on multimodal cell images.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] An unsupervised cell counting and positioning method based on multimodal cell images comprises the following steps:

[0008] Acquire multimodal cell images;

[0009] Analyze cell images;

[0010] Separate cell images of different modalities into single cell images and cell cluster images;

[0011] Calculate the average and standard deviation of the perimeters of all single cells in the single cell image, and perform cell morphology classification on the cells in the single cell image. The classification includes two categories: cells with uniform cell sizes and cells with uneven cell sizes.

[0012] According to the cell characteristics of the cell images with uniform cell sizes and the cell images with uneven cell sizes, a cell cluster segmentation algorithm is selected to segment the cell cluster images;

[0013] Perform cell counting and generate cell localization maps based on dot annotation.

[0014] Preferably, the specific steps of analyzing the cell image include:

[0015] Determine the type of the cell image to determine whether it is a grayscale image. If it has color image features, it is classified as a non-grayscale image class; if it has grayscale image features, it is classified as a grayscale image class.

[0016] Perform differential preprocessing on non-grayscale images and grayscale images respectively;

[0017] Perform segmentation process to obtain binary image;

[0018] The binary image was processed to separate the single cell image and the cell cluster image.

[0019] Preferably, the specific process of performing differential preprocessing on the non-grayscale image and the grayscale image respectively includes:

[0020] Perform image enhancement operations on grayscale cell images;

[0021] For non-grayscale cell images, convert them from RGB color space to HSV color space, select and save only the color of the cell nucleus, save it as a new image and perform grayscale processing.

[0022] Preferably, the specific process of performing the segmentation process to obtain a binary image includes: designing a sliding window on the cell image, setting the sliding window size and the moving step size to certain values, so that the image is divided into several small blocks, the size of the image block is related to the complexity of the image, the smaller the complexity of the image, the larger the image block can be, performing a grayscale clustering algorithm and adaptive threshold processing on each image block, and finally merging all image blocks to form a cell image segmentation result.

[0023] Preferably, the specific process of processing the binary image to separate the single cell image and the cell cluster image is as follows: using distance transformation to generate seed points for the foreground target area, performing seed point detection on the ROI area, if there is a connected ROI area with only one seed point, it proves that this is the area where a single cell is located, and all such areas are saved in one image; if there are multiple seed points simultaneously existing in a connected ROI area, it means that this area is a multi-cell adhesion area, that is, the area where the cell cluster is located, and all such areas are saved as new images; and finally two new binary images are generated, one is a binary image of all single cells, and the other is a cell image of all cell clusters.

[0024] Preferably, the specific process of calculating the mean and standard deviation of the perimeters of all single cells in the single cell image and determining the cell morphology category of the cells in the single cell image is as follows: calculating the perimeter of each ROI area of ​​each newly generated single cell binary image and calculating the perimeter standard deviation; cells with a larger standard deviation are classified as having uneven cell size, and cells with a smaller standard deviation are classified as having uniform cell size.

[0025] Preferably, the specific process of selecting a cell cluster segmentation algorithm to segment the cell cluster image based on the cell features of the cell images with uniform cell sizes and the cell images with uneven cell sizes is as follows: using a watershed algorithm to further segment the cell clusters for the cell images with uneven cell sizes; performing curvature analysis on the cell images with uniform cell sizes, and combining the use of the average cell perimeter and standard deviation calculated for the single cell images to assist in further segmentation, and finally obtaining a segmentation result.

[0026] Preferably, the curvature analysis of the cell image with uniform cell size is combined with the use of the average cell perimeter and standard deviation calculated for the single cell image to assist in further segmentation, and the specific process of finally obtaining the segmentation result is as follows:

[0027] A set of candidate seed points approximate to the cell center is obtained through curvature analysis;

[0028] Determine whether there are other seeds within the distance of the average cell radius around a seed point. If not, they are identified as one cell. If so, they are identified as belonging to the same cell. The average cell radius is determined by the average cell circumference. and standard deviation To calculate, the formula is: , to avoid the existence of adhesions between cells in the cell cluster, which would lead to false seeds in the contact areas between adjacent cells and cause misidentification.

[0029] Preferably, the specific process of performing cell counting and generating a cell localization map based on point annotation is: obtaining the final single cell morphology segmentation result, counting the number of ROI areas, i.e. the number of cells, and generating seed points for the ROI area, i.e. a point-based cell localization map.

[0030] An unsupervised cell counting and positioning system based on multimodal cell images is applied to an unsupervised cell counting and positioning method based on multimodal cell images, comprising:

[0031] Input module: receiving input multimodal cell images as raw image data;

[0032] Image type judgment module: judges the type of all input images and distinguishes them into grayscale images and non-grayscale images;

[0033] Image differentiation preprocessing module: De-noise and enhance contrast operations for grayscale images. For non-grayscale cell images, convert them from RGB color space to HSV color space, select only the color of the cell nucleus and save it as a new image, then grayscale it, and then perform image enhancement operations. The purpose of preprocessing is to correct the illumination of the image and increase the contrast of the cell, so as to maximize the optimal input state of the subsequent algorithm;

[0034] Image segmentation module: Design a sliding window, perform grayscale clustering and adaptive threshold processing on each image block, merge all binarized image blocks, and obtain the segmentation result;

[0035] Cell morphology judgment module: Use distance transformation to generate seed points for each connected ROI area, and divide all connected ROI areas into single cell images and cell cluster images according to the number of seed points in each connected area, so as to facilitate subsequent processing;

[0036] Cell cluster segmentation module: Calculate the average perimeter and standard deviation of all connected ROI areas of the separated single cell map, classify the cell clusters with larger standard deviation as uneven cell size, and further segment them using the watershed algorithm; classify the cell clusters with smaller standard deviation as uniform cell size, and further segment them by combining curvature analysis, average perimeter size and standard deviation of single cells, and finally obtain the cell cluster segmentation result, then merge the single cell and the segmented cell cluster map to obtain the final segmentation result;

[0037] Cell counting and positioning module: count cells in the binary image of single cells and generate a point-based cell positioning map;

[0038] Visualization output module: Visualize the counting results and positioning map, and output the result map.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention divides cell images into two categories based on the characteristics of cell images of different modalities, and selects the corresponding cell cluster segmentation algorithm according to the cell image type to further perform cell cluster segmentation, so that the cell segmentation of multimodal cell images is more accurate and has better generalization performance, which can effectively reduce the burden of annotation work and achieve higher cell counting accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is an overall flow chart of an unsupervised cell counting and positioning method based on multimodal cell images of the present invention;

[0042] Figure 2 Schematic diagram of the original grayscale cell image input for the present invention;

[0043] Figure 3 Schematic diagram of the original fluorescent cell image input for the present invention;

[0044] Figure 4 This is a schematic diagram of a fluorescent cell image after pretreatment according to the present invention;

[0045] Figure 5 This is a schematic diagram of the image sliding window design in the present invention;

[0046] Figure 6 Schematic diagram of the cell image separation of single cells and cell clusters of the present invention; wherein a is a single cell image, and b is a cell cluster image;

[0047] Figure 7 It is a schematic diagram of grayscale cell image segmentation results of the present invention;

[0048] Figure 8 It is a schematic diagram of the cell positioning result based on point annotation of the grayscale cell image of the present invention;

[0049] Fig. 9 This is a schematic diagram of the fluorescent cell image segmentation result of the present invention;

[0050] Fig.10 It is a schematic diagram of the cell localization result based on point annotation of the fluorescent cell image of the present invention;

[0051] Fig.11 Schematic diagram of the structure of an unsupervised cell counting and positioning system for multimodal cell images according to the present invention. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0053] like Figure 1 As shown, in this embodiment, an unsupervised cell counting and positioning method based on multimodal cell images is provided, and the method comprises the following steps:

[0054] Step 1: Input multimodal cell image data, grayscale cell image such as Figure 2 As shown, the fluorescent cell images are Figure 3 shown.

[0055] Step 2: The image type is determined based on the image's own features. The type determination is divided into two categories: one is grayscale image, and the other is non-grayscale image.

[0056] Non-grayscale images are generally fluorescent cell images. Such images often contain other interfering colors, which affect the subsequent processing flow. Therefore, the RGB color space of such images is converted to the HSV color space, and only the color of the cell nucleus is extracted and saved, and other color interference is removed. Grayscale processing is then performed, and image enhancement operations are performed in the same way as grayscale images. The schematic diagram of the fluorescent cell image processed by the method is shown in FIG. Figure 4 As shown, the purpose is to eliminate background interference, remove noise, correct image illumination and increase the contrast of cell images. The implementation methods include but are not limited to background interference elimination processing, noise removal processing, and contrast improvement processing. The purpose is to make the preprocessed image meet the image quality requirements required by the segmentation algorithm and facilitate subsequent processing. Since there are also cases where the quality of the original input image meets the algorithm requirements, this preprocessing process is optional.

[0057] Image preprocessing: To correct for uneven illumination and increase cell contrast in microscopic images, a grayscale correction method was used. This was achieved by taking the logarithm and average of the grayscale values, followed by contrast-limited adaptive histogram equalization (CLAHE) to enhance local contrast, and finally a median filter to mitigate the sharpening effect introduced by CLAHE.

[0058] For the input image The gray level of the correction is relative to The average gray value of all pixels in :

[0059]

[0060] in and is a predefined constant term with a value of and .image After further processing, the grayscale value is compared with The average gray value The relationship between the two is used to correct each pixel, thereby increasing the contrast of the cell. Gray value is greater than of A collection of pixels, The correction method for each pixel is as follows:

[0061]

[0062] in For The average gray value calculated above.

[0063] Step 3: Design a sliding window for the preprocessed image, set the image block size and sliding step size, so that the image is divided into several small blocks. The size of the image block is related to the complexity of the image. The smaller the complexity of the image, the larger the image block can be. Perform grayscale clustering algorithm and adaptive threshold processing on each image block, and finally merge all image blocks to form the cell image segmentation result. The segmented cell image is not limited in size, and images of any size can be applied. It is only necessary to change the sliding window size and moving step size to a suitable size. The sliding step size is usually set to half of the image block, that is, the overlap of the image blocks is 50%, but it can be flexibly adjusted according to actual conditions. The above method is adopted because some images have uneven grayscale distribution, which will lead to poor image segmentation effect. Therefore, the present invention adopts the design concept of sliding window to process the image in blocks. The sliding window design is as follows Figure 5 shown.

[0064] Step 4: Segmentation processing. Perform grayscale clustering and adaptive threshold segmentation operations on each image block divided in step 3. This will obtain a binary mask image of each image block, where the white area is the cell and the black area is the background. Finally, all the image blocks are merged to form a binary image of the cell image. The grayscale cell image segmentation map is shown in the figure. Figure 7 As shown, the fluorescent cell image segmentation diagram is as follows Fig. 9 As shown;

[0065] The adaptive threshold is calculated to segment each local image block, and the formula is as follows:

[0066]

[0067] in: It is The threshold of the same image block, It is a parameter that controls the balance between the threshold of the current image block and the previous image block; It is The median of the grayscale values ​​in the image blocks.

[0068] Step 5: Determine the cell adhesion situation, and perform distance transformation on the binary image of the cell image obtained in step 4 to generate seed points. The main purpose of distance transformation is to calculate the distance from each foreground pixel to the nearest background pixel for each foreground pixel. The local maximum point is the point with the largest distance value, which usually represents the center of the cell. Specifically, the Euclidean distance transformation is used to determine the candidate seed points. These seed points are close to the center point of the cell. The Euclidean distance transformation formula is:

[0069]

[0070] in, are the coordinates of the foreground pixel, is the set of background pixels, yes The pixel coordinates in .

[0071] The distance transform is defined as the change in the Euclidean distance between consecutive points on the cell cluster boundary. This metric helps identify significant deviations on the cell boundary, where two consecutive points and The distance change between can be given by the following formula:

[0072]

[0073] in and are the coordinates of consecutive points.

[0074] After generating seed points through distance transformation, the white area in the binary image generates seed points close to the center of the cell. The number of seed points in each connected ROI area is used to judge the cell morphology and distinguish cell clusters from single cells. If there is a connected ROI area with only one seed point, it proves that this is the area where a single cell is located; if there are multiple seed points in a connected ROI area at the same time, it means that this area is a multi-cell adhesion area, which means that this area is the area where the cell cluster is located.

[0075] Based on this judgment method, two new binary images are finally generated: single cell binary image: all connected areas containing only one seed point are marked as foreground. In this image, only those isolated and independent cell areas are retained. Cell cluster binary image: all connected areas containing multiple seed points are marked as foreground. These areas represent the aggregation of multiple cells. Figure 6 As shown, a is a segmented single cell, and b is a segmented cell cluster;

[0076] Step 6: Determine the cell morphology and calculate the average perimeter of each ROI area in the binary image of all single cells generated in step 5 and standard deviation This process requires saving the calculated average cell perimeter value and standard deviation results, which can help determine the potential cell center in the clustered cell cluster. The formula for calculating the average cell perimeter is:

[0077]

[0078] The formula for calculating standard deviation is:

[0079]

[0080] in: is the circumference of a single cell.

[0081] If the standard deviation is large, it is classified as a category of uneven cell size, and if the standard deviation is small, it is classified as a category of uniform cell size, completing the classification of cell morphology. The classification results are two types: cell images with uniform cell morphology and cell images with uneven cell morphology.

[0082] Step 7: Further segmentation of the cell clusters. Step 6 classifies the cell images by cell morphology, and a differentiated segmentation method is used for the cell clusters based on the classification results. The differentiated cell cluster segmentation refers to: using the watershed algorithm to further segment the cell clusters for the cell images with uneven cell morphology and size. This is because, when processing cell images with uneven cell size and uniform cell size, certain features of the single cell image (in this embodiment, the average cell perimeter and standard deviation) are different. These features can be used as an auxiliary means to help segment the cell clusters.

[0083] The specific operation of further segmenting the cell clusters with uniform cell morphology and size is as follows: generate new seed points through curvature analysis, and find high curvature points by calculating the curvature value on the boundary curve. These points are usually located in the contact area between cells, that is, where the cell boundaries meet. The average cell perimeter size and standard deviation calculated for the single cell image are combined to assist in further segmentation, and finally the segmentation result is obtained. The curvature describes the rate of change of the curve. For a given boundary curve, the curvature analysis formula is:

[0084]

[0085] in: is a point on the boundary curve, is the first-order derivative (indicating the rate of change in the tangent direction); is the second derivative (representing the change in the rate of change of the tangent).

[0086] Since there are adhesions between cells in the cell cluster, the contact area between adjacent cells may generate false seeds. At this time, a distance check is required for further judgment. The judgment basis is: only when no other seed points are detected within the average cell radius distance around the new seed point generated by the curvature analysis, it is identified as a cell. The approximate value of this cell radius is calculated from the average cell perimeter and standard deviation To calculate, the formula is:

[0087]

[0088] this It approximately represents the average radius of the cells and is used to filter out seed points that are too close, thus preventing the same cell from being detected repeatedly.

[0089] If the distance between two seed points is less than this threshold, they are considered to belong to the same cell, otherwise, they are considered to be different cells.

[0090] Through this operation process, good cell cluster segmentation can be achieved for cell images with uniform cell morphology and size, and non-cell areas can be effectively prevented from being misdetected as cells.

[0091] Step 8: Count cells and generate a cell localization map based on point annotation. It mainly counts the number of connected ROI areas, that is, the number of cells. The specific operation is: traverse the matrix corresponding to the image after cell cluster segmentation, mark and count different connected domains in the matrix, and use the variable Num to record the number of connected domain ROI areas. The total number of connected domains Num obtained at last is the number of cells in the image. The binary image finally segmented into single cells is transformed by distance to generate seed points, which is a point-based cell localization map. The cell localization map of grayscale cell image is as follows: Figure 8 As shown, the cell localization map of the fluorescent cell image is as follows Fig.10 shown.

[0092] like Fig.11 As shown, in this embodiment, an unsupervised cell counting and positioning system based on multimodal cell images is also provided, including:

[0093] An input module, used for inputting and reading multimodal cell images;

[0094] An image type determination module, used to determine the type of the input cell image;

[0095] An image differential preprocessing module is used to perform differential preprocessing on different types of input images;

[0096] An image segmentation module, used to segment the preprocessed cell image;

[0097] The cell morphology judgment module is used to judge the cell adhesion in the segmented image, that is, to distinguish single cells from cell clusters, and use the separated single cells to judge whether the cell size is uniform or uneven;

[0098] The cell cluster segmentation module is used to further segment the cell clusters and adopt specific segmentation methods for cell clusters with uniform and uneven cell morphology and size;

[0099] A cell counting and positioning module is used to count the number of cells in the above result graph and generate a point-based cell positioning graph;

[0100] The visualization output module is used to output segmentation results and visualized cell localization maps.

[0101] In summary, the method described in this application can quickly and accurately count and locate cells in multimodal cell images and has good generalization ability.

[0102] Compared with the prior art, the present invention divides cell images into two categories based on the characteristics of cell images of different modalities. For different types of cell images, the cell features are utilized and the corresponding cell cluster segmentation algorithm is selected to perform further cell cluster segmentation, so that the cell segmentation of multimodal cell images is more accurate and has better generalization performance. It can not only effectively reduce the burden of labeling work, but also achieve high-precision cell counting.

[0103] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. An unsupervised cell counting and positioning method based on multimodal cell images, characterized in that: The following steps are involved: Acquire multimodal cell images; Analyze cell images; Separate cell images of different modalities into single cell images and cell cluster images; Analyze and obtain single-cell features based on the single-cell map; Segment the cell cluster map based on single cell features; Perform cell counting and generate cell localization maps based on dot annotation.

2. The unsupervised cell counting and positioning method based on multimodal cell images according to claim 1, characterized in that: The specific steps of analyzing the cell image include: Determine the type of the cell image to determine whether it is a grayscale image. If it has color image features, it is classified as a non-grayscale image class; if it has grayscale image features, it is classified as a grayscale image class. Perform differential preprocessing on non-grayscale images and grayscale images respectively; Perform segmentation process to obtain binary image; The binary image was processed to separate the single cell image and the cell cluster image.

3. The unsupervised cell counting and positioning method based on multimodal cell images according to claim 2, characterized in that: The specific process of performing differential preprocessing on the non-grayscale image and the grayscale image respectively includes: Perform image enhancement operations on grayscale cell images; For non-grayscale cell images, convert them from RGB color space to HSV color space, select and save only the color of the cell nucleus, save it as a new image and perform grayscale processing.

4. The unsupervised cell counting and positioning method based on multimodal cell images according to claim 2, characterized in that: The specific process of performing the segmentation process to obtain a binary image includes: designing a sliding window on the cell image, setting the sliding window size and the moving step size to certain values, so that the image is divided into several small blocks, the size of the image block is related to the complexity of the image, the smaller the complexity of the image, the larger the image block can be, performing a grayscale clustering algorithm and adaptive threshold processing on each image block, and finally merging all image blocks to form a cell image segmentation result.

5. The unsupervised cell counting and positioning method based on multimodal cell images according to claim 2, characterized in that: The specific process of processing the binary image and separating the single cell image and the cell cluster image is as follows: using distance transformation to generate seed points for the foreground target area, performing seed point detection on the ROI area, if there is a connected ROI area with only one seed point, it proves that this is the area where the single cell is located, and all such areas are saved in one image; if there are multiple seed points simultaneously existing in a connected ROI area, it means that this area is a multi-cell adhesion area, that is, the area where the cell cluster is located, and all such areas are saved as new images; and finally two new binary images are generated, one is a binary image of all single cells, and the other is a cell image of all cell clusters.

6. The unsupervised cell counting and positioning method based on multimodal cell images according to claim 1, characterized in that: The specific process of analyzing and obtaining single-cell features according to the single-cell image is as follows: calculating the average and standard deviation of the perimeters of all single cells on the single-cell image, and determining the cell morphology category of the cells in the single-cell image, wherein the category determination includes two categories: cells with uniform cell size and cells with uneven cell size, wherein the specific process of calculating the average and standard deviation of the perimeters of all single cells on the single-cell image and determining the cell morphology category of the cells in the single-cell image is as follows: calculating the perimeter of each ROI area of ​​each newly generated single-cell binary image and calculating the perimeter standard deviation, and the cells with a larger standard deviation are classified as the category of uneven cell size, and the cells with a smaller standard deviation are classified as the category of uniform cell size.

7. The unsupervised cell counting and positioning method based on multimodal cell images according to claim 1, characterized in that: The specific process of segmenting the cell cluster image according to the single cell characteristics is as follows: using the watershed algorithm to further segment the cell clusters for the cell image with uneven cell size; performing curvature analysis on the cell image with uniform cell size, and combining the average cell perimeter and standard deviation calculated for the single cell image to assist in further segmentation, and finally obtaining the segmentation result.

8. The unsupervised cell counting and positioning method based on multimodal cell images according to claim 7, characterized in that: The specific process of performing curvature analysis on the cell image with uniform cell size and combining the average cell perimeter and standard deviation calculated on the single cell image to assist in further segmentation and finally obtaining the segmentation result is as follows: A set of candidate seed points approximate to the cell center is obtained through curvature analysis; Determine whether there are other seeds within the distance of the average cell radius around a seed point. If not, they are identified as one cell. If so, they are identified as belonging to the same cell. The average cell radius is determined by the average cell circumference. and standard deviation To calculate, the formula is: , avoiding the adhesion between cells in the cell cluster, which leads to the generation of false seeds in the contact area between adjacent cells and causes misidentification, thereby better solving the problem of cell cluster segmentation.

9. The unsupervised cell counting and positioning method based on multimodal cell images according to claim 1, characterized in that: The specific process of performing cell counting and generating a cell localization map based on point annotation is as follows: obtaining the final single cell morphology segmentation result, counting the number of ROI regions, i.e., the number of cells, generating seed points for the ROI region, and generating a result at the location of the positioning point in the center of the cell, i.e., a point-based cell localization map.

10. An unsupervised cell counting and positioning system based on multimodal cell images, applied to an unsupervised cell counting and positioning method based on multimodal cell images as claimed in any one of claims 1 to 9, characterized in that: include: Input module: receiving input multimodal cell images as raw image data; Image type judgment module: judges the type of all input images and distinguishes them into grayscale images and non-grayscale images; Image differentiation preprocessing module: De-noise and enhance contrast operations for grayscale images. For non-grayscale cell images, convert them from RGB color space to HSV color space, select only the color of the cell nucleus and save it as a new image, then grayscale it, and then perform image enhancement operations. The purpose of preprocessing is to correct the illumination of the image and increase the contrast of the cell, so as to maximize the optimal input state of the subsequent algorithm; Image segmentation module: Design a sliding window, perform grayscale clustering and adaptive threshold processing on each image block, merge all binarized image blocks, and obtain the segmentation result; Cell morphology judgment module: Use distance transformation to generate seed points for each connected ROI area, and divide all connected ROI areas into single cell images and cell cluster images according to the number of seed points in each connected area, so as to facilitate subsequent processing; Cell cluster segmentation module: Calculate the average perimeter and standard deviation of all connected ROI areas of the separated single cell map, classify the cell clusters with larger standard deviation as uneven cell size, and further segment them using the watershed algorithm; classify the cell clusters with smaller standard deviation as uniform cell size, and further segment them by combining curvature analysis, average perimeter size and standard deviation of single cells, and finally obtain the cell cluster segmentation result, then merge the single cell and the segmented cell cluster map to obtain the final segmentation result; Cell counting and positioning module: count cells in the binary image of single cells and generate a point-based cell positioning map; Visualization output module: Visualize the counting results and positioning map, and output the result map.

Images