Cell detection method based on cell counting plate and target recognition model

Abstract

The invention relates to a cell detection method based on a cell counting plate and a target recognition model, and solves the technical problems of low accuracy and poor universality of the existing cell activity detection method based on cell image processing. And performing scanning optical imaging on the grids of the cell counting plate by using an optical microscope so as to obtain an image, inputting the image into the target identification model, and outputting the image reflecting the cell category and the cell number by the target identification model. The method can be widely applied to the fields of cell counting and activity detection, and a quicker, more objective and more reliable solution is provided for cell biology research and clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of medical engineering integration of cell biology, cell analysis, and image processing. Specifically, it relates to a cell detection method based on a cell counting chamber and a target recognition model. Background Art

[0002] In the fields of cell biology and clinical medicine, research on cell quantity and cell viability is of great significance. On the one hand, cell quantity and cell viability have always been the most basic requirements and evaluation criteria in the cell culture process. On the other hand, processes such as cell processing and analysis all require cell counting and viability detection.

[0003] A cell counting chamber is a commonly used cell counting tool. Through precisely controlled chamber depth and capillary action, it can enable cells to form a monolayer distribution in the microstructure of the cell counting chamber, facilitating cell counting and viability detection under a microscope. The cell counting chamber has a total of nine large squares, each large square is subdivided into 16 small squares, and the No. 5 square is subdivided into 25×16 small squares. However, currently, the commonly used method for cell counting is manual counting. According to the density of cells in the sample, the size of the selected square is determined. If the number of cells is small, then all the cells in a large square are counted; if the number of cells is moderate, then one of the 16 smaller squares is selected; if the cell density is high, then one of the 25 small squares in the center of the No. 5 square is counted; no matter which square is selected for counting, there should be at least 20 - 50 cells in each square for counting. The manual counting method has a large workload, low efficiency, and low accuracy. In addition, when the cell density is relatively high, it is unrealistic to traverse the nine large squares to calculate the accurate value of the cell quantity. Often, the cell quantity of one or four large squares at the four corners is counted, which reduces the counting accuracy.

[0004] In the prior art, viability detection requires pre - treating the cell sample with a cell dye and completing the viability detection of cells by means of the principle of selective permeability of the cell membrane. Most cell dyes are toxic, and the cells treated with cell dyes cannot be used for subsequent cell processing and analysis tasks.

[0005] Cell impedance analysis is a label-free cell viability detection method. Taking adherent cells as an example, the principle is to regard cells, cell culture medium, and cell electrodes as an organic whole, which is named the cell-electrode-culture medium system. By applying a weak alternating voltage excitation signal to this system, and then collecting the returned current signal to calculate the impedance parameters corresponding to the cells, and further determining the impedance analysis of the cells, finally, the detection and quantitative analysis of cell viability can be achieved. However, this method requires an expensive impedance analyzer, and there is a risk of electro-lysis of cells due to the electrical signal, which poses a certain threat to cell viability. The impedance signal of cells is at the nano-volt (nV) level, and complex signal processing and extraction are required.

[0006] With the continuous development of computer technology and digital image processing technology, computer-aided diagnosis (CAD) systems should be applied to the analysis of cell biology. The research method of using a computer to count and detect the viability of cell images is becoming increasingly perfect. Referring to the invention patent applications with the publication numbers of CN116660126A and CN118334105A, the existing solutions for cell counting through image processing generally require the combined processing of multiple image processing algorithms for the image, and many algorithms require manual adjustment of parameters to adapt to different cells, with poor versatility; in addition, the accuracy of detection needs to be improved. Summary of the Invention

[0007] This application aims to solve the technical problems that the existing technology uses a cell counting chamber for manual cell counting, which has a large workload, low efficiency, and low accuracy, and the existing technology for cell viability detection based on cell image processing has low accuracy and poor versatility, and provides a cell detection method based on a cell counting chamber and a target recognition model.

[0008] The cell viability detection method of the present disclosure is based on cell morphology. The principle of cell morphology is that cells will exhibit different morphological characteristics under different physiological states (such as healthy, apoptotic, necrotic, etc.). Healthy cells: usually have regular morphology, intact cell membranes, clear cell nuclei, and uniform cytoplasm. Apoptotic cells: exhibit characteristics such as cell shrinkage, nuclear chromatin condensation, nuclear fragmentation, and cell membrane blebbing (formation of apoptotic bodies). Necrotic cells: have ruptured cell membranes, swollen cytoplasm, and dissolved or fragmented cell nuclei, often accompanied by an inflammatory response. Therefore, quantitative analysis of cell morphology through image processing methods can well distinguish cell viability.

[0009] In the first aspect of the present disclosure, a cell detection method based on a cell counting chamber and a target recognition model is provided, including the following steps:

[0010] The first step is to preprocess the optical microscope;

[0011] Step 2: Construct a cell category dataset for training the target recognition model;

[0012] Step 3: Train the target recognition model with the cell category dataset;

[0013] Step 4: Dilute the original cell suspension to ensure that the concentration of the diluted cell suspension meets the requirements of the cell counting chamber; Disperse the diluted cell suspension, and use a micropipette to aspirate a certain amount of the diluted cell suspension sample and add it to the cell counting chamber;

[0014] Step 5: Let the cell counting chamber stand for a certain period of time to make the cells distribute in a single layer in the squares of the cell counting chamber;

[0015] Step 6: Place the cell counting chamber on the stage of an optical microscope;

[0016] Step 7: Use the optical microscope to scan and optically image the squares of the cell counting chamber to obtain an image;

[0017] Step 8: Input the image into the target recognition model, and the target recognition model outputs an image reflecting the cell category, and the result output by the target recognition model includes the number of cells.

[0018] Preferably, in Step 7, use the optical microscope to separately scan and optically image multiple squares of the cell counting chamber to obtain i images;

[0019] Step 8: Input the i images into the target recognition model in sequence, and the target recognition model outputs i images reflecting the cell category.

[0020] Further preferably, after Step 8:

[0021] Step 9: Calculate the concentration of the original cell suspension according to the cell number data output by the target recognition model.

[0022] Further preferably, in Step 5, make the cells distribute in a single layer in the large squares of the cell counting chamber;

[0023] In Step 7, use the optical microscope to separately scan and optically image multiple large squares of the cell counting chamber;

[0024] The concentration C of the original cell suspension is obtained through the following formula:

[0025] C = X × 10 7 × P

[0026] In the above formula, P represents the dilution factor of the original cell suspension after dilution treatment, and X represents the average value calculated after summing the number of cells in the i images output by the target recognition model.

[0027] Preferably, the target recognition model is a convolutional neural network target detection model.

[0028] Preferably, the second step is implemented through the following process:

[0029] Step 1: Use a culture plate for cell culture and set up live and dead cell control groups.

[0030] Step 2: Use an optical microscope to perform optical imaging on the live and dead cell groups respectively to obtain images of live and dead cells.

[0031] Step 3: Manually label the images of live and dead cells respectively. The labels are divided into two categories, namely live cells and dead cells.

[0032] Step 4: Construct a data set from the labeled images.

[0033] Further preferably, for Step 1, perform trypan blue staining on the live cell group, and the cells that are not stained during imaging are regarded as live cells.

[0034] Further preferably, before constructing the data set from the labeled images in Step 4, perform illumination correction on the obtained images of live and dead cells, and then perform gray-scale processing on the images of live and dead cells after illumination correction.

[0035] In the second aspect of the present disclosure, a cell detection method based on a cell counting plate and a target recognition model is further provided, including the following steps:

[0036] The first step: Preprocess the optical microscope.

[0037] The second step: Construct a cell category data set for training the target recognition model.

[0038] The third step: Train the target recognition model with the cell category data set.

[0039] The fourth step: Dilute the original cell suspension to ensure that the concentration of the diluted cell suspension meets the requirements of the cell counting plate; disperse the diluted cell suspension, and use a micropipette to aspirate a certain amount of the diluted cell suspension sample and add it to the cell counting plate.

[0040] The fifth step: Let the cell counting plate stand for a certain period of time to make the cells distribute in a single layer in the squares of the cell counting plate.

[0041] The sixth step: Place the cell counting plate on the stage of the optical microscope.

[0042] The seventh step: Use the optical microscope to perform optical imaging by scanning the squares of the cell counting plate to obtain an image.

[0043] In the eighth step, the image is input into the target recognition model, and the target recognition model outputs an image reflecting the cell type.

[0044] The beneficial effect of the present disclosure is that it directly recognizes the image obtained by the microscope without the need to preprocess the cells by traditional methods, improving the detection efficiency and accuracy. The convolutional neural network is fast and applicable, and can improve the detection efficiency of the final cell viability detection.

[0045] It can adapt to different cells and has strong versatility. The cell counting is fast, efficient, and accurate. The calculation of the concentration is fast, efficient, and accurate.

[0046] The cell counting plate can provide a monolayer cell sample with uniform distribution, ensuring clear imaging and easy recognition of morphological features, thereby improving the detection accuracy.

[0047] A reliable label-free cell viability detection method is realized. Using the convolutional neural network model is beneficial to improve the precision and accuracy. It can judge the cell viability without relying on the traditional fluorescence staining method, such as automatically judging whether the cells are alive or dead.

[0048] The further features and aspects of the present invention will be clearly recorded in the following description of the specific embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is the basic flowchart of the cell detection method;

[0050] Figure 2 is the flowchart for establishing the target recognition model;

[0051] Figure 3 is the structural schematic diagram of the cell counting plate;

[0052] Figure 4 is the automatic cell counting and classification operation process;

[0053] Figure 5 is the image after manually annotating the cell image, marking the live cells;

[0054] Figure 6 is the image after manually annotating the cell image, marking the dead cells;

[0055] Figure 7 is the output result image of the convolutional neural network target detection model, identifying live cells and dead cells;

[0056] Figure 8 is the number of live cells and dead cells included in the output result of the convolutional neural network target detection model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] The specific embodiments described below are only preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. For those skilled in the art, based on or according to the principles, concepts, and spirits of the present application, some changes or variations can be made, and the technical solutions formed by these changes or variations should all be covered within the protection scope of the present application.

[0059] Reference Figure 1 and 4 , the cell concentration detection method mainly includes the following steps:

[0060] The first step is to preprocess the optical microscope.

[0061] Perform a white balance operation on the optical microscope and record all the parameters after white balance as the basic values. Performing the white balance operation is a key operation to ensure accurate image color. Its main purpose is to eliminate the influence of the light source color temperature deviation on imaging and make the sample color closer to the real observation effect.

[0062] The second step is to construct a cell category dataset for training a convolutional neural network object detection model. Reference Figure 2 .

[0063] Step 1: Use a culture plate for cell culture, set up live and dead cell control groups. The live cell group is cultured according to the normal process, and the dead cell group is inactivated before imaging;

[0064] Step 2: To remove dead cells in the live cell group, perform trypan blue staining on the live cell group. Cells that are not stained during imaging can be regarded as live cells, and stained cells will not be used in subsequent experiments;

[0065] Step 3: Use an optical microscope to perform optical imaging on the live and dead cell groups respectively to obtain images of live and dead cells;

[0066] Step 4: Perform illumination correction on the obtained images of live and dead cells to reduce errors caused by abnormal intensity.

[0067] Step 5: Perform grayscale processing on the images of live and dead cells after illumination correction.

[0068] Step 6: Manually label the images of live and dead cells respectively. The labels are divided into two categories, namely live cells and dead cells; the labeled images are as shown in Figure 5 and Figure 6 . Figure 5 is an image of labeled live cells, Figure 6 is an image of labeled dead cells.

[0069] Step 7: Construct a dataset from the labeled images and divide the live cell and dead cell groups into training sets and test sets at a fixed ratio.

[0070] The third step is to train a convolutional neural network object detection model using the cell category dataset. The specific process can be as follows:

[0071] Input the cell category dataset into the convolutional neural network object detection model for iterative training until the loss function is less than a given threshold, and then end the training. Select the best weight parameters during the training process as the weights for testing.

[0072] Input the manually labeled cell images into the convolutional neural network, and use a 3×3 convolutional kernel to perform convolution operations on the labeled images. The input image is a labeled cell microscopic image, and its annotation information includes cell categories (such as alive, dead) and bounding box coordinates. Before the convolution operation, the original image is normalized to eliminate the interference of illumination differences on feature extraction. The 3×3 convolutional kernel traverses the image in the form of a sliding window, calculates the dot product of the pixel values in the local area and the weights of the convolutional kernel, and initially extracts basic features such as the edges of the cell nucleus and the texture of the cell membrane.

[0073] To match the feature scale size after convolution, it is often necessary to perform padding processing on the boundaries of the input feature map. Usually, 0 elements are added to the boundaries that need to be padded. Since the convolutional kernel cannot completely cover the pixel area at the image edge, to avoid the loss of edge cell information caused by the reduction of the feature map size, a symmetric padding strategy is adopted, and zeros are filled around the input feature map. For example, for an input image of 512×512 pixels, the size after padding is extended to 514×514 to ensure that the output feature map after processing by the 3×3 convolutional kernel still maintains the original resolution of 512×512.

[0074] Through the parallel processing of 64 convolutional kernels, a 64-channel feature map is output, and the spatial resolution is retained. 64 independent 3×3 convolutional kernels are deployed in the convolutional layer of the output feature map, and each convolutional kernel generates a feature map of one channel. These feature maps are stacked in the channel dimension to form a 64-channel composite feature representation.

[0075] Input the output convolutional feature map into the residual layer for processing, enter the first-level 3×3 convolutional kernel processing with a stride of 1 to extract high-order feature information. The first-level convolution of the residual layer further analyzes the basic features, and through the local perception ability of the 3×3 convolutional kernel, combines low-order features such as the edges of the cell nucleus into high-order features.

[0076] Enter the second-level 3×3 convolutional kernel to further refine the features. The second-level convolution optimizes the high-order features at the same resolution and strengthens the expression ability of key features through the weight sharing mechanism.

[0077] The input feature map is directly added element-wise to the output of the main path to preserve the original feature information (such as cell boundary localization). The fused feature map contains local details and global semantic information for subsequent classification. The skip connection adds the feature map input to the residual layer element-wise to the output feature map after two-level convolution. This operation enables the network to utilize both the precise localization information in the shallow layer and the semantic abstraction features in the deep layer simultaneously.

[0078] The residual feature map is input into the max pooling layer for max pooling processing, taking the maximum value within the window to suppress noise and retain significant features.

[0079] The convolution-residual-pooling modules are repeatedly stacked, and finally the cell class probabilities (such as live / dead cells) are output through the fully connected layer. The stacking of the above modules constructs a deep network. After each level of pooling, the resolution of the feature map gradually decreases, while the number of channels increases exponentially to enrich the feature diversity. For example, the second-level module may use a convolutional kernel with 128 channels, and the third level is extended to 256 channels. Finally, the fully connected layer at the end of the network maps the high-dimensional features to the class space and outputs the normalized probabilities through the Softmax function.

[0080] In the fourth step, the original cell suspension is diluted to ensure that the concentration of the diluted cell suspension meets the requirements of the cell counting chamber, and the dilution factor P is recorded; the dilution factor may be 100 or 1000. The diluted cell suspension is dispersed to make the cells as dispersed as possible into single cells; 10 μL of the diluted cell suspension sample is aspirated using a micropipette and added to the loading port of the cell counting chamber, and the 10 μL of the diluted cell suspension sample enters the cell counting chamber and fills the entire grid.

[0081] In the fifth step, the cell counting chamber is left standing for a certain period of time to allow the cells to sink and stabilize within the grid, ensuring that the cells are distributed in a single layer in the squares of the cell counting chamber.

[0082] In the sixth step, the cell counting chamber is placed on the stage of an optical microscope; the cells are observed through the optical microscope. If the number of cells in each large square of the cell counting chamber is less than 5, the diluted cell suspension needs to be redispersed (to reduce the volume of the diluted cell suspension), and the cell suspension sample in the cell counting chamber is disposed of (not recyclable, not reusable, to prevent sample contamination), and a new sample is taken from the redispersed cell suspension using a micropipette and added to the cell counting chamber. If the cells overlap with each other or are difficult to distinguish due to high density, the original cell suspension needs to be diluted to a larger volume or it is necessary to check if cell clumping has occurred (if cell clumping has occurred, the cell suspension needs to be redispersed again) until there is no cell overlap; the dilution factor is recorded.

[0083] Step 7: Use an optical microscope to scan and optically image each of the nine large squares on the cell counting plate as a separate field of view area, number the 9 obtained images, and use i to represent the number, where i = 1, 2, 3... 9.

[0084] Step 8: Input the nine images into the convolutional neural network object detection model in sequence. The convolutional neural network object detection model outputs the detection results, which are images reflecting the cell categories (such as live cells, dead cells). The convolutional neural network object detection model outputs images reflecting the cell categories (such as live cells, dead cells) at the single-cell level, and there are a total of 9 images. As Figure 7 shown, in the recognition results, the red bounding box with the word "live" indicates live cells, and the pink bounding box with the word "dead" indicates dead cells.

[0085] The images output by the convolutional neural network object detection model can automatically determine whether the cells are live or dead cells and identify the cell categories.

[0086] The output results of the convolutional neural network object detection model include the number of live cells and the number of dead cells. For example, Figure 8 as shown, the number of live cells is 38, and the number of dead cells is 14.

[0087] According to the results output by the convolutional neural network object detection model, the number of live cells is L i , where i takes values of 1, 2, 3, 4, 5, 6, 7, 8, 9.

[0088] According to the results output by the convolutional neural network object detection model, the number of dead cells is D i , where i takes values of 1, 2, 3, 4, 5, 6, 7, 8, 9.

[0089] It can be seen that the convolutional neural network object detection model outputs the cell count, completing the cell counting task quickly and accurately. When the cell density is high, it is still possible to traverse the nine large squares to obtain the number of cells in each large square.

[0090] Step 9: Calculate the concentration of the original cell suspension according to the cell count output by the convolutional neural network object detection model.

[0091] As Figure 3 shown, the size specification of one large square on the cell counting plate is 1mm * 1mm * 0.1mm, with a total of 0.1 ul. Therefore, after obtaining the number of cells in one large square, the concentration C of the original cell suspension can be calculated according to the following formula.

[0092] C = X × 10 7 × P

[0093] In the above formula, C represents the concentration of the original cell suspension (the concentration before dilution), P represents the dilution factor, and X represents the average number of cells in i large squares.

[0094] When summing up the number of live cells in 9 large squares and then taking the average, X represents the average number of live cells in 9 large squares, and the finally calculated concentration C of the original cell suspension is the live cell concentration.

[0095] When summing up the number of dead cells in 9 large squares and then taking the average, X represents the average number of dead cells in 9 large squares, and the finally calculated concentration C of the original cell suspension is the dead cell concentration.

[0096] When summing up the number of all cells in 9 large squares and then taking the average, X represents the average number of all cells in 9 large squares, and the finally calculated concentration C of the original cell suspension is the total cell concentration.

[0097] It can be seen that when the cell density is high, it is still possible to traverse nine large squares to find the average value and calculate the concentration quickly and accurately.

[0098] It can be seen that the above method can achieve fast and accurate cell counting tasks, can realize viability analysis without staining the cells, greatly reduces the steps and time of cell experiments, has high practical value and promotion prospects, and can be widely applied to experimental research in the biological field.

[0099] It should be noted that taking 9 large squares of the cell counting plate as the basis is just an example. Several medium squares can also be selected as the basis, or several small squares can be selected as the basis.

[0100] It should be noted that it is not limited to using the convolutional neural network object detection model, and other object detection models can also be used, such as the Yolov5 model.

Claims

1. A cell detection method based on a cell counting chamber and a target recognition model, characterized in that It includes the following steps: First step, preprocess the optical microscope; Second step, construct a cell category dataset for training the target recognition model; Third step, train the target recognition model with the cell category dataset; Fourth step, dilute the original cell suspension to ensure that the concentration of the diluted cell suspension meets the requirements of the cell counting chamber; disperse the diluted cell suspension, and use a micropipette to aspirate a certain amount of the diluted cell suspension sample and add it to the cell counting chamber; Fifth step, let the cell counting chamber stand for a certain period of time so that the cells are distributed in a single layer in the squares of the cell counting chamber; Sixth step, place the cell counting chamber on the stage of the optical microscope; Seventh step, use the optical microscope to scan and optically image the squares of the cell counting chamber to obtain an image; Eighth step, input the image into the target recognition model, and the target recognition model outputs an image reflecting the cell category, and the result output by the target recognition model includes the number of cells.

2. The cell detection method based on a cell counting chamber and a target recognition model according to claim 1, wherein: In the seventh step, use the optical microscope to scan and optically image multiple squares of the cell counting chamber respectively to obtain i images; In the eighth step, input the i images into the target recognition model in sequence respectively, and the target recognition model outputs i images reflecting the cell category.

3. The cell detection method based on a cell counting chamber and a target recognition model according to claim 2, wherein After the eighth step: Ninth step, calculate the concentration of the original cell suspension according to the cell number data output by the target recognition model.

4. The cell detection method based on a cell counting chamber and a target recognition model according to claim 3, wherein: In the fifth step, make the cells be distributed in a single layer in the large squares of the cell counting chamber; In the seventh step, use the optical microscope to scan and optically image multiple large squares of the cell counting chamber respectively; The concentration C of the original cell suspension is obtained by the following formula: C = X × 10 7 × P In the above formula, P represents the dilution multiple of the original cell suspension after dilution treatment, and X represents the average value calculated after summing the number of cells in the i images output by the target recognition model.

5. The cell detection method based on a cell counting chamber and a target recognition model according to claim 1, wherein The target recognition model is a convolutional neural network target detection model.

6. The cell detection method based on a cell counting chamber and a target recognition model according to claim 1, wherein: The second step is implemented through the following process: Step 1, use a culture plate for cell culture, and set up live and dead cell control groups; Step 2, use the optical microscope to perform optical imaging on the live cell and dead cell groups respectively to obtain images of live cells and dead cells; Step 3, perform manual label annotation on the live cell and dead cell images respectively, and the labels are divided into two categories, namely live cells and dead cells; Step 4, construct a dataset with the labeled images.

7. The cell detection method based on a cell counting chamber and a target recognition model according to claim 6, wherein Regarding step 1, perform trypan blue staining on the live cell group, and the cells not stained during imaging are regarded as live cells.

8. The cell detection method based on a cell counting plate and a target recognition model according to claim 7, wherein Before constructing the dataset with the labeled images in step 4, perform light correction on the obtained images of live cells and dead cells, and then perform grayscale processing on the live cell and dead cell images after light correction.

9. A cell detection method based on a cell counting chamber and a target recognition model, characterized in that, It includes the following steps: First step, preprocess the optical microscope; Second step, construct a cell category dataset for training the target recognition model; Third step, train the target recognition model with the cell category dataset; Fourth step, dilute the original cell suspension to ensure that the concentration of the diluted cell suspension meets the requirements of the cell counting chamber; disperse the diluted cell suspension, and use a micropipette to aspirate a certain amount of the diluted cell suspension sample and add it to the cell counting chamber; Fifth step, let the cell counting chamber stand for a certain period of time to make the cells distribute in a single layer in the squares of the cell counting chamber; Sixth step, place the cell counting chamber on the stage of the optical microscope; Seventh step, use the optical microscope to scan and optically image the squares of the cell counting chamber to obtain an image; Eighth step, input the image into the target recognition model, and the target recognition model outputs an image reflecting the cell category.

Images