Tissue removal microscope cell image quantitative analysis method under positioning guidance

By adopting a U-shaped network-based positioning and guiding segmentation model in cell image processing, combined with attention mechanism and feature fusion technology, the problem of insufficient accuracy of edge and severe adhesion cells in cell images is solved, and more accurate cell counting is achieved , improves the reliability of counting results.

CN119006480APending Publication Date: 2024-11-22XIAN UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410957999.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The prior art has insufficient accuracy in identifying image edge cells and severe adhesion cells in cell image localization and segmentation, resulting in inaccurate cell counting results, especially the unsatisfactory segmentation effect of adhesion cells, affecting the reliability of the counting results.

Method used

Using a nuclear localization network model based on U-shaped network structure, combined with convolutional neural network, attention mechanism and feature fusion technology, joint learning is carried out through positioning information-guided segmentation model to improve segmentation performance, especially in the treatment of severe adhesion cells. sex.

Benefits of technology

The segmentation model through the location-guided segmentation model significantly improves the accuracy and reliability of cell counts, and can effectively isolate adhesion cells, reduce errors, provide accurate and rigorous cell counting data, saving human and material resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006480A_ABST
    Figure CN119006480A_ABST
Patent Text Reader

Abstract

The invention discloses a tissue removal microscope cell image quantitative analysis method under positioning guidance, which enables a cell segmentation network to additionally learn positioning information in a training process, utilizes positioning information guidance and balanced segmentation training, and adopts a combined learning strategy to greatly improve the performance of a common segmentation model in segmentation of seriously adhered cells. And the reliability of the microscope cell image counting task is improved. The model combines technologies such as cellular morphology prior, a tissue removal principle, a convolutional neural network, an attention mechanism and feature fusion, and a good result is obtained in a microscope cell image technology task. Through combined learning of cell segmentation and positioning tasks, the defect that the edge cells of the microscope image cannot be positioned is overcome, accurate cell counting and boundary estimation are realized, refined quantitative analysis data are provided for large-scale microscope cell image counting work, and manpower and material resources are effectively saved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of computer vision and image processing, relates to cell image positioning and cell nucleus segmentation tasks, and specifically describes a method for quantitative analysis of tissue clearing microscope cell images under positioning guidance. Background Art

[0002] In traditional medical methods, doctors use artificial means to analyze the texture, size, and morphology of medical cells to obtain various medical parameters for disease diagnosis. With the continuous development of computer technology, artificial intelligence and computer vision technology have also made great progress. In modern medicine, the use of advanced computer image processing technology such as cell segmentation, cell recognition, cell counting, etc. for disease auxiliary diagnosis has been widely used in the field of medical imaging and has become an indispensable part of disease diagnosis.

[0003] When processing and studying cell images, microscopic cell images are a very important branch, which can help biomedical experts understand the basic principles of cell physiology and pathology, and are of great significance for clinical disease diagnosis. The stability of the cell state in human organs and tissues is an important indicator of physical health, and has important research significance in the biomedical field. The diagnosis of many diseases is related to the number and morphology of cells. For example, in cancer treatment, doctors need to test the patient's cells to determine the efficacy and dosage, so as to adjust the treatment plan. When the number of white blood cells in the body increases abnormally, it indicates that the body may suffer from a certain disease. As an important component of blood, the detection and analysis of red blood cells can judge the patient's physiological health status. When the number of red blood cells decreases, the oxygen carrying capacity of the blood will be reduced, and in severe cases, there is a risk of anemia; when the number of red blood cells increases, the blood will become viscous and less fluid, and there is a risk of vascular infarction. In addition, there are many blood diseases that are closely related to red blood cells, including polycythemia vera, sickle cell disease, aplastic anemia, red cell aplastic anemia and hemolysis. The stability of the number of bacteria in the oral environment is an important basis for oral health. When the balance of the number of bacteria is broken, oral diseases may occur, and these diseases in turn will further cause the imbalance of oral flora. In biological research and industrial fields, the detection of yeast cell concentration is very important. By counting yeast cells, its concentration can be accurately counted. In the field of cell biology, cell counting plays an auxiliary role in gene expression analysis. Compared with normal cells, the nucleus of cancerous cells will become larger. In addition, cancerous cells of different degrees also show different nucleus sizes. Cell counting is also an important way to evaluate cell growth and proliferation. By analyzing the number of cells, we can grasp the cell growth rate, cell cycle, cell division and other processes, so as to study the mechanism of cell growth and proliferation. Cell counting can also be used to study the toxicity and effect of drugs. By counting the effect of drugs on cells, we can understand the degree of influence of drugs on cells, so as to evaluate the toxicity and effect of drugs. Therefore, it is necessary to study microscopic cell images in clinical practice.

[0004] Cell counting is a rather complex problem, and requires relatively high counting accuracy. Currently, there are two main methods in clinical practice: manual counting and counting instruments. Manual counting requires researchers with high professional quality and a lot of practical experience to operate, but this method is huge in workload, inefficient, and easily affected by the subjective consciousness of the counting personnel and external objective factors. There is a problem that different doctors get different results, and the counting accuracy is poor. In addition, researchers with this ability are also scarce. Compared with manual counting, the use of instrument counting is efficient and has small errors, but it cannot accurately identify and distinguish cells that are adhered together, and it is easy to mistake adhered cells for one cell. Therefore, the use of automated digital image processing technology to analyze and process medical images has become a new and effective way in clinical medical research. With the rapid development of deep learning technology, convolutional neural networks have shown powerful performance in the field of image processing and have achieved impressive results. Cell segmentation is a very important research direction in the field of medical image processing, which has extremely important practical application value and has a very broad application prospect. Cell segmentation is one of the basic methods to achieve cell counting. It can complete this task quickly, accurately and efficiently, freeing doctors and workers from heavy and boring detection tasks. The accuracy of cell segmentation affects the counting results, so it is particularly important to design an advanced image segmentation network.

[0005] Problems existing in the current domestic technology:

[0006] The key to the research in the positioning task is to identify where the target is, so there is no need to learn too much about the shape information of the target, and a result map of the center of mass of the positioning target is obtained. In the microscope cell image, the center of the cell is intuitively the brightest part of the cell, which means that the pixel value here is relatively large. The principle of the positioning task is to predict the high value of the pixel in the central area of ​​the cell nucleus and generate a small marking box here, so that the adhesion cells can be well identified and located. The marking boxes in the positioning results are independent and non-overlapping. However, since the positioning results cannot accurately identify the cells at the edge of the image, and due to imaging reasons, some cells are brighter overall and have no obvious center of mass, for such cells, the positioning model is likely to generate multiple positioning results for one cell, so the positioning model cannot be directly applied to cell counting alone.

[0007] In the task of cell nucleus segmentation, it is found that due to the influence of adhesion cells, the performance of the network is required to be high. In the simpler case, two cells are adhered, and the case of three or more cells adhering is not common. From the perspective of pixels, the overlapping area between the adhesion cells has a relatively small pixel value, which visually looks darker and more similar to the background. Another point is that the side where the two cells are adhered presents a concave shape, so the usual segmentation result for the segmentation of the adhesion is only a small area in the center of the concave and cannot be separated. Designing a segmentation model with excellent performance can segment cells that are not very adherent, but for cases where adhesion is severe or even overlapping, the ideal effect cannot be obtained. Such shortcomings will have a greater impact on the counting results. Summary of the invention

[0008] The present invention provides a method for quantitative analysis of tissue clearing microscope cell images under positioning guidance, which is used to overcome the influence of multiple problems in cell images, construct a cell nucleus positioning network model based on a U-shaped network structure, so that the cell segmentation network can learn additional positioning information during the training process, and use the positioning information to guide and balance the segmentation training. The joint learning strategy is adopted to greatly improve the performance of the ordinary segmentation model in segmenting severely adhered cells, and increase the reliability of the cell counting task. The model combines convolutional neural networks, attention mechanisms, and feature fusion technologies for medical image processing tasks, and achieves good performance in cell segmentation tasks. Through the joint learning of cell segmentation and positioning tasks, accurate cell counting is achieved, providing accurate and rigorous data for cell counting work, and effectively saving resources such as manpower and material resources.

[0009] To achieve the above object, the technical solution of the present invention is:

[0010] A method for quantitative analysis of tissue clearing microscope cell images under positioning guidance comprises the following steps:

[0011] Step 1: Input a set of cell images, and preprocess each microscope cell image slice: crop the microscope cell images to obtain image blocks, and use different cropping methods for the training set and the test set. The specific implementation on the training set is to randomly select the center point to crop a uniform sub-image block of 572×572 pixels for network training. Before cropping, the number of sub-images to be extracted from each original image will be determined first. For example, it is determined that 10 image blocks are randomly selected from each original image. If there are ten original images in total, then a set of 100 local sample blocks can be obtained. In the test set, the method of sequentially cropping and extracting sub-images is adopted. First, the boundaries of the original image are expanded so that the size of the image is just an integer multiple of 10. This is because the step size when taking in sequence is 10, and the size of the sub-block is consistent with the sub-block in the training set. Then, the sub-block is extracted by sliding on the original image with the step size;

[0012] Step 2: Input the preprocessed image into the positioning model based on the U-shaped network architecture, train and test the model, and finally obtain the output features;

[0013] Step 3: The positioning information of the output features of the positioning model in step 2 represents the position of the cell centroid in the image. Then, this information and the original cell image are used as the input of the segmentation model. The segmentation model adopts a bilateral attention U-type network and receives two inputs, one is the original cell image data, and the other is the image positioning information output by the positioning model in step 2. The positioning-guided segmentation network adds a bilateral attention gating module in the jump connection. The other parts of the model are consistent with the U-type network. The segmentation model learns the positioning information while learning the original image information. The discrete degree of the output result of the segmentation model is improved by the discretized positioning information, so that the concave central area of ​​the adhesion of the adhesion cells is separated, the fine segmentation result is improved, and then the reliability of the counting result is improved;

[0014] Step 4: Finally, perform local maximum detection on the segmentation results output by the network, and proceed as follows:

[0015] a) Define a local area: First, you need to define a local area. Use a rectangle to define the local area. The neighborhood is centered on the current pixel and the size can be fixed or adjustable.

[0016] b) Search for maximum values ​​in local areas: For each local area, detect the maximum or peak point in it. This can usually be achieved by comparing the value in the local area with the values ​​of the surrounding neighborhood. If the value in the local area is greater than or equal to all the values ​​of the surrounding neighborhood, then the point is considered to be a local maximum point;

[0017] c) Filtering out non-maximum points: Usually, the detected local maximum points are further processed to filter out the points that do not meet the conditions;

[0018] d) Output result: The filtered local maximum point is output as the final detection result.

[0019] Compared with the existing cell counting method, the present invention has the following advantages:

[0020] 1) The discrete degree of the segmentation model output result is improved by discretizing the positioning information, so that the concave central area of ​​the adhesion of the adherent cells can be separated, thereby improving the reliability of the counting results.

[0021] 2) The strategy of joint learning based on positioning-guided information-guided segmentation overcomes the disadvantage of being unable to locate cells at the edge of the image. Its output is a cell nucleus segmentation image with a relatively high degree of discreteness. The advantage of this output is that when performing local maximum detection, it can eliminate the interference of some low-pixel information that may exist between severely adhered cells, thereby obtaining accurate counting results.

[0022] 3) After adding the positioning information, the output of the segmentation network is more discrete, and the previously unseparable adhesion cells can be successfully separated or made as discrete as possible while retaining their basic shape. Cell counting based on this will not produce large errors and accurate and reliable data can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a cell localization network model based on a U-shaped network in the present invention;

[0024] Figure 2 It is the result of cell nucleus localization in the present invention;

[0025] Figure 3 It is the overall cell counting model structure of the present invention;

[0026] Figure 4 is a diagram of the bilateral attention gating module;

[0027] Figure 5 The segmentation results of the positioning information model are added, the segmentation results of the positioning guidance information are added, and the counting results of the local maximum detection are added;

[0028] Figure 6 It is a partial visualized counting result in the present invention;

[0029] Figure 7 It is a graphic illustration of the bilinear interpolation method. Specific implementation methods

[0031] The present invention is further described below in conjunction with an embodiment, but the present invention is not limited to the following examples. A method for quantitative analysis of microscopic cell images of tissue clearing under positioning guidance, the specific steps are as follows:

[0032] Step 1: Use the localization model to locate the cell centroid:

[0033] 1.1. Input a set of microscope cell images, each image size is 572×572, and the number of channels is 1;

[0034] 1.2. After two double convolution (DoubleConv) modules, the output result with a size of 568×568 and 64 channels is obtained. The double convolution module consists of two "convolution + batch normalization + ReLU (Conv2d+BatchNorm2d+ReLU)";

[0035] 1.2.1. The two-dimensional convolution (Conv2d) operation performs a filtering operation on the input data to extract features.

[0036] 1.2.2 Batch normalization is a batch normalization operation that is used to accelerate the training process of neural networks and reduce the model's sensitivity to initial weights.

[0037] 1.2.3. ReLU is a rectified linear unit, which is used to introduce nonlinear characteristics to increase the expressiveness and fitting ability of the model.

[0038] 1.2.4. In summary, the role of "convolution + batch normalization + ReLU" is to perform batch normalization on the input data after convolution operation, and then pass the modified linear unit activation function to extract features and introduce nonlinearity, thereby enhancing the performance and generalization ability of the model;

[0039] 1.3. Then, after four downsampling modules, the result is 64×64 in size and 512 in number of channels. The downsampling module consists of a "MaxPool2d+DoubleConv";

[0040] 1.3.1. The maximum pooling operation is used to downsample the input data, reducing the size of the feature map while retaining the most significant features.

[0041] 1.3.2. Double convolution usually refers to a module consisting of two consecutive convolution operations, usually including batch normalization and ReLU activation function. Its function is to increase the depth and complexity of the model in the feature extraction process to enhance the representation ability of the features.

[0042] 1.3.3. In summary, the role of "max pooling + double convolution" is to enhance the feature extraction capability through double convolution operation while performing max pooling downsampling, and gradually reduce the size of the feature map. This is often used in the downsampling path (usually the encoder) in convolutional neural networks to reduce the amount of computation and extract abstract feature representations.

[0043] 1.4. Then, through 4 upsampling (Up) modules, an output result with a size of 388×388 and 64 channels is obtained. The upsampling module is composed of "up conv+copy crop+double convolution". The uplink module on the right side of the upsampling module involves copy cropping, which is slightly more complicated to implement. First, it goes through an upsampling or transposed convolution, and then the same size is cut from the feature map of the same layer on the left path (it is easy to see from the figure that the feature map in the same layer on the left is larger than the size on the right), merged with the feature map on the right, and finally double convolution is performed;

[0044] 1.4.1. Upsampling is completed by interpolation method, and no training parameters are required. The interpolation method is bilinear interpolation. Bilinear interpolation is a linear interpolation extension of the interpolation function with two variables. Its core idea is to perform linear interpolation in two directions respectively, such as Figure 7 As shown, the pixel value of each Q point is multiplied by its own weight, and then added to get the pixel value of the projected point P. The closer the distance between point Q and point P, the greater its weight.

[0045] 1.4.2. "Copy Crop" is used to connect the feature map obtained in the downsampling path (encoder) with the feature map obtained in the upsampling path (decoder) to retain more detail information.

[0046] 1.4.3. Double convolution operation is used to enhance the representation ability of features.

[0047] 1.4.4. In summary, the role of "up-sampling convolution + copy crop + double convolution" is to up-sample the features in the decoder and connect them with the features in the corresponding down-sampling path, and then enhance the representation ability of the features through double convolution operations. This combination is often used in the decoder part of tasks such as semantic segmentation to restore resolution and fuse features of different scales, thereby improving the accuracy and detail representation ability of the model.

[0048] 1.5. Finally, the output is completed after 1 output convolution module. The output result size is 88×388 and the number of channels is 2. Among them, the output convolution module is implemented using 1x1 convolution.

[0049] 1.6. The choice of loss function will have a great impact on the effect of algorithm fitting data.

[0050] Segmenting cell edges is a simple binary classification task, so we choose to use binary cross entropy loss with logits (BCEWithLogitsLoss). BCEWithLogitsLoss is a function provided by Pytorch to calculate binary cross entropy:

[0051] loss(o,t)=-1 / ni∑(t[i]*log(o[i])+(1-t[i])*log(1-o[i]))

[0052] Step 2: The network for localization-guided segmentation learning adopts a bilateral attention U-type network, which receives two inputs, one is the original cell image data, and the other is the image localization information output by the cell localization network. The localization network model is shown in Figure 1 As shown in Figure 2, by learning two different features, the network can output more discrete results without over-segmenting cells. Figure 4 As shown in the figure. Compared with the U-shaped network, the localization guided segmentation network only changes the jump connection and output. The bilateral attention gating module is added to the conventional jump connection to improve the fine segmentation results. At the same time, the output result is no longer just a single nuclear segmentation, but includes three different segmentation results of the cell nucleus, background and boundary. The rest of the model is consistent with the U-shaped network. Therefore, this section only introduces the bilateral attention gating module.

[0053] 2.1. The bilateral attention gating module obtains the corresponding attention coefficient by analyzing the input feature map and the gating signal. It not only readjusts the output features of the encoder, but also considers that the upsampled features can be weighted and readjusted to remove the interference of irrelevant pixels. The core idea is to select the target space area by analyzing the context information provided by the output feature map of the model. The bilateral attention gating mechanism can be used for each pixel X of the output feature map of the network layer L encoder. i l And each pixel g of the upsampled feature map i l Generate an attention coefficient α i ∈[0,1], multiplying the input feature map by the attention coefficient and outputting a semantically relevant and The specific relationship between its input and output is shown in formula (2.1) and formula (2.2).

[0054]

[0055] Where the linear attention coefficient of the bilateral attention gating module is By the input signal X l and gl Calculated. l and g l First, a 1×1 convolution operation is performed and then element-by-element addition is performed. The parameters in formula (2.1) include Bias and σ1 represents the RELU activation function.

[0056] 2.2. The result of the addition is processed by σ1 and then subjected to a 1×1 convolution operation. Corresponding to the sigmoid activation function, sigmoid is chosen instead of softmax in order to achieve better convergence during training, while the output of softmax is generally sparse.

[0057] 2.3 Final and X l and g l and α l The two feature maps output by the bilateral attention gate are then concatenated and convolved. In this way, the bilateral attention gate can locate the area to be paid attention to, rescale the output features of the encoder and upsampling, and only retain the feature responses that are useful for the segmentation task, improving the sensitivity and performance of the model without the need for complex heuristic algorithms and training multiple models and a large number of additional model parameter values.

[0058] Step 3: Local maximum detection:

[0059] 3.1 Define a local area: First, you need to define a local area. Use a rectangle to define the local area. The neighborhood is centered on the current pixel and the size can be fixed or adjustable.

[0060] 3.2 Searching for maximum values ​​in local regions: For each local region, detect the maximum or peak point in it. This can usually be achieved by comparing the value in the local region with the values ​​of the surrounding neighborhood. If the value in the local region is greater than or equal to all the values ​​of the surrounding neighborhood, then the point is considered to be a local maximum point.

[0061] 3.3 Filtering non-maximum points: Usually, the detected local maximum points are further processed to filter out the points that do not meet the conditions. For example, a threshold can be set to filter the local maximum, or post-processing operations such as non-maximum suppression can be performed on the detected peaks.

[0062] 3.4 Output results: The filtered local maximum point is output as the final detection result.

[0063] experiment

[0064] Experiment 1: Cell nuclear localization results

[0065] in Figure 2 (a) represents the results of cell nuclear localization, and the enlarged image of the local details in the green box in the localization result is Figure 2 (b) Display. From the positioning result image, it can be intuitively observed that most cells have been successfully and accurately positioned. However, by observing the local detail image, it can be found that the cells at the edge of the image have not been accurately identified and positioned. In addition, multiple positioning results appear on individual cell nuclei. In view of this, in order to obtain reliable counts, this paper does not choose to directly count the number of cells based on the positioning results, but uses the advantages of positioning to make up for the shortcomings of segmentation, uses positioning to guide cell segmentation, and finally achieves accurate cell counting. The overall cell counting model is as follows Figure 3 shown.

[0066] Experiment 2: Comparison of segmentation results

[0067] like Figure 5 As shown, Figure 5 Column (a) shows the segmentation results of the model without adding positioning information. Figure 5 Column (b) shows the network segmentation results guided by positioning information. The enlarged images of some local details in the red boxes in their images are shown on the right. It can be clearly seen from the segmentation results that compared with the results without adding positioning information, the output of the segmentation network is more discrete after adding positioning information, and the previously unseparable adhesion cells are successfully separated or made as discrete as possible while retaining their basic shape. Based on this, cell counting will not produce large errors, and accurate and reliable data can be obtained, which has important practical significance. However, the segmentation model without the guidance of positioning information cannot separate the concave center area of ​​the cell adhesion or make it more discrete. When this happens, the number of cells counted will have obvious errors, which cannot be used as a reference and has no practical research value.

[0068] Experiment 3: Comparison of cell counting results

[0069] In order to further verify the effectiveness of the design method in this chapter, the cell count results are visualized to illustrate the model effect from a more intuitive and clear perspective. Figure 6Some visual counting results are shown. The counting results are marked with red numbers on each cell in the image. It can be clearly seen that for some adherent cells that are not completely separated, they will not be mistaken for one cell when counting. This is due to the advantages of higher discretization and local maximum detection algorithm between them. The excellent performance of the method in this chapter is vividly illustrated by the visual results. The fundamental orientation of this study is cell counting. All the proposed method models are for obtaining an accurate counting result. Therefore, the accuracy of counting is the ultimate standard for measuring the performance of the proposed method. This section will statistically analyze the counting results of the method proposed in the paper and compare and analyze them with other network models. Due to the limited samples used for testing, the counting results of each model seem to be not much different from the true value, but it cannot objectively and accurately measure the reliability of the counting results. Therefore, in addition to statistical counting results, this paper will also use the error rate to objectively compare and analyze. The error rate is calculated by formula (1.1).

[0070] In summary:

[0071] This article uses the error rate to objectively conduct comparative analysis:

[0072]

[0073] Table 1 lists the counting results of several algorithms.

[0074] Table 1

[0075]

[0076] Technical ideas:

[0077] The technical idea of ​​the present invention is: a network model is used to locate cells, and the positioning information is used as the input of the segmentation network to guide the image segmentation task of adhesion cells. First, by adjusting the U-type network model, a convolution operation with a step size of 2 is used instead of the traditional pooling operation for downsampling, and residual learning is introduced in the convolution connection to reduce the risk of gradient disappearance, and multiple context information features are fused and output in the encoder part to capture information that is more suitable for positioning. Then, the positioning information obtained by the cell positioning network is used as the one-dimensional input of the segmentation network, so that the segmentation network can learn both positioning and segmentation information at the same time. Finally, the effectiveness of this method is verified by experiments, proving that the model can achieve more accurate cell counting tasks.

[0078] The invention discloses a method for quantitative analysis of tissue clearing microscope cell images under positioning guidance, which mainly comprises three stages: using a positioning model to locate the cell centroid position, using a segmentation model to segment the cells, and performing local maximum value detection on the segmentation results.

[0079] For the localization model, the original cell image is first used for model training and testing. The localization information output by the localization network represents the position of the cell centroid in the image. Then, this information and the original cell image are used together as the input of the segmentation model. The segmentation model learns the localization information while learning the original image information. The discrete localization information improves the discreteness of the segmentation model output result. In this way, the concave center area of ​​the adhesion of the adhesion cells is separated, thereby improving the reliability of the counting result. The localization model is based on the U-type network architecture and has made improvements on downsampling and convolution operations. The U-type network is a special fully convolutional neural network model with a symmetrical encoding and decoding branch structure, in which the encoder part extracts a large amount of original information of the image, and the decoder part extracts richer semantic information features. The clever jump connection design connects the corresponding coarse and fine features to form thicker features, reducing the semantic gap between features at different levels, thereby improving the segmentation performance of the network. It should be noted that the jump connection requires that the spatial dimension of the feature must be consistent, so the U-type network has strict restrictions on the size of the input image. The gradient of the deep network can be directly back-propagated to the shallow layer through the jump connection, thus reducing the risk of gradient vanishing during neural network training. In addition, this structure enables the network to capture multi-scale information of the image, retaining the context information of the original input data, and can supplement more detailed spatial information for more advanced semantic features. The specific structure diagram is as follows Figure 1 shown.

[0080] The segmentation model learns the positioning information while learning the original image information. The discrete positioning information is used to improve the discreteness of the segmentation model output results. In this way, the concave central area of ​​the adhesion of the adhesion cells is separated, thereby improving the reliability of the counting results. Figure 3 The network for localization-guided segmentation learning in the paper adopts a bilateral attention U-type network, which receives two inputs, one is the original cell image data, and the other is the image localization information output by the cell localization network. By learning two different feature information, the network can output more discrete results without over-segmenting the cells.

[0081] Bilateral Attention U-Network is a neural network model for image segmentation tasks. It is an improved version of the U-Network architecture, which improves the accuracy and detail retention of segmentation results by introducing a bilateral attention mechanism. In the traditional U-Network, the encoder part gradually downsamples the input image to a low-resolution feature map through convolution and pooling operations, while the decoder part restores the low-resolution feature map to a segmentation result with the same resolution as the input image through upsampling and convolution operations. This structure has certain limitations in processing detail information. Bilateral Attention U-Network solves this problem by introducing a bilateral attention mechanism. Specifically, it adds an attention module between the encoder and the decoder to adaptively adjust the importance of different positions in the feature map. This attention module consists of two branches: the spatial attention branch and the channel attention branch. The spatial attention branch captures the spatial correlation by calculating the similarity between each position in the input feature map and its surrounding pixels. The channel attention branch captures the correlation between channels by calculating the global average and maximum value of each channel in the input feature map. Finally, the outputs of the two branches are multiplied element-wise to obtain the final bilateral attention map, which is used to adjust the importance of different positions and channels in the feature map. By introducing the bilateral attention mechanism, the bilateral attention U-network can better retain detail information and improve the accuracy of segmentation results. It has achieved good results in medical image segmentation, natural image segmentation and other fields.

[0082] Local maximum detection is a commonly used technique in signal processing, image processing, and computer vision to find local maximum points or local peak points in signals or images. The basic idea is to find the maximum values ​​in local areas of signals or images and use these maximum points as potential points of interest.

Claims

1. A method for quantitative analysis of tissue clearing microscopic cell images under positioning guidance, The following steps are involved: Step 1: Input a set of microscope cell images, and preprocess each microscope cell image slice according to the tissue clearing principle, including estimating the size of the mean filter kernel according to the average spacing of manual rough positioning, and then performing uniform filtering of the adaptive cell nucleus scale, performing adaptive brightness stretching on the tissue clearing part and the cell nucleus part, and cropping the image to obtain the image block; Step 2: Combined with the manual rough positioning information, the cell nucleus positioning network of the microscope cell image is constructed, the preprocessed cell image is input into the positioning network, the positioning information feature map is output, and the maximum value analysis method PeakDetectionAnalysis is used to discretize the cell nucleus feature map, filter out the positioning points that do not conform to the cell nucleus morphology prior, and then locate the cell nucleus through the maximum value points in the feature map after the discretization operation, and output the binary positioning information logical value; Step 3: Construction of a cell nucleus boundary segmentation network guided by cell nucleus positioning information. The positioning information of the output feature of the positioning model in step 2 represents the position of the cell nucleus in the image. A composite data set is constructed by combining the cell nucleus positioning information and the microscope cell image. The segmentation model adopts a bilateral attention U-type network, which receives two inputs, one is the original microscope cell image data, and the other is the cell nucleus positioning information output by the positioning model in step 2. The positioning-guided segmentation network adds a bilateral attention gate (BAG) module to the jump connection. The segmentation model learns the positioning information while learning the original image information. The discrete degree of the segmentation model output result is improved by the discretized positioning information, so that the concave center area of ​​the adhesion of the adhesion cells is separated, thereby improving the reliability of the microscope cell image counting results. Step 4: Perform local maximum detection on the two-dimensional segmentation feature map, and use the detection result as the updated cell nucleus positioning information to realize the counting of the microscope cell image; perform threshold segmentation on the two-dimensional segmentation feature map to realize the boundary estimation of the microscope cell image. The specific steps are as follows: a) Define the local area: First, a rectangle is used to define the local area. The neighborhood is centered on the current pixel, and the side length of the rectangular area is set to 1.5 times the average spacing of the manual rough positioning; b) Searching for maximum values ​​in local areas: For each local area, the maximum point or peak point is detected by comparing the value in the local area with the values ​​of the surrounding neighborhood. If the value in the local area is greater than or equal to all the values ​​of the surrounding neighborhood, the point is considered to be a local maximum point; c) Filtering out non-maximum points: setting a threshold based on the average distance estimated by manual localization, thereby filtering out the cell nucleus localization results with a localization distance less than 0.8 times the average distance; d) Output result: The binary logical value of the cell nucleus is output as the final detection result, and the cell count is performed by calculating the sum of the logical values ​​in the area.

2. The method for quantitative analysis of microscopic cell images of tissue clearing under positioning guidance according to claim 1, It is characterized in that Different cropping methods are used for the training set and the test set. The specific implementation on the training set is to randomly select the center point to crop the sub-image block of uniform size of 572×572 pixels for network training. Before cropping, the number of sub-images to be extracted on each original image will be determined first. For example, 10 image blocks will be randomly selected on each original image. If there are ten original images in total, a group of 100 local sample blocks can be obtained. In the test set, the method of sequentially cropping and extracting sub-images is adopted. First, the boundary of the original image is expanded to make the image size an integer multiple of 10. This is because the step size when taking in sequence is 10, and the size of the sub-block is consistent with the training set sub-block, and then the sub-block is extracted by sliding on the original image according to the step size.

3. The method for quantitative analysis of microscopic cell images of tissue clearing under positioning guidance according to claim 1, It is characterized in that The maximum value analysis method is specifically divided into two operations: 1) Discretization operation Discretization refers to converting a continuous feature map into discrete values, usually by converting its pixel values ​​into discrete categories or binary labels according to certain rules. In maximum value analysis, this can be understood as converting each pixel point in the feature map into a discrete value in order to further analyze the characteristics of each discrete point. 2) Maximum value detection After the discretization operation, the maximum analysis method mainly focuses on finding the maximum points in the discretized feature map. These maximum points usually represent the location of the feature, such as the center of the cell nucleus or the boundary of the nucleus; The maximum value detection algorithm selects the threshold-based method. First, you need to select an appropriate threshold, which is determined according to the grayscale level of the image. For each pixel, check the values ​​of its surrounding pixels. If the current pixel value is greater than the value of the adjacent pixels and greater than the set threshold, it is marked as a maximum value.

4. The method for quantitative analysis of microscopic cell images of tissue clearing under positioning guidance according to claim 1, It is characterized in that The segmentation model network adopts a bilateral attention gating network. The positioning guided segmentation network adds a bilateral attention gating module in the jump connection. The network is described as follows: 1) The bilateral attention gating module can output feature maps for each pixel of the Lth layer encoder of the network And each pixel of the upsampled feature map Generate an attention coefficient α i ∈[0,1], multiplying the input feature map by the attention coefficient and outputting a semantically relevant and The specific relationship between its input and output is shown in formula (1) and formula (2); Where the linear attention coefficient of the bilateral attention gating module is By the input signal X l and g l Calculated, X l and g l First, a 1×1 convolution operation is performed and then element-by-element addition is performed. The parameters in formula (1) include Bias and σ 1 Represents the RELU activation function; 2) The result of the addition is passed through σ 1 After processing, a 1×1 convolution operation is performed again. The Corresponding to the sigmoid activation function, sigmoid is chosen instead of softmax to achieve better convergence during training. The output of softmax is generally sparse; 3) Final and X l and g l and α l The two feature maps output by the bilateral attention gate are cascaded and convolved. In this way, the bilateral attention can locate the area where attention is applied, readjust the output features of the encoder and upsampling, and only retain the feature responses that are useful for the segmentation task, thereby improving the performance of the model.