A center point based cell nucleus segmentation method and related device

Abstract

The application discloses a kind of center point-based nuclear segmentation method and related equipment, the method includes: based on the pseudo-label generation algorithm of unsupervised clustering, nuclear image clustering is generated the instance pixel-level pseudo-label required when nuclear segmentation network model training;After obtaining the pseudo-label instantiation result of three classifications, the training of nuclear segmentation network model is carried out using the instance segmentation framework not dependent on non-boundary box;Nuclear center point detection network is trained, and the predicted nuclear center point is obtained by taking local maximum value and filtering operation;The segmentation result and detection result of the nucleus are fused and cut processing, and the corrected nuclear instance segmentation result is obtained.The present application processes pseudo-label into the classification graph of kernel-contour-background three classifications, so that the network can better focus on the contour of the nucleus, predict the nucleus center point, use the watershed algorithm to process the segmentation result, and improve the effect of adherent nuclear instantiation.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a cell nucleus segmentation method, system, terminal, and computer-readable storage medium based on center point. Background Technology

[0002] Using medical images, such as pathological images, for diagnosis is a frequently used diagnostic method in modern medicine. Under a microscope, professional pathologists can assess a patient's condition by observing or statistically analyzing the number, morphology, size, and distribution of cells or cell nuclei in pathological images. However, common pathological slides typically contain thousands of cells, making manual slide reading and diagnosis a very time-consuming and labor-intensive process. The related observation and statistical results are also easily influenced by human factors. In actual diagnostic scenarios, pathologists are prone to drowsiness due to prolonged microscopic reading and excessive attention expenditure, which can even lead to errors. Furthermore, due to the influence of reading experience and other factors, manual slide reading can easily introduce the doctor's personal subjectivity, making the diagnostic results less objective. Therefore, researching robust and efficient auxiliary statistical algorithms for cell-related information in pathological image slides based on technologies such as computer vision, and constructing a computer-aided diagnostic slide reading system, has significant practical implications.

[0003] Nuclei segmentation is a crucial component of the aforementioned computer-aided medical diagnostic systems. It can extract many useful cellular features from pathological images, making it invaluable for nuclear morphometry and computational pathology analysis.

[0004] Traditional cell nucleus segmentation algorithms typically require time-consuming manual feature extraction, often necessitating highly specialized domain knowledge, and often fail to achieve satisfactory results in highly complex and variable pathological images. In recent years, with the development of deep learning, automatic cell nucleus segmentation algorithms based on deep networks have been extensively researched and developed. However, to train a good cell nucleus segmentation network, various fully supervised segmentation algorithms usually require very fine pixel-level annotations, necessitating significant human and material resources from professional pathologists for meticulous manual annotation. Therefore, how to perform cell nucleus segmentation based on weaker-level annotations is of great research significance.

[0005] In addition, in common nuclear pathological images, there are often some adjacent nuclei sticking together, and the edges of the nuclei are often blurred, which poses a great challenge to the model for instance-level segmentation of the nuclei.

[0006] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0007] The main objective of this invention is to provide a center-point-based cell nucleus segmentation method, system, terminal, and computer-readable storage medium, aiming to solve the problem in the prior art that cell nucleus segmentation requires a large amount of manpower and resources, requiring professional pathologists to perform detailed manual annotation, and that adjacent cell nuclei are often stuck together.

[0008] To achieve the above objectives, the present invention provides a cell nucleus segmentation method based on a center point, the cell nucleus segmentation method based on a center point comprising the following steps:

[0009] A pseudo-label generation algorithm based on unsupervised clustering is used to cluster cell nucleus images and generate instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model.

[0010] After obtaining the pseudo-label instantiation results of the three classifications, the cell nucleus segmentation network model is trained using a non-boundary box-dependent instance segmentation framework.

[0011] The cell nucleus center point detection network is trained, and the predicted cell nucleus center points are obtained by taking local maxima and filtering operations;

[0012] The segmentation and detection results of the cell nuclei are fused and processed to obtain the corrected segmentation results of the cell nuclei instances.

[0013] Optionally, in the center-point-based cell nucleus segmentation method, wherein,

[0014] Furthermore, to achieve the above objectives, the present invention also provides a center-point-based cell nucleus segmentation system, wherein the center-point-based cell nucleus segmentation system comprises:

[0015] The pseudo-label generation module is used to cluster cell nucleus images based on an unsupervised clustering pseudo-label generation algorithm, generating instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model;

[0016] The cell nucleus segmentation module is used to train the cell nucleus segmentation network model using a non-boundary box-dependent instance segmentation framework after obtaining the pseudo-label instantiation results of the three classifications.

[0017] The cell nucleus center point detection module is used to train the cell nucleus center point detection network and obtain the predicted cell nucleus center point by taking the local maximum value and filtering operation;

[0018] The cell nucleus instantiation module is used to fuse and cut the cell nucleus segmentation results and detection results to obtain the corrected cell nucleus instance segmentation results.

[0019] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a center-point-based nucleus segmentation program stored in the memory and executable on the processor, wherein when the center-point-based nucleus segmentation program is executed by the processor, it implements the steps of the center-point-based nucleus segmentation method as described above.

[0020] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a center-point-based nucleus segmentation program, which, when executed by a processor, implements the steps of the center-point-based nucleus segmentation method as described above.

[0021] In this invention, a pseudo-label generation algorithm based on unsupervised clustering is used to cluster cell nucleus images, generating pixel-level pseudo-labels for instance instances required for training the cell nucleus segmentation network model. After obtaining the three-class pseudo-label instantiation results, a non-boundary-box-dependent instance segmentation framework is used to train the cell nucleus segmentation network model. A cell nucleus center point detection network is trained, and the predicted cell nucleus center points are obtained through local maxima extraction and filtering operations. The segmentation results and detection results of the cell nucleus are fused and segmented to obtain the corrected cell nucleus instance segmentation results. This invention processes the pseudo-labels into a three-class classification map of kernel-contour-background, allowing the network to better focus on the contour of the cell nucleus and distinguish different cell nucleus instances. An additional center point detection network is used to predict the cell nucleus center points, and a watershed algorithm is used to process the segmentation results, improving the instantiation effect of adherent cell nuclei. Attached Figure Description

[0022] Figure 1 This is a flowchart of a preferred embodiment of the cell nucleus segmentation method based on the center point of the present invention;

[0023] Figure 2 This is a schematic diagram of the overall framework of the cell nucleus segmentation algorithm based on center point supervision in a preferred embodiment of the cell nucleus segmentation method based on center point of the present invention.

[0024] Figure 3 This is a schematic diagram of the pseudo-label generation process based on unsupervised clustering in a preferred embodiment of the cell nucleus segmentation method based on center points of the present invention.

[0025] Figure 4 This is a flowchart of pseudo-labeled cell nucleus instantiation based on the watershed algorithm in a preferred embodiment of the cell nucleus segmentation method based on the center point of the present invention.

[0026] Figure 5This is a flowchart of the pseudo-label-based cell nucleus segmentation model training process in a preferred embodiment of the cell nucleus segmentation method based on center points of the present invention.

[0027] Figure 6 This is a schematic diagram of the cell nucleus center point detection model training in a preferred embodiment of the cell nucleus segmentation method based on center point of the present invention;

[0028] Figure 7 This is a schematic diagram of the cell nucleus center point detection model in a preferred embodiment of the cell nucleus segmentation method based on the center point of the present invention;

[0029] Figure 8 This is a schematic diagram illustrating the fusion process of segmentation results and detection results in a preferred embodiment of the cell nucleus segmentation method based on center point of the present invention;

[0030] Figure 9 This is a schematic diagram illustrating the principle of a preferred embodiment of the cell nucleus segmentation system based on the center point of the present invention;

[0031] Figure 10 This is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0033] For center-point supervised cell nucleus segmentation, the annotations provided during the training phase can only be instance-level center point annotations of the cell nucleus, not pixel-level annotations. However, during the testing phase, the model needs to be able to predict instance-level and pixel-level segmentation results of the cell nucleus. Figure 2 As shown, this method is based on an unsupervised clustering pseudo-label generation algorithm to generate instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model, and then uses the generated pseudo-labels to train the cell nucleus instance segmentation network.

[0034] For the input original image of cell nuclei, K-means clustering is used for unsupervised clustering of foreground and background. Clusters with more overlap with the cell nucleus center point annotations are then identified as cell nucleus foreground. The cell nucleus center point annotations are then used again to instantiate the cell nucleus clustering results. To enable the network model to better distinguish different cell instances, this method processes the pseudo-labels into a three-class classification image of kernel-contour-background, allowing the network to focus more on the outline of the cell nucleus and differentiate between different cell nucleus instances.

[0035] The cell nucleus segmentation network trained based on the aforementioned pseudo-labels still exhibits significant cell nucleus adhesion issues during prediction. To further separate these adhered nuclei and improve the final instance-level performance of cell nucleus segmentation, an additional centroid detection network is used to predict cell nucleus centroids. Then, the watershed algorithm is applied again to process the segmentation results, improving the instantiation effect of adhered cell nuclei.

[0036] The preferred embodiment of the cell nucleus segmentation method based on center point described in this invention, such as... Figure 1 As shown, the center-point-based cell nucleus segmentation method includes the following steps:

[0037] Step S10: The pseudo-label generation algorithm based on unsupervised clustering is used to cluster cell nucleus images to generate instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model.

[0038] Specifically, this invention provides a pseudo-label generation algorithm based on unsupervised clustering, used to generate instance pixel-level pseudo-labels required for training a cell nucleus segmentation network model. The specific algorithm flow is as follows: Figure 3 As shown. Figure 3 The visualization of the center point annotations in the text has been thickened. In the actual annotation, each center point only occupies one pixel. The same applies to the center point annotation visualizations in the following text.

[0039] First, for the input original image of cell nuclei, K-means clustering algorithm is used to perform unsupervised clustering of foreground and background. Clusters with the most overlapping areas with the cell nucleus center point annotations are then identified as the cell nucleus foreground (e.g., ...). Figure 3 In the top left corner of the image, there are two clusters, black and white. The cluster with more overlapping center points is the foreground, and vice versa. At this point, the cell nucleus clustering result only has pixel-level semantic information, not cell nucleus instance-level information (after clustering, white represents the cell nucleus foreground, but the number of cells in the foreground is unknown; instantiation refers to dividing the foreground into each distinct cell nucleus). Then, the cell nucleus center point annotations are used again to instantiate the cell nucleus clustering result. To enable the network model to better distinguish different cell instances, this invention processes the pseudo-labels into a three-class classification map of kernel-contour-background, allowing the network to better focus on the contour of the cell nucleus and ultimately better distinguish different cell nucleus instances.

[0040] This invention uses nucleus center point annotation to segment the results of nucleus foreground clustering, thereby segmenting the nucleus foreground clustering results, which only contain pixel-level semantic information, into individual nucleus instances. The specific process is as follows: Figure 4As shown. By instantiating the clustering results in this way, the instance-level information in the original cell nucleus center point annotations can be integrated into the processed pseudo-labels, and the generated pseudo-labels can also contain more coarse contour information about adjacent cell nuclei.

[0041] If the center point of the cell nucleus is labeled C H×W Where H and W are the height and width of the original image of the cell nucleus, respectively, and C... H×W element c ij (i=1,2,...,H;j=1,2,...,W) takes a value of 1 where there is a nucleus center point marker and a value of 0 where there is no nucleus center point marker.

[0042] First, mark the center point of the cell nucleus in diagram C. H×W Perform an inversion operation to obtain an inverted image of the center point annotations at the same resolution as the processed image:

[0043]

[0044] Among them, I H×W All elements in the middle are 1. In this case, pixels with a value of 1 are defined as foreground pixels, and pixels with a value of 0 are defined as background pixels.

[0045] Next, the center point labels are inverted to obtain the corresponding center points, where the calculation formula for each element is:

[0046]

[0047] in, for The elements are f(m,n), where f(m,n) is the Euclidean distance between two pixels.

[0048] The center point distance diagram D obtained above H×W In the model, the pixel values ​​in the vicinity of the cell nucleus center point are relatively low, while the pixel values ​​in the region farther away from the center points are relatively high. Therefore, the regions with higher pixel values ​​are more likely to be the boundaries of the cell nucleus, providing some coarse information about the cell nucleus boundaries for the watershed algorithm.

[0049] The binary image of the cell nucleus foreground obtained from the above clustering is F. H×W , will F H×W Distance from center point diagram D H×W By merging, we obtain the foreground-distance map:

[0050]

[0051] in, This indicates element-wise multiplication of matrices.

[0052] Next, the watershed algorithm is used to obtain the instantiated segmentation result W. H×W :

[0053] W H×W =w(D′) H×W C H×W );

[0054] Where w(·) represents the watershed algorithm.

[0055] Finally, each instance is subjected to hole-filling post-processing to obtain the final pseudo-tag instantiation result.

[0056] Step S20: After obtaining the pseudo-label instantiation results of the three classifications, the cell nucleus segmentation network model is trained using a non-boundary box-dependent instance segmentation framework.

[0057] Specifically, after obtaining the three-class pseudo-labels, the cell nucleus instance segmentation network can be trained. This invention employs a non-boundary-boundary-dependent instance segmentation framework, which does not rely on cell nucleus bounding boxes but directly performs cell nucleus instance segmentation from the perspective of pixel-level prediction.

[0058] First, the mapping between input cell nucleus images and three-class pseudo-labels is learned using the ResNet-50-based fully convolutional neural network architecture U-Net, such as... Figure 5 As shown. For each input original image of the cell nucleus, the output of the cell nucleus segmentation network model is a three-channel image with the same spatial resolution as the original image, where each channel corresponds to each category of the three-class pseudo-label, namely kernel, contour, and background. In training the above fully convolutional cell nucleus segmentation network, this invention uses cross-entropy loss.

[0059] In the prediction phase, the fully convolutional neural network trained in the above manner predicts the nucleus kernel, contour, and background in the nucleus test image, and retains all predicted nucleus kernels; the nucleus kernels are instantiated according to whether the regions are connected, and morphological dilation is performed on the instantiated nucleus kernels to obtain the nucleus instance-level segmentation results.

[0060] Step S30: Train the cell nucleus center point detection network and obtain the predicted cell nucleus center point by taking the local maximum value and filtering operation.

[0061] Specifically, this invention uses a regression approach to detect the nucleus center point. Since the information about the nucleus center point consists of relatively sparse coordinate values, it is very difficult to learn directly using a neural network. Therefore, these coordinate values ​​usually need to be preprocessed to facilitate network learning.

[0062] like Figure 6 As shown, firstly, a Gaussian convolution kernel is used to convolve with the center point of the cell nucleus to obtain the probability distribution map of the cell nucleus center point. Specifically, the cell nucleus center point is labeled in map C. H×W It can be represented by a δ function with respect to pixel x:

[0063]

[0064] Where Q(x) is a nucleus center point annotation map represented by the delta function of x, and x represents each pixel in the nucleus center point annotation map. i Let A represent the values ​​in the set of center points A, where A = {a1, a2, ..., a...} N} represents the set of N cell nucleus center point labels corresponding to the cell nucleus center point labeling map.

[0065] A Gaussian kernel G with a fixed spread parameter σ. σ Convolution with Q(x) yields the probability distribution of the cell nucleus center points:

[0066]

[0067] Here, * represents the convolution operation.

[0068] The probability distribution map obtained by this method is an image of dimension H×W, defined as U. H×W The probability distribution result map can be used as the learning target of the cell nucleus center point detection network.

[0069] If the original image input to the cell nucleus center point detection network is O H×W The predicted probability map of the center point distribution obtained after passing through the cell nucleus center point detection network Nθ(·) is M. H×W =N θ (O H×W ), where θ is a parameter in the cell nucleus center point detection network;

[0070] To constrain the learning of this cell nucleus center point detection network, a loss function based on Euclidean distance is used:

[0071]

[0072] Among them, u ij m represents the predicted value. ij This represents the actual value.

[0073] Regarding image feature extraction and network structure, this invention uses a ResNet-50-based U-Net to learn the mapping relationship between the input image and the probability distribution map of cell nucleus centers. After training the cell nucleus center detection network based on the above method, the predicted cell nucleus centers can be obtained in the detection phase by taking local maxima and filtering operations, as follows: Figure 7 As shown.

[0074] Specifically, if the probability map of the distribution of center points predicted by the cell nucleus center point detection network is E H×W For E H×W By taking the maximum local probability, a preliminary map of the cell nucleus center point location is obtained:

[0075] E′ H×W =l Δ (E) H×W );

[0076] Among them, l Δ This means that the position corresponding to the local maximum value is calculated based on the minimum pixel distance Δ between the set maximum values, while keeping the probability value of the local maximum value pixel unchanged and setting the probability value of the other pixels to 0.

[0077] Since the network prediction results may contain some noise, a threshold τ is set in the specific implementation process, and local maxima with probability values ​​below the threshold τ in the prediction results are filtered out to obtain the final nucleus center point location map: Local maxima with probability values ​​below the threshold τ can be used to obtain a preliminary nucleus center point location map:

[0078] E″ H×W =S τ (E′) H×W );

[0079] Among them, s τ (·) indicates that all pixel values ​​greater than τ are set to 1, while the values ​​of the remaining pixels are kept to 0.

[0080] Step S40: The segmentation results and detection results of the cell nuclei are fused and cut to obtain the corrected cell nucleus instance segmentation results.

[0081] Specifically, cell nucleus segmentation is essentially an instance segmentation task that requires special attention to different target instances. Based on this, post-processing operations are added to the segmentation and detection results to fuse and cut the segmentation and detection results, ultimately obtaining the corrected cell nucleus instance segmentation results.

[0082] This invention employs a watershed algorithm when instantiating adherent cell nuclei. The fusion process of cell nuclei segmentation results and detection results is as follows: Figure 8 As shown:

[0083] If the predicted probability maps for the three categories—background, kernel, and boundary—in the three channels output by the cell nucleus segmentation network are J, respectively... H×W K H×W and L H×W ; For J H×W K H×W and L H×W The constructed three-channel matrix is ​​subjected to Argmax operation along the channel dimension to obtain the category matrices for background, kernel, and boundary categories; the coordinates corresponding to the kernel category are taken, and the kernel category segmentation result image L is constructed. H×W In this kernel segmentation result, the pixel value corresponding to the core region of the cell nucleus is 1, and the pixel value corresponding to the remaining region is 0. However, some adhered cell nuclei still exist in this kernel segmentation result. Therefore, this invention utilizes the boundary prediction probability map L. H×W The boundary probability information in the graph is used to further divide the nuclei of adherent cells; specifically, the boundary prediction probability map L... H×W Compared with the binarized kernel segmentation result S H×W By fusing the results, we obtain the kernel-boundary probability map:

[0084]

[0085] The probability graph L′ after fusion H×W The data contains both segmentation information of the cell nucleus and probability information for predicting cell nucleus boundaries. Therefore, a watershed segmentation algorithm is applied for further segmentation. To further integrate instance-level information from cell nucleus center point detection, the cell nucleus detection result E″ is... H×W This serves as the seed for the watershed algorithm. Therefore, the instantiated cutting result after watershed cutting and processing can be obtained:

[0086] S″ H×W =w(S′) H×W ,E″ H×W );

[0087] The above instantiated segmentation results are filled with instance-level holes to obtain pseudo-labels for the three categories:

[0088] H H×W =h(S″) H×W );

[0089] Finally, the instance-level dilation operation d(·) is used to post-process the instantiated segmentation results to obtain the final cell nucleus instance segmentation results:

[0090] Z H×W =d(H H×W ).

[0091] This invention proposes an unsupervised algorithm to divide the foreground of the cell nucleus into different cell nucleus instances based on the cell nucleus center point annotation information, providing more cell nucleus instances and contour information for training the cell nucleus segmentation model; it also proposes a method for instantiating adherent cell nuclei by further utilizing the predicted cell nucleus center point information to process the segmentation results.

[0092] Furthermore, such as Figure 9 As shown, based on the above-described center-point-based nucleus segmentation method, the present invention also provides a center-point-based nucleus segmentation system, wherein the center-point-based nucleus segmentation system includes:

[0093] The pseudo-label generation module 51 is used to cluster cell nucleus images based on the unsupervised clustering pseudo-label generation algorithm to generate instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model.

[0094] The cell nucleus segmentation module 52 is used to train the cell nucleus segmentation network model using a non-boundary box-dependent instance segmentation framework after obtaining the pseudo-label instantiation results of the three classifications.

[0095] The cell nucleus center point detection module 53 is used to train the cell nucleus center point detection network and obtain the predicted cell nucleus center point by taking the local maximum value and filtering operation;

[0096] The cell nucleus instantiation module 54 is used to fuse and cut the cell nucleus segmentation results and detection results to obtain the corrected cell nucleus instance segmentation results.

[0097] Furthermore, such as Figure 10 As shown, based on the above-described center-point-based cell nucleus segmentation method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30. Figure 10 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0098] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a center-point-based nucleus segmentation program 40, which can be executed by the processor 10 to implement the center-point-based nucleus segmentation method of this application.

[0099] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the center-point-based cell nucleus segmentation method.

[0100] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0101] In one embodiment, the centroid-based nucleus segmentation step is implemented when the processor 10 executes the centroid-based nucleus segmentation program 40 in the memory 20.

[0102] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a center-point-based nucleus segmentation program, which, when executed by a processor, implements the steps of the center-point-based nucleus segmentation method as described above.

[0103] In summary, this invention provides a cell nucleus segmentation method and related equipment based on center points. The method includes: clustering cell nucleus images using a pseudo-label generation algorithm based on unsupervised clustering to generate instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model; after obtaining the three-class pseudo-label instantiation results, training the cell nucleus segmentation network model using a non-boundary box-dependent instance segmentation framework; training a cell nucleus center point detection network, obtaining predicted cell nucleus center points through local maxima extraction and filtering operations; and fusing and cutting the cell nucleus segmentation results and detection results to obtain corrected cell nucleus instance segmentation results. This invention processes pseudo-labels into a three-class classification map of kernel-contour-background, allowing the network to better focus on the contour of the cell nucleus and distinguish different cell nucleus instances. An additional center point detection network is used to predict cell nucleus center points, and a watershed algorithm is used to process the segmentation results, improving the instantiation effect of adherent cell nuclei.

[0104] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0105] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0106] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A cell nucleus segmentation method based on a center point, characterized in that, The center-point-based cell nucleus segmentation method includes: A pseudo-label generation algorithm based on unsupervised clustering is used to cluster cell nucleus images and generate instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model. After obtaining the pseudo-label instantiation results of the three classifications, the cell nucleus segmentation network model is trained using a non-boundary box-dependent instance segmentation framework. The cell nucleus center point detection network is trained, and the predicted cell nucleus center points are obtained by taking local maxima and filtering operations; The segmentation and detection results of the cell nucleus are fused and processed to obtain the corrected cell nucleus instance segmentation results; The pseudo-label generation algorithm based on unsupervised clustering clusters cell nucleus images to generate instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model. Specifically, it includes: For the input original image of cell nuclei, K-means clustering algorithm is used to perform unsupervised clustering of foreground and background, and the clusters with more overlapping areas with the cell nucleus center point are determined to be the cell nucleus foreground. The cell nucleus clustering results are instantiated using the cell nucleus center point annotation, and the instance pixel-level pseudo-labels are processed into a three-class classification map of kernel-contour-background; The process of instantiating the cell nucleus clustering results using cell nucleus center point annotation, and processing the instance pixel-level pseudo-labels into a three-class classification map of kernel-contour-background, specifically includes: If the cell nucleus center point is labeled as follows ,in, and These represent the height and width of the original image of the cell nucleus, respectively. elements in The value is 1 where the cell nucleus center point is marked, and 0 where the cell nucleus center point is not marked. Nucleus center point annotation diagram Perform an inversion operation to obtain an inverted image of the center point annotations at the same resolution as the processed image: ； in, All elements in the middle are 1. In this case, pixels with a value of 1 are defined as foreground pixels, and pixels with a value of 0 are defined as background pixels. Invert the center point labels to obtain the corresponding center points. The calculation formula for each element is: ； in, for elements, The Euclidean distance between two pixels; The clustered nucleus foreground binarized image is ,Will Distance from center point map By merging, we obtain the foreground-distance map: ； in, This indicates element-wise multiplication of matrices; The watershed algorithm is used to obtain the instantiated segmentation results. : ； in, This represents the watershed algorithm; After performing hole-filling post-processing on each instance, the final pseudo-tag instantiation result is obtained.

2. The cell nucleus segmentation method based on center point according to claim 1, characterized in that, Distance at center point In the image, the pixel values ​​of the cell nucleus center point, which is far from the center point of each annotation, are higher than the pixel values ​​of the surrounding area.

3. The cell nucleus segmentation method based on center point according to claim 1, characterized in that, The watershed algorithm is used to provide coarse information about cell nuclear boundaries.

4. The cell nucleus segmentation method based on center point according to claim 1, characterized in that, After obtaining the pseudo-label instantiation results for the three-class classification, the cell nucleus segmentation network model is trained using a non-boundary box-dependent instance segmentation framework, specifically including: The U-Net architecture, based on ResNet-50, is used to learn the mapping between input cell nucleus images and three-class pseudo-labels. For each input original image of the cell nucleus, the output of the cell nucleus segmentation network model is a three-channel image with the same spatial resolution as the original image of the cell nucleus, wherein each channel corresponds to each category of the three-class pseudo-label. In the prediction phase, the trained fully convolutional neural network is used to predict the nucleus kernel, contour and background in the nucleus test image, and all predicted nucleus kernels are retained. The cell nucleus kernel is instantiated based on whether the region is connected. Then, a morphological expansion operation is performed on the instantiated cell nucleus kernel to obtain the cell nucleus instance-level segmentation result.

5. The cell nucleus segmentation method based on center point according to claim 4, characterized in that, The cell nucleus segmentation network model was trained using cross-entropy loss.

6. The cell nucleus segmentation method based on center point according to claim 4, characterized in that, The training of the cell nucleus center point detection network, which obtains the predicted cell nucleus center points through local maxima extraction and filtering operations, specifically includes: Convolve the Gaussian convolution kernel with the center point of the cell nucleus to obtain the probability distribution map of the center point of the cell nucleus. Nucleus center point annotation diagram , using about the pixels of Represented by functions: ； in, use of A diagram of cell nucleus center point annotations in functional form. Each pixel in the labeling diagram represents the center point of the cell nucleus. This represents the values ​​in the set of center points A. This represents the set of N cell nucleus center point labels corresponding to the cell nucleus center point labeling map. Using fixed propagation parameters Gaussian kernel and Perform convolution to obtain the probability distribution of the cell nucleus center points: ； Where * represents the convolution operation; The resulting probability distribution graph is a graph with dimension 1. The image is defined as The probability distribution map is used as the learning target of the cell nucleus center point detection network; If the original image input to the cell nucleus center point detection network is After passing through the nucleus center point detection network The resulting center point distribution prediction probability map is as follows ,in, Parameters in the network for detecting the center point of the cell nucleus; The learning of the cell nucleus center point detection network is constrained using a loss function based on Euclidean distance: ； in, Indicates the predicted value. Indicates the actual value; The mapping relationship between the input image and the probability distribution map of cell nucleus center points is learned using a U-Net based on ResNet-50; After training the cell nucleus center point detection network, the predicted cell nucleus center point is obtained during the detection phase by taking the local maximum value and filtering operations.

7. The cell nucleus segmentation method based on center point according to claim 6, characterized in that, After training the cell nucleus center point detection network, during the detection phase, the predicted cell nucleus center point is obtained by taking local maxima and filtering operations, specifically including: If the probability distribution map of the nucleus center point prediction network is as follows: ,right By taking the maximum local probability, a preliminary map of the cell nucleus center point location is obtained: ； in, This indicates the minimum pixel distance between the set maximum values. To calculate the location corresponding to the local maximum value, keep the probability value of the local maximum pixel unchanged, and set the probability value of all other pixels to 0; Set a threshold And the probability values ​​in the prediction results that are lower than this threshold will be included. By filtering out local maxima, we obtain the final nuclear center point location map: From the local probability maximum, we can obtain a preliminary nuclear center point location map: ； in, This means all pixel values ​​greater than The values ​​corresponding to the pixels are all set to 1, while the values ​​of the other pixels remain 0.

8. The cell nucleus segmentation method based on center point according to claim 7, characterized in that, The process of fusing and segmenting the cell nucleus segmentation results and detection results to obtain the corrected cell nucleus instance segmentation results specifically includes: If the predicted probability maps for the background, kernel, and boundary categories in the three channels output by the cell nucleus segmentation network are respectively... , and ; right , and The constructed three-channel matrix is ​​subjected to Argmax operation along the channel dimension to obtain the category matrices for the background, kernel, and boundary categories. Extract the coordinates corresponding to the kernel category and construct the kernel category segmentation result image. The pixel value corresponding to the core region of the cell nucleus is 1, and the pixel value corresponding to the remaining region is 0. Using boundary prediction probability graphs The boundary probability information in the data is used to divide the nuclei of adherent cells; Boundary prediction probability map Compared with the binarized kernel segmentation results By fusing the results, we obtain the kernel-boundary probability map: ； Probability graph after fusion The data contains both segmentation information of the cell nucleus and probability information of the cell nucleus boundary prediction. The watershed segmentation algorithm is then applied for segmentation. Results of cell nucleus detection As the seed for the watershed algorithm, we obtain the instantiated cutting result after watershed cutting and processing: ； Instance-level hole filling is performed on the instantiated segmentation results to obtain pseudo-labels for the three categories: ； Apply instance-level dilation to the instantiated segmentation results. Post-processing is performed to obtain the final cell nucleus instance segmentation results: 。 9. A center-point-based cell nucleus segmentation system, characterized in that, The centroid-based nucleus segmentation system is used to implement the centroid-based nucleus segmentation method according to any one of claims 1-8, wherein the centroid-based nucleus segmentation system comprises: The pseudo-label generation module is used to cluster cell nucleus images based on an unsupervised clustering pseudo-label generation algorithm, generating instance pixel-level pseudo-labels required for training the cell nucleus segmentation network model; The cell nucleus segmentation module is used to train the cell nucleus segmentation network model using a non-boundary box-dependent instance segmentation framework after obtaining the pseudo-label instantiation results of the three classifications. The cell nucleus center point detection module is used to train the cell nucleus center point detection network and obtain the predicted cell nucleus center point by taking the local maximum value and filtering operation; The cell nucleus instantiation module is used to fuse and cut the cell nucleus segmentation results and detection results to obtain the corrected cell nucleus instance segmentation results.

10. A terminal, characterized in that, The terminal includes: a memory, a processor, and a centroid-based nucleus segmentation program stored in the memory and executable on the processor, wherein the centroid-based nucleus segmentation program, when executed by the processor, implements the steps of the centroid-based nucleus segmentation method as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a center-point-based nucleus segmentation program, which, when executed by a processor, implements the steps of the center-point-based nucleus segmentation method as described in any one of claims 1-8.

Images