Cell image recognition method and device, electronic equipment, medium and program product

Abstract

The application discloses a cell image recognition method and device, electronic equipment, medium and program product. The method comprises the following steps: inputting a cell original image to be recognized into a first neural network model for global recognition processing to obtain cell region features and unexpected region features; determining dynamic segmentation parameters and selection parameters of a second neural network model based on the cell region features and the unexpected region features; segmenting the cell original image based on the dynamic segmentation parameters to obtain a plurality of cell sub-images; selecting the second neural network model matched with the selection parameters, and inputting the plurality of cell sub-images into the second neural network model for local recognition processing to obtain cell features and cell coordinates; and performing global mapping based on the cell features and the cell coordinates, and simultaneously performing deduplication processing and elimination processing of the unexpected region features to obtain a cell recognition result. The method can solve the problems of accuracy and universality of neural network model in cell recognition.

Description

Technical Field

[0001] This application relates to the field of cell image analysis, specifically to a cell image recognition method, device, electronic device, readable storage medium, and computer program product. Background Technology

[0002] In recent years, cell counting methods under a microscope have shifted from manual counting to automated image recognition. Among these, superior automated image recognition methods employ AI neural network models. The entire high-resolution micrograph is input into the neural network, which extracts image features through multiple convolutional layers. High-dimensional features are then obtained using interest pooling layers and multiple fully connected layers, intelligently identifying cells and determining their quantity and viability.

[0003] In practice, it has been found that when processing high-resolution cell images, this type of method usually requires compressing the image to a fixed size to fit the model. However, this operation often leads to significant compression and deformation of cell features, resulting in severe pixelation, and even the inability to fully recognize the features of small cells.

[0004] It is evident that current methods cause irreversible damage to cellular characteristics, thereby directly limiting the accuracy and versatility of neural network models in cell identification. Summary of the Invention

[0005] In view of the above problems, this application provides a cell image recognition method, device, electronic device, readable storage medium and computer program product, which can solve the problems of accuracy and universality of neural network models in cell recognition.

[0006] In a first aspect, this application provides a cell image recognition method, comprising: Obtain the raw image of the cell to be identified; The original cell image is input into the first neural network model for global recognition processing to obtain cell region features and unexpected region features. Based on the cell region features and the unexpected region features, determine the dynamic segmentation parameters and the selection parameters for the second neural network model; Based on the dynamic segmentation parameters, the original cell image is segmented to obtain multiple cell sub-images; Select the second neural network model that matches the selection parameters, and input the multiple cell sub-images into the second neural network model for local recognition processing to obtain cell features and cell coordinates; Global mapping is performed based on the cell features and cell coordinates, and deduplication and removal of features from unexpected regions are performed simultaneously to obtain cell identification results.

[0007] In the above technical solution, the method can accurately locate the cell region through global recognition of the first neural network model, defining the effective range for subsequent processing. Based on this, the determined dynamic segmentation parameters can be used to directly perform adaptive segmentation on the original high-pixel cell image, thereby replacing the traditional overall image compression operation and avoiding cell feature deformation, pixelation, and the inability to recognize small cell features caused by compression. At the same time, the selection parameters determined based on cell region features can also make the second neural network model more closely match the feature attributes of the cell to be identified, thereby improving the accuracy of local recognition. Finally, through global mapping, deduplication, and removal of unexpected regions, the recognition results can be further optimized to ensure the accuracy and effectiveness of the output data.

[0008] In some embodiments, inputting the original cell image into a first neural network model for global recognition processing includes: The original cell image is filled to obtain a cell-filled image; The cell-filled image is compressed to a preset size to obtain a global preprocessed image; The global preprocessed image is input into the first neural network model for global recognition processing to obtain cell region features and unexpected region features. The preset size is a size adapted to the operation of the first neural network model.

[0009] In the above technical solution, the method can meet the input image size requirements of the first neural network model while minimizing feature distortion caused by direct compression of high-pixel cell original images through a preprocessing method of first filling and then compressing, thus ensuring that the cell region features and unexpected region features obtained by global recognition are more accurate and reliable.

[0010] In some implementations, the first neural network model is a YOLO-Seg model, a U-Net segmentation model, a SegFormer model, a LRASPP model, a BiSeNet model, a CE-Net model, or an nnUNet model.

[0011] In the above technical solution, the method can efficiently and accurately complete the global recognition of the original cell image by selecting mature segmentation neural network models such as YOLO-Seg and U-Net as the first neural network model, and quickly output reliable cell region features and unexpected region features.

[0012] In some embodiments, the cellular region characteristics include cell distribution range, maximum cell diameter, and specific cell state; Wherein, the cell distribution range is the overall boundary range of the cell distribution in the original cell image; The maximum diameter of the cell is the maximum diameter value obtained by the first neural network model after estimating the diameter of all cells in the original cell image; The special cell states include cell clustering and special morphological types, and the special morphological types include non-circular cell morphology.

[0013] In the above technical solution, this method can provide a precise and comprehensive reference for determining the subsequent dynamic segmentation parameters by clearly defining the specific dimensions of cell region characteristics, including cell distribution range, maximum cell diameter, and special cell state. This ensures that the segmentation operation fully matches the actual distribution and morphological characteristics of the cells, while also making the selection of the second neural network model more targeted and improving the effectiveness of the overall recognition process.

[0014] In some implementations, determining the dynamic segmentation parameters and the selection parameters of the second neural network model based on the cell region features and the unexpected region features includes: Based on the cell distribution range, the cell maximum diameter, and the cell specific state, dynamic segmentation parameters are determined; Based on the maximum diameter of the cell, the special state of the cell, and the features of the unexpected region, the selection parameters of the second neural network model are determined.

[0015] In the above technical solution, the method can rely on the dynamic segmentation parameters determined by the cell distribution range, the maximum cell diameter, and the special state of the cell to fit the actual morphology and distribution characteristics of the cell. The selection parameters determined by the maximum cell diameter, the special state of the cell, and the characteristics of the unexpected region can make the selection of the second neural network model more in line with the actual recognition needs, and effectively improve the pertinence of parameter determination.

[0016] In some implementations, determining the dynamic segmentation parameters based on the cell distribution range, the maximum cell diameter, and the specific cell state includes: Based on the cell distribution range, cell-free regions are screened to obtain effective regions to be segmented. The initial segmentation interval is obtained by performing segmentation interval matching based on the maximum diameter of the cell. Based on the specific state of the cells, segmentation and adaptation processing is performed to obtain an adaptation adjustment scheme; By integrating the region to be segmented, the initial segmentation interval, and the adaptation adjustment scheme, dynamic segmentation parameters are obtained.

[0017] In the above technical solution, the method can lock the effective segmentation range by screening cell-free areas, match the initial segmentation interval with the maximum diameter of the cells, and formulate a targeted adjustment plan based on the special state of the cells. The dynamic segmentation parameters formed by multi-dimensional integration can accurately fit the actual distribution, size and morphological characteristics of the cells, making subsequent image segmentation operations more targeted and avoiding the impact of invalid segmentation on recognition efficiency.

[0018] In some implementations, determining the dynamic segmentation parameters based on the cell distribution range, the maximum cell diameter, and the specific cell state includes: When the maximum diameter of the cell is less than a preset diameter threshold, dynamic segmentation parameters are determined based on the preset diameter threshold.

[0019] In the above technical solution, the method can directly determine the dynamic segmentation parameters based on the preset diameter threshold when the maximum diameter of the cell is less than the preset threshold, thereby avoiding the problem of unreasonable segmentation intervals caused by the small cell size and ensuring that the sub-image obtained after segmentation can completely present the features of the small cell.

[0020] In some implementations, determining the selection parameters of the second neural network model based on the maximum cell diameter, the specific cell state, and the features of the unexpected region includes: When the maximum diameter of the cell is less than a preset small cell threshold, the selection parameters of the second neural network model are determined to be the selection parameters adapted for small cell recognition; or When the cell is in a special state of non-circular cell morphology, the selection parameters of the second neural network model are determined to be the selection parameters adapted for the recognition of aberrant cells; or When bubble features are included in the unexpected region features, the selection parameters of the second neural network model are determined to be the selection parameters that adapt to the interference scenario.

[0021] In the above technical solution, the method can determine the selection parameters of the second neural network model according to different situations such as the maximum diameter of the cell, the special state of the cell, and the characteristics of the unexpected region. When identifying small cells, atypical cells, and scenarios with bubble interference, it can match the appropriate model and effectively improve the accuracy of cell recognition in different scenarios.

[0022] In some implementations, the step of segmenting the original cell image based on the dynamic segmentation parameters to obtain multiple cell sub-images includes: The initial segmentation interval is adjusted based on the adaptive adjustment scheme to obtain the target segmentation interval; wherein, the target segmentation interval is used to avoid segmentation of cells in special states; The original cell image is segmented based on the target segmentation interval to obtain multiple initial sub-images; Invalid images are filtered out from the multiple initial sub-images based on the region to be segmented, resulting in multiple cell sub-images.

[0023] In the above technical solution, the method can avoid segmenting cells in special states by adjusting the target segmentation interval through adaptation, and then filter out invalid sub-images by combining the effective regions to be segmented. The final cell sub-image not only fully preserves cell features, but also removes meaningless image parts.

[0024] In some embodiments, the step of inputting the plurality of cell sub-images into the second neural network model for local recognition processing to obtain cell features and cell coordinates includes: Each of the multiple cell sub-images is subjected to normalized padding to obtain multiple local preprocessed images; The multiple pre-processed local images are input into the second neural network model for local recognition processing to obtain cell features and cell coordinates.

[0025] In the above technical solution, the method can perform standardized filling processing on each cell sub-image, so that the size of the local preprocessed image input to the second neural network model is uniform and fits the model's operation requirements, ensuring that the local recognition process is stable and efficient, while preserving the effective features in the cell sub-image to the greatest extent, so that the second neural network model can accurately extract cell features and locate cell coordinates.

[0026] In some implementations, the global mapping based on the cell features and cell coordinates, along with simultaneous deduplication and removal of unwanted region features, yields the cell identification result, including: Global mapping is performed based on the cell features and cell coordinates to obtain the global mapping result; The global mapping result is deduplicated to obtain the deduplicated mapping result; The unexpected region features are removed from the deduplication mapping results to obtain the cell identification results.

[0027] In the above technical solution, the method can restore the position information of cells in the original image through global mapping, and then perform deduplication and removal of unexpected regions in sequence, effectively filtering out duplicate recognition results and invalid data in interference regions, so as to make the final output cell recognition results more accurate and reliable.

[0028] In some implementations, the cell identification results include each cell's global coordinates, estimated cell diameter, roundness, pixel area, average light intensity value, maximum light intensity value, and minimum light intensity value.

[0029] In the above technical solution, the method can output multi-dimensional cell recognition results including global cell coordinates, estimated diameter, roundness, pixel area and light intensity related values, providing comprehensive and accurate data support for cell analysis and meeting the diverse needs for cell information in different scenarios.

[0030] Secondly, this application provides a cell image recognition device, comprising: The acquisition unit is used to acquire the original image of the cell to be identified; A global recognition unit is used to input the original cell image into a first neural network model for global recognition processing to obtain cell region features and unexpected region features. The parameter determination unit is used to determine the dynamic segmentation parameters and the selection parameters of the second neural network model based on the cell region features and the unexpected region features. A dynamic segmentation unit is used to segment the original cell image based on the dynamic segmentation parameters to obtain multiple cell sub-images; A local recognition unit is used to select the second neural network model that matches the selection parameters, and input the multiple cell sub-images into the second neural network model for local recognition processing to obtain cell features and cell coordinates; The global mapping unit is used to perform global mapping based on the cell features and cell coordinates, and simultaneously perform deduplication and removal of features from unexpected regions to obtain cell recognition results.

[0031] In the above technical solution, the device can accurately locate the cell region through global recognition of the first neural network model, defining the effective range for subsequent processing. Based on this, the determined dynamic segmentation parameters can be used to directly perform adaptive segmentation on the original high-pixel cell image, thereby replacing the traditional overall image compression operation and avoiding cell feature deformation, pixelation, and the inability to recognize small cell features caused by compression. At the same time, the selection parameters determined based on the cell region features can also make the second neural network model more closely match the feature attributes of the cell to be identified, thereby improving the accuracy of local recognition. Finally, through global mapping, deduplication, and removal of unexpected regions, the recognition results can be further optimized to ensure the accuracy and effectiveness of the output data.

[0032] Thirdly, this application provides an electronic device including a memory and a processor, the memory storing a computer program, and the processor running the computer program to cause the electronic device to perform the cell image recognition method described in any one of the first aspects.

[0033] Fourthly, this application provides a readable storage medium storing a computer program, which, when executed by a processor, performs the cell image recognition method described in any one of the first aspects.

[0034] Fifthly, this application provides a computer program product comprising a computer program that, when executed by a processor, performs the cell image recognition method described in any one of the first aspects.

[0035] The beneficial effects of this application are as follows: The hybrid recognition method employing a dual neural network model can achieve complementary advantages between global and local recognition. While addressing the limitations of local recognition's field of view and poor model adaptability, it avoids feature loss and insufficient extraction caused by image compression in global recognition. Simultaneously, dynamic mesh partitioning effectively reduces redundant calculations in overlapping areas, significantly improving recognition efficiency. Furthermore, the combination of unexpected region filtering and duplicate recognition deduplication optimization also significantly improves the accuracy of cell recognition. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating the cell image recognition method in some embodiments of this application; Figure 2 These are comparison charts showing the model detection results in some embodiments of this application; Figure 3 This is a diagram of cells with low concentrations and many vacancy sites in some embodiments of this application. Figure 4 This is a schematic diagram illustrating the identification of low concentration distributions in some embodiments of this application; Figure 5 This is a segmentation image with a spacing of 128 pixels in some embodiments of this application; Figure 6 This is a segmentation image with a spacing of 60 pixels in some embodiments of this application; Figure 7 This is a segmentation image with a spacing of 30 pixels in some embodiments of this application; Figure 8 Example diagrams of clustered cells on dividing lines in some embodiments of this application; Figure 9These are global and magnified views of triangular filamentous algal cells in some embodiments of this application; Figure 10 This is a schematic diagram illustrating the effect of air bubbles on cells in some embodiments of this application; Figure 11 Example diagrams illustrating how the global model in some embodiments of this application directly distinguishes cells from large impurities; Figure 12 This is an example diagram showing how the global model in some embodiments of this application directly distinguishes large impurities between cells and large impurities. Figure 13 This is a flowchart illustrating the cell image recognition method in some embodiments of this application; Figure 14 This is a schematic diagram of the structure of the cell image recognition device in some embodiments of this application; Figure 15 This is a schematic diagram of the structure of an electronic device in some embodiments of this application. Detailed Implementation

[0038] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0040] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more (including two), similarly, "multiple sets" refers to two or more sets (including two sets), and "multiple pieces" refers to two or more pieces (including two pieces) unless otherwise explicitly defined.

[0041] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0042] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0043] The three core problems existing in current cell image recognition solutions are: First, due to the limitation of cell size range, when designing overlapping areas using the maximum pixel coverage range, identifying small cells will generate a large amount of duplicate data, leading to a surge in computational load. Secondly, a single network structure is difficult to adapt to all cell types. Optimized configuration parameters for large-sized cells can easily lead to missed detections of small-sized cells and insufficient detection boxes. Furthermore, the model selection requires user judgment, which is cumbersome. Third, the image segmentation method is prone to misidentifying large impurities that fall off during the sample loading process as adhered or aggregated cells, which affects the accuracy of identification.

[0044] To address the aforementioned technical problems, this application provides a cell image recognition method. This method employs a hybrid recognition approach using dual neural network models, achieving complementary advantages between global and local recognition. While resolving the limitations of local recognition's field of view and poor model adaptability, it avoids feature loss and insufficient extraction caused by image compression in global recognition. Furthermore, dynamic mesh partitioning effectively reduces redundant calculations in overlapping areas, significantly improving recognition efficiency. The combination of unexpected region filtering and duplicate recognition deduplication further enhances the accuracy of cell recognition.

[0045] like Figure 1 As shown, some embodiments of this application provide a cell image recognition method, which includes: S101. Obtain the original image of the cell to be identified; S102. Input the original cell image into the first neural network model for global recognition processing to obtain cell region features and unexpected region features; S103. Based on cell region features and unexpected region features, determine the dynamic segmentation parameters and the selection parameters of the second neural network model; S104. Based on dynamic segmentation parameters, the original cell image is segmented to obtain multiple cell sub-images; S105. Select a second neural network model that matches the selection parameters, and input multiple cell sub-images into the second neural network model for local recognition processing to obtain cell features and cell coordinates; S106. Perform global mapping based on cell features and cell coordinates, and simultaneously perform deduplication and removal of features from unexpected regions to obtain cell identification results.

[0046] In some embodiments, raw cell images refer to images of cells taken without any preprocessing before identification and analysis.

[0047] In some embodiments, the first neural network model refers to a neural network model used to perform global recognition processing on the original cell image to extract cell region features and features of unexpected regions.

[0048] In some embodiments, cell region features refer to relevant feature information that reflects the distribution and morphological characteristics of cells in the original image, obtained by the first neural network model.

[0049] In some embodiments, unintended region features refer to the feature information of interfering regions such as impurities and bubbles in the cell image other than the target cell region, which are identified by the first neural network model.

[0050] In some embodiments, dynamic segmentation parameters refer to relevant parameters determined based on cell region features for targeted segmentation of the original cell image.

[0051] In some embodiments, the selection parameters of the second neural network model refer to the parameters used to select the appropriate second neural network model, which are determined based on cell region features and unexpected region features.

[0052] In some embodiments, a cell sub-image refers to a sub-region image containing cells obtained after the original cell image is segmented by dynamic segmentation parameters.

[0053] In some embodiments, the second neural network model refers to a neural network model that is matched with the selection parameters and used for local recognition processing of cell sub-images.

[0054] In some embodiments, cell features refer to information reflecting individual cell attributes obtained through local identification by a second neural network model.

[0055] In some embodiments, cell coordinates refer to the positional information of a cell in a corresponding cell sub-image, which is obtained through local identification by a second neural network model.

[0056] In some embodiments, cell identification results refer to the identification data that contains key information about the cells, obtained after global mapping, deduplication, and removal of unexpected region features.

[0057] In the above embodiments, the method can accurately locate the cell region through global recognition of the first neural network model, defining an effective range for subsequent processing. Based on this, the determined dynamic segmentation parameters can be used to directly perform adaptive segmentation on the original high-pixel cell image, thereby replacing the traditional overall image compression operation and avoiding cell feature deformation, pixelation, and the inability to recognize small cell features caused by compression. At the same time, the selection parameters determined based on cell region features can also make the second neural network model more closely match the feature attributes of the cell to be identified, thereby improving the accuracy of local recognition. Finally, through global mapping, deduplication, and unexpected region removal, the recognition results can be further optimized to ensure the accuracy and effectiveness of the output data.

[0058] In some embodiments, inputting the original cell image into a first neural network model for global recognition processing includes: The original cell image is filled to obtain a cell-filled image; The cell-filled image is compressed to a preset size to obtain a global preprocessed image; The preprocessed global image is input into the first neural network model for global recognition processing to obtain cell region features and unexpected region features. The preset size is the size adapted to the operation of the first neural network model.

[0059] In some embodiments, the first neural network model is input with a fully captured image, such as a 3072×2048 image that is padded to 3072×3072 and then directly compressed to 640×640 pixels; such images are used for both training and inference.

[0060] In some embodiments, the first neural network model tends to use a "segmentation model" for identification, while the second neural network model tends to use an "object detection model." The primary goal of the first neural network is to roughly distinguish cell types and identify unexpected detection areas, such as large impurities. Generally, large impurities have a large span; if an "object detection model" is used, only a rectangular bounding box can be obtained, which cannot define the true extent of the impurity. Furthermore, a large number of normal cells may still exist within the bounding box; directly masking this area would result in the false deletion of real cells. In contrast, the "segmentation model" can output pixel-level precise masks, has a stronger ability to represent irregular boundaries and large-scale structures, and can more completely remove impurity areas. See [link to relevant documentation] for details. Figure 2 The comparison chart showing the model detection results is shown. Figure 2 The left image shows the performance of the object detection model. Figure 2 The right image shows the effect of the segmentation model, with the target object within the box being impurities.

[0061] In some embodiments, in scenarios with less stringent requirements, this method allows two models to use the same model. For example, both models can employ an "object detection model," which, although less accurate than this method, can still perform object recognition. Because its model construction is relatively simple, it can also serve as an extension of this method.

[0062] For example, the output of the first neural network model includes two types of image data: cellular region features and unexpected region features. In cellular regions, the model derives the maximum range of cell distribution, cell diameter, specific cell state, and cell type; in unexpected regions, the model derives regions of bubbles or large impurities. The membership, function, and impact of these outputs are shown in Table 1.

[0063] Table 1 Output of the First Neural Network

[0064] In some embodiments, the training sets and construction methods of the first and second neural network models differ. Because the first neural network focuses more on global-level content, its training set and all training parameters differ from those of the second neural network. In some less demanding special cases, the same dataset can be used for training, but the difference in perspective will affect the model's recognition accuracy.

[0065] In the above embodiments, the method can minimize feature distortion caused by direct compression of high-pixel cell original images while meeting the input image size requirements of the first neural network model through a preprocessing method of first filling and then compressing, thus ensuring that the cell region features and unexpected region features obtained by global recognition are more accurate and reliable.

[0066] In some embodiments, the first neural network model is a YOLO-Seg model, a U-Net segmentation model, a SegFormer model, a LRASPP model, a BiSeNet model, a CE-Net model, or an nnUNet model.

[0067] In some embodiments, the first neural network model can be the YOLO-Seg model, or the U-Net segmentation model commonly used in the field of medical imaging, such as the Transformer-based SegFormer model, LRASPP, BiSeNet and other lightweight segmentation models, or CE-Net, nnUNet and other models commonly used in the field of medical imaging.

[0068] In some embodiments, U-Net employs a symmetrical structure of encoder and decoder, which is widely used in medical images and is particularly suitable for identifying irregular and complex regions.

[0069] SegFormer, through the use of the Transformer architecture, possesses strong context awareness. Its self-attention global receptive field makes it easier to capture the overall shape of impurities across different regions.

[0070] The YOLO-seg model also uses the YOLO series backbone feature extraction network, which is fast inference, simple to deploy, and can perform segmentation in a very short time.

[0071] In the above embodiments, the method can efficiently and accurately complete the global recognition of the original cell image by selecting mature segmentation neural network models such as YOLO-Seg and U-Net as the first neural network model, and quickly output reliable cell region features and unexpected region features.

[0072] In some embodiments, cellular region characteristics include cell distribution range, maximum cell diameter, and specific cell state; Among them, the cell distribution range is the overall boundary range of the cell distribution in the original cell image; The maximum cell diameter is the maximum diameter value obtained by the first neural network model after estimating the diameter of all cells in the original cell image; Special cell states include cell aggregation and special morphological types, with special morphological types including non-circular cell morphology.

[0073] In the above embodiments, this method can provide a precise and comprehensive reference for determining subsequent dynamic segmentation parameters by clearly defining the specific dimensions of cell region features, including cell distribution range, maximum cell diameter, and special cell states. This ensures that the segmentation operation fully matches the actual distribution and morphological characteristics of the cells, while also making the selection of the second neural network model more targeted and improving the effectiveness of the overall recognition process.

[0074] In some embodiments, dynamic segmentation parameters and selection parameters for the second neural network model are determined based on cellular region features and unexpected region features, including: Dynamic segmentation parameters are determined based on cell distribution range, maximum cell diameter, and specific cell states. Based on the maximum cell diameter, special cell states, and features of unexpected regions, the selection parameters for the second neural network model are determined.

[0075] In some embodiments, conventional schemes typically employ fixed segmentation sizes, which can lead to low computational efficiency. For example, overlapping based on the largest cell diameter results in numerous duplicate statistics when detecting small cells, further reducing efficiency. This method, however, can improve segmentation efficiency by pre-judging the first neural network to determine a more efficient segmentation method.

[0076] In the above embodiments, the method can rely on the dynamic segmentation parameters determined by the cell distribution range, the maximum cell diameter, and the special state of the cell to fit the actual morphology and distribution characteristics of the cell. The selection parameters determined by the maximum cell diameter, the special state of the cell, and the features of the unexpected region can make the selection of the second neural network model more in line with the actual recognition needs, and effectively improve the pertinence of parameter determination.

[0077] In some embodiments, dynamic segmentation parameters are determined based on cell distribution range, maximum cell diameter, and specific cell states, including: Cell-free regions are screened based on cell distribution range to obtain effective regions to be segmented. The initial segmentation interval is obtained by matching the segmentation interval based on the maximum cell diameter. Based on the specific state of cells, segmentation and adaptation processing is performed to obtain an adaptation adjustment scheme; By integrating the region to be segmented, the initial segmentation interval, and the adaptation adjustment scheme, dynamic segmentation parameters are obtained.

[0078] In some embodiments, in low-concentration scenarios or with specific distributions, areas may be devoid of cells; see [reference needed]. Figure 3 The image shows cells at low concentrations with a large number of vacancy sites. (Through...) Figure 4 The diagram illustrating the identification of low-concentration distributions shows that, during segmentation, there are clearly five regions devoid of any cells. Therefore, these regions can be directly excluded from the statistical counting process, reducing the computational load on the algorithm.

[0079] In some embodiments, the method can determine how the spacing during segmentation needs to be adjusted based on the estimated maximum diameter of the cells in the first neural network model, such as... Figure 5 , Figure 6 , Figure 7 As shown. Figure 5 A segmentation image with a spacing of 128 pixels is shown. Figure 6 A segmentation image with a spacing of 60 pixels was generated. Figure 7 A segmentation image with a spacing of 30 pixels is shown.

[0080] In some embodiments, when identifying cells, the cells themselves may not be large, but they may exhibit numerous clusters, and these clusters may lie precisely on the dividing lines, such as... Figure 8 As shown ( Figure 8 An example diagram of clustered cells on the dividing line is shown. In this case, the cluster can be identified and judged as a whole cell, and the dividing region can be adjusted according to the overall size of the cluster.

[0081] In some embodiments, dynamically adjusting the cutting method through the above-described determination can help maximize the efficiency of detection and judgment.

[0082] In the above embodiments, the method can identify the effective segmentation range by screening cell-free regions, match the initial segmentation interval with the maximum diameter of the cells, and formulate a targeted adjustment plan based on the special state of the cells. The dynamic segmentation parameters formed by multi-dimensional integration can accurately fit the actual distribution, size and morphological characteristics of the cells, making subsequent image segmentation operations more targeted and avoiding the impact of invalid segmentation on recognition efficiency.

[0083] In some embodiments, dynamic segmentation parameters are determined based on cell distribution range, maximum cell diameter, and specific cell states, including: When the maximum cell diameter is less than a preset diameter threshold, dynamic segmentation parameters are determined based on the preset diameter threshold.

[0084] In some embodiments, since the overlapping area is generally within a certain range, it is not necessary to count the entire image. It is only necessary to count the cells in the potentially overlapping areas, observe the maximum diameter of the cells in this area, and ensure that the maximum diameter of the cells in the overlapping area is included.

[0085] In some embodiments, for extreme cases, the cells may be small, and the cell features may be completely lost after the large image is compressed. If it is determined that there is no cell overlap, it will result in no overlapping area. To deal with this kind of problem, a fixed minimum overlap pixel value can be given. For example, if the overlap pixel value is less than 10 pixels, the default is 10 pixels of overlap.

[0086] In some embodiments, image compression may cause a slight deviation in the maximum diameter, and directly using the maximum diameter may not be sufficient to reflect the true diameter. In such cases, a safety factor can be multiplied by the identified diameter; for example, after obtaining the maximum cell diameter, it can be multiplied by 1.2 as an interval parameter.

[0087] In the above embodiments, the method can directly determine the dynamic segmentation parameters based on the preset diameter threshold when the maximum diameter of the cell is less than the preset threshold, thereby avoiding the problem of unreasonable segmentation intervals caused by the small cell size and ensuring that the sub-image obtained after segmentation can completely present the features of the small cell.

[0088] In some embodiments, the selection parameters of the second neural network model are determined based on the maximum cell diameter, special cell states, and features of unexpected regions, including: When the maximum diameter of the cell is less than the preset small cell threshold, the selection parameters of the second neural network model are determined to be the selection parameters suitable for small cell recognition; or When the cell is in a special state of non-circular cell morphology, the selection parameters of the second neural network model are determined to be the selection parameters adapted for the recognition of aberrant cells; or When bubble features are included in the unexpected region features, the selection parameters of the second neural network model are determined to be the selection parameters that adapt to the interference scenario.

[0089] In some embodiments, the output of the first neural network model influences the type selection of the second neural network model.

[0090] For example, in the first neural network model, the determination of cell diameter, special cell state (clustering, chain-like or special shape), and whether there are large air bubbles can affect the selection of the second neural network model or the fine-tuning of the model parameters.

[0091] Therefore, there can be multiple categories for the second neural network model, which are influenced by the output of the first neural network model.

[0092] In some embodiments, cell diameter, specific cell state, and large air bubbles can all affect the model selection or parameter fine-tuning of the second neural network. The difference in diameter is mainly reflected in the fact that if small cells are the focus of detection, applying a large cell model may easily lead to them being treated as small particulate impurities, resulting in missed detections.

[0093] When counting multiple cell types simultaneously, you may encounter cells with unusual morphologies. For example, most cells are spherical, but some algal cells are triangular or filamentous. In such cases, different models need to be used for identification, such as... Figure 9 As shown. Among them, Figure 9 The image shows a global view (left) and a magnified view (right) of a triangular filamentous algal cell.

[0094] In some embodiments, the impact of bubbles on the model is mainly due to the fact that bubbles may be a strong source of background interference, and nearby cells may be affected by the bubbles, resulting in them not being fully visible. Therefore, an enhanced model is needed to avoid potential cell undercounting, such as... Figure 10 As shown. Among them, Figure 10 A schematic diagram illustrating the effect of air bubbles on cells is shown.

[0095] In the above embodiments, the method can determine the selection parameters of the second neural network model according to different situations such as the maximum diameter of the cell, the special state of the cell, and the characteristics of the unexpected region. When identifying small cells, atypical cells, and scenarios with bubble interference, the method can match the appropriate model, effectively improving the accuracy of cell recognition in different scenarios.

[0096] In some embodiments, the original cell image is segmented based on dynamic segmentation parameters to obtain multiple cell sub-images, including: The initial segmentation interval is adaptively adjusted based on the adaptive adjustment scheme to obtain the target segmentation interval; wherein, the target segmentation interval is used to avoid segmentation of cells in special states; The original cell image is segmented based on the target segmentation interval to obtain multiple initial sub-images; Invalid images are filtered out from multiple initial sub-images based on the region to be segmented, resulting in multiple cell sub-images.

[0097] In the above embodiments, the method can avoid segmenting cells in special states by adjusting the target segmentation interval through adaptation, and then filter out invalid sub-images by combining the effective regions to be segmented. The final cell sub-images not only fully preserve cell features, but also remove meaningless image parts.

[0098] In some embodiments, multiple cell sub-images are input into a second neural network model for local recognition processing to obtain cell features and cell coordinates, including: Each cell sub-image in multiple cell sub-images is normalized and padded to obtain multiple local preprocessed images; Multiple pre-processed local images are input into a second neural network model for local recognition processing to obtain cell features and cell coordinates.

[0099] In some embodiments, the input to the second neural network model is a cropped image, typically within 640×640 pixels (normalized size), which is then normalized to 640×640 pixels by padding. Note that the image is not compressed and its original features are not affected. Such images are used for both training and inference.

[0100] In some embodiments, the second neural network model can use the YOLO series of deep learning object detection models for cell identification and counting. Among them, YOLO (You Only Look Once) is a representative object detection model in the development of deep learning. It is characterized by end-to-end training, single-stage prediction, high speed, and high accuracy, and is widely used in image recognition and medical analysis.

[0101] In some embodiments, the second neural network model may be developed based on the YOLOv8 model, but is not limited to this model.

[0102] In some embodiments, the second neural network model is used for local image operations. Its input is a segmented image, typically 640×640 pixels in size, which is then padded and normalized to 640×640 pixels. These images are characterized by a small field of view, but because they are not compressed, they have high image clarity, making them suitable for identifying cell features and accurately estimating cell diameter. During training and inference, cell contour features are accurately identified. These images may also contain both cellular and non-cellular features; in this case, the non-cellular features are small impurities and foreign objects that can be directly excluded.

[0103] After cell identification in the local image, the cells are mapped to the original image, and all duplicate cell features are excluded. The remaining cell coordinates and feature information are the final full-image cell identification result.

[0104] In the above embodiments, the method can perform standardized filling processing on each cell sub-image, so that the size of the local preprocessed image input to the second neural network model is uniform and fits the model's computational requirements, ensuring that the local recognition process is stable and efficient, while preserving the effective features within the cell sub-image to the greatest extent, so that the second neural network model can accurately extract cell features and locate cell coordinates.

[0105] In some embodiments, global mapping is performed based on cell features and cell coordinates, and deduplication and removal of features from unwanted regions are performed simultaneously to obtain cell identification results, including: Global mapping is performed based on cell features and cell coordinates to obtain the global mapping result; The global mapping result is deduplicated to obtain the deduplicated mapping result; Unexpected region features are removed from the deduplication mapping results to obtain cell identification results.

[0106] In some embodiments, since the main factors affecting the screening statistics are unexpected feature recognition regions, interference-type bubbles or large impurities can be marked as unexpected regions to be removed using a masking method after their shape is identified. See also... Figure 11 , Figure 11 An example diagram is shown illustrating how a global model directly distinguishes cells from large impurities; Figure 12 An example diagram is shown of a global model that directly distinguishes large impurities between cells and large impurities.

[0107] In some embodiments, when the method is passed to the next model, it can know that the location of this region needs to be excluded, without performing cell counting statistics or directly excluding the already counted parameter locations.

[0108] In some embodiments, the method can also identify other features, such as grids or tick marks. When passed to the next model, the model knows that the way this region is identified may be slightly different from that of the normal model and needs to be processed with a special exclusion method.

[0109] In the above embodiments, the method can restore the position information of cells in the original image through global mapping, and then perform deduplication and removal of unwanted regions in sequence, effectively filtering out duplicate recognition results and invalid data in interference regions, so as to make the final output cell recognition results more accurate and reliable.

[0110] In some embodiments, the cell identification results include the global coordinates of each cell, the estimated cell diameter, roundness, pixel area, average light intensity value, maximum light intensity value, and minimum light intensity value.

[0111] In some embodiments, the second neural network outputs cell recognition features including: estimated cell diameter, roundness, pixel area, average light intensity value, maximum light intensity value, minimum light intensity value, and coordinate position, indicating the location of each cell center in the global image.

[0112] In the above embodiments, the method can output multi-dimensional cell recognition results including global cell coordinates, estimated diameter, roundness, pixel area, and light intensity related values, providing comprehensive and accurate data support for cell analysis and meeting the diverse needs for cell information in different scenarios.

[0113] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below. In some embodiments, such as Figure 13 As shown, the cell image recognition method includes: S201. Perform global image operations on the original image. The copied file of the original image is directly compressed to 640×640 pixels and imported into the first neural network model.

[0114] S202, the first neural network model identifies the cellular region and the unexpected region.

[0115] S203, Cell Region Section: Acquire and identify cell characteristics, mainly including cell distribution, cell diameter, and special cell states.

[0116] S204, Unexpected Regions: Acquire and identify unexpected feature types, mainly including larger bubbles, larger impurities, etc.

[0117] In some embodiments, the method may also incorporate rulers, grid lines, etc.

[0118] S205. Collect and summarize the parameter information that affects dynamic segmentation, identify the more efficient dynamic segmentation method based on various cell characteristics, and pass it to the dynamic segmentation step.

[0119] In some embodiments, various cell characteristics may include cell distribution, cell diameter, and special cell states.

[0120] In some embodiments, the method may not search the entire image, but instead find the largest cell in the overlapping area as the interval parameter; or in extreme cases, if the cells are small, a minimum interval pixel may be reserved; or to avoid compression loss, the diameter of the largest cell obtained may be multiplied by a safety factor to ensure that it can be included.

[0121] S206. Collect and summarize the parameter information that affects the selection of the second neural network model. Based on the cell diameter, special conditions, and unexpected large bubbles, identify the more suitable second neural network model and pass it to the local operation steps.

[0122] S207. Collect and summarize the information data that affects the screening statistics, mainly based on the identification of unexpected areas to form a mask, and pass this range to the data removal step.

[0123] In some embodiments, the unintended area may include larger air bubbles, larger impurities, etc.

[0124] In some embodiments, the objects to be filtered may also include influencing factors such as rulers and grid lines, which require further elimination.

[0125] S208. Take the original image and the dynamic segmentation parameters obtained in S205, and segment the image.

[0126] S209. After cutting, N sub-images will be obtained and imported into the local operation steps.

[0127] S210. Normalize the sub-images to 640×640 pixels and then import them into the second neural network model. During this normalization import process, the images are not compressed, and the original image features are completely maintained.

[0128] In some embodiments, the method can also perform some beneficial optimization processing on the image. For example, denoising or image sharpening. The purpose of such methods is to improve the recognition of the cropped image, but the goal is to make the cropped sub-images easier to recognize.

[0129] S211. In the second neural network model, the image is identified, the effective features and coordinate positions of cells are extracted, and cells on the segmentation line are excluded.

[0130] S212. Convert all extracted features to the global coordinate system and summarize the cell extraction data of all sub-graphs.

[0131] In some embodiments, the cell coordinates obtained in S211 are obtained by a rectangular box identified by the model, and the center of the rectangular box is determined as the center of the cell, which is used as the coordinate position of the cell. During the statistical summary, the coordinate positions of all sub-image cells need to be converted to their global coordinate positions within the original image. The conversion method is related to the dynamic segmentation, with the segmentation coordinate line being the relative origin of the sub-image coordinates. In addition to cell coordinate conversion, the statistical cell characteristics will also be listed simultaneously, including the estimated cell diameter, roundness, pixel area, average light intensity value, maximum light intensity value, and minimum light intensity value.

[0132] S213. Based on the screening statistics of the unexpected region obtained in S207, compare all cell coordinate positions obtained in S212 and remove all cell position data belonging to the unexpected region.

[0133] S214. Based on S213, perform deduplication screening of all cell data. The role of the segmentation time interval parameter is to provide an effective overlapping area and avoid counting data on the segmentation line. Therefore, in this step, it is necessary to "deduplicate" cells that have been counted repeatedly in the global coordinates to prevent cells from being counted multiple times.

[0134] S215. After the filtering operations of S213 and S214, the cell coordinates and corresponding cell feature information obtained are finally summarized, which is the final output result.

[0135] In the above embodiments, this method can identify multiple features of cells and unexpected regions through a globally acting first neural network, and dynamically adjust the segmentation method based on the output parameters to determine the applicable model to be called by the second neural network, as well as identify the unexpected regions that need to be excluded. Furthermore, a local neural network model is introduced to accurately identify cell features while preserving the detailed information of the original image, thereby achieving rapid screening and reliable determination of abnormal cells.

[0136] Furthermore, this method incorporates a deduplication strategy for overlapping regions, avoiding the analysis of cells along the segmentation lines, thereby achieving stable and accurate cell location identification. The overall method demonstrates significant stability, reliability, and practical value in engineering applications.

[0137] Figure 14 A schematic diagram of a cell image recognition device is shown. It should be understood that this device is related to... Figure 1The method executed in the middle corresponds to the steps involved in the aforementioned method. The specific functions and effects of the device can be found in the description above. To avoid repetition, detailed descriptions are omitted here.

[0138] The cell image recognition device includes: Acquisition unit 310 is used to acquire the original image of the cell to be identified; The global recognition unit 320 is used to input the original cell image into the first neural network model for global recognition processing to obtain cell region features and unexpected region features. The parameter determination unit 330 is used to determine the dynamic segmentation parameters and the selection parameters of the second neural network model based on cell region features and unexpected region features. The dynamic segmentation unit 340 is used to segment the original cell image based on dynamic segmentation parameters to obtain multiple cell sub-images; The local recognition unit 350 is used to select a second neural network model that matches the selection parameters, and input multiple cell sub-images into the second neural network model for local recognition processing to obtain cell features and cell coordinates. The global mapping unit 360 is used to perform global mapping based on cell features and cell coordinates, and simultaneously perform deduplication and removal of features from unexpected regions to obtain cell recognition results.

[0139] In some embodiments, the global identification unit 320 includes: The first filling subunit 321 is used to fill the original cell image to obtain a cell-filled image; Compression subunit 322 is used to compress the cell-filled image to a preset size to obtain a global preprocessed image; The first input subunit 323 is used to input the global preprocessed image into the first neural network model for global recognition processing to obtain cell region features and unexpected region features. The preset size is the size adapted to the operation of the first neural network model.

[0140] In some embodiments, the first neural network model is a YOLO-Seg model, a U-Net segmentation model, a SegFormer model, a LRASPP model, a BiSeNet model, a CE-Net model, or an nnUNet model.

[0141] In some embodiments, cellular region characteristics include cell distribution range, maximum cell diameter, and specific cell state; Among them, the cell distribution range is the overall boundary range of the cell distribution in the original cell image; The maximum cell diameter is the maximum diameter value obtained by the first neural network model after estimating the diameter of all cells in the original cell image; Special cell states include cell aggregation and special morphological types, with special morphological types including non-circular cell morphology.

[0142] In some embodiments, the parameter determination unit 330 is specifically used to determine dynamic segmentation parameters based on cell distribution range, maximum cell diameter and special cell state. The parameter determination unit 330 is also used to determine the selection parameters of the second neural network model based on the maximum cell diameter, special cell states, and features of unexpected regions.

[0143] In some embodiments, the parameter determination unit 330 includes: The screening subunit 331 is used to screen cell-free regions based on the cell distribution range to obtain effective regions to be segmented. Matching subunit 332 is used to perform segmentation interval matching based on the maximum cell diameter to obtain the initial segmentation interval; The adapter subunit 333 is used to perform segmentation and adaptation processing based on the specific state of the cell to obtain an adaptation adjustment scheme. Integration subunit 334 is used to integrate the region to be segmented, the initial segmentation interval, and the adaptation adjustment scheme to obtain dynamic segmentation parameters.

[0144] In some embodiments, the parameter determination unit 330 is further configured to determine dynamic segmentation parameters based on a preset diameter threshold when the maximum diameter of the cell is less than a preset diameter threshold.

[0145] In some embodiments, the parameter determination unit 330 is specifically used to determine the selection parameters of the second neural network model as selection parameters suitable for small cell recognition when the maximum diameter of the cell is less than a preset small cell threshold; or When the cell is in a special state of non-circular cell morphology, the selection parameters of the second neural network model are determined to be the selection parameters adapted for the recognition of aberrant cells; or When bubble features are included in the unexpected region features, the selection parameters of the second neural network model are determined to be the selection parameters that adapt to the interference scenario.

[0146] In some embodiments, the dynamic segmentation unit 340 includes: The adjustment subunit 341 is used to adjust the initial segmentation interval based on the adaptive adjustment scheme to obtain the target segmentation interval; wherein, the target segmentation interval is used to avoid segmentation of special cell states; Segmentation subunit 342 is used to segment the original cell image based on the target segmentation interval to obtain multiple initial sub-images; The screening subunit 343 is used to screen out invalid images from multiple initial sub-images based on the region to be segmented, thereby obtaining multiple cell sub-images.

[0147] In some embodiments, the local identification unit 350 includes: The second filling subunit 351 is used to perform standardized filling processing on each of the multiple cell sub-images to obtain multiple local preprocessed images. The second input subunit 352 is used to input multiple local preprocessed images into the second neural network model for local recognition processing to obtain cell features and cell coordinates.

[0148] In some embodiments, the global mapping unit 360 includes: Mapping subunit 361 is used to perform global mapping based on cell features and cell coordinates to obtain global mapping results; Deduplication subunit 362 is used to remove duplicates from the global mapping result to obtain the deduplication mapping result; Subunit 363 is used to remove unexpected region features from the deduplication mapping results to obtain cell recognition results.

[0149] In some embodiments, the cell identification results include the global coordinates of each cell, the estimated cell diameter, roundness, pixel area, average light intensity value, maximum light intensity value, and minimum light intensity value.

[0150] like Figure 15 As shown, this application provides an electronic device 400, which includes a processor 401 and a memory 402. The processor 401 and the memory 402 are interconnected and communicate with each other through a communication bus 403 and / or other forms of connection mechanism (not shown). The memory 402 stores a computer program that can be executed by the processor 401. When the computing device is running, the processor 401 executes the computer program to perform the method in any of the aforementioned optional implementations.

[0151] This application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the method in any of the aforementioned optional implementations.

[0152] The computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0153] This application provides a computer program product, which includes a computer program that, when run by a processor, executes the method in any of the aforementioned optional implementations.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application. In particular, as long as there is no conflict, the various technical features mentioned in the embodiments can be combined in any way. This application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A cell image recognition method, characterized in that, The method includes: Obtain the raw image of the cell to be identified; The original cell image is input into a first neural network model for global recognition processing to obtain cell region features and unexpected region features. The cell region features include cell distribution range, maximum cell diameter, and special cell states. The unexpected region features include impurity feature information and bubble feature information. The special cell states include cell clustering state and special morphological types. The special morphological types include non-circular cell morphology. Based on the cell distribution range, cell-free regions are screened to obtain effective regions to be segmented. The initial segmentation interval is obtained by performing segmentation interval matching based on the maximum diameter of the cell. Based on the specific state of the cells, segmentation and adaptation processing is performed to obtain an adaptation adjustment scheme; By integrating the region to be segmented, the initial segmentation interval, and the adaptation adjustment scheme, dynamic segmentation parameters are obtained. When the maximum diameter of the cell is less than a preset small cell threshold, the selection parameters of the second neural network model are determined to be the selection parameters for small cell recognition; or when the cell is in a special state of non-circular cell morphology, the selection parameters of the second neural network model are determined to be the selection parameters for aberrant cell recognition; or when the features of the unexpected region include bubble features, the selection parameters of the second neural network model are determined to be the selection parameters for the interference scenario; wherein, one selection parameter corresponds to one second neural network model; Based on the dynamic segmentation parameters, the original cell image is segmented to obtain multiple cell sub-images; each cell sub-image is a sub-region image containing cells obtained after the original cell image is segmented by the dynamic segmentation parameters. A second neural network model matching the selection parameters is selected, and the multiple cell sub-images are input into the second neural network model for local recognition processing to obtain cell features and cell coordinates; the cell features refer to relevant information reflecting individual cell attributes obtained through local recognition by the second neural network model. Global mapping is performed based on the cell features and cell coordinates to obtain the global mapping result; The global mapping result is deduplicated to obtain the deduplicated mapping result; The unexpected region features are removed from the deduplication mapping result to obtain the cell recognition result; the cell recognition result includes the global coordinates of each cell, the estimated cell diameter, roundness, pixel area, average light intensity value, maximum light intensity value, and minimum light intensity value.

2. The cell image recognition method according to claim 1, characterized in that, The step of inputting the original cell image into the first neural network model for global recognition processing includes: The original cell image is filled to obtain a cell-filled image; The cell-filled image is compressed to a preset size to obtain a global preprocessed image; The global preprocessed image is input into the first neural network model for global recognition processing to obtain cell region features and unexpected region features. The preset size is a size adapted to the operation of the first neural network model.

3. The cell image recognition method according to claim 1, characterized in that, The first neural network model is a YOLO-Seg model, a U-Net segmentation model, a SegFormer model, a LRASPP model, a BiSeNet model, a CE-Net model, or an nnUNet model.

4. The cell image recognition method according to claim 1, characterized in that, The cell distribution range is the overall boundary range of the cell distribution in the original cell image; The maximum cell diameter is the maximum diameter value obtained by the first neural network model after estimating the diameter of all cells in the original cell image.

5. The cell image recognition method according to claim 1, characterized in that, The method further includes: When performing segmentation interval matching based on the maximum cell diameter, if the maximum cell diameter is less than a preset diameter threshold, dynamic segmentation parameters are determined based on the preset diameter threshold.

6. The cell image recognition method according to claim 1, characterized in that, The original cell image is segmented based on the dynamic segmentation parameters to obtain multiple cell sub-images, including: The initial segmentation interval is adjusted based on the adaptive adjustment scheme to obtain the target segmentation interval; wherein, the target segmentation interval is used to avoid segmentation of cells in special states; The original cell image is segmented based on the target segmentation interval to obtain multiple initial sub-images; Invalid images are filtered out from the multiple initial sub-images based on the region to be segmented, resulting in multiple cell sub-images.

7. The cell image recognition method according to claim 1, characterized in that, The step of inputting the multiple cell sub-images into the second neural network model for local recognition processing to obtain cell features and cell coordinates includes: Each of the multiple cell sub-images is subjected to normalized padding to obtain multiple local preprocessed images; The multiple pre-processed local images are input into the second neural network model for local recognition processing to obtain cell features and cell coordinates.

8. A cell image recognition device, characterized in that, The cell image recognition device includes: The acquisition unit is used to acquire the original image of the cell to be identified; A global recognition unit is used to input the original cell image into a first neural network model for global recognition processing to obtain cell region features and unexpected region features. The cell region features include cell distribution range, maximum cell diameter, and special cell states. Unexpected region features include impurity feature information and bubble feature information. The special cell states include cell clustering state and special morphological types. The special morphological types include non-circular cell morphology. The parameter determination unit is used to determine the dynamic segmentation parameters and the selection parameters of the second neural network model based on the cell region features and the unexpected region features. A dynamic segmentation unit is used to segment the original cell image based on the dynamic segmentation parameters to obtain multiple cell sub-images; the cell sub-image is a sub-region image containing cells obtained after the original cell image is segmented by the dynamic segmentation parameters. A local recognition unit is used to select the second neural network model that matches the selection parameters, and input the multiple cell sub-images into the second neural network model for local recognition processing to obtain cell features and cell coordinates; the cell features refer to relevant information reflecting individual cell attributes obtained through local recognition by the second neural network model. A global mapping unit is used to perform global mapping based on the cell features and cell coordinates, and simultaneously perform deduplication and removal of features from unexpected regions to obtain cell recognition results. The cell recognition results include the global coordinates of each cell, the estimated cell diameter, roundness, pixel area, average light intensity value, maximum light intensity value, and minimum light intensity value. The parameter determination unit includes: The screening subunit is used to screen cell-free regions based on the cell distribution range to obtain effective regions to be segmented. The matching subunit is used to perform segmentation interval matching based on the maximum cell diameter to obtain the initial segmentation interval; The adapter subunit is used to perform segmentation and adaptation processing based on the specific state of the cell to obtain an adaptation adjustment scheme. The integration subunit is used to integrate the region to be segmented, the initial segmentation interval, and the adaptation adjustment scheme to obtain dynamic segmentation parameters; Specifically, the parameter determination unit is used to determine the selection parameters of the second neural network model as selection parameters for small cell recognition when the maximum diameter of the cell is less than a preset small cell threshold; or to determine the selection parameters of the second neural network model as selection parameters for aberrant cell recognition when the cell's special state is a non-circular cell shape; or to determine the selection parameters of the second neural network model as selection parameters for interference scenarios when the unexpected region features include bubble features; wherein, one selection parameter corresponds to one second neural network model; The global mapping unit includes: The mapping subunit is used to perform global mapping based on cell features and cell coordinates to obtain the global mapping result. The deduplication sub-unit is used to remove duplicates from the global mapping result and obtain the deduplication mapping result. The sub-unit removal is used to remove unexpected region features from the deduplication mapping results to obtain cell identification results.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the cell image recognition method according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which, when executed by a processor, performs the cell image recognition method according to any one of claims 1 to 7.

11. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, performs the cell image recognition method according to any one of claims 1 to 7.

Images