A deep learning-based blood cell image automatic classification system

Abstract

The present application relates to the technical field of image recognition, and particularly relates to a blood cell image automatic classification system based on deep learning, which solves the technical problems of excessive dependence on morphological features and lack of function state perception in the prior art. The system comprises: a data acquisition platform for obtaining original blood cell images and extracting single-cell image blocks containing single blood cells; a blood cell intelligent classification unit for performing multi-modal feature extraction and multi-modal feature fusion on the single-cell image blocks to obtain comprehensive feature indexes; the multi-modal features include cell dynamic behavior features containing morphological features and motion features, and interactive features of cells in a microenvironment; and a classification output unit for inputting the comprehensive feature indexes into a pre-trained classification model to obtain an automatic classification result of the single blood cells.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and particularly relates to a blood cell image automatic classification system based on deep learning. BACKGROUND

[0002] Blood cell classification is one of the core tasks of clinical blood tests, and its results directly provide key basis for diagnosis, treatment detection and prognosis evaluation of diseases such as anemia, infection and leukemia. With the growth of medical needs and technology iteration, the traditional blood cell classification method gradually shows limitations, and the rise of deep learning technology provides a core solution for building an efficient and accurate automatic classification system.

[0003] The existing blood cell image automatic classification system excessively relies on the morphological features of cells in the model training and inference process, and cannot effectively perceive and fuse dynamic information related to the functional state and life activities of cells. This inherent defect leads to the lack of classification model in the interpretation ability of the physiological and pathological function context of cells, which significantly restricts the performance of the system in complex clinical scenarios, and there is a risk of missed detection and misjudgment in the automatic classification of rare cell types with important clinical significance and cells in a specific functional state. SUMMARY

[0004] In order to solve the technical problems of excessive dependence on morphological features and lack of functional state perception in the prior art, the purpose of the present application is to provide a blood cell image automatic classification system based on deep learning, and the technical solution adopted is as follows:

[0005] In the first aspect, a blood cell image automatic classification system based on deep learning is provided, comprising: a data acquisition platform, a blood cell intelligent classification unit and a classification output unit; the data acquisition platform is used to obtain original blood cell images and extract single cell image blocks containing single blood cells; the blood cell intelligent classification unit is used to extract and fuse multi-modal features of single cell image blocks, and obtain comprehensive feature indexes; the multi-modal features include: cell dynamic behavior features including morphological features and motion features, and interactive features of cells in microenvironment; the classification output unit is used to input the comprehensive feature indexes into a pre-trained classification model to obtain the automatic classification results of single blood cells.

[0006] Based on the above technical scheme, in the blood cell image automatic classification system based on deep learning provided by the application, the limitations of traditional methods which only rely on morphological features are broken through, and the dynamic behavior features of cells combining morphological features and motion features, and the interactive features of cells in microenvironments are innovatively extracted, and a multi-modal feature system is constructed. Among them, the dynamic behavior features can reflect the functional related attributes such as cell activity and motion directionality, and the microenvironment interactive features can reflect the functional roles of cells in the colony, which enriches the feature dimension and makes the features more fully fit the physiological and pathological states of cells. Then, the classification result is output by combining the pre-training classification model, which can utilize the learning ability of the model to multiple dimensions of features to improve the reliability of classification decision.

[0007] In combination with the first aspect, in a possible implementation manner, the blood cell intelligent classification unit is specifically configured to: obtain a single cell motion trajectory from a sequence of multiple frames of continuous single cell image blocks, and analyze the single cell motion trajectory to obtain an active chemotaxis index; extract a cell contour feature of each frame of single cell image block in the sequence of multiple frames of continuous single cell image blocks to obtain a single cell contour deformation degree; and generate a dynamic function fingerprint based on the active chemotaxis index and the single cell contour deformation degree.

[0008] In combination with the first aspect, in a possible implementation manner, the blood cell intelligent classification unit is specifically configured to: obtain a single cell motion trajectory point set from a sequence of multiple frames of continuous single cell image blocks by using an optical flow method and feature matching; analyze a final displacement and a moving distance of the single cell from the single cell motion trajectory point set; and determine the active chemotaxis index according to a ratio of a modulus of the final displacement to the moving distance.

[0009] In combination with the first aspect, in a possible implementation manner, the blood cell intelligent classification unit is specifically configured to: for each frame of single cell image block, extract a contour feature corner point by using a corner point detection algorithm, and output a corner point set of each frame of single cell image block; calculate an Euclidean distance between any two contour feature corner points, and take an average value of the Euclidean distances of multiple groups of corner point pairs in the corner point set as a contour stretch degree of each frame of single cell image block; and determine a variance of multiple contour stretch degrees corresponding to the sequence of multiple frames of continuous single cell image blocks as the single cell contour deformation degree.

[0010] In combination with the first aspect, in a possible implementation manner, the blood cell intelligent classification unit is specifically configured to: perform normalization processing on the active chemotaxis index and the single cell contour deformation degree; and determine a product of the normalized results of the active chemotaxis index and the single cell contour deformation degree as the dynamic function fingerprint.

[0011] With reference to the first aspect, in a possible implementation manner, the blood cell intelligent classification unit is specifically configured to: construct a spatial topology network based on the single-cell bounding box coordinates of the single-cell image block; analyze the motion direction of the single cell from the single-cell motion trajectory point set; calculate the similarity of the motion direction of any two single cells in the spatial topology network, and the distance change rate of the center coordinate distance with time; correct the similarity of the motion direction by the distance change rate to obtain an interaction synergy coefficient, and statistically obtain the mean value of the interaction synergy coefficient of each single cell and the adjacent cells as an adjacent cell interaction synergy factor.

[0012] With reference to the first aspect, in a possible implementation manner, the blood cell intelligent classification unit is specifically configured to: construct a spatial topology network based on the single-cell bounding box coordinates of the single-cell image block; analyze the motion direction of the single cell from the single-cell motion trajectory point set; calculate the similarity of the motion direction of any two single cells in the spatial topology network, and the distance change rate of the center coordinate distance with time; correct the similarity of the motion direction by the distance change rate to obtain an interaction synergy coefficient, and statistically obtain the mean value of the interaction synergy coefficient of each single cell and the adjacent cells as an adjacent cell interaction synergy factor.

[0013] With reference to the first aspect, in a possible implementation manner, the blood cell intelligent classification unit is specifically configured to: calculate the Pearson correlation coefficient of the dynamic function fingerprint and the adjacent cell interaction synergy factor; and generate a comprehensive feature index based on the dynamic function fingerprint, the adjacent cell interaction synergy factor, and the Pearson correlation coefficient.

[0014] With reference to the first aspect, in a possible implementation manner, the data acquisition platform comprises: an optical imaging unit, a motion control unit, and an image acquisition unit; the optical imaging unit is configured with a high-resolution charge coupled device (CCD) image sensor, a 100-fold oil immersion objective lens, and an automatic stage; the motion control unit is configured to drive the optical imaging unit to perform Z-axis multi-layer scanning to obtain a blood image sequence, and to synthesize the blood image sequence into an original blood cell image by using an image fusion algorithm; and the image acquisition unit is configured to control the optical imaging unit to acquire the blood image sequence at a preset frame rate.

[0015] With reference to the first aspect, in a possible implementation manner, the data acquisition platform further comprises: an image processing unit; the image processing unit is configured to perform image denoising and image enhancement on the original blood cell image to obtain a standardized image; and the image processing unit is further configured to locate and crop the image region of a single blood cell in the standardized image, and perform normalization processing on the cropped image region to obtain a single-cell image block with uniform size.

[0016] In a second aspect, an automatic blood cell image classification device based on deep learning is provided, comprising a processor and a storage medium; the storage medium comprises instructions, and the processor is configured to execute the instructions to implement the actions described in the first aspect and any possible implementation manner of the first aspect. The automatic blood cell image classification device based on deep learning can be an electronic device or a chip in an electronic device.

[0017] In a third aspect, a computer-readable storage medium is provided, and the computer-readable storage medium stores instructions, which, when executed on an automatic blood cell image classification device based on deep learning, cause the automatic blood cell image classification device based on deep learning to perform the actions described in the first aspect and any possible implementation manner of the first aspect.

[0018] In a fourth aspect, a computer program product containing instructions is provided, which, when executed on an automatic blood cell image classification device based on deep learning, cause the automatic blood cell image classification device based on deep learning to perform the actions described in the first aspect and any possible implementation manner of the first aspect.

[0019] The present application has the following beneficial effects:

[0020] The present application breaks through the limitation of traditional morphological features, innovatively extracts morphological features and motion features combined with cell dynamic behavior features, and interaction features of cells in a microenvironment, and constructs a multi-modal feature system. Among them, the dynamic behavior features can reflect the functional related attributes such as cell activity and motion directionality, and the microenvironment interaction features can reflect the functional role of cells in the colony, which enriches the feature dimension and makes the features more comprehensive to the physiological and pathological state of cells. Then, combining with the output classification result of the pre-training classification model, the learning ability of the model to multi-dimensional features can be utilized to improve the reliability of classification decision. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0022] Figure 1 A system architecture diagram of an automatic blood cell image classification system based on deep learning provided by an embodiment of the present application is shown in the figure;

[0023] Figure 2 A system architecture diagram of a data acquisition platform provided by an embodiment of the present application is shown in the figure;

[0024] Figure 3 A schematic diagram of a raw blood cell image provided by one embodiment of the present application;

[0025] Figure 4 A schematic diagram of an image region for locating a single blood cell provided by one embodiment of the present application;

[0026] Figure 5 A schematic diagram of a hardware structure of a blood cell image automatic classification device based on deep learning provided by one embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined purposes, the following describes in detail the specific implementation, structure, features and effects of a blood cell image automatic classification system based on deep learning according to the present application, with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0028] 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 the present application belongs.

[0029] The specific scheme of the blood cell image automatic classification system based on deep learning provided by the present application is specifically described below with reference to the accompanying drawings.

[0030] Please refer to Figure 1 , which shows a system architecture diagram of a blood cell image automatic classification system based on deep learning provided by one embodiment of the present application. The blood cell image automatic classification system based on deep learning includes a data acquisition platform, a blood cell intelligent classification unit and a classification output unit.

[0031] The data acquisition platform is used to acquire raw blood cell images and extract single-cell image blocks containing single blood cells.

[0032] In some implementations, in combination with Figure 1 , as shown in Figure 2 , the above data acquisition platform specifically includes an optical imaging unit, a motion control unit and an image acquisition unit.

[0033] The optical imaging unit is configured with a high-resolution charge-coupled device image sensor (with a resolution of no less than 2048*2048 pixels), a 100-fold oil immersion objective lens, and an automatic stage, and can clearly capture the subtle morphological features (such as cell membrane texture and cell nucleus lobulation structure) of blood cells in the collection tube, thereby providing high-quality original data for subsequent cell morphological feature extraction and ensuring the accuracy of image acquisition.

[0034] The motion control unit is used to drive the optical imaging unit to perform Z-axis multi-layer scanning (such as with a step size of 0.5 pm) to obtain a blood image sequence, and to synthesize the original blood cell image by means of an image fusion algorithm, so that the image sequence of blood cells at different focal planes can be obtained, and a clear image of the whole field of view can be synthesized by means of the image fusion algorithm (as shown in FIG. 1), thereby solving the problem of out-of-focus of part of the cell structure in single focal plane imaging and ensuring that all cell structures can be clearly presented. Figure 3

[0035] The image acquisition unit is used to control the optical imaging unit to acquire the blood image sequence at a preset frame rate (for example, 30 frames / s), so that the dynamic motion process of blood cells (such as the migration of white blood cells and the aggregation of platelets) can be captured, thereby providing continuous time series data for extracting cell motion features and breaking through the limitation that static imaging cannot reflect the dynamic behavior of cells.

[0036] The optical imaging unit, the motion control unit, and the image acquisition unit are organically coordinated to build a complete technical chain from optical imaging to dynamic data acquisition, and realize multi-dimensional blood cell information acquisition combining high-resolution static morphology and continuous dynamic behavior. Such coordination not only solves the problem that the traditional imaging system cannot simultaneously consider static details and dynamic processes, but also provides comprehensive data support for subsequent multi-modal feature extraction (such as fusion of morphological features and motion features), and finally improves the perception ability of the blood cell automatic classification system to the functional state of cells and the classification accuracy.

[0037] Further, as shown in FIG. 1, the data acquisition platform described above further includes an image processing unit. Figure 2

[0038] The image processing unit is used to perform image denoising and image enhancement on the original blood cell image to obtain a standardized image.

[0039] In some implementations, an automatic focusing evaluation function and a sharpness detection algorithm (such as a Tenengrad function, a Brenner function, or a Laplacian variance algorithm) are used to eliminate image frames that are severely out of focus, have motion blur, or are excessively contaminated. An adaptive threshold segmentation algorithm can also be applied to remove background noise and retain the cell region.

[0040] ​​Based on the above actions, low-quality data can be effectively filtered out, making the cell region clearly distinguishable from the background, providing a clean image basis for subsequent feature extraction, and reducing the interference of invalid data on the analysis results.

[0041] The image processing unit is also used to locate and crop the image region of a single blood cell in the standardized image, and to normalize the cropped image region to obtain a single-cell image block of uniform size.

[0042] In some implementations, such as Figure 4 As shown, deep learning-based detection models (such as the YOLO algorithm) or traditional image processing algorithms (such as edge detection) are used to locate and crop image regions containing complete single target blood cells (white blood cells, red blood cells, or platelets), and generate corresponding coordinate bounding box data. Then, the images are cropped and normalized to generate single-cell image patches of uniform size and clean background, which constitute the dataset used for model training and classification.

[0043] Based on the above actions, image differences caused by different acquisition batches and different equipment parameters can be eliminated, ensuring that all samples are analyzed at the same scale, and improving the stability and comparability of feature extraction.

[0044] The blood cell intelligent classification unit is used to extract and fuse multimodal features from single-cell image blocks to obtain comprehensive feature indicators.

[0045] Among them, multimodal features include: cell dynamic behavior features that include morphological and kinematic features, and cell interaction features in the microenvironment.

[0046] In some implementations, morphological feature extraction may include: extracting 12 morphological parameters such as cell area, perimeter, roundness, and concavity / convexity based on single-cell image patches; extracting cell membrane contours using edge detection algorithms and calculating the rate of change of contour curvature; and detecting cell nucleus morphological features, including nucleocytoplasmic ratio and number of nuclear lobes (for leukocytes), through Hough transform.

[0047] Motion feature extraction may include: calculating cell displacement vectors between consecutive frames using optical flow, and calculating instantaneous velocity, average velocity, and chemotactic index based on the displacement vectors; extracting cell membrane feature points through corner detection, and calculating contour deformation.

[0048] Interaction feature extraction may include: constructing a Delaunay triangulation based on the cell center coordinates to determine the neighboring cells of each cell (distance <50μm); constructing a cell-cell association matrix to record the spatial positional relationship between cells; and calculating motion correlation index, distance change rate, interaction coordination coefficient, etc. based on the spatial positional relationship.

[0049] a classification output unit configured to input the comprehensive feature index into a pre-trained classification model to obtain an automatic classification result of the single blood cell.

[0050] In some implementations, based on the comprehensive feature index of each single cell and its associated features (cell segmentation mask morphology, size, etc.), as key spatiotemporal dynamic features, the pre-trained machine learning model (such as gradient boosting decision tree, temporal convolutional network or graph neural network) is input: the model learns the mapping relationship between the comprehensive feature index and the single cell type on the blood cell image through offline training, and generates the automatic classification result of the blood cell image in the current time window.

[0051] In some implementations, an improved temporal convolutional network (TCN) is used, which includes 3 convolutional blocks and 2 fully connected layers. The training data set contains at least 100,000 labeled samples, covering 8 types of cells such as red blood cells, white blood cells (5 subtypes), and platelets, and is divided into training set, validation set and test set according to 7:2:1.

[0052] The training process includes: inputting the comprehensive feature index sequence (extracting 1 feature value every 30 frames); using Adam optimizer, initial learning rate 0.001, and 32 samples selected for one training; training for 50 rounds, using early stopping strategy (stopping if the validation set accuracy does not improve for 3 rounds).

[0053] The classification process includes: real-time extraction of the comprehensive feature index F of the cell to be classified, inputting F into the pre-trained TCN model, and outputting the prediction probability of each category; using the maximum probability principle to determine the cell category, and setting the confidence threshold to 0.7; marking low-confidence samples (confidence <0.7) as pending review and triggering the manual verification process.

[0054] The output cell category, confidence, and key feature parameters (such as activity chemotaxis index and interaction synergy coefficient) are output; a classification result report is generated, including cell quantity statistics and abnormal cell labeling; and the system is connected to the laboratory information system to realize automatic data uploading and sharing.

[0055] Based on the above technical solution, the limitations of traditional methods that rely only on morphological features are overcome, and the dynamic behavior features of cells combined with morphological features and interaction features of cells in the microenvironment are innovatively extracted to build a multi-modal feature system. The dynamic behavior features can reflect the functional properties of cell activity and directional movement, and the microenvironment interaction features can reflect the functional role of cells in the colony, enriching the feature dimension and making the features more comprehensive and consistent with the physiological and pathological state of cells. Then, the pre-trained classification model outputs the classification result, which can utilize the learning ability of the model on multi-dimensional features to improve the reliability of classification decision.

[0056] In a possible implementation manner, the blood cell intelligent classification unit is specifically configured to perform the following S21 to S23, which are specifically described as follows.

[0057] S21, obtain a single cell motion trajectory from a sequence of multiple frames of continuous single cell image blocks, and analyze the single cell motion trajectory to obtain an active chemotaxis index.

[0058] The motion activity and directionality of cells are core signs of their execution of immune defense, chemotactic response and other functions, and pathological states (such as infection, inflammation) can significantly change the motion mode of cells (such as from random patrol to directional chemotaxis).

[0059] In some implementation manners, for a single cell in each frame of image, a flow method (Lucas-Kanade algorithm) is used to calculate a displacement vector of the single cell in adjacent frames, and an ORB (Oriented FAST and Rotated BRIEF) feature point matching (emphasizing matching cell membrane edge stable feature points) is used to correct displacement deviation, so as to realize accurate tracking across frames.

[0060] S22, extract a cell contour feature of each frame of single cell image block in the sequence of multiple frames of continuous single cell image blocks to obtain a single cell contour deformation degree.

[0061] The dynamic stability of cell morphology is closely related to the function execution (such as pseudopod extension of activated platelets and morphological deformation of phagocytic macrophages), and cells with abnormal functions under pathological conditions often have dramatic morphological dynamic changes.

[0062] S23, generate a dynamic function fingerprint based on the active chemotaxis index and the single cell contour deformation degree.

[0063] The function state of a cell needs to be reflected through the synergistic effect of its motion characteristics and morphological changes, and therefore, based on the active chemotaxis index and the single cell contour deformation degree, a dynamic function fingerprint is generated, which can quantify the comprehensive characteristics of the function state of the cell.

[0064] Based on the above technical solution, the traditional static image analysis is upgraded to dynamic behavior analysis, the cell activity state is captured through the motion trajectory and the contour change, and a technical leap from morphological observation to functional perception is realized. Then, the abstract cell activity is converted into a calculable digital index through the active chemotaxis index and the contour deformation degree, and a technical difficulty that the cell function state is difficult to be quantified by traditional methods is solved. The dynamic function fingerprint integrates the dual information of cell motion directionality and morphological stability, can effectively distinguish cells with similar morphology but different function states, and improves the feature discrimination ability.

[0065] In a possible implementation manner, the blood cell intelligent classification unit is specifically configured to perform the following S31 to S33, which are specifically described as follows.

[0066] S31, for a plurality of frames of continuous single-cell image block sequence, the single-cell motion trajectory point set is obtained by optical flow method and feature matching.

[0067] In some implementations, the geometric center of the first frame single-cell is taken as the initial coordinate , the displacement vector of the corresponding single-cell in the adjacent frame is calculated by the optical flow method, and the matching result of the ORB feature point (cell membrane edge) is combined to obtain the single-cell center coordinate is the total number of frames in the sequence); and the trajectory point set of each single-cell is output: .

[0068] The optical flow method can calculate the displacement information of the cells between adjacent frames based on the gray level change of the pixels in the image, and the ORB feature point matching can extract stable feature points from key positions such as cell membrane edges, and the displacement calculation result is further verified and optimized through the matching of the feature points. Combined with both, the continuity of the optical flow method is used to track the cell movement trend, and the accuracy of the ORB feature points is used to ensure the precision of the trajectory points, so as to accurately obtain the motion trajectory point set of the single-cell in the continuous frames, provide basic data support for subsequent analysis of the displacement and distance of the cell movement, and solve the problem that the motion trajectory of the single-cell in the dynamic image sequence is difficult to accurately capture.

[0069] S32, the final displacement and moving distance of the single-cell are analyzed from the single-cell motion trajectory point set.

[0070] In some implementations, the final displacement of each cell in the overall image sequence data is respectively obtained and the total path length , is the number of adjacent frames in the sequence).

[0071] The final displacement reflects the straight-line distance from the initial position to the final position of the cell, and embodies the overall position change of the cell movement.

[0072] The total path length is the sum of the moving distances between the consecutive adjacent frames, and represents the path length actually moved by the cell.

[0073] S33, according to the ratio of the modulus of the final displacement and the moving distance, the activity chemotaxis index is determined.

[0074] In some implementations, the final displacement is associated with the total path length, and the activity chemotaxis index of each cell is output :

[0075] ; ​

[0076] When the cell movement direction is stable and has strong directionality, the ratio of the total path length to the length of the final displacement tends to 1, and at this time tends to 0; when the cell movement has no obvious directionality and is relatively random, the total path length is much greater than the length of the final displacement, tends to 1. In this way, the directionality of cell movement, which is an abstract functional characteristic, is converted into a quantifiable numerical indicator. Through the size, it can be directly judged whether the cell has a strong chemotactic response, and the smaller the value, the more stable the cell movement direction and the stronger the directionality, which is a sign of a strong chemotactic response of the cell.

[0077] Based on the above technical solution, a complete cell dynamic behavior quantification system is constructed, the cell functional state is accurately described by converting the cell movement trajectory into a calculable active chemotaxis index, and the limitation of traditional static morphological analysis is broken through, thereby providing a key dynamic characteristic basis for blood cell functional classification.

[0078] In a possible implementation manner, the blood cell intelligent classification unit is specifically configured to execute the following S41 to S43, which are described in detail as follows.

[0079] S41, for each single cell image block, the contour feature corner points are extracted by an angle point detection algorithm, and a corner point set of each single cell image block is output.

[0080] In some implementation manners, for each single cell image block in the image sequence, a Shi-Tomasi corner point detection algorithm is used to extract contour feature corner points, and a corner point set of each mask is output. , is the number of corner points of the current mask.

[0081] The Shi-Tomasi corner point detection algorithm can preferentially retain the corner points with significant curvature changes on the cell membrane edge, and these corner points are key feature points of the cell contour, which can accurately capture the shape features of the cell contour.

[0082] S42, the Euclidean distance between any two contour feature corner point pairs is calculated, the Euclidean distances of multiple corner point pairs in the corner point set are averaged, and the contour stretch degree of each single cell image block is obtained.

[0083] In some implementation manners, for each corner point set, the Euclidean distances between all corner point pairs are calculated, and the arithmetic mean value reflects the overall stretch degree of the cell contour of the current frame.

[0084] The spatial distribution information of the corner points is converted into a value reflecting the overall stretching degree of the cell contour. In this way, the shape features of the cell contour (reflected by the distribution of the corner points) are quantified, so that the contour stretching conditions of different cells or the same cell at different times can be compared numerically, and a single-frame quantitative index is provided for subsequent analysis of the dynamic changes of the cell contour.

[0085] S43, the variance of the contour stretching degrees corresponding to the plurality of continuous single-cell image block sequences is determined as the single-cell contour deformation degree.

[0086] The variance of the value set corresponding to each image segmentation mask in the sequence data is calculated , and is defined as the single-cell contour deformation degree.

[0087] The variance reflects the dispersion degree of the data, and can quantify the dynamic stability of the cell contour. The greater the value, the more intense the contour jitter and deformation of the cell in a short period of time, which is a sign of active function. It can make up for the deficiency that the dynamic functional activity of the cell cannot be judged by only using the single-frame contour features, and provides an important dynamic quantitative basis for analysis and classification based on the functional state of the cell.

[0088] Based on the above technical solutions, the traditional static morphological analysis is expanded to dynamic morphological change analysis, and a quantitative correlation between cell morphological dynamic change and functional activity is established, which provides a key dynamic morphological dimension for multi-modal feature fusion. It can be applied to scenarios that require functional activity evaluation (such as immune cell activation detection, tumor cell malignancy judgment, etc.), and significantly improves the perception ability of the blood cell classification system to the functional state of the cell.

[0089] In one possible implementation, the blood cell intelligent classification unit is specifically configured to perform the following S51 to S52, which will be described in detail as follows:

[0090] S51, the active chemotaxis index and the single-cell contour deformation degree are normalized.

[0091] In some implementations, the active chemotaxis index and the cell contour deformation degree corresponding to each single cell on the blood cell image are standardized, including: the values of and are mapped to the interval by a normalization function .

[0092] The purpose is to eliminate the incomparability caused by the difference in dimensions or the large difference in numerical range of the two indicators. For example, the active chemotaxis index may fluctuate in a large numerical range, while the contour deformation degree may change in a relatively small range. After normalization, they can be at the same weight level in subsequent calculation and analysis, facilitating fusion and other operations, and laying a foundation for subsequent generation of dynamic function fingerprints that can comprehensively reflect the function state of cells.

[0093] S52, the product of the normalized results of the active chemotaxis index and the single cell contour deformation degree is determined as the dynamic function fingerprint.

[0094] The dynamic function fingerprint of each single cell , which comprehensively reflects the cell's motion directionality (reflected by the active chemotaxis index) and morphological activity (reflected by the contour deformation degree), is used to quantify the active and directional comprehensive characteristics of the cell's function state. Among them, and respectively represent the normalized results of the active chemotaxis index and the cell contour deformation degree .

[0095] When the value of is high, it indicates that the cell has both directional motion characteristics and relatively active morphological changes, which is consistent with the performance of functional active cells such as activated immune cells. By this way of product, the function characteristics of two dimensions of cell motion and morphology are fused into a single quantitative indicator, which can more simply and intuitively reflect the comprehensive function state of cell activity and directionality, providing a key quantitative basis for subsequent classification, recognition, and other applications based on cell function state, and enabling the judgment of cell function to expand from a single dimension (motion or morphology) to a comprehensive dimension.

[0096] Based on the above technical solution, the fusion problem caused by the difference in multi-dimensional characteristics is solved, the digital fingerprint of the cell function state is constructed, the precise quantification of the function characteristics is realized, and the high-discriminative core input feature is provided for the subsequent classification model. The blood cell analysis is upgraded from morphology recognition to function perception, which is suitable for scenarios that require function state evaluation (such as infectious disease immune state monitoring, tumor microenvironment analysis, etc.), significantly expanding the clinical application value of the blood cell classification system.

[0097] In one possible implementation, the blood cell intelligent classification unit described above is further configured to perform the following S61, which will be described in detail as follows:

[0098] S61, determining the instantaneous speed of the single cell motion according to the set of single cell motion trajectory points and the time interval of consecutive adjacent frames.

[0099] For each single cell, the instantaneous speed of consecutive adjacent frames in its image sequence is calculated based on the set of trajectory points of the single cell The single-cell motion trajectory point set records the position information of the cell in consecutive frames, and the instantaneous speed of the cell in each adjacent frame interval can be calculated by combining the time interval of the consecutive adjacent frames The instantaneous speed reflects the speed of the cell at a certain moment, and through this calculation, the dynamic change of the speed of the cell can be captured, such as whether the cell is accelerating, decelerating or maintaining uniform motion, etc., which provides the most basic speed data for subsequent analysis of the motion activity of the cell.

[0100] S62, based on the instantaneous speed, optimizing the active chemotaxis index.

[0101] and then obtaining the arithmetic mean value of the sequence , reflecting the overall activity of the cell (the larger the value, the stronger the activity of the cell). The arithmetic mean value is a comprehensive statistics of the instantaneous speed of the cell in the entire image sequence, which can eliminate the influence of possible local fluctuations of the instantaneous speed, and reflects the activity of the cell motion as a whole. Because if the cell motion is more active, the value of the instantaneous speed will be larger, and the corresponding arithmetic mean value will also be larger; on the contrary, if the cell motion is not active, the arithmetic mean value will be smaller. In this way, the activity of the cell motion is quantified, so that the activity of the motion of different cells or the same cell in different states can be compared through this value, which provides a key quantitative basis for subsequent analysis and classification based on the activity of the cell motion.

[0102] The speed variance of the single-cell motion is calculated , reflecting the dynamic stability of the speed, and a large variance indicates strong speed fluctuation, which may be a marker of active cell function.

[0103] The average speed and the speed variance are normalized (mapped to the interval), and then fused into the motion vigor index . Among them, and respectively represent the normalization results of the instantaneous speed and the speed variance .

[0104] The motion vigor index and the active chemotaxis index are weighted and fused to generate the optimized active chemotaxis index . The weight coefficient can be determined through experiment or model training, for example, the weight coefficient of the active chemotaxis index in the clinical sample is 0.6, which gives priority to the directional core meaning of the active chemotaxis index , while supplementing the vigor information of the motion vigor index .

[0105] Based on the above technical scheme, by accurately quantifying the instantaneous motion speed of the cells, the dynamic characteristics of the cell motion are captured, which provides important basic data for subsequent cell motion activity analysis, abnormal cell recognition and multi-modal feature fusion, is one of the key links to realize the function from cell morphology analysis to function state perception, and significantly improves the perception ability of the dynamic function state of the blood cells.

[0106] In one possible implementation manner, the blood cell intelligent classification unit is specifically configured to perform the following S71 to S74, which are described below.

[0107] S71, constructing a spatial topology network based on the single cell boundary box coordinates of the single cell image block.

[0108] The function state of the cells in the blood is affected by the local cell colony (for example, the neutrophils gather to enhance the bactericidal efficiency during inflammation, and the platelets and red blood cells adhere to each other during blood coagulation). The isolated analysis of a single cell cannot reflect its function role in the colony, and the dynamic relationship of the adjacent cells needs to be used to reveal the function synergy or antagonism state.

[0109] In some implementation manners, based on the single cell boundary box coordinates, a spatial topology network of each frame of blood cell image is constructed, the nodes in the spatial topology network are cells, the node positions are determined by the geometric center coordinates of the single cell boundary box coordinates (i.e., the center point of the boundary box), and the connection line between any two nodes is the intercellular distance. It should be noted that the adjacent cells need to satisfy the intercellular distance less than the distance threshold, and are connected by edges in the topology network, so as to ensure that only the cells adjacent in space are connected. The distance threshold can be set according to the cell density in the collection tube and the research target, for example, for white blood cell analysis, the distance threshold can be set to 2 times the average diameter of the cells.

[0110] The spatial topology network is constructed based on the single cell boundary box coordinates, and a single cell is analyzed in the spatial relationship of the cell colony. The function of the cells in the blood is often affected by the local colony, such as the aggregation of neutrophils to kill bacteria during inflammation, and the adhesion of platelets and red blood cells during blood coagulation, and the isolated analysis of a single cell cannot reflect its function role in the colony. The spatial topology network takes the cells as nodes and the intercellular distance as edges, provides a spatial structure framework for subsequent analysis of the dynamic relationship (such as synergy and antagonism) between the cells, and enables the research on the function of the cells to expand from the single cell level to the cell colony level.

[0111] S72, analyzing the motion direction of the single cell from the single cell motion trajectory point set.

[0112] The motion trajectory point set of each single cell in the blood cell image sequence data is obtained, the motion direction vector of each cell in each frame is calculated. The vector is normalized to a unit vector , representing the moving direction of the cell. Wherein, is the center coordinate of the cell in the frame , and is the center coordinate of the cell in the next frame . This provides a basis for subsequent analysis of the similarity of the moving direction between cells, and can clearly determine whether the cells are moving towards or away from a certain area, or showing random moving direction, which is a key prerequisite for studying the motion coordination of cell populations.

[0113] S73, calculating the similarity of the moving direction of any two single cells in the spatial topological network, and the distance change rate of the center coordinate distance over time.

[0114] For all connected cells of the cell in the spatial topological network , the cosine similarity of the moving direction vector of the two cells on the time series data is calculated: , defined as the motion correlation index between the two cells, with a value range . The closer to 1 indicates that the moving direction of the cells is consistent (such as neutrophils moving towards the infection focus during inflammation), close to -1 indicates that the moving direction of the cells is opposite (such as normal cells avoiding abnormal cells), and close to 0 indicates that there is no significant correlation between the motion of the cells (such as randomly distributed red blood cells).

[0115] Further, the same moving direction coordination between cells may correspond to different functional patterns: for example, cells moving in the same direction continue to approach, which represents adhesion aggregation (platelet aggregation after activation); if the distance is maintained, it represents co-migration (such as neutrophil directional chemotaxis). Therefore, combined with the distance change trend between different cells, the interaction type between high coordination moving cells is judged, which specifically includes:

[0116] Based on the center coordinate distance between the cell and the corresponding adjacent cell in each frame image and the image sequence acquisition time interval , the distance change rate between the cells is calculated, wherein is the adjacent frame in the sequence.

[0117] When the distance change rate is negative, it indicates that the two cells are approaching each other, which may be adhesion or chemotaxis behavior under coordination; when the value is positive, it indicates that the two cells are moving away from each other, which may be the separation or mutual repulsion of cells that no longer cooperate.

[0118] S74, the similarity of the motion direction is corrected by the distance change rate, an interaction synergy coefficient is obtained, and the average of the interaction synergy coefficients of each single cell and adjacent cells is taken as an adjacent cell interaction synergy factor.

[0119] The motion correlation index between two cells is corrected by the distance change rate, and the final cell interaction synergy coefficient is output: .

[0120] is a symbol function, When the value is negative, the function value is -1; when the value is positive, the function value is 1.

[0121] The motion direction similarity is corrected by the distance change rate, the relationship between the motion direction and the position change between cells is integrated, and the functional interaction state between cells is quantified. The greater the value, the higher the intensity and heterogeneity of the functional interaction (synergy or pathology) between cells, corresponding to the pathological phenomenon of efficient synergistic migration or strong abnormal invasion.

[0122] Further, the average of the cell interaction synergy coefficients between each single cell and all connected cells on the blood cell image is taken as the adjacent cell interaction synergy factor of the corresponding single cell, which can reflect the overall interaction behavior abnormality of the cell in the microenvironment. The greater the value, the more abnormal the interaction behavior of the corresponding single cell in the microenvironment, which will be used as an important feature input of the classification model.

[0123] Based on the above technical solutions, a quantitative system of intercellular functional interaction is constructed, and the functional state of the cell colony is accurately described through the fusion analysis of the spatial topological network and the motion characteristics, which breaks through the limitation of traditional single cell analysis that cannot reflect the synergistic / antagonistic relationship between cells, and provides core technical support for blood cell microenvironment function analysis.

[0124] In one possible implementation, the blood cell intelligent classification unit is specifically configured to perform the following S81 to S82, which will be described in detail below.

[0125] S81, calculate the Pearson correlation coefficient of the dynamic function fingerprint and the adjacent cell interaction synergy factor.

[0126] Generally, the functional state of a cell depends not only on its own dynamic behavior, but also on the influence of other cells in its microenvironment.

[0127] In some implementations, for each cell in the image sequence, the dynamic function fingerprint and the corresponding adjacent cell interaction synergy factor are extracted every 30 frames to construct a one-to-one data set and Z-score data normalization is performed.

[0128] Computing data set Pearson correlation coefficient between , the value range is The value quantifies the linear correlation strength between the cell's own functional state and the microenvironment interaction.

[0129] The value is closer to 1, indicating that the cell's own functional activity (such as high motility, morphological variability) and strong microenvironment interaction (such as cooperative migration or aggregation) occur synchronously. Commonly seen in activated immune cells (such as the directional chemotaxis and aggregation behavior of neutrophils near the infection site), which is a typical manifestation of their efficient immune response;

[0130] The value is closer to -1, indicating that the cell's own function is stable (low motility, morphological stability) but is in a strong interactive microenvironment (such as being actively repelled or invaded by other cells). Commonly seen in passive avoidance and strong antagonism of normal blood cells when disturbed by cancer cells or abnormal cells, which is a sign of pathological microenvironment invasion of normal tissues;

[0131] The value is close to 0, indicating that there is no significant linear correlation between the cell's own functional state and the microenvironment interaction behavior, commonly seen in red blood cells randomly moving in physiological state or platelets in resting state, whose behavior is mainly affected by blood flow dynamics rather than active functional regulation.

[0132] S82, based on dynamic functional fingerprint, adjacent cell interaction synergy factor and Pearson correlation coefficient, generate comprehensive feature index.

[0133] To further improve the classification system's understanding of the cell's functional state and microenvironment context, the aforementioned multi-dimensional features are fused to construct a comprehensive feature index that reflects the cell's own and environmental synergy state, and input into the classifier for final decision.

[0134] In some implementations, the system inputs each cell Dynamic functional fingerprint of the last frame of the dynamic image sequence and adjacent cell interaction synergy factor , outputs the classification model important feature index of each cell :

[0135] ;

[0136] ​Add the cell's own dynamic activity and the microenvironment interaction intensity to construct a preliminary comprehensive functional activity index. The higher the additive value, the more active the cell shows in its own behavior and environmental interaction, and it may be in a significant functional execution state (such as immune response, blood clotting activation, tumor invasion, etc.).

[0137] Further multiplied by the linear correlation strength between the cell's own functional state and the microenvironment interaction behavior to effectively distinguish true functional activity from passive abnormal response, and avoid the misjudgment risk brought by the additive value alone.

[0138] The value of is a larger positive value, indicating that the cell is highly active and has strong synergy with the microenvironment, consistent behavior, and is in a high-efficiency functional execution state (neutrophils in inflammation, activated platelet aggregation);

[0139] When the value of is a smaller negative value, it indicates that the cell is stable but is strongly disturbed or excluded by the microenvironment, and is in a passive abnormal state (normal cells surrounded by tumor cells, immune cells attacking self cells);

[0140] When the value of is close to 0, the cell's own behavior has no significant correlation with environmental interaction, and is mostly in a resting or random motion state (mature red blood cells, resting lymphocytes, non-functional activated platelets).

[0141] Based on the above technical solution, through functional and interaction correlation analysis and multi-dimensional feature fusion, a comprehensive feature index reflecting the synergistic relationship between the cell's own functional state and the microenvironment interaction is constructed, breaking through the limitations of traditional single feature or independent feature analysis, providing a more clinically practical function and environment dual-dimensional judgment basis for blood cell function classification, and significantly improving the precision perception and classification reliability of the cell's functional state.

[0142] It should be noted that the above-mentioned order of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or can be advantageous.

[0143] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other, and each embodiment mainly describes the differences from other embodiments.

[0144] The embodiment of the present application can divide the function units of the blood cell image automatic classification device based on deep learning according to the above method examples. For example, each function unit can be divided according to each function, or two or more functions can be integrated in one processing unit. The integrated unit can be realized in the form of hardware or in the form of a software function unit. It should be noted that the division of the unit in the embodiment of the present application is illustrative, and is only a logical function division. In actual implementation, another division mode can be used.

[0145] The embodiment of the present application also provides a hardware structure diagram of a blood cell image automatic classification device based on deep learning. Referring to Figure 5 , the blood cell image automatic classification device 500 based on deep learning includes a processor 501, and optionally, a memory 502 connected with the processor 501.

[0146] In the first possible implementation mode, referring to Figure 5 , the blood cell image automatic classification device 500 based on deep learning further includes a transceiver 503. The processor 501, the memory 502 and the transceiver 503 are connected through a bus. The transceiver 503 is used for communicating with other devices or communication networks. Optionally, the transceiver 503 can include a transmitter and a receiver. The device for realizing the receiving function in the transceiver 503 can be regarded as a receiver, and the receiver is used for executing the receiving steps in the embodiment of the present application. The device for realizing the sending function in the transceiver 503 can be regarded as a transmitter, and the transmitter is used for executing the sending steps in the embodiment of the present application.

[0147] Based on the first possible implementation mode, Figure 5 , the structure diagram shown can be used to illustrate the structure of the blood cell image automatic classification device based on deep learning involved in the above embodiment.

[0148] , the structure diagram shown can be used to illustrate the structure of the blood cell image automatic classification device based on deep learning involved in the above embodiment. Figure 5 The system chip in the blood cell image automatic classification device based on deep learning can also be illustrated. In this case, the actions performed by the blood cell image automatic classification device based on deep learning can be realized by the system chip, and the specific actions performed can be referred to in the above, and will not be described here.

[0149] In the implementation process, each step in the method provided by the embodiment of the present application can be completed by the integrated logic circuit of hardware in the processor or the instruction in the form of software. The steps of the method disclosed in the embodiment of the present application can be directly embodied as hardware processor execution completion, or executed by the combination of hardware and software modules in the processor.

[0150] The processor in the present application can include, but is not limited to, at least one of the following: a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a micro controller unit (MCU), or an artificial intelligence processor, and the like, each of which is a computing device running software, and each of which can include one or more cores for executing software instructions to perform operations or processing. The processor can be a separate semiconductor chip, or can be integrated with other circuits as a semiconductor chip, for example, can be integrated with other circuits (such as coding and decoding circuits, hardware acceleration circuits, or various bus and interface circuits) to form a SoC (system on chip), or can be integrated as a built-in processor in an ASIC. The ASIC integrated with the processor can be packaged separately or packaged together with other circuits. In addition to including cores for executing software instructions to perform operations or processing, the processor can further include necessary hardware accelerators, such as field programmable gate arrays (FPGAs), PLDs (programmable logic devices), or logic circuits implementing special logic operations.

[0151] The memory in the embodiments of the present application can include at least one of the following types: a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM). In some scenarios, the memory can also be a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer, but is not limited thereto.

[0152] The embodiments of the present application also provide a computer readable storage medium including instructions, which, when executed on a computer, cause the computer to perform any of the above methods.

[0153] The embodiments of the present application also provide a computer program product including instructions, which, when executed on a computer, cause the computer to perform any of the above methods.

[0154] The embodiment of the present application further provides a chip, comprising a processor and an interface circuit, the interface circuit being coupled with the processor, the processor being used for running computer programs or instructions to realize the method, and the interface circuit being used for communicating with other modules outside the chip.

[0155] In the above embodiments, the implementation can be achieved by software, hardware, firmware or any combination thereof, entirely or partially. When implemented by software, the implementation can be achieved in the form of a computer program product, entirely or partially. The computer program product comprises one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the flow or function described in the embodiments of the present application is generated, entirely or partially. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable apparatus. The computer instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (for example, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (for example, infrared, wireless, microwave, etc.) mode. The computer readable storage medium can be any available medium that can be accessed by a computer or data storage device such as one or more servers, data centers, etc. integrated with one or more media. The available medium can be a magnetic medium (for example, floppy disk, hard disk, magnetic tape), an optical medium (for example, DVD) or a semiconductor medium (for example, solid state disk (SSD)) and the like.

[0156] Although the present application is described herein in conjunction with various embodiments, other variations of the disclosed embodiments can be understood and implemented by those skilled in the art with reference to the drawings, the disclosure, and the specific examples. In the present application, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality. A single processor or other unit can implement several functions listed in the present application.

[0157] Although the present application has been described in connection with the preferred embodiments thereof with reference to the specific content thereof, it will be apparent to those skilled in the art that various modifications and changes can be made thereto without departing from the spirit and scope of the application. Accordingly, the description and drawings are to be regarded as illustrative in nature and are not restrictive. It will be apparent that those skilled in the art can modify and adapt the application without departing from the spirit and scope of the application. Accordingly, such modifications and variations are intended to fall within the scope of the application and its equivalents.

Claims

1. A deep learning-based blood cell image automatic classification system, characterized in that, The application relates to a blood cell intelligent classification system. The data acquisition platform is used for acquiring original blood cell images and extracting single-cell image blocks containing single blood cells. The blood cell intelligent classification unit is used for multi-modal feature extraction and multi-modal feature fusion of the single-cell image blocks to obtain comprehensive feature indexes; the multi-modal features include cell dynamic behavior features containing morphological features and motion features and cell interaction features in a microenvironment; the cell dynamic behavior features refer to features capable of reflecting cell activity and motion directional properties, and the cell interaction features in the microenvironment refer to features capable of reflecting the functional role of cells in a colony. The classification output unit is used for inputting the comprehensive feature indexes into a pre-trained classification model to obtain automatic classification results of the single blood cells. The blood cell intelligent classification unit is specifically used for: constructing a spatial topology network based on single-cell boundary box coordinates of the single-cell image blocks; analyzing the motion direction of single cells from a single-cell motion trajectory point set; calculating the similarity of the motion directions of any two single cells in the spatial topology network and the distance change rate of the center coordinates distance with time; correcting the similarity of the motion directions through the distance change rate to obtain an interaction synergy coefficient, and taking the average value of the interaction synergy coefficients of each single cell and adjacent cells as an adjacent cell interaction synergy factor, which is used for reflecting the interaction features of cells in a microenvironment. The blood cell intelligent classification unit is specifically used for:

2. The blood cell image automatic classification system according to claim 1, wherein acquiring single-cell motion trajectories from a sequence of multiple frames of continuous single-cell image blocks and analyzing the single-cell motion trajectories to obtain an activity chemotaxis index; extracting cell contour feature of each frame of single-cell image blocks in the sequence of multiple frames of continuous single-cell image blocks to obtain single-cell contour deformation degree; generating a dynamic function fingerprint based on the activity chemotaxis index and the single-cell contour deformation degree, and the dynamic function fingerprint is used for reflecting the cell dynamic behavior features containing morphological features and motion features. The blood cell intelligent classification unit is specifically used for:

3. The blood cell image automatic classification system according to claim 2, wherein, acquiring a single-cell motion trajectory point set through an optical flow method and feature matching for the sequence of multiple frames of continuous single-cell image blocks; analyzing the final displacement and moving distance of single cells from the single-cell motion trajectory point set; determining the activity chemotaxis index according to the ratio of the modulus of the final displacement to the moving distance. The blood cell intelligent classification unit is specifically used for:

4. The blood cell image automatic classification system according to claim 3, wherein, extracting contour feature corners through a corner detection algorithm for each frame of single-cell image blocks to output a corner set of the single-cell image blocks; calculating the Euclidean distance between any two contour feature corner pairs, and taking the average value of the Euclidean distances of multiple contour feature corner pairs in the corner set as the contour stretch degree of each frame of single-cell image blocks; determining the variance of multiple contour stretch degrees corresponding to the sequence of multiple frames of continuous single-cell image blocks as the single-cell contour deformation degree. The blood cell intelligent classification unit is specifically used for:

5. The blood cell image automatic classification system according to claim 4, wherein, normalizing the activity chemotaxis index and the single-cell contour deformation degree. ​ The product of the normalized results of the active chemotaxis index and the single-cell contour deformation degree is determined as the dynamic function fingerprint.

6. The blood cell image automatic classification system according to claim 2, wherein, The blood cell intelligent classification unit is further configured to: determine the instantaneous velocity of the single cell motion according to the set of single-cell motion trajectory points and the time interval of consecutive adjacent frames; optimize the active chemotaxis index based on the instantaneous velocity.

7. The blood cell image automatic classification system according to claim 1, wherein, The blood cell intelligent classification unit is specifically configured to: calculate a Pearson correlation coefficient between the dynamic function fingerprint and the adjacent cell interaction synergy factor, the Pearson correlation coefficient being used to reflect the interaction characteristics of the cell in the microenvironment; generate the comprehensive feature index based on the dynamic function fingerprint, the adjacent cell interaction synergy factor and the Pearson correlation coefficient.

8. The blood cell image automatic classification system according to claim 1, wherein, The data acquisition platform comprises an optical imaging unit, a motion control unit and an image acquisition unit. The optical imaging unit is configured with a high-resolution charge-coupled device image sensor, a 100-fold oil immersion objective lens and an automatic stage. The motion control unit is configured to drive the optical imaging unit to perform Z-axis multi-layer scanning to obtain a blood image sequence, and to synthesize the blood image sequence into the original blood cell image through an image fusion algorithm. The image acquisition unit is configured to control the optical imaging unit to acquire the blood image sequence at a preset frame rate.

9. The blood cell image automatic classification system according to claim 8, wherein, The data acquisition platform further comprises an image processing unit. The image processing unit is configured to perform image denoising and image enhancement on the original blood cell image to obtain a standardized image. The image processing unit is further configured to locate and crop an image region of a single blood cell in the standardized image, and to perform normalization processing on the cropped image region to obtain a single-cell image block with uniform size.

Images