Multispectral microscope blood cell automatic classification and counting system

Abstract

The invention discloses a multispectral microscope blood cell automatic classification and counting system. The system is composed of a multispectral illumination and microscopic imaging module, a spectrum and geometric calibration module, a multispectral preprocessing and cell segmentation module, a multispectral discrimination index and spectral band weight adaptive updating module, a cell graph structure classification module, a man-machine interaction module, an online adaptive learning module and the like. The method comprises the following steps: acquiring a blood smear image by a system under a narrow-band multi-spectrum condition, constructing cellular spectrum-morphological characteristics by combining a multi-scale segmentation result after noise suppression, flat field correction and background deduction, generating a discrimination index with a self-adaptive spectrum band weight, and completing joint classification and counting on a cell map; and meanwhile, carrying out constrained increment updating on the spectral band weight and the classification model by utilizing an artificial correction result of the low-confidence-coefficient cells. Compared with a traditional single-channel microscopic imaging and static classification method, the method has higher classification accuracy and counting stability under different dyeing and imaging conditions, and the workload of manual recheck can be reduced.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis and intelligent diagnostic technology, specifically to an automatic classification and counting system for blood cells using a multispectral microscope. Background Technology

[0002] In clinical hematology examinations, blood cell differential and counting are essential for assessing the body's immune status, infection severity, and hematopoietic function. Current blood cell analysis methods primarily employ electrical impedance tomography, laser scattering, and automated identification methods based on ordinary microscope images.

[0003] Among them, electrical impedance and light scattering methods are suitable for large-scale routine testing, but their ability to subdivide cell morphology and identify abnormal cells is limited. Automatic identification methods based on ordinary microscopes generally use grayscale or RGB color images and perform cell segmentation and classification through traditional image processing and deep learning, but they are heavily dependent on staining conditions, lighting stability and equipment differences, making it difficult to maintain consistency of results between different laboratories.

[0004] To improve cell differentiation capabilities, some studies have introduced multispectral or hyperspectral microscopy, increasing spectral dimensions by acquiring images across multiple spectral bands. However, existing methods often employ fixed spectral band combinations and fixed feature weights, typically determined only once during model training. This makes it difficult to adaptively adjust based on the spectral distribution and cell composition of different samples. Furthermore, most existing methods classify each cell as an independent sample, failing to fully utilize the spatial distribution relationships and population spectral patterns of cells within the same field of view. Under complex cases and long-term operational conditions, classification accuracy and robustness remain insufficient.

[0005] Therefore, based on existing multi-band microscopic acquisition, how to utilize the statistical characteristics of the current sample and cell population context information to synergistically and adaptively adjust the band weights and classification model, thereby significantly improving the accuracy and long-term stability of blood cell classification and counting in complex clinical samples under the conditions of limited band number and limited acquisition time, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] Technical Objective: To address the shortcomings of existing technologies, this invention discloses an automated blood cell classification and counting system using multispectral microscopy. It constructs a multispectral discriminant index that combines spectral and morphological features, utilizes an adaptive update mechanism for spectral weights based on the statistical characteristics of the current sample, and builds a cell-level graph structure around this discriminant index. Graph neural network classification is then performed on this graph structure. Furthermore, an online adaptive learning mechanism for jointly updating spectral weights and graph model parameters is introduced. This enables the system to achieve high-precision automated classification and counting of blood cells in complex clinical samples under limited spectral band counts and acquisition time, while maintaining stable and reliable diagnostic performance in long-term operation and cross-laboratory applications.

[0007] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0008] An automated blood cell classification and counting system using multispectral microscopy includes the following steps:

[0009] A multispectral illumination module is used to generate illumination light with multiple narrow bands within a predetermined wavelength range for multispectral illumination of blood smear samples;

[0010] The microscopic imaging module is used to acquire multi-band microscopic images of blood cells under various narrow-band illumination conditions;

[0011] The object-carrying and autofocusing module is used to carry blood smear samples and move them in a two-dimensional plane as well as to automatically focus along the optical axis.

[0012] The spectral and geometric calibration module is used for wavelength calibration, spatial registration, and plan field correction of multi-band microscopic images;

[0013] The data acquisition and synchronization control module is used to coordinate the working sequence of the multispectral illumination module, the microscopic imaging module, and the object loading and autofocus module, and outputs the acquired multispectral microscopic images to the multispectral preprocessing module.

[0014] The multispectral preprocessing module is used to perform noise suppression, background subtraction, and intensity normalization on multispectral micro images to obtain preprocessed multispectral micro images.

[0015] The cell segmentation and multi-scale feature extraction module is used to segment individual cell regions in preprocessed multi-band microscopic images and extract morphological features, spectral features and frequency domain features for each cell region.

[0016] The multispectral discrimination index generation and band weight adaptive update module is used to generate a multispectral discrimination index representing the joint discrimination ability of spectrum and shape for each cell region based on spectral features and background gray level, combined with morphological features. The module also adaptively iterates and updates the band weights according to the statistical characteristics of each band in the current sample and its correlation with cell discrimination, so that the bands with greater cell discrimination can obtain higher weights in subsequent feature fusion and classification.

[0017] The cell graph construction and graph structure classification module is used to construct weighted edges based on the spatial proximity relationship between cells and the difference of the multispectral discriminant index, using each cell region as a node in the graph structure, multispectral discriminant index, morphological features and frequency domain features as node features, and multispectral discriminant index difference to obtain cell-level graph structure. Then, graph-based neural network calculation is performed on the cell-level graph structure to jointly predict the category of the cell corresponding to each node and obtain cell category label.

[0018] The counting and result quality control module is used to summarize the number of cells of various types according to cell category labels, output blood cell classification and counting results, and provide quality assessment information based on node classification confidence distribution and abnormal cell aggregation areas;

[0019] The human-computer interaction and online adaptive learning module is used to receive the correction results of some cell category labels by the inspectors, and to combine the correction results with the original cell category labels to form incremental data. Under the preset stability constraints, the module also incrementally updates the band weights and graph structure neural network model parameters to compensate for the long-term drift of sample distribution and equipment status.

[0020] Preferably, the multispectral discrimination index generation and band weight adaptive update module is used to calculate enhanced spectral contrast for each cell region i and each selected band k. and multispectral discriminant index ,in:

[0021] ;

[0022] ;

[0023] in, is the average gray value of cell region i on spectral band k, in gray levels; is the average gray value of the background region on spectral band k, in gray levels; To prevent the denominator from being too small and causing numerical instability, a positive constant is used; K is the selected number of spectral bands; is the spectral band weighting coefficient, and is a real constant; is the morphological roughness index of cell region i, and is a dimensionless value; represents the ellipticity index of cell region i, which is a dimensionless value; and is the morphological feature weighting coefficient, and is a real constant.

[0024] Preferably, the multispectral discrimination index generation and band weight adaptive update module is used to iteratively update the band weight coefficients based on the enhanced spectral contrast distribution of all cells in the current sample on each band, and the update rule is as follows:

[0025] ,

[0026] in, and These are the weighting coefficients of spectral band k at the t-th and t+1-th iterations, respectively; The step size parameter is a real constant. Enhanced spectral contrast for all cells s in the current sample The variance; K is the sum of the variances of all bands; K is the number of selected bands, and j is the band index.

[0027] Preferably, the statistical characteristics related to cell differentiation in the multispectral discriminant index generation and band weight adaptive update module include at least one of the following: the inter-class variance between cell categories based on the multispectral discriminant index, the sample variance of each band enhancing spectral contrast, and the signal-to-noise ratio inside and outside the band; the stability constraints in the human-computer interaction and online adaptive learning module include at least one of the following: the learning rate does not exceed a preset upper limit, the parameter change amplitude does not exceed a preset range, and the performance degradation of the model on the validation set does not exceed a preset threshold.

[0028] Preferably, the cell graph construction and graph structure classification module includes a cell graph construction unit, which is used to establish an edge between two cell nodes when the following conditions are met simultaneously: the spatial distance between cell centers is less than a preset distance threshold and the corresponding multispectral discriminant index difference does not exceed a preset difference threshold, and the combined function determined by the spatial distance, the multispectral discriminant index difference and the spectral feature similarity is used as the edge weight.

[0029] Preferably, the human-computer interaction and online adaptive learning module limits the learning rate to no more than a preset upper limit when performing incremental parameter updates, and limits the parameter change range of each update to no more than a preset range. When the performance of the model on the validation set decreases beyond a preset threshold, it rolls back to the previous stable version of the parameters.

[0030] An automated blood cell classification and counting method using multispectral microscopy, applied to the aforementioned automated blood cell classification and counting system using multispectral microscopy, specifically includes the following steps:

[0031] Step S1: Load the blood smear sample and autofocus it. Use the multispectral illumination module and the microscopic imaging module to acquire multispectral microscopic images under multiple narrowband illumination conditions.

[0032] Step S2: Perform noise suppression, background subtraction and intensity normalization on the multi-band microscopic image to obtain a preprocessed multi-band microscopic image;

[0033] Step S3: Segment individual cell regions in the preprocessed multi-band microscopic image and extract the morphological features, spectral features and frequency domain features of the cell regions;

[0034] Step S4: Based on spectral features and background grayscale, combined with morphological features, generate a multispectral discrimination index for each cell region to characterize the joint spectral-morphological discrimination ability. Adaptively iteratively update the weights of the spectral bands according to the statistical characteristics of the correlation between each spectral band and cell discrimination in the current sample, so that the spectral bands with greater cell discrimination can obtain higher weights in subsequent feature fusion and classification.

[0035] Step S5: Using each cell region as a node, the multispectral discriminant index, morphological features, and frequency domain features are used as node features. A weighted cell graph structure is constructed based on the spatial proximity relationship between cells and the difference in the multispectral discriminant index. A graph-based neural network classification is then performed on the cell graph structure to jointly predict the category of the cells corresponding to each node and obtain the cell category label.

[0036] Step S6: Summarize the number of cells of each type according to the cell category label, output the blood cell classification and counting results, and generate quality assessment information based on classification confidence and abnormal distribution.

[0037] Step S7: When the inspector corrects some cell category labels, the correction results and the original cell category labels constitute incremental data. Under the preset stability constraints, the band weights and graph structure neural network model parameters are updated incrementally at the same time. The stability constraints include at least one of the following: the learning rate does not exceed the preset upper limit, the parameter change range does not exceed the preset range, and the performance decline of the model on the validation set does not exceed the preset threshold.

[0038] Preferably, the construction of the weighted cell graph structure in step S5 includes: when the spatial distance between two cells is less than a preset distance threshold and the difference of the corresponding multispectral discriminant index does not exceed a preset difference threshold, an edge is established between the corresponding nodes, and the edge weight is a combination function determined by the spatial distance, the difference of the multispectral discriminant index and the spectral feature similarity.

[0039] Preferably, the incremental update in step S7 limits the learning rate of each parameter update to no more than a preset upper limit, and monitors the model performance during the cumulative update process. When the performance drops below a preset threshold, the model is rolled back to the previous stable version.

[0040] Beneficial effects: The multispectral microscope automatic blood cell classification and counting system provided by this invention has the following beneficial effects:

[0041] 1. This invention utilizes a multispectral discrimination index generation module to nonlinearly combine spectral contrast information between cells and background under different spectral bands with morphological indicators such as morphological roughness and ellipticity to form a single index for comprehensively characterizing the spectral-morphological joint discrimination ability of cells. Compared to schemes that only use grayscale or RGB color features, this invention can maintain high cell class separation under complex conditions such as variations in staining depth, slight cell distortion, and mild background contamination, reducing dependence on fixed staining conditions.

[0042] 2. This invention analyzes the statistical characteristics (such as inter-class variance, sample variance, and signal-to-noise ratio) of the multispectral discriminant index and enhanced spectral contrast in the current sample, and iteratively updates the weights of each band, so that bands that contribute more to cell differentiation are assigned higher weights in subsequent graph structure classification. Compared with schemes that fix band weights or determine weights in a single training iteration, this invention can adaptively highlight spectral segments with higher diagnostic value for different samples and different staining schemes without increasing the number of bands or acquisition time, significantly improving classification accuracy and cross-sample robustness.

[0043] 3. This invention constructs a cell-level graph structure with a multispectral discriminant index as the core. The node features simultaneously contain spectral, shape, and frequency information. The edge weights combine spatial proximity and discriminant index differences. Joint classification is performed using a graph neural network. It not only considers the local features of single cells but also introduces the contextual information of cell populations. It has a stronger ability to suppress misclassification of abnormal proportion samples, local cell clusters, and edge single cells, which helps to improve the stability of the overall counting.

[0044] 3. Through human-computer interaction and an online adaptive learning module, this invention can utilize a small amount of high-quality manually corrected data to perform small-step joint updates of the spectral weights and graph neural network model parameters under preset stability constraints. This allows the system to gradually adapt to the sample distribution and operating habits of specific laboratories. In long-term operation and cross-laboratory applications, compared to traditional models that only determine parameters during the training phase, this invention's approach can better mitigate the performance degradation over time, resulting in higher consistency of classification and counting results across different times and devices. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0046] Figure 1 This is a schematic diagram of the overall structure of the multispectral microscope blood cell automatic classification and counting system of the present invention;

[0047] Figure 2 This is a structural block diagram of the multispectral illumination module and the microscopic imaging module of the present invention;

[0048] Figure 3 This is a schematic diagram of the multispectral preprocessing and cell segmentation process of the present invention;

[0049] Figure 4 This is a schematic diagram of the functional structure of the multispectral discrimination index generation and band weight adaptive update module of the present invention;

[0050] Figure 5 This is a schematic diagram of the functional structure of the cell diagram construction and graph structure classification module of the present invention;

[0051] Figure 6 This is a schematic diagram comparing the classification accuracy and the proportion of low-confidence cells under different collection conditions between the embodiments of the present invention and the control embodiments. Detailed Implementation

[0052] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0053] An automated blood cell classification and counting system using multispectral microscopy includes the following steps:

[0054] The multispectral illumination module is used to generate illumination light with multiple narrow bands within a predetermined wavelength range (e.g., visible to near-infrared). The band switching is achieved through a tunable narrow band filter unit to illuminate blood smear samples with multiple bands.

[0055] The microscopic imaging module is used to acquire multi-band microscopic images of blood cells under various narrow-band illumination conditions. The multi-band microscopic images have a strict spatial correspondence within the same field of view.

[0056] The object-carrying and autofocusing module is used to carry blood smear samples and move them in a two-dimensional plane, as well as to automatically focus along the optical axis to ensure clear images in each field of view.

[0057] The spectral and geometric calibration module is used to perform wavelength calibration, spatial registration, and flat field correction on multi-spectral microscopic images, eliminating the effects of inconsistencies in imaging geometry and illumination between different spectral bands.

[0058] The data acquisition and synchronization control module is used to coordinate the working sequence of the multispectral illumination module, the microscopic imaging module, and the object loading and autofocus module, and outputs the acquired multispectral microscopic images to the multispectral preprocessing module.

[0059] The multispectral preprocessing module is used to perform noise suppression, background subtraction, and intensity normalization on multispectral microscopic images to obtain preprocessed multispectral microscopic images, providing stable input for subsequent feature extraction.

[0060] The cell segmentation and multi-scale feature extraction module is used to segment individual cell regions in preprocessed multi-band microscopic images and extract morphological features, spectral features, and frequency domain features for each cell region.

[0061] The multispectral discrimination index generation and band weight adaptive update module is used to generate a multispectral discrimination index representing the joint discrimination ability of spectrum and shape for each cell region based on spectral features and background gray level, combined with morphological features. The module also adaptively iterates and updates the band weights according to the statistical characteristics of each band in the current sample and its correlation with cell discrimination, so that the bands with greater cell discrimination can obtain higher weights in subsequent feature fusion and classification.

[0062] The cell graph construction and graph structure classification module is used to construct weighted edges based on the spatial proximity relationship between cells and the difference of the multispectral discriminant index, using each cell region as a node in the graph structure. The module then performs graph-based neural network computation on the cell-level graph structure to jointly predict the category of the cells corresponding to each node and obtain the cell category label.

[0063] The counting and result quality control module is used to summarize the number of cells of various types according to cell category labels, output blood cell classification and counting results, and provide quality assessment information based on node classification confidence distribution and abnormal cell aggregation areas.

[0064] The human-computer interaction and online adaptive learning module provides a user interface for receiving the correction results of some cell category labels by the inspectors. The correction results and the original cell category labels are combined to form incremental data. Under the preset stability constraints, the spectral weights and graph structure neural network model parameters are updated incrementally to compensate for the long-term drift of sample distribution and equipment status.

[0065] An automated blood cell classification and counting method using multispectral microscopy, applied to the aforementioned automated blood cell classification and counting system using multispectral microscopy, specifically includes the following steps:

[0066] Step S1: Load the blood smear sample and autofocus it. Use the multispectral illumination module and the microscopic imaging module to acquire multispectral microscopic images under multiple narrowband illumination conditions.

[0067] Step S2: Perform noise suppression, background subtraction and intensity normalization on the multi-band microscopic image to obtain a preprocessed multi-band microscopic image;

[0068] Step S3: Segment individual cell regions in the preprocessed multi-band microscopic image and extract the morphological features, spectral features and frequency domain features of the cell regions;

[0069] Step S4: Based on spectral features and background grayscale, combined with morphological features, generate a multispectral discrimination index for each cell region to characterize the joint spectral-morphological discrimination ability. Adaptively iteratively update the weights of the spectral bands according to the statistical characteristics of the correlation between each spectral band and cell discrimination in the current sample, so that the spectral bands with greater cell discrimination can obtain higher weights in subsequent feature fusion and classification.

[0070] Step S5: Using each cell region as a node, the multispectral discriminant index, morphological features, and frequency domain features are used as node features. A weighted cell graph structure is constructed based on the spatial proximity relationship between cells and the difference in the multispectral discriminant index. A graph-based neural network classification is then performed on the cell graph structure to jointly predict the category of the cells corresponding to each node and obtain the cell category label.

[0071] Step S6: Summarize the number of cells of each type according to the cell category label, output the blood cell classification and counting results, and generate quality assessment information based on classification confidence and abnormal distribution.

[0072] Step S7: When the inspector corrects some cell category labels, the correction results and the original cell category labels constitute incremental data. Under the preset stability constraints, the band weights and graph structure neural network model parameters are updated incrementally at the same time. The stability constraints include at least one of the following: the learning rate does not exceed the preset upper limit, the parameter change range does not exceed the preset range, and the performance decline of the model on the validation set does not exceed the preset threshold.

[0073] Example 1

[0074] like Figure 1 As shown, the multispectral microscopy blood cell automatic classification and counting system of this embodiment includes:

[0075] The system includes modules for multispectral illumination, microscopic imaging, object loading and autofocus, spectral and geometric calibration, data acquisition and synchronization control, multispectral preprocessing, cell segmentation and multi-scale feature extraction, multispectral discriminant index generation and adaptive band weighting update, cell graph construction and graph structure classification, counting and result quality control, and human-computer interaction and online adaptive learning.

[0076] A multispectral illumination module and a microscopic imaging module are optically coupled to form a multispectral band microscopic imaging system. The object loading and autofocus module supports two-dimensional scanning and autofocus of blood smears. The spectral and geometric calibration module calibrates wavelength response and spatial registration during system installation and maintenance. The data acquisition and synchronization control module controls band switching, camera exposure, and stage movement, and sends the image stream to the back-end processing module. The multispectral preprocessing module, cell segmentation and multi-scale feature extraction module, multispectral discriminant index generation and adaptive band weight update module, cell map construction and graph structure classification module, counting and result quality control module, and human-computer interaction and online adaptive learning module can be implemented using the same processor or multiple collaborative processing units.

[0077] like Figure 2 As shown, the multispectral illumination module includes a broadband light source, a tunable narrowband filter unit, a light-collecting and homogenizing unit, and an optical path coupling unit. The broadband light source can be a supercontinuum light source or a broadband LED array. The tunable narrowband filter unit outputs multiple narrowband illumination lights with center wavelengths ranging from 420 nm to 720 nm, and the full width at half maximum (FWHM) of each band can be no greater than 20 nm. The light-collecting and homogenizing unit is used to homogenize the narrowband light, and the optical path coupling unit couples the narrowband light into the illumination optical path of the microscopic imaging module.

[0078] The microscopic imaging module includes an objective lens, a tube / imaging lens assembly, and an image sensor. The numerical aperture of the objective lens is preferably not less than 0.65 to ensure resolution of cell nuclei and cytoplasmic details. The image sensor can be a high-sensitivity CMOS or CCD image sensor, supporting rapid acquisition of multi-spectral images.

[0079] The object carrier and autofocus module includes a three-dimensional movable object carrier stage and a focus control unit. It can use an autofocus algorithm based on image sharpness evaluation to ensure that images of each spectral band are on the same focal plane.

[0080] Example 2

[0081] like Figure 3 As shown, the multispectral preprocessing module performs the following steps on the acquired raw multispectral microscopic images:

[0082] Step 1, Noise Suppression: Spatial filtering or temporal filtering methods are used to suppress sensor noise and random noise.

[0083] Step 2, Flat Field Correction: Perform flat field correction on each spectral band using the pre-measured flat field image to eliminate illumination inhomogeneities.

[0084] Step 3, Background Estimation and Subtraction: Estimate the background gray level of each spectral band based on the cell-free regions in the image, which will be used for subsequent discriminant index calculation.

[0085] Step 4, Intensity Normalization: Normalize the gray values ​​of images with different spectral bands to a unified range to facilitate comparison during feature construction.

[0086] Step 5, Candidate Cell Region Detection: The cell segmentation and multi-scale feature extraction module first uses threshold segmentation and morphological operations to obtain candidate cell regions.

[0087] Step 6, Fine-grained segmentation (deep network): Combine a deep learning-based segmentation network to perform fine-grained segmentation of the candidate regions, and obtain pixel-level masks for single cells.

[0088] For each cell region, the cell segmentation and multi-scale feature extraction module extracts the following features:

[0089] Morphological characteristics, including cell area, perimeter, length of major and minor axes, ellipticity, and morphological roughness;

[0090] Spectral characteristics, including the average gray value, maximum value, and minimum value of the cellular region in each spectral band;

[0091] Frequency domain features include performing two-dimensional frequency transformation on the cellular region or its neighborhood to extract indicators such as energy distribution and directionality.

[0092] Example 3

[0093] This embodiment focuses on the construction of the multispectral discrimination index and the band weight update strategy.

[0094] like Figure 4 As shown, the multispectral discrimination index generation and band weight adaptive update module includes: an enhanced spectral contrast calculation unit, a multispectral discrimination index calculation unit, a sample statistical characteristic analysis unit, and a band weight update unit.

[0095] For each cell region i and each selected spectral band k, enhance spectral contrast. and multispectral discriminant index The calculation formula is as follows:

[0096]

[0097]

[0098] in, is the average gray value of cell region i on spectral band k, in gray levels; is the average gray value of the background region on spectral band k, in gray levels; To prevent the denominator from being too small and causing numerical instability, a positive constant is used; K is the selected number of spectral bands; is the spectral band weighting coefficient, and is a real constant; is the morphological roughness index of cell region i, and is a dimensionless value; represents the ellipticity index of cell region i, which is a dimensionless value; and is the morphological feature weighting coefficient, and is a real constant.

[0099] The sample statistical characteristic analysis unit is used to statistically analyze the enhanced spectral contrast and multispectral discrimination index of all cells within the current sample, and calculate the indicators of each spectral band in terms of cell differentiation, for example:

[0100] Multispectral discriminant index categorized by cell type The inter-class variance;

[0101] Enhance spectral contrast in each band The sample variance;

[0102] The gray signal-to-noise ratio of each spectral band in the cellular region and the background region.

[0103] The band weight update unit iteratively updates the weights of each band based on one or more of the above indicators. The update process can employ... The normalized variance-driven update formula is as follows: and These are the weighting coefficients of spectral band k at the t-th and t+1-th iterations, respectively; The step size parameter is a real constant. Enhanced spectral contrast for all cells s in the current sample The variance; The sum of variances of all bands; K is the number of selected bands, and j is the band index; alternatively, other monotonically increasing functions based on the proportion of inter-class variance or the signal-to-noise ratio can be used, and the formula given in the specification is one specific implementation. In this embodiment, the weight update is performed at the sample batch level, that is, after each batch of samples is processed, the band weights are iteratively updated, so that the system can gradually adapt to the current sample characteristics and staining scheme of the laboratory.

[0104] The multispectral discriminant index is not only used for spectral band weight updates, but also serves as one of the key node features in the cell graph construction and graph structure classification modules. It also participates in edge weight calculation and online learning decision-making. Therefore, it plays a core role in the overall system of this invention by connecting multispectral band feature construction, graph structure modeling, and model adaptive updates.

[0105] Example 4

[0106] This embodiment emphasizes the role of the multispectral discriminant index in graph structure construction.

[0107] like Figure 5 As shown, the cell map construction unit in the cell map construction and graph structure classification module first uses all cell regions in the same field of view as nodes, and the node features include multispectral discriminant indices. In addition, morphological and frequency domain characteristics are considered. Subsequently, weighted edges are established between nodes according to the following rules:

[0108] If the spatial center distance between two cells is less than the distance threshold If the two cells are considered to be spatially adjacent, then they are considered to be adjacent to each other.

[0109] If the difference in the multispectral discriminant index between the two cells is less than the difference threshold If so, it is considered that the two cells have similar patterns in spectral-morphological features;

[0110] When both of the above conditions are met, an edge is established between the corresponding nodes, and its edge weight can be defined as a combination function of spatial distance, discriminant index difference and spectral feature similarity.

[0111] The graph structure classification unit performs graph neural network computation on the aforementioned cell graph, updating node features and fusing information through multi-layer graph convolution or graph attention operations. Since node features contain multispectral discriminant indices, and edge weights depend on the difference in discriminant indices, the graph neural network can fully utilize the interrelationships between cells in the spectral-shape feature space during propagation, thereby achieving more robust joint classification in complex samples.

[0112] Example 5

[0113] This embodiment focuses on explaining the meaning and implementation of stability constraints.

[0114] The human-computer interaction and online adaptive learning module provides an interactive interface displaying cell images, category labels, and confidence levels. This allows laboratory personnel to manually verify and correct cells with low confidence levels or those flagged as potentially abnormal by the system. The module then combines the manually corrected labels with the system's original predictions to create incremental data, which is used to update the band weights and graph neural network model parameters online.

[0115] To avoid model instability caused by online learning, this invention pre-sets stability constraints, including but not limited to:

[0116] Learning rate upper limit constraint: the learning rate used during online updates does not exceed the preset upper limit to avoid parameters changing too quickly;

[0117] Parameter variation range constraint: In each update, the variation range of spectral weights and model parameters is limited to a preset range, for example, by adding L2 norm penalty to parameter changes or setting a pruning threshold.

[0118] Performance monitoring constraints are implemented by maintaining a representative set of samples internally. When online updates cause the performance of the validation set to degrade beyond a preset threshold, the system rolls back to the previous stable version of parameters or reduces the learning rate to continue updating.

[0119] Under this stability constraint, the present invention can gradually absorb information corrected by humans while avoiding overfitting the model to a small amount of incremental data, thereby maintaining a steady improvement in classification performance and overall stability during long-term operation.

[0120] Example 6

[0121] This embodiment provides a qualitative comparison process between the present invention and three control schemes to illustrate the overall technical effect of the present invention.

[0122] Comparison Scheme 1 is RGB + single-cell CNN classification:

[0123] Only ordinary RGB microscopic images are used; single cells are classified using convolutional neural networks, without constructing graph structures, updating spectral weights, or performing online learning.

[0124] Comparative scheme two is multi-band fixed weight + single-cell classification:

[0125] The same multi-band microscopic acquisition and preprocessing procedure as the present invention is used; a fixed spectral weight combination is adopted to input the multi-band features into a convolutional neural network for single-cell classification.

[0126] Scheme 3 uses multi-band + graph structure classification but lacks adaptive band weights and online learning:

[0127] Multi-band acquisition and preprocessing were used; a graph structure with cells as nodes was constructed, and a graph neural network was used for classification; however, the band weights were fixed during the training phase and no online adaptive learning was performed.

[0128] The proposed solution combines a multispectral discriminant index, adaptive band weighting, cell diagram structure, and online learning.

[0129] The invention utilizes the multispectral discriminant index generation, adaptive update of spectral band weights, construction of cell-level graph structures and joint classification with graph neural networks, as well as the online adaptive learning mechanism.

[0130] The four methods were tested on the same publicly available blood smear dataset, and the tests were repeated under simulated staining changes and different collection equipment conditions. The results show that:

[0131] Compared with the control scheme 1, the scheme of the present invention significantly improves the overall classification accuracy in complex samples, especially in the case of mild staining unevenness and obvious background contamination, and the identification of abnormal cells is more stable.

[0132] Compared with control scheme 2, under the condition of the same number of bands, the present invention's scheme, through adaptive updating of band weights, significantly reduces performance fluctuations under different samples and different staining schemes.

[0133] Compared with the control scheme 3, the present invention suppresses the performance degradation caused by changes in time and device status through an online adaptive learning mechanism, and can maintain a high classification accuracy and counting consistency in long-term operation scenarios.

[0134] Figure 6 The above comparison results are illustrated schematically, intuitively reflecting the advantages of the present invention over the control scheme in terms of classification accuracy and the proportion of low-confidence cells.

[0135] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multispectral microscope blood cell automatic classification and counting system, characterized in that, Specifically comprising the following steps: A multi-spectral illumination module for generating illumination light of multiple narrow-band spectral bands in a predetermined wavelength range to illuminate the blood smear sample in a multi-spectral manner; A microscopic imaging module for acquiring multi-spectral microscopic images of blood cells under the illumination of each narrow-band spectral band; A sample carrying and auto-focusing module for carrying the blood smear sample and moving in a two-dimensional plane and auto-focusing along the optical axis direction; A spectral and geometric calibration module for wavelength calibration, spatial registration and flat field correction of the multi-spectral microscopic images; A data acquisition and synchronous control module for coordinating the working timing of the multi-spectral illumination module, the microscopic imaging module and the sample carrying and auto-focusing module, and outputting the acquired multi-spectral microscopic images to the multi-spectral preprocessing module; A multi-spectral preprocessing module for noise suppression, background subtraction and intensity normalization of the multi-spectral microscopic images to obtain preprocessed multi-spectral microscopic images; A cell segmentation and multi-scale feature extraction module for segmenting single cell regions in the preprocessed multi-spectral microscopic images and extracting morphological features, spectral features and frequency domain features for each cell region; A multi-spectral discriminant index generation and spectral band weight adaptive update module for generating a multi-spectral discriminant index representing the spectral-morphological joint discrimination ability for each cell region based on the spectral features and background gray level, and combining the morphological features, and adaptively and iteratively updating the spectral band weight according to the statistical characteristics of the correlation between each spectral band and the cell discrimination in the current sample, so that the spectral bands with greater cell discrimination degree obtain higher weight in subsequent feature fusion and classification; A cell graph construction and graph structure classification module for taking each cell region as a node in the graph structure, taking the multi-spectral discriminant index, the morphological feature and the frequency domain feature as the node feature, constructing a weighted edge according to the spatial proximity relationship between cells and the difference value of the multi-spectral discriminant index, obtaining a cell-level graph structure, and performing graph-based neural network calculation on the cell-level graph structure to jointly predict the class of the cell corresponding to each node, and obtaining the cell class label; A counting and result quality control module for summarizing the number of cells of each class according to the cell class label, outputting the blood cell classification and counting result, and giving quality evaluation information according to the node classification confidence distribution and abnormal cell aggregation area; A human-computer interaction and online adaptive learning module for receiving the correction result of part of the cell class labels by the inspection personnel, and constituting the incremental data with the original cell class label, and simultaneously updating the spectral band weight and the graph structure neural network model parameters under the preset stability constraint to compensate for the long-term drift of the sample distribution and the equipment state.

2. The multispectral microscopic blood cell automatic classification and counting system according to claim 1, wherein, The multispectral discriminant index generation and spectral band weight adaptive update module is configured to compute an enhanced spectral contrast for each cell region i and each selected spectral band k and a multispectral discriminant index wherein: ； ； wherein, is the average gray value of the cell region i on the band k, unit is gray level; is the average gray value of the background region on the band k, unit is gray level; is a normal number to prevent the denominator from being too small to cause numerical instability; K is the number of selected bands; is the band weight coefficient, which is a real constant; is the morphological roughness index of the cell region i, which is a dimensionless value; is the ellipticity index of the cell region i, which is a dimensionless value; and is the morphological feature weight coefficient, which is a real constant.

3. The multispectral microscopic blood cell automatic classification and counting system according to claim 1, wherein, The multi-spectral discriminant index generation and spectral band weight adaptive update module is used for iteratively updating the spectral band weight coefficient according to the enhanced spectral contrast distribution of all cells in each spectral band in the current sample, and the update rule is: ， in, and These are the weighting coefficients of spectral band k at the t-th and t+1-th iterations, respectively; The step size parameter is a real constant. Enhanced spectral contrast for all cells s in the current sample The variance; K is the sum of the variances of all bands; K is the number of selected bands, and j is the band index.

4. The multispectral microscopic blood cell automatic classification and counting system according to claim 1, wherein, The statistical properties related to cell differentiation in the multispectral discriminative index generation and spectral band weight adaptive update module include at least one of the following: the inter-class variance between each cell category based on the multispectral discriminative index, the sample variance of the spectral band enhanced spectral contrast, and the signal-to-noise ratio inside and outside the spectral band; the stability constraints in the human-computer interaction and online adaptive learning module include at least one of the following: the learning rate does not exceed the preset upper limit, the parameter variation amplitude does not exceed the preset range, and the performance of the model on the validation set does not decrease by more than the preset threshold.

5. The multispectral microscopic blood cell automatic classification and counting system according to claim 1, wherein, The cell graph construction and graph structure classification module includes a cell graph construction unit, which is used to establish an edge between two cell nodes when the following conditions are met simultaneously: the spatial distance between the cell centers is less than the preset distance threshold, and the corresponding multispectral discriminative index difference value does not exceed the preset difference threshold, and the combination function determined by the spatial distance, the multispectral discriminative index difference value, and the spectral feature similarity is used as the edge weight.

6. The multispectral microscopic blood cell automatic classification and counting system according to claim 1, wherein, The human-computer interaction and online adaptive learning module limits the learning rate to not exceed the preset upper limit when performing parameter incremental update, and limits the parameter variation amplitude of each update to not exceed the preset range, and rolls back to the last stable version of the parameter when the performance of the model on the validation set decreases by more than the preset threshold.

7. A multispectral microscopic blood cell automatic classification and counting method, characterized in that, A multispectral microscope blood cell automatic classification and counting system according to any one of claims 1-6, specifically comprising the following steps: Step S1, loading and autofocusing the blood smear sample, and using the multispectral illumination module and the microscopic imaging module to collect multispectral band microscopic images under multiple narrow-band spectral band illumination conditions; Step S2, performing noise suppression, background subtraction, and intensity normalization on the multispectral band microscopic images to obtain preprocessed multispectral band microscopic images; Step S3, segmenting single cell regions in the preprocessed multispectral band microscopic images, and extracting morphological features, spectral features, and frequency domain features of the cell regions; Step S4, generating a multispectral discriminative index representing the spectral-morphological joint discrimination ability for each cell region based on the spectral features and background gray level, and combining the morphological features, and adaptively and iteratively updating the spectral band weights according to the statistical properties of each spectral band related to cell differentiation in the current sample, so that spectral bands with greater cell differentiation degree obtain higher weights in subsequent feature fusion and classification; Step S5, taking each cell region as a node, taking the multispectral discriminative index, the morphological features, and the frequency domain features as node features, constructing a weighted cell graph structure according to the spatial proximity relationship between cells and the multispectral discriminative index difference value, and performing graph-based neural network classification on the cell graph structure to jointly predict the category of the corresponding cell of each node, and obtaining the cell category label; Step S6, aggregating the number of cells of each category according to the cell category label, outputting the blood cell classification and counting result, and generating quality evaluation information according to the classification confidence and abnormal distribution. Step S7, when the inspector corrects the partial cell category label, the correction result and the original cell category label constitute incremental data, and the spectral band weight and the graph structure neural network model parameter are updated incrementally under the preset stability constraint, wherein the stability constraint includes at least one of the following: the learning rate is not more than the preset upper limit, the parameter variation range is not more than the preset range, and the performance of the model on the validation set is not more than the preset threshold.

8. The multispectral microscopic method for automatic classification and counting of blood cells according to claim 7, wherein, The construction of the weighted cell graph structure in step S5 includes: when the spatial distance between two cells is less than a preset distance threshold and the corresponding multispectral discriminant index difference is not more than a preset difference threshold, an edge is established between the corresponding nodes, and a combination function determined by the spatial distance, the multispectral discriminant index difference and the spectral feature similarity is taken as the edge weight.

9. The multispectral microscopic method for automatic classification and counting of blood cells according to claim 7, wherein, The incremental update in step S7 limits the learning rate of each parameter update to be not more than a preset upper limit, and monitors the model performance in the cumulative update process, and rolls back to the previous stable version when the performance decreases by more than a preset threshold.

Images