An artificial intelligence-based mature neutrophil granule recognition system

Abstract

The application discloses a kind of mature neutrophil granule identification systems based on artificial intelligence, comprising: data preprocessing module, multidimensional feature joint extraction module, robust feature screening module, model construction and training module and multicenter verification and output module.The application initiates microscopic dynamics digital characteristic quantity system, breaks the algorithm "black box": for the pain point that traditional computer vision lacks quantitative physical topology in cell morphology Interpretation, initiates GDI, GSHI, GSDI three major customer mathematics index.Through the visual diagnosis experience of senior blood morphology expert is abstracted into topological equation, successfully realize the high sensitivity detection to microscopic particle morphology pathological variation (such as toxic granule color is extremely deep, size is different, local aggregation), greatly improve the clinical interpretability of model.

Description

Technical Field

[0001] This invention belongs to the field of cell particle recognition technology, and in particular relates to a mature neutrophil particle recognition system based on artificial intelligence. Background Technology

[0002] Neutrophils are the most abundant immune cells in human peripheral blood. The number, density, and distribution characteristics of granules in their cytoplasm are core morphological indicators characterizing cell maturity and functional status. In clinical cell morphology testing, the identification of the granule status of mature neutrophils is a crucial step in the analysis of peripheral blood and bone marrow samples, and it has significant technical value in identifying dysplastic hematopoiesis caused by myelodysplastic syndromes (MDS), assessing severe infections, and evaluating sample quality.

[0003] Currently, the interpretation of granule status in mature neutrophils mainly relies on manual observation by pathologists using a microscope. This method has inherent drawbacks such as high subjectivity, inconsistent judgment standards, low detection efficiency, and poor result consistency. Differences in experience and visual judgment among different testers can easily lead to significant errors in the interpretation of granule status, making it difficult to achieve standardized and quantitative objective assessment and failing to meet the accuracy and efficiency requirements of large-scale clinical sample analysis.

[0004] With the development of artificial intelligence and medical image processing technology, technologies such as deep learning and convolutional neural networks have been gradually applied to peripheral blood cell classification and morphological recognition, achieving coarse-grained automatic classification of cell types. However, existing technical solutions have extremely obvious limitations when it comes to "microscopic particle feature analysis": First, existing image preprocessing usually uses direct grayscale conversion or simple macroscopic RGB analysis, which seriously destroys the highly diagnostic multidimensional biochemical staining characteristics in clinical stained microscopic images, making the algorithm insensitive to subtle color changes such as the basophilic deep staining of poisoning particles, and losing important multidimensional spectral information; Second, existing large AI models (such as pure CNN architecture) can only extract black-box macroscopic texture features, lacking in-depth exploration of the "physical dynamics" of individual microscopic particles inside cells. For abnormal particle manifestations of high clinical concern (such as changes in staining and density, increased heterogeneity in particle size, local aggregation, or uneven distribution), existing technologies completely lack precise quantitative indicators based on mathematical topological mechanisms, directly leading to low specificity in identifying microscopic pathological variations. Third, the traditional image features extracted by the models lack robust screening mechanisms for microscopic particles. Under common clinical image perturbations such as brightness, contrast, and batch differences in staining, the feature stability is extremely poor, and the recognition results are easily interfered with. Fourth, for the problem of blurred morphological boundaries between normal particles and samples with reduced particle size, existing simple classification networks lack a concatenation optimization scheme for macro and micro features, resulting in low discrimination efficiency between these two types of samples. Fifth, existing models are mostly trained on small sample data from a single center and do not adopt a strict multi-center independent external dataset validation mode. Their generalization and adaptation capabilities across devices and institutions are seriously insufficient, making it difficult to truly implement them in actual clinical laboratory procedures.

[0005] In summary, there is currently a lack of a standardized, quantitative, robust, and highly generalizable mature intelligent neutrophil particle recognition technology, which fails to address the challenges of feature loss under complex staining backgrounds and the difficulty in quantifying microscopic spatial topological features. To overcome the numerous shortcomings of existing technologies and meet the needs of automated clinical cell morphology analysis, there is an urgent need to develop a mature neutrophil particle recognition method and system based on artificial intelligence. Summary of the Invention

[0006] The purpose of this invention is to address the technical problems existing in the prior art, such as loss of multidimensional color information, difficulty in quantifying microscopic spatial topological features, poor feature robustness, and blurred boundaries of multi-classification, and to provide a mature neutrophil particle recognition system based on artificial intelligence.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a mature neutrophil granule recognition system based on artificial intelligence, comprising:

[0008] The data preprocessing module is used to acquire stained microscopic images of mature neutrophils, perform multidimensional color space decoupling and filtering on the images and extract multispectral features, perform global pixel intensity normalization preprocessing, and label the cytoplasmic regions of the cells as the analysis objects.

[0009] The multidimensional feature joint extraction module is used to jointly extract traditional texture features and deep convolution features of images, and constructs a multidimensional neutrophil particle micro-dynamic features extraction based on the decoupling results of multidimensional color space and biological morphological dynamics equations.

[0010] The robust feature selection module is used to perform feature dimensionality reduction on the extracted micro-dynamic features through LASSO regression and to select a robust feature set by combining the robust separation index RSS.

[0011] The model building and training module is used to build and train convolutional neural network models and random forest models based on the selected robust feature set and image feature data, and to perform three-class classification training for mature neutrophils with reduced, normal, and increased granules.

[0012] The tandem diagnostic optimization module is used to integrate random forest and convolutional neural network models to build a tandem diagnostic model and optimize the accurate identification of normal cells and samples with reduced or abnormal particles.

[0013] The multicenter validation and output module is used to perform cross-center generalization validation of the tandem diagnostic model using independent external datasets, and finally outputs the particle identification and classification results of mature neutrophils.

[0014] Furthermore, normalization preprocessing involves calculating the mean, standard deviation, minimum, and maximum values ​​of the global pixel intensity of the image after decoupling the multidimensional color space of the image, serving as a unified numerical benchmark.

[0015] Furthermore, traditional texture features include first-order statistical features, gradient features, Haralick co-occurrence matrix features, local binary pattern (LBP) features, frequency domain features, topological features, and morphological features; deep convolutional features are extracted based on ImageNet pre-trained convolutional neural network models.

[0016] Furthermore, the multidimensional neutrophil granule microdynamic characteristics include the granule density color index (GDI), the granule size uniformity index (GSHI), and the granule spatial distribution divergence index (GSDI).

[0017] The particle density color index GD is obtained using equation (1):

[0018] GDI = [Σ( I(p_i) × w_i ) ] / Area_cytoplasm (1)

[0019] Where N is the total number of particles identified in a single cell, I(p_i) is the local pixel intensity of the i-th particle, w_i is the local adaptive weighting factor, and Area_cytoplasm is the effective cytoplasmic area.

[0020] The particle size uniformity index GSHI is obtained using equation (2):

[0021] GSHI = exp [ - σ_area / (μ_area + ε) ] (2)

[0022] Where μ_area is the statistical mean of the pixel area of ​​a single particle, σ_area is the standard deviation of the pixel area of ​​a single particle, and ε is a smoothing constant to prevent the denominator from being zero;

[0023] The particle spatial distribution divergence index GSDI is obtained using equation (3):

[0024] GSDI = μ_NND / (σ_NND + ε) (3)

[0025] Where μ_NND is the mean of the nearest neighbor distances of all particle nodes, σ_NND is the standard deviation of the nearest neighbor distances of all particle nodes, and ε is the smoothing constant.

[0026] Furthermore, the robust separation index RSS is obtained using equation (4):

[0027] RSS(f)=σinter(f) / [σˉintra(f)+ϵ] (4)

[0028] Where f is the feature, σinter(f) is the standard deviation of the mean of the center of different particle state categories, σˉintra(f) is the average standard deviation of the same category under multi-source perturbation, and ϵ is a smoothing term to prevent the denominator from being zero.

[0029] Furthermore, the robust feature selection module is also used to conduct perturbation experiments on micro-dynamic features with brightness, contrast, scaling, Gaussian noise, and Gaussian blur, and quantitatively evaluate the stability of the features through the coefficient of variation (CV) and paired t-test.

[0030] Furthermore, the convolutional neural network model is ResNet50, DenseNet121, or EfficientNet-B0.

[0031] Furthermore, during training, the microscopic images for both the convolutional neural network model and the random forest model were uniformly adjusted to 224×224 pixels and divided into training, validation, and test sets in an 8:1:1 ratio; the random forest model was trained using hierarchical 5-fold cross-validation.

[0032] Furthermore, the independent external datasets were obtained from different medical institutions, and the validation metrics included precision, recall, F1 score, AUC, and confusion matrix.

[0033] Furthermore, the cascaded diagnostic optimization module is also used to perform LDA dimensionality reduction projection and MDS probability prediction based on robust texture feature sets to determine the optimal classification threshold for normal and particle-reduced samples.

[0034] A method for identifying mature neutrophil granules based on artificial intelligence is provided, including the following steps: Step S1, acquiring stained microscopic images of mature neutrophils, performing multidimensional color space decoupling filtering on the images and extracting multispectral features, grayscale conversion, global pixel intensity normalization preprocessing, and marking the cytoplasmic region of interest as the analysis object;

[0035] Step S2: Jointly extract traditional texture features and depth convolution features of the image; and based on the color space decoupling results and the biomorphic dynamics equation, construct and extract multidimensional neutrophil particle microdynamic features, which include at least the particle density color index (GDI), particle size uniformity index (GSHI), and particle spatial distribution divergence index (GSDI).

[0036] Step S3: LASSO regression is used for feature dimensionality reduction, and robust texture feature set is obtained by screening based on the robust separation index RSS.

[0037] Step S4: Construct ResNet50, DenseNet121, EfficientNet-B0 convolutional neural network models and random forest models, and complete three-class classification training with reduced, normal, and increased granules.

[0038] Step S5: Combine random forest and EfficientNet-B0 to build a cascaded model and optimize the discrimination efficiency between normal particles and particles with reduced particle size.

[0039] Step S6: Use a multi-center independent dataset for external generalization verification and output the particle state recognition result.

[0040] Furthermore, the normalization process involves decoupling the multidimensional color space of the image and calculating the mean, standard deviation, minimum, and maximum values ​​of the global pixel intensity of the image as a unified numerical benchmark; the region of interest is annotated by morphology experts and the cytoplasm is precisely extracted.

[0041] Furthermore, in step S2, the traditional texture features include first-order statistical features, gradient features, Haralick co-occurrence matrix features, local binary pattern (LBP) features, frequency domain features, topological features, and morphological features; the deep convolutional features are extracted based on ImageNet pre-trained ResNet50, DenseNet121, and EfficientNet-B0 networks.

[0042] Furthermore, the specific extraction and calculation method of the multidimensional neutrophil particle microdynamic features in step S2 is as follows: Extracting the particle density color index (GDI): In the decoupled multidimensional color space, the local pixel intensity of all detected particles is obtained and summed with an additional spatial weight matrix. The ratio of this to the extracted total cytoplasmic pixel area is used as GDI, which is used to quantify abnormal changes in particle density and color depth. The formula is: GDI = [ Σ( I(p_i) × w_i )] / Area_cytoplasm; where N is the total number of particles identified in a single cell, I(p_i) is the local pixel intensity of the i-th particle, w_i is the local adaptive weight factor, and Area_cytoplasm is the effective cytoplasmic area. Extracting the particle size uniformity index (GSHI): The mean and standard deviation of the area of ​​all detected monomeric particles within the region of interest in the cytoplasm are calculated. A negative exponential decay distribution function based on the area variation coefficient is constructed to calculate GSHI, which is used to quantify pathological variations in particle morphology and size. The formula is: GSHI = exp [ - σ_area / (μ_area + ε) ]; where μ_area is the statistical mean of the pixel area of ​​a monomeric particle, σ_area is the standard deviation of the pixel area of ​​a monomeric particle, and ε is a smoothing constant to prevent the denominator from being zero. Extracting the Spatial Dispersion Index (GSDI): Extracting the central topological coordinates of all detected particles, calculating the nearest neighbor Euclidean distance between each particle and its nearest neighbor of the same type, and obtaining the GSDI by calculating the ratio of the mean to the standard deviation of the nearest neighbor Euclidean distances. This is used to assess the uniformity of intracellular particle arrangement or focal aggregation state; the formula is: GSDI = μ_NND / (σ_NND + ε); where μ_NND is the mean of the nearest neighbor distances of all particle nodes, σ_NND is the standard deviation of the nearest neighbor distances of all particle nodes, and ε is a smoothing constant.

[0043] Furthermore, the robust separation index RSS in step S3 is calculated as follows: RSS(f) = σ_inter(f) / [σ_intra(f) + ε]; where f is the feature, σ_inter(f) is the standard deviation of the mean of different particle state categories, σ_intra(f) is the average standard deviation of the same category under multi-source perturbation, and ε is a smoothing term to prevent the denominator from being zero.

[0044] Furthermore, step S3 also includes performing perturbation experiments on the features with brightness, contrast, scaling, Gaussian noise, and Gaussian blur, and quantitatively evaluating the stability of the features through the coefficient of variation (CV) and paired t-test.

[0045] Furthermore, in step S4, the images are uniformly adjusted to 224×224 pixels and divided into training set, validation set, and test set in an 8:1:1 ratio; a hierarchical 5-fold cross-validation is used to train the random forest model.

[0046] Furthermore, in step S5, the cascaded model is used to reduce the false positive rate of normal samples being misclassified as having reduced particles, while maintaining a high recall rate for samples with reduced particles.

[0047] Furthermore, in step S6, the multi-center independent datasets are derived from different medical institutions, and the validation metrics include precision, recall, F1 score, AUC, and confusion matrix.

[0048] Furthermore, it also includes LDA dimensionality reduction projection and MDS probability prediction based on robust texture feature sets to determine the optimal classification threshold for normal and particle-reduced samples.

[0049] The beneficial effects of this invention are:

[0050] (1) Pioneering a digital feature quantification system for microscopic dynamics, breaking the "black box" of algorithms: Addressing the pain point of traditional computer vision lacking quantitative physical topological interpretation in cell morphology, this invention pioneered three objective mathematical indices: GDI, GSHI, and GSDI. By abstracting the visual diagnostic experience of senior blood morphology experts into topological equations, it successfully achieved high-sensitivity detection of pathological variations in microscopic particle morphology (such as extremely dark staining, varying sizes, and focal aggregation of poisoned particles), significantly improving the clinical interpretability of the model.

[0051] (2) Decoupling of multidimensional color space to restore real biochemical staining characteristics: It overcomes the defect of pathological color loss caused by traditional image analysis relying solely on grayscale conversion. By preserving and analyzing multispectral feature dimensions, the algorithm can keenly capture the subtle basophilic staining differences unique to microscopic images under myelodysplastic syndrome (MDS) or severe infection.

[0052] (3) Constructing a robust separation network with strong resistance to clinical interference: Based on the innovative design of the robust separation index (RSS) and multi-source perturbation experiment, it accurately screens out extremely unstable false features that are easily affected by sample batch, lighting and staining differences, so that the identification model still shows extremely high stability in complex clinical multi-source microscopic real-world data.

[0053] (4) Serial architecture optimization and multi-center generalization verification: To address the long-standing problem of overlapping boundaries between normal and particle-reduced samples in clinical testing, a serial diagnostic architecture was constructed by integrating the advantages of different classifiers. This architecture minimizes false positives and has undergone rigorous verification on multi-center independent external datasets, thus completely bridging the last mile from basic research and development to the implementation of intelligent algorithms in complex cross-institutional clinical environments. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of texture feature extraction and LASSO regression dimensionality reduction of a mature neutrophil particle recognition system based on artificial intelligence proposed in this invention.

[0055] Figure 2 This is a schematic diagram illustrating the feature robustness assessment and LDA dimensionality reduction of a mature neutrophil particle recognition system based on artificial intelligence proposed in this invention.

[0056] Figure 3 This is a schematic diagram of the multi-model external validation confusion matrix + ROC of a mature neutrophil particle recognition system based on artificial intelligence proposed in this invention.

[0057] Figure 4 This is a schematic diagram comparing the performance of a cascaded model of a mature neutrophil particle recognition system based on artificial intelligence proposed in this invention.

[0058] Figure 5 This is a schematic diagram of the module composition of a mature neutrophil granule recognition system based on artificial intelligence proposed in this invention. Detailed Implementation

[0059] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0060] In the description of this application, it should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. For ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0061] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and are not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0062] It should be noted that in the description of this application, the directional terms such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description. Unless otherwise stated, these directional terms do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the scope of protection of this application. The directional terms "inner" and "outer" refer to the inner and outer contours relative to the outline of each component itself.

[0063] It should be noted that, in this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0064] A mature neutrophil granule identification method based on artificial intelligence includes the following steps:

[0065] Step S1: Acquire stained microscopic images of mature neutrophils, perform multidimensional color space decoupling filtering on the images and extract multispectral features, perform grayscale conversion and global pixel intensity normalization preprocessing, and mark the regions of interest in the cytoplasm as the analysis objects.

[0066] Step S2: Jointly extract traditional texture features and depth convolution features of the image; and based on the color space decoupling results and the biomorphic dynamics equation, construct and extract multidimensional neutrophil particle microdynamic features, which include at least the particle density color index (GDI), particle size uniformity index (GSHI), and particle spatial distribution divergence index (GSDI).

[0067] Step S3: LASSO regression is used for feature dimensionality reduction, and robust texture feature set is obtained by screening based on the robust separation index RSS.

[0068] Step S4: Construct ResNet50, DenseNet121, EfficientNet-B0 convolutional neural network models and random forest models, and complete three-class classification training with reduced, normal, and increased granules.

[0069] Step S5: Combine random forest and EfficientNet-B0 to build a cascaded model and optimize the discrimination efficiency between normal particles and particles with reduced particle size.

[0070] Step S6: Use a multi-center independent dataset for external generalization verification and output the particle state recognition result.

[0071] This application establishes a complete, standardized, and highly clinically interpretable AI-driven neutrophil particle recognition process, automating the entire process from multidimensional image preprocessing to final microscopic state recognition output. It not only completely resolves the drawbacks of subjective human interpretation and the coarse-grained recognition (black box effect) of existing single CNN models, but also overcomes the major clinical challenges of distinguishing between normal and reduced-particle samples due to extremely blurred morphological boundaries, as well as poor cross-center generalization. Ultimately, it achieves accurate and objective recognition of three states: reduced, normal, and increased (including abnormal accumulation during severe infection).

[0072] In a preferred embodiment of this application, the normalization process involves calculating the mean, standard deviation, minimum, and maximum values ​​of the global pixel intensity of the image after decoupling from the multidimensional color space, serving as a unified numerical benchmark. The region of interest is marked using the diagnostic visual habits of experienced hematologists, and the cytoplasm is intelligently extracted. Its advantages are: it overcomes the bottleneck of color loss in microscopic biochemical diagnosis caused by traditional single grayscale conversion; it achieves unified image normalization through global multispectral pixel intensity statistics, eliminating invalid background and red blood cell overlap interference, accurately focusing on the core analytical object, and significantly improving the extraction quality of subtle staining differences.

[0073] In a preferred embodiment of this application, the traditional texture features in step S2 include first-order statistical features, gradient features, Haralick co-occurrence matrix features, local binary pattern (LBP) features, frequency domain features, topological features, and morphological features; the deep convolutional features are extracted based on ImageNet pre-trained ResNet50, DenseNet121, and EfficientNet-B0 networks. More importantly, this example creatively constructs three major micro-dynamic features: GDI, GSHI, and GSDI. Their advantages are: traditional texture combined with deep convolution can adaptively capture the macroscopic distribution contours of particles; while the novel three micro-dynamic features map subjective pathological morphologies (such as deep coloring of poisoned particles, increased size heterogeneity, and local mesh aggregation) that are extremely difficult to quantify clinically into highly sensitive scalar exponents through mathematical topology, achieving multi-dimensional and rapid decoding of the microscopic pathological evolution within a single cell, fundamentally improving the basic discriminative performance of the classification model.

[0074] In a preferred embodiment of this application, the robust separation index RSS in step S3 is calculated as follows: RSS(f) = σ_inter(f) / [σ_intra(f) + ε]; where f is a feature, σ_inter(f) is the standard deviation of the mean of the center of different particle state categories, σ_intra(f) is the average standard deviation of the same category under multi-source perturbation, and ε is a smoothing term to prevent the denominator from being zero. Its advantages are: quantitatively verifying the specificity and anti-interference redundancy of features through the RSS formula; scientifically screening out fragile features severely affected by local slide preparation, lighting, and batch staining differences; balancing high inter-class separation performance with high intra-class anti-perturbation capability; and providing a high-purity, high-quality core vector for subsequent multi-source data models.

[0075] In a preferred embodiment of this application, step S3 further includes perturbation experiments on the features with brightness, contrast, scaling, Gaussian noise, and Gaussian blur, and quantitatively assessing the feature stability using the coefficient of variation (CV) and paired t-tests. Its advantages are: by comprehensively simulating various severe clinical microscopic optical deviations and fabrication defects in the real world, the model is rigorously screened and validated, ensuring that the core algorithm maintains an extremely low misdiagnosis rate even under extremely complex clinical real-world multi-source image conditions.

[0076] In a preferred embodiment of this application, in step S4, the images are uniformly adjusted to 224×224 pixels and divided into training, validation, and test sets in an 8:1:1 ratio; a hierarchical 5-fold cross-validation model is used to train the random forest model. Standardizing the dataset partitioning ratio ensures fairness and unbiased evaluation during model training; hierarchical 5-fold cross-validation can deeply extract feature distribution information from a limited number of medical samples, effectively avoiding model overfitting and ensuring the absolute robustness of model training convergence.

[0077] In a preferred embodiment of this application, the tandem model in step S5 is used to reduce the false positive rate of normal samples being misclassified as neutropenia, while maintaining a high recall rate for neutropenic samples. Its advantages are: it specifically addresses the most extreme challenge in bone marrow cell identification—the high visual overlap between normal mature neutrophils and mild neutropenia (early stages of pathological hematopoiesis). While preserving more accurate capture of early lesions (avoiding missed detections), it minimizes the false positive rate of over-alarming healthy samples, thus balancing the safety of clinical testing with high diagnostic specificity.

[0078] In the preferred example of this application, the multi-center independent dataset in step S6 comes from medical institutions of different levels, and the validation metrics include precision, recall, F1 score, AUC, and confusion matrix. Its advantages are: it breaks the limitations of isolated training data from single laboratories, and uses the "gold standard" of multi-center independent testing to ensure the model's excellent adaptability across medical terminal devices and geographical imaging practices. The objective and multi-dimensional evaluation parameter system provides strong, high-level evidence-based medical support for the instrument's final registration and real-world commercial clinical application.

[0079] In a preferred embodiment of this application, the method further includes LDA dimensionality reduction projection and MDS probability prediction based on a robust texture feature set to determine the optimal classification threshold for normal and granulation-reduced samples. Its advantages are: dimensionality reduction enables topological visualization of multidimensional sample point clusters in Euclidean space, clearly presenting the biological drift trajectory of neutrophils under different pathological states; combined with a probability prediction model, it accurately delineates objective classification tangents, providing for the first time a standardized and accurately reproducible quantitative reference scale for the early microscopic pathological hematopoietic detection of myelodysplastic syndrome (MDS), which currently relies heavily on subjective judgment in clinical practice.

[0080] like Figure 5 As shown, an artificial intelligence-based mature neutrophil granule recognition system is applied to an artificial intelligence-based mature neutrophil granule recognition method, comprising:

[0081] The data preprocessing module is used to acquire stained microscopic images of mature neutrophils and perform grayscale conversion, pixel intensity normalization, and region of interest annotation in the cytoplasm.

[0082] The feature extraction module is used to jointly extract traditional texture features and depth convolution features from an image;

[0083] The robust feature selection module is used to select robust texture feature sets through LASSO regression dimensionality reduction and RSS exponent calculation.

[0084] The model training module is used to build and train convolutional neural network models and random forest models;

[0085] The cascade optimization module is used to fuse random forest and EfficientNet-B0 to build a cascade model and optimize the identification of normal and granular reduced samples;

[0086] The multi-center validation module is used to perform generalization validation using independent external datasets and output recognition results.

[0087] The system's various functional modules work together to automate the entire process of identification, enabling the engineering implementation of the method. The system is stable and efficient, and can be directly adapted to clinical pathology testing equipment to meet the needs of automated identification of clinical cell states.

[0088] like Figures 1-4 As shown, the mature neutrophil granule identification method based on artificial intelligence described in this invention has been experimentally verified to achieve excellent levels in feature stability, classification accuracy, generalization ability, and clinical identification efficacy. Specific verification results are as follows:

[0089] Feature selection and robustness verification:

[0090] Feature filtering results:

[0091] LASSO regression was used to perform dimensionality reduction and screening of high-dimensional texture features. Figure 1 (A~1B), 26 key texture features were obtained through screening, as shown in Table 1:

[0092] Table 1

[0093]

[0094] Based on the Robust Separability Index (RSS), six core robust features were further selected, as shown in Table 2:

[0095] Table 2

[0096]

[0097] The metrics used include sparse_area_ratio, haralick_5, intensity_mean, gabor_std_f0, intensity_std, intensity_skew, grain density color index (GDI), grain size uniformity index (GSHI), and grain spatial distribution divergence index (GSDI). The classification model built based on these features achieved an AUC of 0.96 on the training set and 0.921 on the validation set, demonstrating excellent feature discrimination ability and generalization.

[0098] Based on the RSS metric, 26 features extracted from seven physical dimensions (intensity, gradient, texture Haralick, texture LBP, frequency domain, topology, and granularity) were evaluated. The six texture features with the highest RSS scores were then selected (sparse_area_ratio, haralick_5, intensity_mean, gabor_std_f0, intensity_std, intensity_skew). As the brightness coefficient increased from 1.0 to 1.4... Figure 2Although intensity_mean, intensity_std, and haralick_5 show a certain upward trend, the numerical gradients among the three groups (Increased, Normal, Reduced) remain highly consistent, without cross-confusion. In Blur ( Figure 2 GL) and Resize Figure 2 Under MR perturbation, sparse_area_ratio and intensity_mean exhibit extremely high stationarity, almost unaffected by parameter changes. While gabor_std_f0 and gradient_energy show some numerical decay under strong ambiguity, they can still clearly define the three classes of samples across the entire parameter range. Figure 2 The images in the text are arranged from left to right and top to bottom, in alphabetical order from A to T.

[0099] Further LDA dimensionality reduction projection was performed based on the nine best robust features selected. Figure 2 The results showed that the three types of samples formed well-defined clusters in two-dimensional space. The contribution (discretion) of the first discriminant axis (LD1) was as high as 91.16%, perfectly reproducing the biological continuum of granularity from "reduced" to "normal" and then to "increased", in which the normal group (cyan) was accurately distributed between the two pathological groups.

[0100] By performing multidimensional scaling analysis on the core features and constructing an MDS probability prediction model, we evaluated its potential application value in clinical screening. Figure 2 (T). In the comparison of probability distributions of Normal Population and Reduced Population (often indicating pathological conditions such as MDS), the two sets of curves showed excellent separation. The probability scores of the normal group were concentrated around 0, while those of the reduced group shifted significantly towards 1. The optimal cut-off point determined by the model was 0.53. At this threshold, the model can efficiently identify samples with abnormal granulation. In addition, by setting a 2 / 3 normal reference interval (0.12), the definition of the distribution boundary of normal samples was further strengthened, providing an objective and quantitative criterion for the interpretation of neutrophil granules in clinical practice.

[0101] Comparison of classification performance of convolutional neural network architectures with different depths:

[0102] To further explore the potential of artificial intelligence in end-to-end feature extraction and to compare the performance differences between adaptive feature learning and physical dimension features in particle recognition tasks, this application introduces deep convolutional neural network (CNN) technology. We comprehensively compare three mainstream CNN architectures (DenseNet121, ResNet50, and EfficientNet-B0) with a Random Forest (RF) model built based on texture features selected in the first two stages.

[0103] Overall performance:

[0104] The evaluation results are shown in Table 3:

[0105] Table 3

[0106]

[0107] In the classification task targeting neutrophil granule status (decreased, increased, normal), all models demonstrated high prediction accuracy. Among them, DenseNet121 performed best across all core metrics, achieving a macro-average and weighted average precision, recall, and F1 score of 0.95. This indicates that DenseNet's feature reuse mechanism, achieved through dense connections, can more effectively capture subtle granule morphological changes within neutrophils.

[0108] Differences in particle category identification among different models:

[0109] Particle-enhanced group: All models in this group exhibited extremely high recognition rates, with ResNet50 achieving a recall of 0.99 and an F1 score of 0.99. This indicates that the increased particle size resulted in significant cell image features, making it easy for the model to capture their morphological characteristics.

[0110] Reduced Particle Group: This group is slightly more difficult to identify. DenseNet121 maintained balanced performance in this group (F1: 0.93), while ResNet50 had a lower recall (0.84), indicating that ResNet50 may have a risk of missing diagnoses when handling sparse particle features.

[0111] In the normal group, DenseNet121 still maintains its lead with an F1 score of 0.94; the Random Forest model performed slightly worse in this group (F1: 0.89), suggesting that adaptive deep features are more advantageous than simple physical dimensionality reduction features in distinguishing between normal particle distribution and abnormal fluctuations.

[0112] A comparison between deep learning and physical feature models:

[0113] Notably, the RandomForest model, built upon the 26 high-dimensional physical features selected by LASSO in the first two stages, also demonstrated highly competitive performance (MacroF1: 0.93). While its overall performance was slightly lower than DenseNet121, it outperformed or matched EfficientNet-B0, which uses adaptive feature extraction. This finding demonstrates that the physical dimension features selected in the first two stages (such as sparse_area_ratio and intensity_mean) have extremely high biological relevance and statistical interpretability, and can highly summarize the granule dynamics of mature neutrophils.

[0114] External validation and multi-center performance evaluation:

[0115] To evaluate the model's generalization ability on unseen cross-center datasets, this application applied four internally trained models to an independent external validation set. The classification metrics were analyzed comprehensively, as shown in Table 4.

[0116] Table 4

[0117]

[0118] and confusion matrix ( Figure 3 AD) and ROC curve ( Figure 3 EfficientNet-B0 model exhibits different performance characteristics across all classes, achieving the most balanced precision and recall. (Confusion matrix) Figure 3 (B) shows that it is most accurate in identifying "Normal" samples (F1: 0.74), with a relatively even distribution of misclassifications. Its ROC curve ( Figure 3 In F), the AUC for all three categories remained at 0.89 or higher, indicating that it effectively balanced the overall classification boundary. The ResNet50 model demonstrated extremely high accuracy (0.98) when identifying the "Increased" category, with a confusion matrix of ( Figure 3 C) also confirms that it has almost no false positives in this category. However, its area under the curve (AUC: 0.86) when identifying the "Reduced" group is relatively small. Figure 3 The recall rate (G) was relatively lower than other models. The DenseNet121 model exhibited significant recognition bias when processing complex "Normal" class images, with a recall rate of only 0.43. From the confusion matrix ( Figure 3A) It is evident that a large number of normal samples (116 cases) were incorrectly classified into the "Reduced" group by the model, directly leading to a decrease in the overall macro-mean score. The RF model achieved the highest recall rate among all tested models in both the "Reduced" and "Increased" abnormal groups (both 0.92). Further analysis using ROC curves and AUC values ​​showed that the RF model achieved an AUC of 0.94 in identifying the "Reduced" category. Figure 3 H), surpassing all CNN models; its AUC in the "granularity increased" and "normal" groups were also as high as 0.99 and 0.90, respectively. Combined with the confusion matrix ( Figure 3 D) Objectively speaking, compared with the table indicators, the RF model produced more false positives when capturing the "normal" category (classifying 92 normal samples as granulation reduction), which affected its recall rate (0.52) in the normal group. However, this also proves that these 6 geometric and texture features based on image intensity have a high degree of directional sensitivity in capturing the pathological change of "granulation loss".

[0119] Performance evaluation of the tandem model for high-difficulty identification (normal particle group and reduced particle group):

[0120] Given that all models demonstrated extremely high recognition accuracy for the "increased granulation" group in preliminary validation, the core challenge in clinical diagnosis has focused on the blurred morphological boundaries of the "normal" and "reduced granulation" sample categories. To further improve the differentiation efficiency between these two states, this application conducted a deep comparison between the Random Forest (RF) model and the pure convolutional neural network (PureCNN, i.e., EfficientNet-B0), which exhibit different performance in these two categories. Confusion matrix ( Figure 4 A) shows that the RF model exhibits a strong ability to capture the abnormal pathological condition of "reduced granulation," successfully identifying 211 out of 229 real samples (a recall rate of 92.1%). However, this high sensitivity comes at the cost of specificity, resulting in up to 92 "normal" samples being overdiagnosed as having "reduced granulation." In contrast, PureCNN (i.e., EfficientNet-B0) is more conservative and global in its feature extraction. Figure 4 B). It effectively protected the "normal" category, significantly reducing the number of normal samples misidentified as "reduced particle size" to 39; however, its sensitivity to "reduced particle size" was insufficient, resulting in 50 missed diagnoses (mispredicted as "normal"). To neutralize the "high false positive rate" of RF and the "high false negative rate" of pure CNN, this application combines the two to construct a cascade model. The most prominent contribution of the cascade model is its accurate protection of "normal" samples. For example, the confusion matrix ( Figure 4As shown in C), while maintaining a high recognition rate for "reduced particles" (177 cases), the model drastically reduced the number of cases where "normal" samples were misclassified as "reduced" to a manageable 24 cases (a nearly 74% decrease in false positives compared to RF's 92 cases), successfully identifying 191 normal samples. This significant reduction in misclassification directly translates into an improvement in overall diagnostic efficacy. Figure 4 As shown in Figure D, the area under the curve (AUC) of the pure CNN and the simple RF model in distinguishing these two states are 0.841 and 0.878, respectively; while the AUC of the cascaded model climbs to 0.887, achieving the best discrimination among the three strategies. This indicates that by using a cascaded mechanism to complement each other's advantages, noise caused by feature overlap can be filtered out to the maximum extent without seriously sacrificing the recall rate of pathological samples, providing the most robust and safe comprehensive judgment for clinical practice.

[0121] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A mature neutrophil granule recognition system based on artificial intelligence, characterized in that... include: The data preprocessing module is used to acquire stained microscopic images of mature neutrophils, perform multidimensional color space decoupling and filtering on the images and extract multispectral features, perform global pixel intensity normalization preprocessing, and label the cytoplasmic regions of the cells as the analysis objects. The multidimensional feature joint extraction module is used to jointly extract traditional texture features and deep convolution features of images, and constructs a multidimensional neutrophil particle micro-dynamic features extraction based on the decoupling results of multidimensional color space and biological morphological dynamics equations. The robust feature selection module is used to perform feature dimensionality reduction on the extracted micro-dynamic features through LASSO regression and to select a robust feature set by combining the robust separation index RSS. The model building and training module is used to build and train convolutional neural network models and random forest models based on the selected robust feature set and image feature data, and to perform three-class classification training for mature neutrophils with reduced, normal, and increased granules. The tandem diagnostic optimization module is used to integrate random forest and convolutional neural network models to build a tandem diagnostic model and optimize the accurate identification of normal cells and samples with reduced or abnormal particles. The multicenter validation and output module is used to perform cross-center generalization validation of the tandem diagnostic model using independent external datasets, and finally outputs the particle identification and classification results of mature neutrophils.

2. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: The normalization preprocessing involves calculating the mean, standard deviation, minimum, and maximum values ​​of the global pixel intensity of the image after decoupling the multidimensional color space, serving as a unified numerical benchmark.

3. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: The traditional texture features include first-order statistical features, gradient features, Haralick co-occurrence matrix features, local binary pattern (LBP) features, frequency domain features, topological features, and morphological features; deep convolutional features are extracted based on a pre-trained convolutional neural network model from ImageNet.

4. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: The multidimensional neutrophil granule microdynamic features include the granule density color index (GDI), the granule size uniformity index (GSHI), and the granule spatial distribution divergence index (GSDI). The particle density color index GD is obtained using equation (1): GDI = [Σ( I(p_i) × w_i ) ] / Area_cytoplasm (1) Where N is the total number of particles identified in a single cell, I(p_i) is the local pixel intensity of the i-th particle, w_i is the local adaptive weighting factor, and Area_cytoplasm is the effective cytoplasmic area. The particle size uniformity index GSHI is obtained using equation (2): GSHI = exp [ - σ_area / (μ_area + ε) ] (2) Where μ_area is the statistical mean of the pixel area of ​​a single particle, σ_area is the standard deviation of the pixel area of ​​a single particle, and ε is a smoothing constant to prevent the denominator from being zero; The particle spatial distribution divergence index GSDI is obtained using equation (3): GSDI = μ_NND / (σ_NND + ε) (3) Where μ_NND is the mean of the nearest neighbor distances of all particle nodes, σ_NND is the standard deviation of the nearest neighbor distances of all particle nodes, and ε is the smoothing constant.

5. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: The robust separation index RSS is obtained using equation (4): RSS(f)=σinter(f) / [σˉintra(f)+ϵ] (4) Where f is the feature, σinter(f) is the standard deviation of the mean of the center of different particle state categories, σˉintra(f) is the average standard deviation of the same category under multi-source perturbation, and ϵ is a smoothing term to prevent the denominator from being zero.

6. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: The robust feature selection module is also used to conduct perturbation experiments on micro-dynamic features with brightness, contrast, scaling, Gaussian noise, and Gaussian blur, and to quantitatively evaluate feature stability through the coefficient of variation (CV) and paired t-test.

7. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: The convolutional neural network model is ResNet50, DenseNet121, or EfficientNet-B0.

8. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: During training, the microscopic images of the convolutional neural network model and the random forest model were uniformly adjusted to 224×224 pixels and divided into training set, validation set and test set in an 8:1:1 ratio; the random forest model was trained using hierarchical 5-fold cross-validation.

9. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: The independent external datasets are from different medical institutions, and the validation metrics include precision, recall, F1 score, AUC, and confusion matrix.

10. The mature neutrophil granule recognition system based on artificial intelligence according to claim 1, characterized in that: The cascaded diagnostic optimization module is also used to perform LDA dimensionality reduction projection and MDS probability prediction based on robust texture feature sets to determine the optimal classification threshold for normal and particle-reduced samples.

Images