A method and system for detecting basophil activation based on microfluidics and AI image recognition

Abstract

The application discloses a kind of based on microfluidic and AI image recognition basophil activation detection method and system.The method comprises: the peripheral blood sample to be measured is injected into the microfluidic chip device containing conical microporous filter membrane and enriched basophil;In reaction cavity, pre-mixed dry CD63-FITC and CD203c-PE fluorescently labeled antibody and allergen stimulant are added;After incubation at 37 DEG C, 512×512 pixel high-resolution image is obtained by digital microscopic imaging system;Improved target detection model is used to identify activated cells, the model uses lightweight FasterNet backbone network and dynamic attention mechanism, generates cell activation heat map by Gaussian kernel density estimation, combines fluorescence intensity weighted calculation activation index and outputs visual report.System includes microfluidic chip module, fluorescently labeled module, temperature control incubation module, image acquisition module, AI analysis module and man-machine interaction module.The application realizes high sensitivity, high specificity and automatic detection, and is suitable for clinical diagnosis of allergic diseases and blood diseases.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical detection, and particularly relates to a basophil activation detection method and system based on microfluidics and AI image recognition. BACKGROUND

[0002] The basophil activation test (BAT) is an important functional detection method for clinically diagnosing allergic diseases. The expression changes of surface markers (such as CD63 and CD203c) of basophils under the stimulation of allergens are detected to evaluate the allergic state of the body. The basophil activation detection has important significance in the clinical diagnosis of allergic diseases, inflammatory reactions and blood system diseases. The traditional detection method mainly relies on flow cytometry or manual microscope observation, and has the following technical bottlenecks:

[0003] (1) The operation process is complex, and multiple centrifugation and washing steps are needed, which can easily cause cell loss; the sample processing efficiency is low: the existing technology needs to separate cells through density gradient centrifugation, which is time-consuming and easy to cause loss of target cells; (2) Professional personnel are needed for compensation adjustment, and the standardization degree of operation is low; the degree of automation is low: manual microscopy relies on experience judgment, and the evaluation of cell activation state is highly subjective; (3) The detection sensitivity is significantly affected by the sample processing process; (4) The calculation resource is limited: the calculation amount of the existing AI model (such as YOLOv5) on a 512x512 resolution image reaches 14 GFLOPs, which is difficult to realize real-time analysis. The equipment is expensive and the maintenance cost is high, which is difficult to popularize in primary medical institutions.

[0004] In recent years, although the microfluidic technology can enrich cells through microstructures (such as 8 μm filter membranes), there are generally problems such as filter membrane blockage and cell activity decline; and the segmentation accuracy of the detection model based on CNN for closely adjacent cells is less than 85%. In addition, the existing heat map generation method does not fuse the fluorescence intensity and spatial distribution information, resulting in an error rate of more than 10% in the calculation of the activation index.

[0005] Therefore, it is urgent to develop an automatic detection scheme integrating efficient cell enrichment, high-precision AI recognition and multi-parameter quantitative analysis to break through the technical bottlenecks of the existing technology. SUMMARY

[0006] The present application provides a basophil activation detection method and system based on microfluidics and AI image recognition, which can enrich basophils through a microfluidic chip and improve the accuracy and automation of detection by combining AI image recognition technology.

[0007] In a first aspect, the present application provides a basophil activation detection method based on microfluidics and AI image recognition, which comprises:

[0008] S1, inject the peripheral blood sample to be tested into a microfluidic chip device, and enrich basophilic granulocytes through a tapered micropore filter membrane structure with a pore size of 6-8 μm, wherein the microfluidic chip device comprises a sample injection port, a cell enrichment area, a reaction chamber preloaded with freeze-dried reagents, and a microscopic observation window;

[0009] S2, add premixed dry CD63-FITC and CD203c-PE fluorescent labeled antibodies and an allergen stimulant into the microfluidic chip reaction chamber;

[0010] S3, incubate at 37°C for 15-30 minutes to fully activate the cells;

[0011] S4, acquire a 512x512 pixel high-resolution image of the basophilic granulocyte sample using a digital microscopic imaging system;

[0012] S5, identify and classify the activated basophilic granulocytes in the image through an improved target detection model, wherein the model uses a lightweight backbone network and fuses a dynamic attention mechanism, and wherein:

[0013] The lightweight backbone network reduces the calculation amount to below 5.9 GFLOPs through selective channel processing technology of partial convolution operation;

[0014] The dynamic attention mechanism automatically identifies the key areas of the image for feature enhancement;

[0015] S6, generate a cell activation heat map based on the classification results and calculate a basophilic granulocyte activation index, and output a detection report.

[0016] Optionally, in an implementation form of the first aspect of the present application, the microfluidic chip device in step S1 comprises:

[0017] The sample injection port is used to introduce the peripheral blood sample to be tested, and is connected to a syringe pump to control the flow rate to be 10-50 μL / min;

[0018] The cell enrichment area integrates a tapered micropore filter membrane with a pore size of 6-8 μm, and enriches basophilic granulocytes through physical entrapment, while filtering out red blood cells and plasma components;

[0019] The reaction chamber is preloaded with freeze-dried CD63-FITC and CD203c-PE fluorescent antibodies and an allergen stimulant, and has a chamber volume of 50 μL. After injecting the buffer, the reagents are dissolved and the cell activation reaction is triggered;

[0020] The microscopic observation window is covered with a transparent PDMS film with a thickness of 150 μm, and is adapted to a 20x objective lens digital microscopic imaging system for real-time image acquisition.

[0021] Optionally, in an implementation form of the first aspect of the present application, the improved target detection model in the step S5 is a lightweight detection model based on a Transformer architecture, which comprises the following modules connected in turn:

[0022] The feature extraction module adopts a four-level hierarchical FasterNet backbone network, each level comprising: a partial convolution layer for selectively processing 1 / 4 input channels, and the remaining channels being directly passed; a reverse residual block for realizing channel expansion and feature interaction through a 1×1 convolution; and an embedding layer for performing spatial down-sampling and channel expansion; wherein the overall computational complexity of the feature extraction module is controlled to be below 5.9 GFLOPs.

[0023] The attention optimization module comprises a bidirectional routing attention mechanism, which comprises the following units: a dynamic window screening unit for selecting the top k key areas based on query correlation; a sparse attention calculation unit for processing only the attention weights of the screened areas; and a GPU acceleration module for optimizing the calculation process by using dense matrix multiplication; the attention optimization module realizes a real-time inference speed of 145 FPS.

[0024] The multi-task output module comprises: a cell positioning branch for generating a detection frame and a segmentation mask with a resolution of 512×512; and an activation analysis branch for outputting a classification heat map comprising the following parameters: a proportion of CD63+ / CD203c+ double-positive cells, an activation index score based on fluorescence intensity, and a spatial distribution heat map of abnormally activated cells.

[0025] The performance optimization module comprises: a channel pruning unit for compressing the model volume; a mixed precision calculation unit for accelerating the inference process; and an ONNX runtime interface for realizing cross-platform deployment capability.

[0026] Optionally, in an implementation form of the first aspect of the present application, the FasterNet backbone network adopts a hierarchical design and comprises four stages, each stage being provided with an embedding or merging layer before it for performing spatial down-sampling and channel expansion; and a FasterNet block is used as the core calculation unit of each stage, which comprises a partial convolution layer and two point-wise convolution layers in turn, and the input features are passed through a shortcut connection to enhance the gradient flow and feature reuse capability.

[0027] Optionally, in an implementation form of the first aspect of the present application, the model is adapted to the activated basophil cell recognition task through transfer learning, specifically comprising:

[0028] The model training stage: based on a pre-trained target detection model architecture, fine-tuning training is performed using a medical microscopic image dataset, the dataset comprising labeled activated basophil cells and their morphological features, including granule distribution features and cell membrane integrity features.

[0029] a feature extraction stage: multi-scale cell features are extracted by a FasterNet backbone network; a dynamic sparse attention mechanism of a BRA module is combined to focus on significant features of activated cell regions, including degranulation phenomenon features and membrane protrusion features;

[0030] an output optimization stage: an output layer is adjusted to target category detection and classification results, the target categories including activated basophil cells and non-activated basophil cells; an intersection over union threshold is used to optimize the positioning accuracy of a bounding box, so as to accurately distinguish closely adjacent cell clusters;

[0031] The FasterNet backbone network adopts a four-level hierarchical structure, each level including part of a convolutional layer, a reverse residual block and an embedding layer, and the overall computational complexity is controlled below 5.9 GFLOPs; the BRA module includes a dynamic window screening unit, a sparse attention calculation unit and a GPU acceleration module, and realizes a real-time inference speed of 145 FPS.

[0032] Optionally, in an implementation form of the first aspect of the present application, the S6, based on the classification result, generates a cell activation heat map and calculates a basophil cell activation index, and outputs a detection report, including:

[0033] generating a cell activation heat map based on the classification result: using the detection frame confidence score and the category probability output by the model, a two-dimensional heat map is generated by an interpolation algorithm, which intuitively displays the distribution density and activation degree of activated basophil cells in the image;

[0034] calculating a basophil cell activation index: the ratio of the number of activated cells to the total number of basophil cells in the field of view is calculated, and a quantitative activation index is obtained by weighted calculation combined with cell morphological parameters;

[0035] outputting a structured detection report: automatically generating a visual report containing an activation heat map, an activation index, a cell count and an abnormal cell label, supporting DICOM / PDF format output, and interfacing with a hospital information system.

[0036] Optionally, in an implementation form of the first aspect of the present application, the two-dimensional heat map generated by the interpolation algorithm includes:

[0037] performing Gaussian kernel density estimation on the detection frame output by the model to generate a two-dimensional heat distribution map,

[0038] color gradient rendering is performed on the heat map according to the fluorescence labeling intensity;

[0039] superimposed on the original microscopic image to form a composite visualization result.

[0040] Optionally, in an implementation form of the first aspect of the present application, the generating the two-dimensional heat map specifically comprises the following steps:

[0041] S6.1, input data preprocessing: receiving a set of bounding boxes output by a target detection model wherein each bounding box b i =(x i ,y i ,w i ,h i ,s i ) contains the center coordinates (x i ,y i ), width and height w i ,h i and confidence score s i ; and obtaining the fluorescence intensity value associated with each bounding box

[0042] S6.2, Gaussian kernel density estimation: based on the center coordinates (x i ,y i ) of the bounding box, the discrete points are smoothed and interpolated by a Gaussian kernel function to generate a continuous probability density distribution:

[0043]

[0044] wherein σ represents the bandwidth of the kernel, controlling the smoothness of the heat distribution, which is proportional to the average size of the bounding box, I i represents the fluorescence intensity value associated with the i-th bounding box, represents the fluorescence signal intensity of the CD63 / CD203c marker on the cell surface, and is used as a weighted coefficient of confidence and fluorescence intensity, (x,y) represents the coordinates of any point in the image, which is used to calculate the heat value of the point, (x i ,y i ) represents the center coordinates of the i-th bounding box, represents the center position of the activated basophilic granulocyte, s i represents the confidence score of the i-th bounding box, reflecting the confidence degree of the model that the bounding box is an activated cell;

[0045] S6.3, density field calculation and normalization: generating grid points (x g ,y g ) in the image domain WxH, and calculating the superimposed density value of each grid point:

[0046]

[0047] D(x g ,y g ) is normalized to the range [0, 1] to obtain the heat density field D norm , wherein (x g,y g ) represents the coordinates of the image grid points for generating the discretized heat map, N represents the total number of bounding boxes, i.e., the number of activated basophilic granulocytes identified in the image, D(x g ,y g ) represents the superimposed density value at the grid point (x g ,y g ), reflecting the cell distribution density and activation intensity of the region, D norm represents the normalized heat density field for standardized display;

[0048] S6.4, fluorescence signal enhancement processing: according to the fluorescence intensity I i Adjust the color saturation as a weight:

[0049]

[0050] where Saturation(x,y) represents the color saturation of the heat map based on the fluorescence intensity I i Adjust the color saturation as a weight: i (x g ,y g ) represents the Gaussian kernel density contribution value of the i-th bounding box at the image grid point (x g ,y g ), which is used to quantify the influence of the bounding box on the current grid point (x g ,y g ) in the heat map;

[0051] S6.5, heat map and image fusion: superimpose the generated heat map Color(x,y) and the original microscopic image Image(x,y) through Alpha blending:

[0052] Composite(x,y) = a.Color(x,y) + (1-a).Image(x,y),

[0053] where a represents the transparency adjustment parameter for controlling the superimposition ratio of the heat map and the original image, Color(x,y) represents the color value of the generated heat map, and Image(x,y) represents the pixel value of the original microscopic image;

[0054] S6.6, target edge enhancement: superimpose contour markers on the edges of the bounding box to improve the recognition of activated cells;

[0055] S6.7, output visualization results: generate a composite heat map containing activated cell distribution density, fluorescence intensity gradient, and edge markers for visualization of the detection report.

[0056] In a second aspect, the embodiments of the present application provide a basophil activation detection system based on microfluidics and AI image recognition, which is applied to the basophil activation detection method based on microfluidics and AI image recognition as described in the first aspect, and the system comprises:

[0057] a microfluidic chip module for injecting a peripheral blood sample to be tested into a microfluidic chip device, enriching basophils through a tapered micropore filter structure with a pore size of 6-8 pm, wherein the microfluidic chip device comprises a sample injection port, a cell enrichment area, a reaction chamber preloaded with freeze-dried reagents, and a microscopic observation window;

[0058] a fluorescence labeling module for adding premixed dry CD63-FITC and CD203c-PE fluorescent labeled antibodies and an allergen stimulant in the microfluidic chip reaction chamber;

[0059] a temperature-controlled incubation module for incubating at 37°C for 15-30 minutes to fully activate the cells;

[0060] an image acquisition module for acquiring a 512x512 pixel high-resolution image of the basophil sample using a digital microscopic imaging system;

[0061] an AI analysis module for identifying and classifying activated basophils in the image through an improved target detection model, wherein the model uses a lightweight backbone network and fuses a dynamic attention mechanism, and wherein:

[0062] the lightweight backbone network reduces the calculation amount to below 5.9 GFLOPs through a selective channel processing technology of partial convolution operation;

[0063] the dynamic attention mechanism automatically identifies the key areas of the image for feature enhancement;

[0064] a human-computer interaction module for generating a cell activation heat map and calculating a basophil activation index based on the classification results, and outputting a detection report.

[0065] In a third aspect, the embodiments of the present application provide an electronic device, which comprises:

[0066] a processor;

[0067] a memory for storing processor-executable instructions;

[0068] wherein the processor is configured to implement the basophil activation detection method based on microfluidics and AI image recognition as described in the first aspect when executing the instructions.

[0069] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium, which stores a program. The program instructs a device to execute the method for detecting basophil activation based on microfluidics and AI image recognition as described in the first aspect.

[0070] The application discloses a method and system for detecting basophil activation based on microfluidics and AI image recognition. The method comprises: injecting a peripheral blood sample to be tested into a microfluidic chip device containing a conical microporous filter membrane to enrich basophils; adding premixed dry CD63-FITC and CD203c-PE fluorescent labeled antibodies and an allergen stimulant in a reaction cavity; after incubation at 37 DEG C, obtaining a 512*512 pixel high-resolution image through a digital microscopic imaging system; using an improved target detection model to identify activated cells, the model uses a lightweight FasterNet backbone network and a dynamic attention mechanism, generates a cell activation heat map through Gaussian kernel density estimation, combines fluorescence intensity weighted calculation to calculate an activation index and outputs a visual report. The system comprises a microfluidic chip module, a fluorescent labeling module, a temperature control incubation module, an image acquisition module, an AI analysis module and a man-machine interaction module. The application realizes high sensitivity, high specificity and automatic detection, and is suitable for clinical diagnosis of allergic diseases and blood diseases.

[0071] Advantages:

[0072] 1. High sensitivity and specificity detection. The basophils are precisely enriched by a 6-8 μm conical microporous filter membrane, and the capture efficiency of target cells and the specificity of activation marker detection are significantly improved by combining CD63 / CD203c double fluorescent labeling, which effectively avoids the interference of red blood cells.

[0073] 2. Rapid automatic analysis. The microfluidic chip and the pre-frozen reagent are integrated, the sample processing time is shortened to 30 minutes; the lightweight AI model (calculation amount <5.9 GFLOPs) is used to realize 145 FPS real-time image analysis, which is more than 10 times more efficient than traditional manual microscopy.

[0074] 3. Precise quantitative evaluation. Based on the heat map generation algorithm of Gaussian kernel density estimation, the detection frame confidence (s_i) and the fluorescence intensity (I_i) are fused as double weights, which can quantify the spatial distribution density and activation degree of activated cells, and the error rate is less than 5%.

[0075] 4. Strong clinical practicability. The DICOM / PDF report output is supported and connected with the hospital information system, and the detection results include the activation index, the positioning of abnormal cells and other key parameters, which provide standardized data support for the diagnosis of allergic diseases and blood diseases.

[0076] 5. Low cost and easy to promote. Microfluidic chips can be mass-produced, reducing the cost of single detection by 60%; the model supports ONNX cross-platform deployment, adapting to the hardware conditions of primary medical institutions. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 The flowchart of the basophil activation detection method based on microfluidics and AI image recognition provided by an embodiment of the present application is shown.

[0078] Figure 2 The improved target detection model structure diagram provided by an embodiment of the present application is shown.

[0079] Figure 3 The flowchart of generating a two-dimensional heat map provided by an embodiment of the present application is shown.

[0080] Figure 4 The system architecture diagram of a basophil activation detection system based on microfluidics and AI image recognition provided by an embodiment of the present application is shown.

[0081] Figure 5 The schematic diagram of an electronic device provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, not all the embodiments of the present application.

[0083] It should be noted that in the embodiments of the present application, "at least one" means one or more, and more means two or more. Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing the specific embodiments and are not intended to limit the present application.

[0084] It should be noted that in the embodiments of the present application, the terms "first", "second", etc. are used only for the purpose of distinguishing description, and cannot be understood as indicating or implying relative importance, nor can it be understood as indicating or implying sequence. The features limited by "first", "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the terms "exemplary" or "for example" are used to indicate an example, illustration or description. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the use of "exemplary" or "for example" is intended to present the relevant concept in a specific manner.

[0085] All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0086] Embodiment one

[0087] Figure 1 A microfluidic and AI image recognition-based basophil activation detection method flowchart is provided for an embodiment of the present application.

[0088] As Figure 1 shown, a microfluidic and AI image recognition-based basophil activation detection method includes:

[0089] S1, injecting a to-be-tested peripheral blood sample into a microfluidic chip device, enriching basophils through a tapered microporous filter membrane structure with a pore size of 6-8 pm, wherein the microfluidic chip device comprises a sample injection port, a cell enrichment area, a reaction chamber preloaded with freeze-dried reagents, and a microscopic observation window.

[0090] The microfluidic chip device is used to enrich basophils through a tapered microporous filter membrane structure with a pore size of 6-8 pm. Microfluidic chip technology has the characteristics of high throughput, high sensitivity, and fast response in biomedical detection, and is suitable for cell separation and enrichment. Efficient separation and enrichment of cells are achieved through the microstructure of the microfluidic chip, providing a basis for subsequent activation detection. In step S1, basophils are enriched through a tapered microporous filter membrane structure with a pore size of 6-8 pm, while red blood cells and plasma components are filtered out. This physical entrapment method is a highly efficient cell separation method.

[0091] Specifically, in the present embodiment, the microfluidic chip device in step S1 comprises:

[0092] Sample injection port: used to introduce the to-be-tested peripheral blood sample, connected to a syringe pump to control the flow rate to be 10-50 pL / min. The microfluidic chip is usually composed of two layers of chip bases (such as PDMS, glass, etc.), including microchannels, microstructures, sample inlets, and detection windows. In this device, the sample injection port is used to introduce the to-be-tested peripheral blood sample, and is connected to a syringe pump to control the flow rate to be 10-50 pL / min.

[0093] Cell enrichment area: integrated tapered microporous filter membrane with a pore size of 6-8 pm, enriching basophils through physical entrapment while filtering out red blood cells and plasma components.

[0094] Reaction chamber: preloaded with freeze-dried CD63-FITC and CD203c-PE fluorescent antibodies and allergen stimulants, with a chamber volume of 50 pL. After injecting the buffer, the reagents are dissolved and the cell activation reaction is triggered.

[0095] Microscope observation window: covered with a transparent PDMS film with a thickness of 150 μm, and a digital microscope imaging system with a 20x objective lens was used for real-time image acquisition.

[0096] The microfluidic chip device in step S1 covers multiple functional modules such as sample processing, cell enrichment, reaction and imaging, and embodies the multifunctional integration capability of microfluidic chips in biomedical research.

[0097] S2, adding premixed dry CD63-FITC and CD203c-PE fluorescently labeled antibodies and allergen stimulants into the microfluidic chip reaction chamber. S3, incubating at 37°C for 15-30 minutes to activate the cells. Add CD63-FITC and CD203c-PE fluorescently labeled antibodies and allergen stimulants in the reaction chamber, and activate the basophils by incubation. CD63 and CD203c are important markers of basophil activation, and their expression levels can reflect the activation state of the cells. Through the combination of fluorescent labeling and stimulants, quantitative detection of basophil activation is achieved. Incubate at 37°C, 5% CO2 in an incubator for 15-30 minutes to promote cell uptake of dyes and metabolic reactions.

[0098] Specifically, in this embodiment, the preparation of the tapered microporous filter membrane includes the following steps: forming a microchannel mold on a glass slide by a photolithography process, with an exposure time of 15 seconds; mixing PDMS main agent and curing agent at a mass ratio of 10:1, degassing, and pouring into the mold, and curing at 70°C for 2 hours; cutting and punching after peeling off the PDMS layer, and bonding with the glass slide through oxygen plasma treatment to form a closed microfluidic structure.

[0099] Specifically, in this embodiment, the premixed dry reagents in step S2 include:

[0100] S2.1, fluorescently labeled antibody combination: containing CD63-FITC antibody and CD203c-PE antibody, wherein the concentration of CD63-FITC antibody is 0.1-1.0 μg / μL, and the concentration of CD203c-PE antibody is 0.1-1.0 μg / μL.

[0101] S2.2, allergen stimulant: selected from at least one of anti-IgE antibody or specific allergen protein, with a concentration of 1-10 μg / μL.

[0102] S2.3, freeze-drying protectant: containing polyvinyl alcohol with a mass fraction of 1-5% and sucrose with a mass fraction of 0.5-2%; wherein the premixed dry reagents are prepared by the following method:

[0103] a) mixing the fluorescently labeled antibodies, allergen stimulants and freeze-drying protectants in proportion.

[0104] b) Spotting the mixed solution into the reaction chamber of the microfluidic chip at a volume of 1-5 μL / well.

[0105] c) Drying the mixed solution at -40℃ to -80℃ for 12-24 hours by vacuum freeze-drying process to form a freeze-dried reagent layer.

[0106] S4, acquiring a 512x512 pixel high-resolution image of the basophil sample using a digital microscopic imaging system.

[0107] The ultra-high sensitivity laser confocal live cell imaging analysis system can achieve an imaging speed of 13 frames per second at a resolution of 512x512, capable of capturing rapidly changing sample processes. A high-resolution image (512x512 pixels) is acquired using a digital microscopic imaging system, and image recognition is performed using an improved target detection model. This model uses a lightweight backbone network and a dynamic attention mechanism to reduce computational load and improve recognition accuracy. This method combines microfluidic and AI image recognition technology to achieve high-precision detection of basophil activation.

[0108] S5, identifying and classifying activated basophils in the image through an improved target detection model, which uses a lightweight backbone network and integrates a dynamic attention mechanism, wherein: the lightweight backbone network reduces the computational load to below 5.9 GFLOPs through selective channel processing technology using partial convolution operations; the dynamic attention mechanism automatically identifies key image regions for feature enhancement. A cell activation heat map is generated from the classification results and the basophil activation index is calculated, and a detection report is output. Not only does it provide a quantitative activation index, but it also visually displays the cell activation state through the heat map, providing a reliable basis for clinical diagnosis.

[0109] Figure 2 The improved target detection model structure provided by an embodiment of the present application is shown in the figure. Specifically, as shown in Figure 2 In this embodiment, the improved target detection model in step S5 is a lightweight detection model based on the Transformer architecture, which includes the following modules connected in turn:

[0110] Feature extraction module: uses a four-level hierarchical FasterNet backbone network, each level containing: a partial convolution layer for selective processing of 1 / 4 input channels, with the remaining channels directly passed through; a reverse residual block that expands the channel and interacts with the features through 1x1 convolution; an embedding layer for spatial down-sampling and channel expansion; wherein the overall computational complexity of the feature extraction module is controlled to be below 5.9 GFLOPs.

[0111] The structure aims to reduce redundant calculations by selectively processing input channels with partial convolution, while expanding features and spatially downsampling with reverse residual blocks and embedding layers. Partial convolution is an efficient convolution operation that reduces computational load and memory access by only convolving a portion of the input channels, while the remaining channels are passed directly, thereby improving computational efficiency and reducing computational complexity. Reverse residual blocks and embedding layers are used for feature expansion and spatial downsampling to enhance the model's expressive power. The overall structure achieves extraction and fusion of features at different scales through multi-level hierarchical design, thereby improving the model's performance and efficiency.

[0112] Attention optimization module: includes a bidirectional routing attention mechanism consisting of the following units: dynamic window screening unit, selects the top k key areas based on query relevance; sparse attention calculation unit, only processes the attention weights of the screened areas; GPU acceleration module, adopts dense matrix multiplication to optimize the calculation process; the attention optimization module achieves a real-time inference speed of 145FPS.

[0113] Among them, the bidirectional routing attention mechanism: this mechanism is realized through the dynamic window screening unit and the sparse attention calculation unit. The dynamic window screening unit selects the top k key areas based on query relevance, while the sparse attention calculation unit only processes the attention weights of these key areas, thereby reducing the amount of calculation and improving efficiency.

[0114] GPU acceleration module: this module adopts dense matrix multiplication to optimize the calculation process, in order to improve the computational efficiency. This optimization method can significantly improve the calculation speed on GPU, especially when dealing with large-scale data.

[0115] Performance indicators: this module achieves a real-time inference speed of 145FPS, indicating that it has high performance in actual application. The attention optimization module combines sparse attention, GPU acceleration and efficient calculation strategies, aiming to improve the computational efficiency and inference speed of the attention mechanism.

[0116] Multi-task output module: includes: cell localization branch, used to generate 512x512 resolution detection boxes and segmentation masks; activation analysis branch, used to output classification heat maps containing the following parameters: CD63+ / CD203c+ double positive cell proportion, activation index score based on fluorescence intensity, spatial distribution heat map of abnormal activation cells. Multi-task output module: includes cell localization branch, activation analysis branch and related parameter output, used to generate detection boxes, segmentation masks, classification heat maps, etc.

[0117] The multi-task output module includes three main branches: cell localization branch, activation analysis branch, and related parameter output. These modules aim to generate output results such as detection boxes, segmentation masks, and classification heatmaps to support cell detection and analysis. In the cell localization branch, the system generates 512x512 resolution detection boxes and segmentation masks to identify and locate cells. This high-resolution output helps accurately identify the location and morphological features of cells. The activation analysis branch outputs various parameters, including the proportion of CD63+ / CD203c+ double-positive cells, activation index scores based on fluorescence intensity, and spatial distribution heatmaps of abnormal activation cells. These parameters help evaluate the activation state and functional status of cells. The multi-task output module provides comprehensive support for cell detection and analysis by integrating cell localization and activation analysis functions. The design of this multi-task output module helps improve the accuracy of detection and the depth of analysis.

[0118] Performance optimization module: includes channel pruning unit for model volume compression, mixed precision calculation unit for inference acceleration, and ONNX runtime interface for cross-platform deployment capability. Performance optimization module: includes channel pruning unit, mixed precision calculation unit and ONNX runtime interface for model compression, inference acceleration and cross-platform deployment.

[0119] The model combines the lightweight design of the Transformer architecture and the multi-task processing capability, suitable for high-precision and high-efficiency detection tasks.

[0120] Specifically, the FasterNet backbone network adopts hierarchical design and contains four stages. An embedding or merging layer is set before each stage to perform spatial downsampling and channel expansion. The core computing unit of each stage is the FasterNet block, which includes a partial convolution layer and two pointwise convolution layers in turn, and transmits input features through a shortcut connection to enhance gradient flow and feature reuse capabilities.

[0121] The model adapts to the activated basophil granulocyte recognition task through transfer learning, specifically including:

[0122] Model training phase: based on the pre-trained target detection model architecture, fine-tuning training is performed using a medical microscopic image dataset. The dataset contains labeled activated basophil granulocytes and their morphological features, including granule distribution characteristics and cell membrane integrity characteristics. Through transfer learning, the model can utilize the feature extraction capabilities of the pre-trained model to quickly adapt to new tasks.

[0123] Feature extraction phase: multi-scale cell features are extracted through the FasterNet backbone network.

[0124] FasterNet adopts a four-level hierarchical structure, each level containing part of the convolutional layer, the reverse residual block and the embedding layer, and the overall computational complexity is controlled below 5.9 GFLOPs. Combined with the dynamic sparse attention mechanism of the BRA module, the significant features of the activated cell activation area are focused, including the degranulation phenomenon feature and the membrane protrusion feature.

[0125] Output optimization stage: adjust the output layer to the target category detection and classification result, the target category includes activated basophilic granulocytes and non-activated basophilic granulocytes. Use the intersection over union threshold to optimize the bounding box positioning accuracy, and realize the accurate differentiation of closely adjacent cell clusters. The dynamic window screening unit, sparse attention calculation unit and GPU acceleration module in the BRA module realize the real-time inference speed of 145FPS.

[0126] Among them, the FasterNet backbone network adopts a four-level hierarchical structure, each level containing part of the convolutional layer, the reverse residual block and the embedding layer, and the overall computational complexity is controlled below 5.9 GFLOPs; the BRA module includes a dynamic window screening unit, a sparse attention calculation unit and a GPU acceleration module, and realizes a real-time inference speed of 145FPS.

[0127] S6, based on the classification result, generate a cell activation heat map and calculate the basophilic granulocyte activation index, and output a detection report. Through the classification result, a cell activation heat map is generated and the basophilic granulocyte activation index is calculated, and a detection report is output. Not only provide quantitative activation index, but also directly show the cell activation state through the heat map, provide reliable basis for clinical diagnosis.

[0128] Specifically, in the present embodiment, the S6, based on the classification result, generates a cell activation heat map and calculates the basophilic granulocyte activation index, and outputs a detection report, comprising:

[0129] Based on the classification result, generate a cell activation heat map: use the detection frame confidence score and class probability output by the model to generate a two-dimensional heat map through interpolation algorithm, and directly display the distribution density and activation degree of activated basophilic granulocytes in the image;

[0130] Calculate the basophilic granulocyte activation index: count the ratio of the number of activated cells to the number of total basophilic granulocytes in the field of view, and calculate the quantitative activation index by combining the cell morphological parameters;

[0131] Output structured detection report: automatically generate a visual report containing activation heat map, activation index, cell count and abnormal cell label, support DICOM / PDF format output, and interface with hospital information system.

[0132] The generating the two-dimensional heat map by the interpolation algorithm comprises: performing Gaussian kernel density estimation on the detection frame output by the model to generate a two-dimensional heat distribution map, and performing color gradient rendering on the heat map according to the fluorescence labeling intensity; and superimposing and displaying on the original microscopic image to form a composite visualization result.

[0133] Figure 3 A flowchart for generating a two-dimensional heat map is provided for an embodiment of the present application. Further, as shown in Figure 3 The generating the two-dimensional heat map specifically comprises the following steps:

[0134] S6.1, input data preprocessing: receiving a detection frame set output by a target detection model Wherein each detection frame b i =(x i ,y i ,w i ,h i ,s i ) contains center coordinates (x i ,y i ), width and height w i ,h i , and confidence score s i ; and the fluorescence intensity value associated with each detection frame is obtained

[0135] S6.2, Gaussian kernel density estimation: based on the center coordinates (x i ,y i ) of the detection frame, the discrete points are smoothed and interpolated by the Gaussian kernel function to generate a continuous probability density distribution:

[0136]

[0137] Wherein σ represents the bandwidth of the kernel, which controls the smoothness of the heat distribution and is proportional to the average size of the detection frame, I i represents the fluorescence labeling intensity value associated with the i-th detection frame, represents the fluorescence signal intensity of the CD63 / CD203c marker on the cell surface, and is used as a weighted coefficient of confidence and fluorescence intensity, (x,y) represents the coordinates of any point in the image, which is used to calculate the heat value of the point, (x i ,y i ) represents the center coordinates of the i-th detection frame, represents the center position of the activated basophil granulocyte, and s i represents the confidence score of the i-th detection frame, reflecting the confidence degree of the model that the detection frame is an activated cell. In this step, the bandwidth controls the smoothness of the kernel, which is proportional to the average size of the detection frame; the fluorescence intensity value is used as a weighted coefficient of confidence and fluorescence intensity to calculate the heat value of any point in the image.

[0138] S6.3, Density field calculation and normalization: Generate grid points (x g ,y g ) on the image domain WxH, calculate the superimposed density value of each grid point:

[0139]

[0140] Normalize D(x g ,y g ) to the range [0, 1] to obtain the heat density field D norm , where (x g ,y g ) represents the coordinates of the image grid points, used to generate the discretized heat distribution, N represents the total number of detection boxes, i.e. the number of activated basophilic granulocytes identified in the image, D(x g ,y g ) represents the superimposed density value at grid point (x g ,y g ), reflecting the cell distribution density and activation intensity of the region, and D norm represents the normalized heat density field, used for standardized display. Generate grid points on the image domain, calculate the superimposed density value of each grid point, and normalize the result to the range to obtain the heat density field, which is used for standardized display.

[0141] S6.4, Fluorescence signal enhancement processing: Adjust the color saturation according to the fluorescence intensity I i :

[0142]

[0143] where Saturation(x, y) represents the heat map color saturation adjusted based on the fluorescence intensity I i , the higher the value, the stronger the fluorescence signal, K i (x g ,y g ) represents the Gaussian kernel density contribution value of the i-th detection box at image grid point (x g ,y g ), used to quantify the influence of the detection box on the current grid point (x g ,y g ) in the heat map. Adjust the color saturation according to the fluorescence intensity to enhance the visual effect of the heat map.

[0144] S6.5, Heat map and image fusion: Superimpose the generated heat map Color(x, y) and the original microscopic image Image(x, y) through Alpha blending:

[0145] Composite(x, y) = a Color(x, y) + (1 - a) Image(x, y),

[0146] wherein a represents a transparency adjustment parameter for controlling the superimposition ratio of the heat map and the original image, Color(x, y) represents the color value of the generated heat map, and Image(x, y) represents the pixel value of the original microscopic image. The generated heat map and the original microscopic image are superimposed through Alpha blending to enhance the visualization effect.

[0147] S6.6, target edge enhancement: superimpose contour markers on the detection frame edge to improve the recognition of activated cells.

[0148] S6.7, output visualization results: generate a composite heat map containing activated cell distribution density, fluorescence intensity gradient and edge markers for visualization of the detection report. This process combines target detection, density estimation, image processing and visualization techniques to generate a heat map reflecting the distribution of activated basophilic granulocytes and fluorescence signals.

[0149] This method combines microfluidic technology, fluorescence labeling, AI image recognition and activation index calculation, and has the characteristics of high throughput, high sensitivity and fast response. Compared with traditional flow cytometry, this method has advantages in cell enrichment, activation detection and image recognition, and is suitable for rapid diagnosis and treatment monitoring of allergic diseases.

[0150] Example Two

[0151] As shown in Figure 4 The present application provides a basophilic granulocyte activation detection system architecture based on microfluidics and AI image recognition, which is applied to the basophilic granulocyte activation detection system based on microfluidics and AI image recognition as described in Example One, and includes a microfluidic chip module 11, a fluorescence labeling module 12, a temperature-controlled incubation module 13, an image acquisition module 14, an AI analysis module 15, and a human-computer interaction module 16.

[0152] The microfluidic chip module 11 is used to inject the peripheral blood sample to be tested into the microfluidic chip device, and to enrich basophilic granulocytes through a tapered micropore filter membrane structure with a pore size of 6-8 μm. The microfluidic chip device includes a sample injection port, a cell enrichment area, a reaction chamber with pre-loaded freeze-dried reagents, and a microscopic observation window.

[0153] The fluorescence labeling module 12 is used to add pre-mixed dry CD63-FITC and CD203c-PE fluorescent labeling antibodies and allergen stimulants in the microfluidic chip reaction chamber.

[0154] The temperature-controlled incubation module 13 is used to incubate the cells at 37°C for 15-30 minutes to fully activate the cells.

[0155] An image acquisition module 14 is configured to acquire a 512*512 pixel high-resolution image of the basophil sample using a digital microscopic imaging system.

[0156] An AI analysis module 15 is configured to identify and classify the activated basophils in the image by using an improved target detection model, which uses a lightweight backbone network and a dynamic attention mechanism, wherein: the lightweight backbone network reduces the calculation amount to below 5.9 GFLOPs by using a partial convolution operation and a selective channel processing technology; and the dynamic attention mechanism automatically identifies the key regions of the image for feature enhancement.

[0157] A human-computer interaction module 16 is configured to generate a cell activation heat map and calculate a basophil activation index based on the classification result, and output a detection report.

[0158] Figure 5 An electronic device is provided in an embodiment of the present application. As shown in Figure 5 The electronic device at least includes the following parts: a processor 101 and a memory 100, a communication interface 103, and a bus 102.

[0159] In an embodiment of the present application, the memory 100 is configured to store processor 101 executable instructions, and the processor 101 is configured to implement the method of the first aspect when executing the instructions.

[0160] In an embodiment of the present application, a computer readable storage medium includes instructions, and the instructions instruct the device to execute the method of the first aspect. For example, the instructions instruct the device to execute the method shown in the flow steps in the first aspect. Figure 1

[0161] The program that works in the electronic device related to an embodiment of the present application can be a program (a program that makes a computer function) that controls a central processing unit (CPU) and the like to realize the functions of the above-described embodiments related to one solution of the present application. Then, the information processed by these devices is temporarily stored in a random access memory (RAM) when it is processed, and then stored in various ROMs such as a read-only memory (Flash ROM), a hard disk drive (HDD), and the like, and read out, corrected, and written by the CPU as needed.

[0162] It should be noted that a part of the electronic device of the above-described embodiments can also be realized by a computer. In this case, a program for realizing the control function can be recorded in a computer readable recording medium, and the program recorded in the recording medium is read into a computer and executed to realize.​

[0163] Note that the "computer" referred to here means a computer built into an electronic device, using a computer including hardware such as an OS, peripheral devices, and the like. Further, the "computer-readable recording medium" refers to removable media such as a floppy disk, a magneto-optical disk, a ROM, a CD-ROM, and the like, storage devices built into computers such as hard disks, and the like.

[0164] Further, the "computer-readable recording medium" can include a medium that dynamically stores a program for a short period of time, such as a communication line in the case of transmitting the program via a network such as the Internet or a communication line such as a telephone line, and a medium that stores a program for a fixed period of time, such as a volatile memory inside a computer that is a server or a client in this case. Further, the above program can be a program for realizing a part of the above functions, and can also be a program that can realize the above functions by being combined with a program already recorded in a computer.

[0165] Further, the electronic device in the above embodiments can also be realized as an assembly (device group) composed of a plurality of devices. Each device constituting the device group can have each function or a part or all of each functional block of the electronic device of the above embodiments. As the device group, all of each function or each functional block of the electronic device can be possessed.

[0166] Those skilled in the art will recognize that the above embodiments are merely illustrative of the application and should not be taken as limiting the scope of the application, as defined by the appended claims, since modification and changes can be made thereto without departing from the spirit of the application.

Claims

1. A method for detecting basophil activation based on microfluidics and AI image recognition, characterized in that, The method includes: S1. Inject the peripheral blood sample to be tested into the microfluidic chip device, and enrich basophils through the cone-shaped microporous filter membrane structure with a pore size of 6-8 μm. The microfluidic chip device includes a sample injection port, a cell enrichment area, a reaction chamber with pre-prepared lyophilized reagents, and a microscopic observation window. S2. Add premixed and dried CD63-FITC and CD203c-PE fluorescently labeled antibodies and allergen stimulants into the microfluidic chip reaction chamber; S3. Incubate at 37°C for 15-30 minutes to fully activate the cells; S4. A 512×512 pixel high-resolution image of a basophil sample was obtained using a digital microscopic imaging system. S5. Activated basophils in the image are identified and classified using an improved target detection model. The model employs a lightweight backbone network and incorporates a dynamic attention mechanism, wherein: The lightweight backbone network reduces computation to below 5.9 GFLOPs through selective channel processing of partial convolution operations; The dynamic attention mechanism automatically identifies key regions of the image and performs feature enhancement. S6. Generate a cell activation heatmap based on the classification results and calculate the basophil activation index, then output a detection report; The improved target detection model in step S5 is a lightweight detection model based on the Transformer architecture, which includes the following modules connected in sequence: Feature extraction module: A four-level hierarchical FasterNet backbone network is adopted, each level includes: a partial convolutional layer, which is used to selectively process 1 / 4 of the input channels, and the remaining channels are directly passed; an inverse residual block, which realizes channel expansion and feature interaction through 1×1 convolution; and an embedding layer, which is used to perform spatial downsampling and channel expansion. The overall computational complexity of the feature extraction module is controlled below 5.9 GFLOPs. Attention optimization module: includes a bidirectional routing attention mechanism, which consists of the following units: a dynamic window filtering unit, which selects the top k key regions based on query relevance; a sparse attention calculation unit, which only processes the attention weights of the filtered regions; and a GPU acceleration module, which uses dense matrix multiplication to optimize the calculation process; the attention optimization module achieves a real-time inference speed of 145 FPS. Multi-task output module: includes: cell localization branch, used to generate a 512×512 resolution detection box and segmentation mask; activation analysis branch, used to output a classification heatmap containing the following parameters: percentage of CD63+ or CD203c+ double positive cells, activation index score based on fluorescence intensity, and spatial distribution heatmap of abnormally activated cells. Performance optimization module: includes: channel pruning unit for compressing model size; mixed precision calculation unit for accelerating inference process; ONNX runtime interface for enabling cross-platform deployment; The model is adapted to the activated basophil recognition task through transfer learning, specifically including: Model training phase: Based on the pre-trained target detection model architecture, fine-tuning training is performed using a medical microscopic image dataset, which contains labeled activated basophils and their morphological features, including granule distribution features and cell membrane integrity features. Feature extraction stage: Multi-scale cell features are extracted through the FasterNet backbone network; combined with the dynamic sparse attention mechanism of the BRA module, the salient features of cell activation regions are focused, including degranulation phenomenon features and membrane protrusion features; Output optimization stage: Adjust the output layer to the target category detection and classification results, where the target category includes activated basophils and inactivated basophils; optimize the bounding box localization accuracy using the cross-union threshold to achieve accurate differentiation of closely adjacent cell clusters; The FasterNet backbone network adopts a four-level hierarchical structure, with each level containing some convolutional layers, inverse residual blocks, and embedding layers, and the overall computational complexity is controlled below 5.9 GFLOPs. The BRA module includes a dynamic window filtering unit, a sparse attention computing unit, and a GPU acceleration module, achieving a real-time inference speed of 145 FPS.

2. The method for detecting basophil activation based on microfluidics and AI image recognition according to claim 1, characterized in that, The microfluidic chip device in step S1 includes: Sample injection port: used to introduce the peripheral blood sample to be tested, connected to the injection pump to control the flow rate at 10-50 μL / min; Cell enrichment zone: An integrated cone-shaped microporous filter membrane with a pore size of 6-8 μm enriches basophils through physical retention while filtering out erythrocytes and plasma components; Reaction chamber: Pre-contains lyophilized CD63-FITC and CD203c-PE fluorescent antibodies and allergen stimulants. The chamber volume is 50 μL. After injecting buffer, the reagents are dissolved and the cell activation reaction is triggered. Microscopic observation window: Covered with a 150μm thick transparent PDMS film, and equipped with a digital microscopic imaging system with a 20x objective lens for real-time image acquisition.

3. The method for detecting basophil activation based on microfluidics and AI image recognition according to claim 2, characterized in that, The FasterNet backbone network adopts a hierarchical design, consisting of four stages. Each stage is preceded by an embedding or merging layer to perform spatial downsampling and channel expansion. The core computational unit of each stage is a FasterNet block, which sequentially includes a partial convolutional layer and two pointwise convolutional layers. Input features are passed through shortcut connections to enhance gradient flow and feature reuse capabilities.

4. The method for detecting basophil activation based on microfluidics and AI image recognition according to claim 3, characterized in that, S6, based on the classification results, generates a cell activation heatmap and calculates the basophil activation index, outputting a detection report, including: Cell activation heatmap generated based on classification results: Using the detection box confidence score and class probability output by the model, a two-dimensional heatmap is generated through an interpolation algorithm to intuitively display the distribution density and activation degree of activated basophils in the image; Calculate the basophil activation index: The ratio of the number of activated cells to the total number of basophils in the statistical field of view is used for weighted calculation in combination with cell morphology parameters to obtain a quantitative activation index; Output structured detection reports: Automatically generate visual reports containing activation heatmaps, activation indices, cell counts, and abnormal cell markers, supporting DICOM or PDF format output, and interfacing with hospital information systems.

5. The method for detecting basophil activation based on microfluidics and AI image recognition according to claim 4, characterized in that, The generation of the two-dimensional heat map using the interpolation algorithm includes: Gaussian kernel density estimation is performed on the detection boxes output by the model to generate a two-dimensional thermal distribution map. Color gradient rendering of the heatmap is performed based on the intensity of the fluorescent markers; The images are overlaid on the original microscopic images to form a composite visualization.

6. The method for detecting basophil activation based on microfluidics and AI image recognition according to claim 5, characterized in that, The generation of the two-dimensional heat map specifically includes the following steps: S6.1 Input Data Preprocessing: Receive the set of detection boxes output by the target detection model. , where each detection box Including center coordinates ,Width ,high and confidence score Simultaneously, acquire the fluorescence intensity value associated with each detection frame. ; S6.2 Gaussian kernel density estimation: based on the center coordinates of the detection box A continuous probability density distribution is generated by smoothly interpolating discrete points using a Gaussian kernel function. ， in, This represents the bandwidth of the control kernel, which controls the smoothness of the thermal distribution and is proportional to the average size of the detection frame. Indicates the first The fluorescence label intensity value associated with each detection box represents the fluorescence signal intensity of CD63 or CD203c labels on the cell surface, and is used as a weighting coefficient between confidence level and fluorescence intensity. This represents the coordinates of any point in the image, used to calculate the thermal value at that point. Indicates the first The center coordinates of each detection box indicate the central location of the activated basophil. Indicates the first The confidence score of each detection box reflects the model's confidence that the detection box represents an active cell; S6.3 Density Field Calculation and Normalization: In the Image Domain Generate grid points Calculate the stacking density value for each grid point: ， Will Normalization to The range is used to obtain the thermal density field. ,in, The coordinates of the image grid points are used to generate a discretized thermal distribution. This represents the total number of detection boxes, i.e., the number of activated basophils identified in the image. Represents grid points The superposition density value at a given location reflects the cell distribution density and activation intensity in that region. This represents the normalized thermodynamic density field, used for standardized display. S6.4 Fluorescence signal enhancement processing: based on fluorescence intensity Adjust color saturation as a weight: ， in, Indicated based on fluorescence intensity Adjusting the color saturation of the heatmap; a higher value indicates a stronger fluorescence signal. Indicates the first A detection box at image grid points The Gaussian kernel density contribution value at a given location is used to quantify the impact of the detection frame on the current grid point in the heatmap. The extent of the impact; S6.5, Heatmap and Image Fusion: The generated heatmap... Compared with the original microscopic image Alpha mixing and superposition: ， in, This represents the transparency adjustment parameter, used to control the overlay ratio between the heatmap and the original image. This represents the color values ​​of the generated heatmap. Represents the pixel values ​​of the original microscopic image; S6.6 Target Edge Enhancement: Overlay contour markers at the edges of the detection box to improve the recognition of activated cells; S6.7 Output visualization results: Generate a composite heatmap containing the distribution density of activated cells, fluorescence intensity gradient and edge markers for visualization of the test report.

7. A basophil activation detection system based on microfluidics and AI image recognition, applied to the basophil activation detection method based on microfluidics and AI image recognition as described in any one of claims 1 to 6, characterized in that, The system includes: A microfluidic chip module is used to inject a peripheral blood sample to be tested into a microfluidic chip device, and enrich basophils through a cone-shaped microporous filter membrane structure with a pore size of 6-8 μm. The microfluidic chip device includes a sample injection port, a cell enrichment area, a reaction chamber with pre-prepared lyophilized reagents, and a microscopic observation window. The fluorescent labeling module is used to add premixed and dried CD63-FITC and CD203c-PE fluorescently labeled antibodies and allergen stimulants into the reaction chamber of the microfluidic chip; The temperature-controlled incubation module is used to incubate at 37°C for 15-30 minutes to fully activate the cells; The image acquisition module is used to acquire 512×512 pixel high-resolution images of basophil samples using a digital microscopic imaging system; The AI ​​analysis module is used to identify and classify activated basophils in images using an improved target detection model. This model employs a lightweight backbone network and incorporates a dynamic attention mechanism, wherein: The lightweight backbone network reduces computation to below 5.9 GFLOPs through selective channel processing of partial convolution operations; The dynamic attention mechanism automatically identifies key regions of the image and performs feature enhancement. The human-computer interaction module is used to generate cell activation heatmaps based on classification results, calculate the basophil activation index, and output a detection report.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that instructs the device to perform the basophil activation detection method based on microfluidics and AI image recognition as described in any one of claims 1 to 6.

Images