A system and method for detecting and screening activity abnormalities of frozen-thawed sperm

Abstract

This invention discloses a system and method for detecting and screening abnormal sperm activity in frozen and revived sperm, relating to the fields of big data processing and intelligent control technology. The invention involves injecting revived sperm into a microfluidic chip and acquiring a continuous phase-difference microscopic image sequence; extracting features such as sperm centroid coordinates, dynamic motion state, and internal refractive index gradient; using an improved dual-flow spatiotemporal graph convolutional neural network to process these features. This network introduces a spatial attention mechanism to amplify minute damage, utilizes an adaptive distortion time module to capture abnormal trajectories, and combines a multi-center contrast-focus joint loss function to overcome class imbalance, outputting sperm quality labels and coordinates; and integrates sperm instantaneous velocity for precise compensation of hydrodynamic delay time, triggering real-time microfluidic physical field switching to execute screening. This invention is entirely label-free and damage-free, achieving highly sensitive detection and automated, precise sorting of ice crystal damage in revived sperm.

Description

Technical Field

[0001] This invention belongs to the field of big data processing and intelligent control technology, and specifically relates to a system and method for detecting and screening abnormal sperm motility in frozen and thawed sperm. Background Technology

[0002] In the field of assisted reproductive technology (ART), sperm cryopreservation and thawing are crucial for ensuring fertility reserves. However, sperm are highly susceptible to damage from physical puncture by ice crystals and osmotic stress during freezing and thawing, leading to serious problems such as plasma membrane damage, decreased acrosome integrity, reduced motility, and increased DNA fragmentation after thawing. These issues directly affect the success rate of in vitro fertilization (IVF) or intracytoplasmic sperm injection (ICSI).

[0003] Currently, clinical testing and screening of frozen and thawed sperm primarily relies on two existing technologies. The first is manual microscopic morphological observation, which depends on the embryologist's subjective experience. This method is not only inefficient and has extremely low throughput, but also makes it very difficult to distinguish minute membrane damage in transparent sperm under a regular optical microscope, thus failing to accurately assess the degree of cryopreservation damage within the cells. The second is flow cytometry sorting based on fluorescent staining. While this method can accurately detect sperm viability and DNA fragmentation using specific fluorescent probes, fluorescent dyes (such as PI and Hoechst) are often cytotoxic or mutagenic, and high-energy laser irradiation can cause irreversible oxidative damage to sperm DNA. Sperm that have undergone fluorescent labeling and flow cytometry sorting have lost their physiological function for subsequent clinical fertilization and can only be used for scientific research or diagnostic evaluation.

[0004] This shows that there is an irreconcilable contradiction between non-destructiveness and accuracy in existing technologies; there is still a need for an intelligent detection and control system that is completely label-free and non-destructive, and can identify microstructural damage and abnormal movement trajectories of frozen and revived sperm with high throughput and objective accuracy, and achieve automated sorting. Summary of the Invention

[0005] (a) Technical problems to be solved To address the problems in related technologies, this invention provides a system and method for detecting and screening abnormal sperm motility in frozen and thawed sperm, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: S1. The frozen and thawed sperm sample is injected into a microfluidic chip, and microscopic image sequences and scattering signals are continuously acquired using wide-field optical imaging combined with phase difference microscopy detection technology. S2. Extract a set of sperm multidimensional features from the microscopic image sequence and scattering signal in S1, including centroid coordinate time series, dynamic motion state features, single-frame phase difference image and internal microstructure features. S3. Train an improved dual-stream spatiotemporal graph convolutional neural network using a historical sperm feature dataset to obtain a sperm activity correlation model; The sperm multidimensional feature set output by S2 is input into the sperm activity association model to obtain the quality grading label of each sperm and the pixel coordinates of the current frame; S4. Receive the quality grading label and pixel coordinates output by S3, and call the dynamic motion state features output by S2 to calculate the hydrodynamic delay time of the target sperm. Based on the quality grading label and the fluid dynamic delay time, the physical field switch of the microfluidic sorting actuator is triggered in real time to change the fluid motion trajectory of the target sperm, thereby obtaining the screened frozen and revived sperm. This invention combines wide-field phase difference imaging with microfluidic technology to achieve label-free detection of frozen and revived sperm, avoiding irreversible damage to sperm DNA caused by traditional fluorescent staining. Targeting the unique microstructural damage and nonlinear motion distortions of revived sperm, an improved dual-flow spatiotemporal graph convolutional neural network is used for deep feature mining, overcoming the limitations of low efficiency and high subjectivity of traditional manual microscopic examination. By combining AI-output classification labels with precise hydrodynamic delay time calculations and linking them with the microfluidic physical field, closed-loop control from microscopic feature recognition to macroscopic physical screening is achieved, improving the reliability and automation level of high-quality sperm screening.

[0007] Preferably, step S1 includes the following steps: S11. The frozen and thawed sperm sample is pushed into the detection area of ​​a microfluidic chip with a fluid focusing structure at a constant flow rate using a precision injection pump, so that the sperm are distributed in a single-layer fluid plane. S12. Turn on the wide field light source and phase difference annular aperture of the microscope, and use a high-speed CMOS camera to continuously acquire a sequence of continuous dynamic microscopic images of sperm populations within the single-layer fluid plane field of view. S13. The continuous dynamic microscopic image sequence acquired in S12 is divided into video stream blocks with fixed time windows through the edge computing gateway. An absolute timestamp and spatial coordinate system origin information are added to each frame image and written in parallel to the Hadoop distributed file system to generate a large dataset of sperm optical features for subsequent steps. This invention acquires high-resolution, label-free, dynamic images of sperm by using hydrodynamic focusing and high-frame-rate wide-field phase difference imaging. It also utilizes edge computing to achieve distributed, structured storage of the data, laying a high-quality data foundation for subsequent feature extraction.

[0008] Preferably, step S2 includes the following steps: S21. Read the continuous dynamic microscopic image sequence from the sperm optical feature big data set in S1, apply the Gaussian mixture background modeling algorithm to separate the foreground moving target, and use the multi-target tracking algorithm to obtain the centroid coordinate time series of each sperm in the continuous frame. S22. Read the centroid coordinate time series in S21, and calculate the dynamic motion state features based on the coordinate difference between adjacent frames and the timestamp; the dynamic motion state features include curvilinear motion speed, linear motion speed and average path speed. S23. Based on the centroid coordinate time series in S21, extract a single-frame phase difference image of each sperm within its centroid coordinate bounding box from the original image sequence; calculate the pixel-level second-order spatial derivative of the single-frame phase difference image to quantify the refractive index gradient distribution inside the cell. S24. Set the refractive index gradient threshold, count the area and distribution of connected regions of pixels within the bounding box that are higher than the refractive index gradient threshold, and extract the combination of the top body light scattering intensity attenuation feature and the film micro-damage feature as the internal microstructure feature. This invention transforms unstructured massive image sequences into structured mathematical features such as sperm centroid coordinates, multidimensional motion velocity, and quantified plasma membrane / acrosomal refractive index anomalies through multi-target tracking and second-order spatial derivative calculation; it eliminates the dependence on chemical staining and achieves non-destructive, objective, and numerical characterization of sperm membrane integrity and internal microstructural damage.

[0009] Preferably, step S3 includes the following steps: S31. Based on the data characteristics of weak contrast and trajectory distortion of frozen and revived sperm, the structure and loss function of the traditional dual-stream graph convolutional neural network model are improved to obtain an improved dual-stream spatiotemporal graph convolutional neural network. S32. Collect images of frozen and thawed sperm from historical batches and the corresponding clinical activity test results, extract features and perform manual classification and annotation to obtain a historical sperm feature dataset; S33. Use the historical feature dataset to perform backpropagation training on the improved dual-stream spatiotemporal graph convolutional neural network, adjust the node weight parameters of the improved dual-stream spatiotemporal graph convolutional neural network, and when the loss function value on the validation set meets the preset convergence condition, solidify the network parameters to obtain the sperm activity association model. S34. Input the sperm multidimensional feature set in S2 into the sperm activity association model for forward propagation inference to obtain the quality grading label of each sperm and the pixel coordinates of the current frame; This invention addresses the unique characteristics of weak contrast and trajectory distortion in frozen and revived sperm by improving the dual-stream spatiotemporal graph convolutional neural network. It solves the model collapse problem caused by the extremely high proportion of dead sperm in the revived samples by objectively calculating and setting a multi-center contrast-focus joint loss function. Furthermore, by accurately mapping the purely mathematical features extracted by S2 to quality labels and absolute coordinates to guide physical sorting, it not only improves the detection rate of high-quality sperm but also ensures the interpretability of classification decisions.

[0010] Preferably, step S31 includes the following steps: S311. Improve the standard convolutional layer in the spatial feature extraction stream of the traditional dual-flow graph convolutional neural network model by adding a phase difference-enhanced spatial attention mechanism module to amplify the high-frequency micro-variation features of refractive index. S312. Improve the fixed receptive field temporal convolutional layer in the temporal feature extraction stream of the traditional dual-flow graph convolutional neural network model and replace it with an adaptive distortion temporal graph convolutional module, which is used to dynamically adjust the dilation rate parameter of the temporal graph convolutional kernel according to the instantaneous acceleration. S313. The cross-entropy loss function of the traditional dual-flow graph convolutional neural network model is deleted and replaced, and a multi-center contrast-focus joint loss function is added to resist the noise of massive dead sperm impurities and amplify the feature weights of high-quality sperm. This invention addresses the pain points of weak contrast, motion distortion, and class imbalance in revived sperm by reconstructing the spatial flow, temporal flow, and loss function of traditional neural networks. This enables the network to adaptively focus on high-frequency micro-damage and abnormal trajectories, enhancing the model's ability to capture underlying damaged features.

[0011] Preferably, the training of the multi-center contrast-focus joint loss function in step S33 includes the following steps: S331. In the feature space, set a learnable cluster center for each quality grading category, calculate the Euclidean distance between the sample feature vector and the center of its category as the intra-class cohesion, calculate the distance with the center of the non-category as the inter-class repulsion, and force the feature vectors of dead sperm and impurity samples to be pushed away from the feature domain of excellent sperm, thus obtaining the contrast loss term. S332. By introducing a dynamic adjustment factor, the focus loss term is calculated, and the excellent sperm samples, which account for a very small proportion, are given a penalty weight that is higher than that of dead sperm samples by a preset multiple. The comparison loss term and the focus loss term are weighted and summed to calculate the total loss, and the network weight parameters are updated through the Adam optimizer. S333. Set the maximum number of training iterations; when the total loss decreases for η10 consecutive training cycles is less than the total loss decrease threshold of 0.001, or when the maximum number of training iterations is reached, the preset convergence condition is satisfied, and the sperm activity correlation model is obtained. This invention constructs a multi-center comparison-focus joint loss function based on objective data distribution, which effectively overcomes the serious class imbalance problem caused by the massive amount of dead sperm impurities in the resuscitation fluid, forces the network to amplify the feature weights of a very small number of high-quality sperm, and improves the model's classification robustness under extremely high background noise.

[0012] Preferably, step S34 includes the following steps: S341, Phase Difference Enhancement Spatial Attention Mechanism Module, reads a single-frame phase difference image from the sperm multidimensional feature set output by S2, and generates a refractive index anomaly gradient map by calculating the second-order phase derivative of adjacent pixels through the Laplacian operator. The refractive index anomaly gradient map is input into the channel perception network to generate a spatial attention mask, and then multiplied element-wise with the original convolutional feature map to amplify the high-frequency micro-variation features at the top edge and the damaged part of the plasma membrane, thus obtaining the microstructure feature vector. S342, the adaptive distortion time-map convolution module reads the centroid coordinate time series from the sperm multidimensional feature set in S2, calculates the instantaneous acceleration of sperm between adjacent frames; and dynamically maps the absolute value of the instantaneous acceleration to generate the dilation rate parameter of the time-map convolution kernel to obtain the motion trajectory feature vector. Specifically: if the absolute value of acceleration is lower than the micro-tremor threshold, the expansion rate is kept at 1; if it is between the micro-tremor threshold and the floating threshold, the expansion rate is set to 2; if it is higher than the floating threshold, the expansion rate parameter is set to a discrete integer such as 4. S343. The microstructure feature vector and the motion trajectory feature vector are concatenated and input into the fully connected classification layer of the sperm activity association model; The fully connected classification layer uses the Softmax activation function for classification mapping, outputting quality grade labels including excellent, viable but abnormal, dead sperm, and impurities, and extracts the center point of the bounding box of sperm in the image space, and simultaneously outputs the pixel coordinates of sperm in the current frame; This invention accurately integrates the static microscopic damage and dynamic trajectory distortion features of sperm through the collaborative reasoning of spatial attention mechanism and adaptive temporal graph convolution, eliminating the feature extraction blind spots of traditional algorithms and realizing high-confidence intelligent grading and coordinate positioning of revived sperm activity.

[0013] Preferably, step S4 includes the following steps: S41. The central control computer receives the sperm quality grading label and its pixel coordinates in the current frame from the output of S3 in real time. S42. Based on the pre-calibrated physical space mapping matrix between the optical field of view and the microfluidic chip, convert the pixel coordinates received in S41 into absolute physical coordinates within the microfluidic chip. S43. Call the dynamic motion state features in the sperm multidimensional feature set output by S2, and combine them with the absolute physical coordinates to calculate the hydrodynamic delay time required for the target sperm to move from the current position to the microfluidic sorting branch with the laminar liquid. S44. When the target sperm with the received quality grading label of "excellent" reaches the trigger warning line, the central control computer sends a control level to the radio frequency signal generator of the microfluidic sorting actuator within the preset advance response period before the end of the countdown of the hydrodynamic delay time calculated in S43, to excite the sorting physical field and deflect the target sperm to the collection channel. This invention transforms virtual grading labels and coordinates, combined with precisely calculated fluid dynamic delay time, into high-frequency control commands for the physical field. Through precise triggering of surface acoustic waves, it achieves non-destructive physical deflection and automatic collection of highly active sperm, thus connecting the intelligent detection and physical screening processes and forming a high-throughput, automated closed-loop control method.

[0014] Preferably, the calculation process of the hydrodynamic delay time in S43 includes the following steps: S431. Extract the laminar parabolic velocity profile distribution equation of the microfluidic chip, and determine the liquid velocity of the streamline layer in which the target sperm is located based on the Y-axis lateral coordinate of the absolute physical coordinates of the target sperm. S432. The basic arrival time is obtained by dividing the straight-line distance between the current longitudinal coordinate of the target sperm on the X-axis and the center coordinate of the sorting branch by the fluid velocity of the streamline layer. S433: Read the linear motion velocity in the dynamic motion state characteristics of S2, add its projection on the X-axis as a compensation term to the liquid velocity for time compensation correction, and output the fluid dynamics delay time. This invention breaks away from the traditional, crude method of relying solely on average flow velocity to estimate time in sorting. By integrating the laminar flow parabolic velocity profile with the instantaneous linear velocity of the sperm itself, it achieves dynamic and precise compensation for fluid dynamic delay time, thereby improving the accuracy of predicting the time when the target sperm arrives at the sorting junction.

[0015] A system for detecting and screening abnormal sperm motility in frozen and thawed sperm, used to implement the aforementioned method for detecting and screening abnormal sperm motility in frozen and thawed sperm, the system comprising an optical data acquisition module, a feature extraction module, an AI discrimination and grading module, and a microfluidic execution module; Optical data acquisition module: includes microfluidic sample pump, wide-field phase difference microscope and high-speed CMOS camera, used to acquire continuous dynamic microscopic image sequences and build a large dataset of sperm optical features; Feature extraction module: Deployed on a preset computing architecture, it is used to read a large dataset of sperm optical features, extract and output a set of multidimensional sperm features including single-frame phase difference images and dynamic motion state features; AI discrimination and grading module: Used to train an improved dual-stream spatiotemporal graph convolutional neural network using historical data to obtain a sperm activity correlation model, and receive a multi-dimensional sperm feature set for forward inference, outputting sperm quality grading labels and physical-level absolute pixel coordinates; Microfluidic execution module: Includes a central control computer and a surface acoustic wave microfluidic chip, used to receive tags and coordinates and call S2 features to calculate fluid delay time, and execute physical field switching to complete sperm screening.

[0016] (III) Beneficial Effects The present invention has the following beneficial effects: This invention achieves label-free and non-destructive detection throughout the entire process. By combining wide-field optical imaging with phase difference microscopy, and using Gaussian mixture background modeling and Laplace second-order spatial derivative calculation, it can directly extract the dynamic trajectory of sperm and the intracellular refractive index gradient distribution from a single layer of fluid. It eliminates the dependence of traditional flow cytometry on chemical fluorescent dyes, avoids the toxic damage of markers to sperm DNA and physiological functions, and ensures that the selected high-quality sperm can be directly and safely used in subsequent clinical assisted reproductive procedures such as ICSI.

[0017] This invention improves the sensitivity and noise robustness in identifying cryogenic damage to revived sperm. Addressing the data characteristics of revived sperm, including weak contrast, motility distortion, and a high proportion of dead sperm, this invention constructs an improved dual-stream spatiotemporal graph convolutional neural network. Specifically, a phase difference-enhanced spatial attention mechanism precisely amplifies the high-frequency features of micro-damage to the plasma membrane; an adaptive distortion temporal graph convolutional module dynamically adapts to the nonlinear trajectories of sperm twitching or sudden movement; and a multi-center contrast-focus joint loss function derived from objective data effectively overcomes the problem of class imbalance, achieving high-confidence intelligent discrimination of rare, high-quality sperm against a background of massive amounts of dead sperm impurities.

[0018] This invention achieves highly efficient automated closed-loop control from microscopic feature recognition to macroscopic physical screening. By integrating laminar flow parabolic velocity profiles with the instantaneous linear velocity of sperm, this method breaks through the limitations of traditional estimation based solely on average flow velocity, achieving precise compensation for fluid dynamic delay time. Through industrial control bus and surface acoustic wave microfluidic technology, the system can trigger physical field switches in real time based on the quality tags and high-precision time commands output by AI, achieving precise trajectory deflection of target sperm. This reduces the subjectivity and labor intensity of manual microscopic examination, providing assisted reproductive laboratories with a high-throughput, high-reliability intelligent sorting solution.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to the present invention. Figure 2 This is a flowchart illustrating the process of obtaining sperm classification and real-time pixel coordinates in a method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to the present invention. Figure 3 This is a flowchart illustrating the improvement of the dual-flow graph convolutional neural network model in the present invention's method for detecting and screening abnormal sperm motility in frozen and thawed sperm; Figure 4 This is a schematic diagram of the sperm screening process in a method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to the present invention. Figure 5 This is a schematic diagram of a module of a system for detecting and screening abnormal sperm motility in frozen and thawed sperm according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0023] Please see Figure 1 This invention discloses a method for detecting and screening abnormal sperm motility in frozen and thawed sperm, comprising the following steps: S1. The frozen and thawed sperm sample is injected into a microfluidic chip, and microscopic image sequences and scattering signals are continuously acquired using wide-field optical imaging combined with phase difference microscopy detection technology. S1 includes the following steps: S11. The frozen and thawed sperm sample is pushed into the detection area of ​​a microfluidic chip with a fluid focusing structure at a constant flow rate using a precision injection pump, so that the sperm are distributed in a single-layer fluid plane. In specific implementation, the above embodiment S11 is as follows: a 0.5 ml sperm cryopreservation tube stored in liquid nitrogen is rapidly thawed in a 37°C water bath and dripped into the sample injection cell of the microfluidic chip. The piezoelectric ceramic micropump system is started; in order to ensure that the sperm do not overlap and block each other during imaging, the fluid must be in a strictly laminar flow state and the sperm must be distributed in a monolayer. Furthermore, to ensure the fluid is in a strictly laminar flow state and the sperm are distributed in a monolayer, the following measures were taken: the main sheath fluid flow rate was set at 20 μL / min, and the sample flow rate was set at 2 μL / min; the microchannel was designed with a width of 50 μm and a depth of 20 μm; and the Reynolds number formula Re=(ρ) was applied. v D h ) / μ, where ρ represents the fluid density, v D represents fluid velocity. h The diameter represents the hydraulic diameter, and the viscosity represents the fluid dynamic viscosity. The calculated Reynolds number in the channel is approximately 0.015, which is much smaller than the critical value of 2000 for laminar flow, proving that the fluid is in an extremely stable laminar flow state. The sheath fluid squeezes the sample flow from both sides, focusing it into a central flow band with a width of 5 μm, forcing all sperm to line up in a single row and pass through the focal plane of the microscope in sequence. S12. Turn on the wide field light source and phase difference annular aperture of the microscope, and use a high-speed CMOS camera to continuously acquire a sequence of continuous dynamic microscopic images of sperm populations within the single-layer fluid plane field of view. In specific implementation, the above embodiment S12 is as follows: Traditional bright-field microscopes cannot observe the internal structure of transparent sperm; in this embodiment, a wide-field LED cold light source with a wavelength of 550nm is turned on, and the light illuminates the sample after passing through the annular aperture of the condenser lens; images are continuously acquired using a high-speed CMOS camera; according to the Nyquist sampling theorem, the sampling frequency needs to be at least 60fps to avoid aliasing. In order to accurately capture the unique high-frequency weak tremors of revived sperm, the camera sampling frequency in this embodiment is redundantly set to 200fps (frames / second), the image resolution is 2048×2048 pixels, and the physical spatial resolution is calibrated to 0.22μm / pixel; the wagging frequency of the sperm tail flagella is usually around 10 30Hz, which is 200fps; S13. The continuous dynamic microscopic image sequence acquired in S12 is divided into video stream blocks with fixed time windows through the edge computing gateway. An absolute timestamp and spatial coordinate system origin information are added to each frame image and written in parallel to the Hadoop distributed file system to generate a large dataset of sperm optical features for subsequent steps. In specific implementation, the above embodiment S13 is as follows: The amount of high-definition image data generated per second is approximately 800MB / s, which would cause severe latency if processed by a single machine; the edge computing gateway divides every 100 frames (corresponding to a 0.5-second time window) into a video stream data block; a globally unique absolute timestamp (accurate to milliseconds) and spatial coordinate system origin offset are assigned to each frame of image; using "timestamp-device ID-batch number" as a composite primary key, the structured data stream containing the image matrix is ​​written in parallel to the Hadoop Distributed File System (HDFS), reducing latency; The above embodiments acquire high-resolution, label-free, dynamic images of sperm by using hydrodynamic focusing and high-frame-rate wide-field phase difference imaging, and utilize edge computing to achieve distributed structured storage of data, laying a high-quality data foundation for subsequent feature extraction. S2. Extract a real-time sperm multidimensional feature set from the microscopic image sequence and scattering signal in S1, including the centroid coordinate time series, dynamic motion state features, single-frame phase difference image, and internal microstructure features. S2 includes the following steps: S21. Read the continuous dynamic microscopic image sequence from the sperm optical feature big data set in S1, apply the Gaussian mixture background modeling algorithm to separate the foreground moving target, and use the multi-target tracking algorithm to obtain the centroid coordinate time series of each sperm in the continuous frame. In specific implementation, S21 of the above embodiment is as follows: The Spark computing cluster reads the sperm optical feature dataset output by S1, applies the Gaussian mixture background modeling algorithm (GMM, learning rate set to 0.01, variance threshold set to 16) to separate the moving sperm foreground; uses Kalman filtering combined with the Hungarian matching algorithm to perform multi-target tracking, and outputs the centroid coordinate time series of each sperm (e.g., identifier Sperm ID 001) within the N=100 frame time window. P ={ p 1, p 2,..., p i ,..., p N},in p i =( x i , y i ), No. i The centroid coordinates of sperm in the frame image. x i Is the sperm in the first... i The horizontal coordinate of the centroid in a frame image (usually in pixels or micrometers). y i Is the sperm in the first...i The centroid ordinate in a frame image; S22. Read the centroid coordinate time series in S21, and calculate the dynamic motion state features based on the coordinate difference between adjacent frames and the timestamp; the dynamic motion state features include curvilinear motion speed, linear motion speed and average path speed. In specific implementation, the above embodiment S22 specifically involves: reading the centroid coordinate time series output by S21. P Calculate the velocity of the curvilinear motion: ; where the time interval between adjacent frames Δ t =0.005s; x i+1 Indicates that the sperm was in the first stage i+ The x-coordinate of the centroid in a single frame of an image. y i+1 Indicates that the sperm was in the first stage i+ The centroid's ordinate in a single frame image; calculate the linear velocity: ;in, x N Indicates that the sperm was in the first stage N The horizontal coordinate of the centroid in the frame image, y N Indicates that the sperm was in the first stage N The centroid ordinate in a frame image; S23. Based on the centroid coordinate time series in S21, extract a single-frame phase difference image of each sperm within its centroid coordinate bounding box from the original image sequence; calculate the pixel-level second-order spatial derivative of the single-frame phase difference image to quantify the refractive index gradient distribution inside the cell. S24. Set the refractive index gradient threshold, count the area and distribution of connected regions of pixels within the bounding box that are higher than the refractive index gradient threshold, and extract the combination of the top body light scattering intensity attenuation feature and the film micro-damage feature as the internal microstructure feature. All the features extracted in S21 to S24 together constitute a real-time sperm multidimensional feature set; In specific implementation, the above embodiments S23 and S24 are as follows: ice crystal damage caused by freezing and thawing can cause leakage of sperm acrosome enzymes or micro-ruptures of the plasma membrane; in the phase difference image, the refractive index of the healthy acrosome region is much different from that of the surrounding medium due to the density of the material, presenting a high-contrast "bright halo"; while the refractive index of the damaged acrosome decreases, and the bright halo becomes darker. By extracting a 40×40 pixel bounding box image of the sperm head I ( x , y ); Calculate the second derivative using the Laplace operator: 2 I =( 2I / x 2 )+( 2 I / y 2 );in, I ( x , y ) represents the bounding box image of the sperm head at pixel coordinates ( x , y The grayscale value at () 2 I / x 2 , 2 I / y 2 These represent the images in x direction and y The second partial derivative of the direction; the second derivative can amplify small changes in gray level (refractive index); setting a refractive index gradient threshold. Th g =45 (grayscale gradient value, based on 256 grayscale levels); Traverse the anterior two-thirds region of the sperm head (acromion region), if the condition in this region satisfies | 2 I |> Th g If the area of ​​the connected region of consecutive pixels is less than 30 square pixels, the quantized extracted feature is "attenuation of the light scattering intensity of the top body", which marks the damage to its internal structure. Through this purely mathematical gradient operator, the system gets rid of the dependence on fluorescent staining and realizes the numerical evaluation of membrane integrity without damage. The above embodiments transform unstructured massive image sequences into structured mathematical features such as sperm centroid coordinates, multidimensional motion velocity, and quantified plasma membrane / acrosomal refractive index anomalies through multi-target tracking and second-order spatial derivative operations; thus eliminating the dependence on chemical staining and achieving non-destructive, objective numerical characterization of sperm membrane integrity and internal microstructural damage. S3. Train an improved dual-stream spatiotemporal graph convolutional neural network using a historical sperm feature dataset to obtain a sperm activity correlation model; Input the real-time sperm multidimensional feature set in S2 into the sperm activity association model to obtain the quality grading label of each sperm and the physical absolute pixel coordinates of the current frame; Please see Figure 2 S3 includes the following steps: S31. Based on the data characteristics of weak contrast and trajectory distortion of frozen and revived sperm, the structure and loss function of the traditional dual-stream graph convolutional neural network model are improved to obtain an improved dual-stream spatiotemporal graph convolutional neural network. Please see Figure 3 S31 includes the following steps: S311. Improve the standard convolutional layer in the spatial feature extraction stream of the traditional dual-flow graph convolutional neural network model by adding a phase difference enhancement spatial attention mechanism module to amplify the high-frequency micro-variation features of refractive index. Traditional convolutional neural networks tend to ignore small damage to the plasma membrane when processing low-contrast phase difference images, so a phase difference enhancement spatial attention mechanism is introduced. S312. The fixed receptive field temporal convolutional layer in the temporal feature extraction stream of the traditional dual-flow graph convolutional neural network model is improved and replaced with an adaptive distortion temporal graph convolutional module, which is used to dynamically adjust the expansion rate parameter of the temporal graph convolutional kernel according to the instantaneous acceleration. After revival, due to the damage to energy metabolism, the movement trajectory of sperm is no longer smooth, but fluctuates and even twitches in place. The traditional temporal convolutional receptive field is fixed and cannot extract this distortion feature. Therefore, the fixed receptive field temporal convolutional layer is improved and replaced with an adaptive distortion temporal graph convolutional module. S313. The cross-entropy loss function of the traditional dual-flow graph convolutional neural network model is deleted and replaced, and a multi-center contrast-focus joint loss function is added to resist the noise of massive dead sperm impurities and amplify the feature weights of high-quality sperm. S32. Collect images of frozen and thawed sperm from historical batches and the corresponding clinical activity test results, extract features and perform manual classification and annotation to obtain a historical sperm feature dataset; In specific implementation, the above embodiment S32 specifically involves: collecting microscopic video clips of 78,000 revived sperm accumulated over the past 3 months, and manually reverse-annotating them in conjunction with subsequent clinical ICSI fertilization results; dividing them into a training set (70%), a validation set (15%), and a test set (15%) to obtain a historical sperm feature dataset with clear labels of "excellent", "surviving but abnormal", and "necrospermia / impurities"; S33. Use the historical feature dataset to perform backpropagation training on the improved dual-stream spatiotemporal graph convolutional neural network, adjust the node weight parameters of the improved dual-stream spatiotemporal graph convolutional neural network, and when the loss function value on the validation set meets the preset convergence condition, solidify the network parameters to obtain the sperm activity association model. In specific implementation, the above embodiment S33 is as follows: the resuscitation fluid contains the vast majority of dead sperm and cryoprotectant impurities (>80%), while truly usable "good" sperm for ICSI is extremely rare (<5%). To address this class imbalance, this embodiment constructs a joint loss function; total loss Loss = λ 1* L Focal + λ 2* L Con ;in, L Focal Indicates the focal loss term. L Con Indicates the comparative loss term; During the model warm-up and training phase, the mean absolute value of the focus loss gradient of the first 100 batches is extracted. L Focal Compare the mean absolute value of the loss gradient L Con The ratio of the two losses was calculated. Tests showed that the gradient magnitude generated by the contrastive loss was approximately twice that of the focal loss. To ensure a dynamic balance between the contributions of the two types of losses to weight updates during backpropagation and to prevent one loss term from dominating gradient descent, a weight coefficient was set. λ 1 = 1.0 λ 2 = 0.5; The training of the multi-center contrast-focus joint loss function in S33 includes the following steps: S331. In the feature space, set a learnable cluster center for each quality grading category, calculate the Euclidean distance between the sample feature vector and the center of its category as the intra-class cohesion, calculate the distance with the center of the non-category as the inter-class repulsion, and force the feature vectors of dead sperm and impurity samples to be pushed away from the feature domain of excellent sperm, thus obtaining the contrast loss term. In specific implementation, the above embodiment S331 specifically refers to: the comparison loss term L Con The specific calculation rule is as follows: In the multidimensional feature space, a dynamic cluster center is set for the "absent sperm" class, and the Euclidean distance between the current sample feature vector and the center of the "absent sperm" class is calculated. D bad If the true label of the current sample is "excellent", then the exclusion rule is applied, and its loss term includes max(0, margin). D bad ); Furthermore, principal component analysis (PCA) was used to reduce the dimensionality of the feature vectors in the historical feature dataset of S32. The average initial Euclidean distance between the center of the "good" sperm cluster and the center of the "dead sperm" cluster in the feature space was calculated to be 2.08. To ensure the absolute clarity of the classification boundary, approximately 2.4 times this average distance was taken as the upper limit of the rejection threshold, and the margin was calculated to be 2.08 × 2.4 ≈ 5.0. This forced the network to push the good sperm away from the dead sperm cluster by at least 5.0 units in the feature space, thus enhancing the model's robustness against noise in a background full of impurities. S332. By introducing a dynamic adjustment factor, the focus loss term is calculated, and the excellent sperm samples, which account for a very small proportion, are given a penalty weight that is higher than that of dead sperm samples by a preset multiple. The comparison loss term and the focus loss term are weighted and summed to calculate the total loss, and the network weight parameters are updated through the Adam optimizer. In specific implementation, the above embodiment S332 specifically refers to: the focus loss term L Focal The calculation rules are as follows: L Focal = α t (1 p t ) γ log( p t );in, α t As a category weight adjustment factor, p t To predict the probability that a sample belongs to its true class for the model. γ For focusing parameters; Furthermore, based on the inverse class frequency mapping of historical datasets, the proportion of dead sperm and impurities in the historical S32 data was approximately 88.2%, while the proportion of high-quality sperm was approximately 11.8%; according to the inverse class frequency formula... α i =(1 / Freq i ) / ∑(1 / Freq i ) calculation, original α good1 ≈0.88, original α bad1 ≈0.12; where, Freq i Indicates the first i The frequency of a sample class in the historical dataset (i.e., the proportion of that class in the total number of samples); to prevent gradient explosion in a single iteration, a smoothing decay factor of 0.85 is introduced, and the final calculation is as follows. αgood =0.88×0.85≈0.75, α bad =0.12×0.85≈0.1; Focusing parameters γ Set to 2.0; adjustment factor for the "Excellent" category. α good =0.75, the moderating factor for the "necrospermia" category. α bad =0.1; When the network misclassifies a rare, high-quality sperm, the resulting penalty error gradient will be amplified by 0.75 / 0.1=7.5 times, forcing the model to pay close attention to rare, high-quality samples; S333, Set the maximum number of training iterations; when continuous η When the total loss decrease over 10 training cycles is less than the total loss decrease threshold of 0.001, or when the maximum number of training iterations is reached, the preset convergence condition is met, and the sperm activity correlation model is obtained. In specific implementation, the above embodiment S333 is as follows: the Adam optimizer is used with an initial learning rate of 0.001; the total loss on the validation set is calculated after each epoch; when the total loss decreases by less than 0.001 for 10 consecutive epochs, or when the maximum number of training iterations of 100 is reached, it is determined that the model has reached Nash equilibrium, training is stopped and all node weight parameters are fixed, and the sperm activity association model is clearly output. S34. Input the sperm multidimensional feature set in S2 into the sperm activity association model for forward propagation inference to obtain the quality grading label of each sperm and the physical absolute pixel coordinates of the current frame. S34 includes the following steps: S341, Phase Difference Enhancement Spatial Attention Mechanism Module, reads a single-frame phase difference image from the sperm multidimensional feature set output by S2, and generates a refractive index anomaly gradient map by calculating the second-order phase derivative of adjacent pixels through the Laplacian operator. The refractive index anomaly gradient map is input into the channel perception network to generate a spatial attention mask, and then multiplied element-wise with the original convolutional feature map to amplify the high-frequency micro-variation features at the top edge and the damaged part of the plasma membrane, thus obtaining the microstructure feature vector. In specific implementation, the above embodiment S341 is as follows: Traditional convolutional neural networks tend to overlook minor damage to the plasma membrane when processing low-contrast phase difference images; the calculation rule for the phase difference enhancement spatial attention mechanism defined in this embodiment is: obtain the absolute value matrix of the second derivative of the input image in S23. M =| 2 I | and normalize Mnorm =[ M min( M )] / [max( M ) min( M ]], to obtain the basic mask matrix; then M norm The input channel perceptron (consisting of one 1×1 convolutional layer, one ReLU layer, and another 1×1 convolutional layer) is processed by a sigmoid activation function, and the output is an attention mask. A mask ; Furthermore, the Sigmoid function maps the weight values ​​of high-gradient regions (membrane damage sites with abrupt refractive index changes) to the interval [0.8, 1.0], while mapping the weight values ​​of low-gradient regions (uniform cytoplasm or background fluid) to the interval [0.0, 0.2]; feature map multiplication is then performed. ;in, F en This represents the enhanced microstructure feature matrix. F or This represents the spatial feature matrix output by the original convolutional layer. A mask This represents the calculated spatial attention mask; the above steps enable the neural network to focus its computational resources on assessing the location of structural damage to cells. S342, the adaptive distortion time-map convolution module reads the centroid coordinate time series from the sperm multidimensional feature set in S2, calculates the instantaneous acceleration of sperm between adjacent frames; and dynamically maps the absolute value of the instantaneous acceleration to generate the dilation rate parameter of the time-map convolution kernel to obtain the motion trajectory feature vector. Specifically: if the absolute value of acceleration is lower than the micro-tremor threshold, the expansion rate is kept at 1; if it is between the micro-tremor threshold and the floating threshold, the expansion rate is set to 2; if it is higher than the floating threshold, the expansion rate parameter is set to a discrete integer such as 4. In specific implementation, the above embodiment S342 specifically involves: calculating the first coordinate sequence... t Instantaneous velocity of a frame v t and instantaneous acceleration a t =( v t - v t-1 ) / Δ t Set the microtremor threshold Th low =10μm / s 2 Th high =50μm / s 2 ; Rule 1: If | a t |≤Th low (Sperm twitches in place), set expansion rate d =1; At this point, the convolution kernel closely samples adjacent frames to specifically analyze high-frequency micro-tremor features. Rule 2: If Th low <| a t |≤Th high Set the expansion rate d=2; Rule 3: If | a t |>Th high (Sperm suddenly swims rapidly), set the expansion rate d=4; at this time, the convolution kernel skips sampling across frames, the receptive field expands by 4 times, and it is specifically used to capture macroscopic nonlinear distortion trajectories; through the above explicit mathematical mapping rules, the temporal graph convolutional network can, like a human expert, adaptively zoom in to see twitching and zoom out to see swimming. S343. The microstructure feature vector and the motion trajectory feature vector are concatenated and input into the fully connected classification layer of the sperm activity association model; The fully connected classification layer uses the Softmax activation function for classification mapping, outputting quality grade labels including excellent, viable but abnormal, dead sperm, and impurities, and extracts the center point of the bounding box of sperm in the image space, and simultaneously outputs the physical absolute pixel coordinates of sperm in the current frame; In specific implementation, the above embodiment S343 is as follows: the feature vectors output from the spatial flow and the temporal flow are concatenated and then fed into a fully connected layer, where they are mapped to a probability distribution using the Softmax function; finally, the quality grading label of the sperm is clearly output (0 represents excellent, directly used for ICSI; 1 represents viable but abnormal, discarded; 2 represents dead sperm / impurities), and the center of the sperm bounding box is extracted, and the physical absolute pixel coordinates of the sperm in the current frame are output simultaneously. X pix , Y pix ); The above embodiments improve the dual-stream spatiotemporal graph convolutional neural network to address the unique characteristics of weak contrast and trajectory distortion in frozen and revived sperm. By objectively calculating and setting a multi-center contrast-focus joint loss function, the problem of model collapse caused by the extremely high proportion of dead sperm in the revived samples is solved. By accurately mapping the pure mathematical features extracted by S2 into quality labels and absolute coordinates to guide physical sorting, not only is the detection rate of rare and high-quality sperm improved, but the interpretability of classification decisions is also guaranteed. S4 receives the quality grading label and physical absolute pixel coordinates output by S3, and calls the dynamic motion state features in S2 to calculate the hydrodynamic delay time of the target sperm. Based on the quality grading label and the fluid dynamic delay time, the physical field switch of the microfluidic sorting actuator is triggered in real time to change the fluid motion trajectory of the target sperm, thereby obtaining the screened frozen and revived sperm. S4 includes the following steps: S41. The central control computer receives the sperm quality grading label and its physical absolute pixel coordinates in the current frame from the output of S3 in real time. S42. Based on the pre-calibrated physical space mapping matrix between the optical field of view and the microfluidic chip, convert the physical-level absolute pixel coordinates received in S41 into absolute physical coordinates within the microfluidic chip. In specific implementation, S41 and S42 of the above embodiments are as follows: The central control computer receives the tag (assumed to be "0-excellent") and pixel coordinates output by S3 in real time via gigabit Ethernet; it performs conversion according to the pre-calibrated mapping matrix: the calibration parameter is 1 pixel = 0.22 μm; if the longitudinal pixel coordinates of sperm are received... X pix =1000, then the absolute physical coordinates X phy =1000×0.22=220μm; The physical coordinates of the center of the sorting branch (physical execution point) of the microfluidic chip are fixed as follows: X target =2000μm; then the remaining straight-line physical distance between the target sperm and the sorting fork is... D re =2000 220 = 1780 μm; S43. Call the dynamic motion state features in the sperm multidimensional feature set output by S2, and combine them with the absolute physical coordinates to calculate the hydrodynamic delay time required for the target sperm to move from the current position to the microfluidic sorting branch with the laminar liquid. The calculation process of the hydrodynamic delay time in S43 includes the following steps: S431. Extract the laminar parabolic velocity profile distribution equation of the microfluidic chip, and determine the liquid velocity of the streamline layer in which the target sperm is located based on the Y-axis lateral coordinate of the absolute physical coordinates of the target sperm. S432. The basic arrival time is obtained by dividing the straight-line distance between the current longitudinal coordinate of the target sperm on the X-axis and the center coordinate of the sorting branch by the fluid velocity of the streamline layer. S433: Read the linear motion velocity in the dynamic motion state characteristics of S2, add its projection on the X-axis as a compensation term to the liquid velocity for time compensation correction, and output the fluid dynamics delay time. In specific implementation, the above embodiment S43 is as follows: Traditional sorting systems often directly divide the distance by the average flow velocity, resulting in a large deviation in sorting timing; this embodiment explicitly calls the dynamic motion characteristics in S2 for precise compensation; and extracts the simplified laminar parabolic velocity profile equation from the Navier-Stokes equations. V flow ( Y )= V max *[1 (Y W / 2) 2 / (W / 2) 2 Given a channel width W = 50 μm, the maximum flow at the center... V max =100μm / s; Current x-coordinate of the target sperm Y phy =25μm (located at the center of the channel), calculate the liquid traction velocity in its streamline layer. V flow ( Y =100μm / s; At this point, explicitly call the linear velocity (VSL) of the sperm extracted in S2, and take its projection on the X-axis. V sperm =20μm / s as a compensation term; calculate the actual resultant velocity. V total = V flow ( Y )+ V sperm =120μm / s; Output precise hydrodynamic delay time T delay = D re / V total =1780 / 120=14.83ms; S44. When the target sperm with the received quality grading label of "excellent" reaches the trigger warning line, the central control computer sends a control level to the radio frequency signal generator of the microfluidic sorting actuator within the preset advance response period before the end of the countdown of the hydrodynamic delay time calculated in S43, to excite the sorting physical field and deflect the target sperm to the collection channel. For specific implementation details, please refer to [link / reference]. Figure 4 In the above embodiment S44, specifically: when the "superior" tagged sperm crosses the software-defined trigger warning line, the FPGA-based central controller starts a 14.83ms high-precision hardware countdown timer; the advance response period for the surface acoustic wave generator to establish an effective sound field is set to 1.5ms; when the countdown reaches 14.83ms... At 13.33ms, the controller outputs a 5V TTL high-level trigger signal to the RF generator via the industrial control bus. The RF generator instantly outputs an alternating electrical signal with a voltage amplitude of 40V and a frequency of 30MHz, which is applied to the interdigital transducer (IDT) on the sidewall of the microfluidic chip. The piezoelectric substrate instantly excites traveling surface acoustic waves (SAW). When the superior sperm arrives at the sorting fork at exactly 14.83ms, the acoustic radiation force generated by the surface acoustic wave in the fluid (calculated to be about 15pN) acts on the side of the sperm, deflecting its lateral trajectory by about 15μm and accurately pushing it into the superior sperm collection channel above. If the label output by S3 is "1 (abnormal)" or "2 (absence of sperm)," the controller will not send a trigger signal, the IDT will remain silent, and the abnormal sample will enter the waste liquid channel ahead in a straight line under the influence of fluid inertia and the main injection flow. The above embodiments combine virtual grading labels and coordinates with precisely calculated fluid dynamic delay time to transform them into high-frequency control commands for the physical field. Through precise triggering of surface acoustic waves, non-destructive physical deflection and automatic collection of highly active sperm are achieved, opening up the link from intelligent detection to physical screening and forming a high-throughput, automated closed-loop control method.

[0024] For further details, please refer to Figure 2 A system for detecting and screening abnormal sperm motility in frozen and thawed sperm is provided to implement the aforementioned method for detecting and screening abnormal sperm motility in frozen and thawed sperm. The system includes an optical data acquisition module, a feature extraction module, an AI discrimination and grading module, and a microfluidic execution module. Optical data acquisition module: includes microfluidic sample pump, wide-field phase difference microscope and high-speed CMOS camera, used to acquire continuous dynamic microscopic image sequences and build a large dataset of sperm optical features; Feature extraction module: Deployed on a preset computing architecture, it is used to read a large dataset of sperm optical features, extract and output a set of multidimensional sperm features including single-frame phase difference images and dynamic motion state features; AI discrimination and grading module: Used to train an improved dual-stream spatiotemporal graph convolutional neural network using historical data to obtain a sperm activity correlation model, and receive a multi-dimensional sperm feature set for forward inference, outputting sperm quality grading labels and physical-level absolute pixel coordinates; Microfluidic execution module: Includes a central control computer and a surface acoustic wave microfluidic chip, used to receive tags and coordinates and call S2 features to calculate fluid delay time, and execute physical field switching to complete sperm screening.

[0025] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0026] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for detecting and screening abnormal sperm motility in frozen and thawed sperm, characterized in that, Includes the following steps: S1. The frozen and thawed sperm sample is injected into a microfluidic chip, and microscopic image sequences and scattering signals are continuously acquired using wide-field optical imaging combined with phase difference microscopy detection technology. S2. Extract a set of sperm multidimensional features from the microscopic image sequence and scattering signal in S1, including centroid coordinate time series, dynamic motion state features, single-frame phase difference image and internal microstructure features. S3. Train an improved dual-stream spatiotemporal graph convolutional neural network using a historical sperm feature dataset to obtain a sperm activity correlation model; Input the sperm multidimensional feature set in S2 into the sperm activity association model to obtain the quality grading label of each sperm and the pixel coordinates of the current frame; S4. Receive the quality grading label and pixel coordinates from S3, and call the dynamic motion state features from S2 to calculate the hydrodynamic delay time of the target sperm. Based on the quality grading label and the fluid dynamic delay time, the physical field switch of the microfluidic sorting actuator is triggered in real time to change the fluid motion trajectory of the target sperm, thereby obtaining the screened frozen and revived sperm.

2. The method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to claim 1, characterized in that, S1 includes the following steps: S11. The frozen and thawed sperm sample is pushed into the detection area of ​​a microfluidic chip with a fluid focusing structure at a constant flow rate using a precision injection pump, so that the sperm are distributed in a single-layer fluid plane. S12. Turn on the wide field light source and phase difference annular aperture of the microscope, and use a high-speed CMOS camera to continuously acquire a sequence of continuous dynamic microscopic images of sperm populations within the single-layer fluid plane field of view. S13. The continuous dynamic microscopic image sequence acquired in S12 is divided into video stream blocks with fixed time windows through the edge computing gateway. An absolute timestamp and spatial coordinate system origin information are added to each frame image and written in parallel to the Hadoop distributed file system to generate a large dataset of sperm optical features for subsequent steps.

3. The method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to claim 2, characterized in that, S2 includes the following steps: S21. Read the continuous dynamic microscopic image sequence from the sperm optical feature big data set in S1, apply the Gaussian mixture background modeling algorithm to separate the foreground moving target, and use the multi-target tracking algorithm to obtain the centroid coordinate time series of each sperm in the continuous frame. S22. Read the centroid coordinate time series in S21, and calculate the dynamic motion state features based on the coordinate difference between adjacent frames and the timestamp; the dynamic motion state features include curvilinear motion speed, linear motion speed and average path speed. S23. Based on the centroid coordinate time series in S21, extract a single-frame phase difference image of each sperm within its centroid coordinate bounding box from the original image sequence; calculate the pixel-level second-order spatial derivative of the single-frame phase difference image to quantify the refractive index gradient distribution inside the cell. S24. Set the refractive index gradient threshold, count the area and distribution of connected regions of pixels within the bounding box that are higher than the refractive index gradient threshold, and extract the combination of the top body light scattering intensity attenuation feature and the film micro-damage feature as the internal microstructure feature. All the features extracted in S21 to S24 together constitute a real-time sperm multidimensional feature set.

4. The method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to claim 3, characterized in that, S3 includes the following steps: S31. Based on the data characteristics of weak contrast and trajectory distortion of frozen and revived sperm, the structure and loss function of the traditional dual-stream graph convolutional neural network model are improved to obtain an improved dual-stream spatiotemporal graph convolutional neural network. S32. Collect images of frozen and thawed sperm from historical batches and the corresponding clinical activity test results, extract features and perform manual classification and annotation to obtain a historical sperm feature dataset; S33. Use the historical feature dataset to perform backpropagation training on the improved dual-stream spatiotemporal graph convolutional neural network, adjust the node weight parameters of the improved dual-stream spatiotemporal graph convolutional neural network, and when the loss function value on the validation set meets the preset convergence condition, solidify the network parameters to obtain the sperm activity association model. S34. Input the sperm multidimensional feature set in S2 into the sperm activity association model for forward propagation inference to obtain the quality grading label of each sperm and the pixel coordinates of the current frame.

5. The method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to claim 4, characterized in that, S31 includes the following steps: S311. Improve the standard convolutional layer in the spatial feature extraction stream of the traditional dual-flow graph convolutional neural network model by adding a phase difference-enhanced spatial attention mechanism module. S312. Improve the fixed receptive field temporal convolutional layer in the temporal feature extraction stream of the traditional dual-flow graph convolutional neural network model and replace it with an adaptive distortion temporal graph convolutional module. S313. The cross-entropy loss function of the traditional dual-flow graph convolutional neural network model is deleted and replaced, and a multi-center contrast-focus joint loss function is added.

6. The method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to claim 4, characterized in that, S33 includes the following steps: S331. In the feature space, set a learnable cluster center for each quality grading category, calculate the Euclidean distance between the sample feature vector and the center of its category as the intra-class cohesion, calculate the distance with the center of the non-category as the inter-class repulsion, and force the feature vectors of dead sperm and impurity samples to be pushed away from the feature domain of excellent sperm, thus obtaining the contrast loss term. S332. By introducing a dynamic adjustment factor, the focus loss term is calculated, and a penalty weight higher than that of dead sperm samples is assigned to the excellent sperm samples; the comparison loss term and the focus loss term are weighted and summed to calculate the total loss, and the network weight parameters are updated through the Adam optimizer. S333. Set the maximum number of training iterations; when the total loss decreases for η consecutive training cycles is less than the total loss decrease threshold, or when the maximum number of training iterations is reached, the preset convergence condition is met, and the sperm activity correlation model is obtained.

7. The method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to claim 4, characterized in that, S34 includes the following steps: S341, Phase Difference Enhancement Spatial Attention Mechanism Module, reads a single-frame phase difference image from the sperm multidimensional feature set output by S2, and generates a refractive index anomaly gradient map by calculating the second-order phase derivative of adjacent pixels through the Laplacian operator. The refractive index anomaly gradient map is input into the channel perception network to generate a spatial attention mask, and then multiplied element-wise with the original convolutional feature map to amplify the high-frequency micro-variation features at the top edge and the damaged part of the plasma membrane, thus obtaining the microstructure feature vector. S342, the adaptive distortion time-map convolution module reads the centroid coordinate time series from the sperm multidimensional feature set in S2, calculates the instantaneous acceleration of sperm between adjacent frames; and dynamically maps the absolute value of the instantaneous acceleration to generate the dilation rate parameter of the time-map convolution kernel to obtain the motion trajectory feature vector; specifically: if the absolute value of the acceleration is lower than the micro-tremor threshold, the dilation rate is kept at 1; if it is between the micro-tremor threshold and the swimming threshold, the dilation rate is set to 2; if it is higher than the swimming threshold, the dilation rate parameter is set to a discrete integer such as 4. S343. The microstructure feature vector and the motion trajectory feature vector are concatenated and input into the fully connected classification layer of the sperm activity association model; The fully connected classification layer uses the Softmax activation function for classification mapping, outputting quality grade labels including excellent, viable but abnormal, dead sperm, and impurities, and extracts the center point of the bounding box of sperm in the image space, and simultaneously outputs the pixel coordinates of sperm in the current frame.

8. The method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to claim 4, characterized in that, S4 includes the following steps: S41. The central control computer receives the sperm quality grading label and its pixel coordinates in the current frame from the output of S3 in real time. S42. Based on the pre-calibrated physical space mapping matrix between the optical field of view and the microfluidic chip, convert the pixel coordinates received in S41 into absolute physical coordinates within the microfluidic chip. S43. Call the dynamic motion state features in the sperm multidimensional feature set output by S2, and combine them with the absolute physical coordinates to calculate the hydrodynamic delay time required for the target sperm to move from the current position to the microfluidic sorting branch with the laminar liquid. S44. When the target sperm with the received quality grading label of "excellent" reaches the trigger warning line, the central control computer sends a control level to the radio frequency signal generator of the microfluidic sorting actuator within the preset advance response period before the end of the countdown of the hydrodynamic delay time calculated in S43, to excite the sorting physical field and deflect the target sperm to the collection channel.

9. A method for detecting and screening abnormal sperm motility in frozen and thawed sperm according to claim 8, characterized in that, The calculation process of the hydrodynamic delay time in S43 includes the following steps: S431. Extract the laminar parabolic velocity profile distribution equation of the microfluidic chip, and determine the liquid velocity of the streamline layer in which the target sperm is located based on the Y-axis lateral coordinate of the absolute physical coordinates of the target sperm. S432. The basic arrival time is obtained by dividing the straight-line distance between the current longitudinal coordinate of the target sperm on the X-axis and the center coordinate of the sorting branch by the fluid velocity of the streamline layer. S433: Read the linear motion velocity in the dynamic motion state characteristics of S2, add its projection on the X-axis as a compensation term to the liquid velocity for time compensation correction, and output the fluid dynamics delay time.

10. A system for detecting and screening abnormal sperm motility in frozen and thawed sperm, characterized in that, The system, which implements the method for detecting and screening abnormal sperm motility in frozen and revived sperm as described in any one of claims 1-9, includes an optical data acquisition module, a feature extraction module, an AI discrimination and grading module, and a microfluidic execution module.

Images