A real-time monitoring method and system for surface acoustic wave microfluidic particle sorting

Abstract

This invention provides a real-time monitoring method and system for surface acoustic wave (SAW) microfluidic particle sorting. Addressing the problems of small particle size, blurred imaging, difficulty in stable identification, and lack of real-time monitoring methods in microfluidic processes, it proposes a deep learning image recognition network based on a normalized channel-spatial attention mechanism and a particle counting method based on lifecycle management. The image recognition network, by introducing a normalized channel-spatial attention mechanism into a convolutional neural network, effectively enhances the feature representation ability of small-sized and defocused particles, improving the accuracy of particle detection. The particle counting method, through lifecycle management of particle trajectories, achieves suppression of transient noise targets and reliable counting of stable trajectories, thereby enabling accurate identification of sorting events. This invention can realize real-time monitoring of the SAW microfluidic particle sorting process, and its applications include, but are not limited to, microfluidic particle sorting, rare cell enrichment, and single-cell omics in precision medicine.

Description

Technical Field

[0001] This invention relates to the fields of microfluidic particle sorting and computer vision technology, specifically to a real-time monitoring method and system for surface acoustic wave microfluidic particle sorting. Background Technology

[0002] Surface acoustic wave (SAW) microfluidic particle sorting systems are microfluidic systems that integrate SAW generators onto microfluidic chips to manipulate particles in microfluidics, thereby sorting out target particles. SAW particle manipulation methods offer advantages such as being non-contact, gentle, and pollution-free, making them suitable for sorting highly active, fragile cells. However, the accurate sorting of SAW sorting chips relies on the adjustment of acoustic modulation parameters and signal trigger delays. Traditional parameter control is limited by the purity of target particles in the samples before and after sorting, lacking monitoring of sorting events during the process, thus failing to achieve real-time parameter adjustment and sorting status early warning. With the development of imaging and computer vision technologies, the application of image processing methods to microfluidic sorting systems has gained increasing attention, providing a potential technical approach for monitoring the SAW sorting process and real-time parameter control. However, technical challenges remain in accurate identification and real-time counting at the single-particle scale. Summary of the Invention

[0003] Traditional surface acoustic wave (SAW) particle sorting systems assess sorting efficiency by calculating the recovery rate of sorted samples, thus guiding parameter adjustment. However, they lack information on individual sorting events during the process, leading to a lack of timeliness in parameter adjustment. While computer vision methods can acquire image data of the sorting process, the small particle size and indistinct morphological differences in the single-particle sorting process of SAW make feature extraction difficult, resulting in low target detection accuracy and high detection noise, which in turn affects counting accuracy. Furthermore, the high particle velocity in the flow channel makes single-frame particle counting methods based on target recognition within the Region of Interest (ROI) prone to missing particles. Setting the ROI too large can lead to repeated particle counting. Therefore, counting methods relying solely on single-frame target recognition have technical limitations.

[0004] To address the aforementioned technical problems, this invention provides a real-time monitoring method and system for surface acoustic wave microfluidic particle sorting, comprising the following contents: S1. Deep Learning Image Recognition Network Enhanced by Normalized Channel-Spatial Attention Mechanism: In surface acoustic wave (SAW) microfluidic particle sorting systems, particles are typically small in size, have weak texture, and exhibit minimal morphological differences. Furthermore, they are prone to defocusing, low light, and noise interference during microscopic imaging, leading to insufficient feature extraction capabilities of traditional convolutional networks for small targets and resulting in missed detections and false detections. In addition, existing channel attention mechanisms (such as SE or CBAM) typically generate channel weights through additional fully connected or convolutional layers, which can introduce unnecessary parameter noise in small target recognition scenarios, causing unstable attention distribution.

[0005] To address the aforementioned difficulty in small particle recognition, this invention proposes a target recognition network based on the Normalized Channel-Spatial Attention (NBAM) mechanism. This mechanism eliminates the need for additional parameters; instead, it directly utilizes the γ parameter of the batch normalization (BN) layer in the convolutional network to construct stable channel attention weights, and then combines this with spatial attention to enhance the response in small particle regions, thereby improving the recognition ability of weakly textured small particles. Specifically, it includes the following steps: S1-1, Normalization-based Channel Attention Module: Addressing the issues of weak feature signals and indistinct channel differences in small particles, such as... Figure 2 As shown, this invention utilizes the existing γ parameter in the batch normalization (BN) layer to construct channel weights. The γ parameter naturally reflects the network's sensitivity to different channels. A stable channel weight map can be generated by normalizing the γ value without requiring additional parameters, thus avoiding the noise introduced by fully connected layers in traditional channel attention (such as SE). The normalized channel weights ω are obtained through the following calculation method, where γ characterizes the importance of each channel to the stretching of the feature distribution:

[0006] The meaning of each letter in the formula is as follows: B in B represents the input tensor of a certain layer in the network. out To output the result after processing by the BN layer, μ B and σ B The input training batch data represents the mean and variance, with γ and β being scaling and shift factors, respectively. This means that the feature maps of a batch of training data are used as input, and the mean and variance are used to normalize the batch data to obtain a 0-1 distribution. ε is a small positive number to avoid division by zero. After scaling with γ and baseline shifting with β, the distribution of feature map values ​​is made to fall within the region of maximum gradient of the loss function. γ and β are parameters learned by the network during training, typically with initial values ​​of 1 and 0. Therefore, the scaling factor γ for each channel... iThis naturally reflects the network's sensitivity to different channels. By normalizing the γ value, a stable channel weight map can be generated without additional parameters, thus avoiding the noise introduced by the fully connected layer in traditional channel attention (such as SE). After normalizing γi, a scaling factor ωi for each channel can be obtained.

[0007] This channel attention mechanism can significantly enhance the channel response associated with particle regions, making the weak texture features of small particles more prominent and improving the recognition robustness under complex lighting conditions.

[0008] S1-2. Enhanced Particle Region Saliency Based on Spatial Attention: Due to the complex background of microfluidic imaging, particle positions are easily affected by noise, making it difficult to accurately locate particle regions using only channel attention. This invention compensates for the insufficient spatial localization capability of channel attention through a spatial attention module. For example... Figure 2 As shown, the spatial attention module performs average pooling and max pooling operations on the input features to aggregate information from all channels into two spatial context descriptors. These descriptors are then concatenated into the convolutional layer to model spatial relationships, and finally activated by a sigmoid function to obtain the spatial attention map. This mechanism can highlight salient regions of particles in the spatial dimension, enhancing the detectability of particle regions and improving localization accuracy in situations with complex backgrounds, uneven lighting, and local blurring.

[0009] S1-3, Normalized Channel – Fusion of Spatial Attention: such as Figure 2 As shown, this invention applies channel attention maps and spatial attention maps sequentially to the input feature map, achieving fusion through element-wise multiplication. This allows the model to simultaneously enhance key features in the channel dimension and improve regional saliency in the spatial dimension. This fusion mechanism significantly improves the feature responses of small particles, out-of-focus particles, and weakly textured particles, providing high-quality feature input for subsequent detection heads and enhancing the overall stability and accuracy of recognition.

[0010] Image recognition network construction and attention mechanism embedding: such as Figure 3 As shown, this invention inserts the aforementioned normalized channel-spatial attention mechanism into predetermined positions in the backbone and / or neck layers of the YOLO architecture. By enhancing the particle saliency of mid-to-low-level features, the detection network achieves stronger small target recognition capabilities while maintaining a lightweight design. Experimental results demonstrate that this attention mechanism significantly improves the recall rate of small particle detection and reduces the false detection rate, providing more reliable input data for subsequent lifecycle management counting algorithms, thereby enhancing the overall performance of the entire real-time monitoring system.

[0011] S2. Particle counting algorithm based on lifecycle management: In response to the characteristics of high-speed particle flow, large inter-frame displacement, and significant velocity fluctuations in surface acoustic wave microfluidic sorting systems, traditional counting methods based on single-frame detection cannot simultaneously ensure the accurate capture of high-velocity particles and the effective differentiation of low-velocity particles. When the detection area is set too large, low-velocity particles will repeatedly appear in multiple consecutive frames and be counted repeatedly. When the detection area is set too small, high-speed particles may only pass through the ROI in a very short time and thus cannot be captured, resulting in missed counts. In addition, single-frame detection results are easily affected by imaging noise, producing false particles and introducing false counts.

[0012] To address the aforementioned technical problems, this invention proposes a multi-objective counting method based on a "prediction-correlation-lifecycle management" strategy, achieving stable and robust real-time counting of high-speed particle streams. For example... Figure 4 As shown, this method takes the detection results of consecutive image frames as input, constructs a stable target trajectory through motion prediction, cross-frame matching, and lifecycle constraints, and counts particles based on the trajectory lifecycle. Specifically, it includes the following steps: Target detection and initial set construction: The deep learning image recognition network described in claim 1 is used to perform particle detection on the current frame image acquired by the high-speed camera. This network enhances the feature representation ability of small-sized particles through a normalized channel-spatial attention mechanism, improving the detection accuracy and recall of the candidate set, and providing more stable input data for the subsequent counting module.

[0013] S2-1. Motion Prediction Based on Kalman Filter: For particles whose tracking trajectories have been established in the previous frame, this invention uses Kalman filtering to predict their position and state. Since particle displacement is large under high-speed flow, this step can utilize the particle's motion trend to constrain the matching range, thereby reducing the risk of mismatch caused by matching solely based on Euclidean distance and improving the robustness of cross-frame association.

[0014] S2-2, Cross-frame association matching mechanism: A matching cost matrix is ​​constructed based on the prediction results and the current detection results. The global optimal match is obtained through the Hungarian algorithm to achieve trajectory tracking of multiple targets in the detection area. This invention sets a matching distance threshold to prevent abnormal offsets caused by illumination fluctuations, temporary occlusion, or detection noise from being incorrectly associated, thereby ensuring the stability and continuity of trajectory IDs.

[0015] S2-3, Lifecycle Management and Counting Rules: For successfully matched trajectories, update their position and lifecycle count; for unmatched detection results, create new trackers.

[0016] This invention sets a lifecycle threshold, ensuring that particle targets are counted only after they have remained stably present for several consecutive frames. This mechanism simultaneously solves the following problems: (1) Missed counts due to the brief appearance of high-speed particles; (2) Repeated counting caused by trajectory interruption due to temporary occlusion of particle trajectory; (3) False counts caused by noise in single-frame detection; When the lifecycle of a trajectory reaches a threshold, the counter for its corresponding category is incremented once, and the trajectory is automatically cleared.

[0017] S2-4. Real-time cyclic execution: In the real-time monitoring system described in claim 1, the above steps are continuously and cyclically executed along with the high-speed acquired image frames to realize real-time particle counting and sorting event monitoring of the inlet, target outlet and waste liquid outlet areas.

[0018] S3. Establish a real-time monitoring system based on image recognition and particle counting: The sorting efficiency of surface acoustic wave (SAW) microfluidic particle sorting systems is affected by factors such as fluid velocity changes, acoustic modulation parameters, and trigger delay. In traditional systems, offline evaluation relies solely on changes in sample concentration before and after sorting, which cannot obtain real-time information for a single sorting event. Therefore, it is impossible to achieve adaptive parameter adjustment or timely warning when abnormal sorting occurs. Furthermore, acoustic excitation causes velocity disturbances and spatial drift of particles in the inlet and outlet areas, making it extremely difficult to accurately obtain the real-time sorting status through manual observation or monitoring based on a single scanning point.

[0019] To address the aforementioned problems, this invention proposes a real-time monitoring system based on image recognition and particle counting algorithms, such as... Figure 1 As shown, by constructing three typical monitoring areas—the inlet, the target outlet, and the waste liquid outlet—the entire sorting process can be visualized and monitored, and events can be identified, thereby enabling real-time status assessment of the sorting system. Specifically, this includes the following steps: S3-1. Constructing Typical Monitoring Areas and Continuously Acquiring Images: Based on the microscopic imaging range of the surface acoustic wave microfluidic sorting chip, this invention constructs independent monitoring areas in the inlet area, target outlet area, and waste liquid outlet area. A high-speed camera continuously captures video streams from these three areas as real-time monitoring data sources. This step ensures that monitoring covers the entire process from particle entry into the flow channel to final entry into the sorting outlet, providing complete trajectory data for each sorting event.

[0020] S3-2. Particle Detection and Classification Based on Image Recognition Network: For each continuously acquired image frame, the deep learning image recognition network described in S1 of this invention is used for particle detection and classification, identifying four types of particles: target particles, non-target particles, sorted target particles, and sorted non-target particles. The normalized channel-spatial attention mechanism improves the small particle recognition performance under complex lighting conditions, enabling the real-time monitoring system to obtain particle detection results with high recall and low false detection rate, ensuring the reliability of monitoring data from the source.

[0021] S3-3. Particle Counting and Trajectory Stabilization Based on Lifecycle Management: For the detection results within each monitoring area, the lifecycle management-based counting algorithm described in S2 of this invention is input to perform real-time tracking, association, and counting. Kalman spectroscopy compensates for the displacement instability of particles under high-speed motion, the Hungarian algorithm ensures the consistency of trajectory IDs, and lifecycle management avoids erroneous counting caused by temporary occlusion and detection noise. This enables the system to achieve stable, multi-target, and lossless real-time counting under high-speed video streams.

[0022] S3-4. Real-time sorting event identification and system status update: The system automatically analyzes real-time sorting events based on particle count changes in the three monitoring areas. (1) If the number of target particles in the entrance area increases but the number of target particles at the exit does not increase accordingly, it is determined to be a missed sorting event; (2) If target particles appear in the waste liquid area, it is determined to be a missorting event; (3) If the counts in the three regions are inconsistent or abnormal sequence changes occur, an early warning or guidance parameter adjustment can be triggered.

[0023] This invention generates sorting status data streams in real time, enabling status assessment, fault detection, and anomaly tracking of the sorting system.

[0024] Compared with existing technologies, the technical solution of this application has the following beneficial effects: By counting the particles flowing into the sample and through the sorting outlet, automated monitoring of the sorting process is achieved, reducing the workload of experimental personnel. A deep learning image recognition network enhanced by a normalized channel-spatial attention mechanism improves the ability to extract small-scale features, achieving high-precision classification and recognition of small-sized particles under complex flow channel backgrounds and lighting conditions. A lifecycle management-based counting method improves upon the problem of easy omission or duplication in counting high-speed moving particles by traditional single-frame target recognition counting methods, achieving high-precision and stable counting with trajectory consistency. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of a preferred embodiment of a real-time monitoring method and system for microfluidic particle sorting using surface acoustic waves provided by the present invention. Figure 2 is a structural diagram of the normalized channel-spatial attention mechanism; Figure 3 is a schematic diagram of a deep learning image recognition network enhanced by a normalized channel-spatial attention mechanism; Figure 4 is a flowchart of the particle counting algorithm based on lifecycle management. Detailed Implementation

[0026] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 4 .

[0027] A real-time detection method for microfluidic particle sorting using surface acoustic waves includes the following steps: S1. Image Recognition: Acquire continuous image frames of the microfluidic chip sorting process, and use a deep learning image recognition network containing a normalized channel-spatial attention module to detect and classify particles in the image frames.

[0028] This step specifically includes data collection and annotation, image recognition network construction, detection, and classification.

[0029] Data Acquisition and Labeling: A high-speed microscopic imaging system (650 fps, 1024×1024 resolution) was used to acquire continuous images of the surface acoustic wave (SAW) microfluidic chip during the sorting process. A total of 1874 images were acquired, forming the dataset. Based on particle size and focal plane shift, the particles were divided into four categories: target particles (TP), non-target particles (NP), sorted target particles (ST), and sorted non-target particles (SN). Figure 4 In the target particle recognition section of this dataset, smaller particles on the focal plane are classified as target particles, while larger particles offset from the focal plane are classified as non-target particles. Smaller particles offset from the focal plane are considered sorted target particles, and larger particles offset from the focal plane are considered sorted non-target particles. The dataset is divided into training and validation sets in an 8:2 ratio, with 1499 images used for training and 375 for validation. The LabelImg 1.8.1 tool was used to annotate all particles with rectangular bounding boxes, generating YOLO format annotation files.

[0030] Specific implementation of the image recognition network: In this embodiment, YOLOv11n is used as the base network structure, such as... Figure 2As shown, a normalized channel-spatial attention module (NBAM) is inserted before and after Layer 8 of the Backbone and after Layer 16 of the Neck. This module includes: a channel attention part that calls the γ parameter of the BatchNorm layer in the network to form a channel weight vector; and uses the normalization formula ω... i = γ i / Σγ i The weights are obtained by normalizing γ; the input feature maps are recalibrated through channel-wise multiplication. For spatial attention, average pooling and max pooling are performed on the input features along the channel dimension to obtain two H×W feature maps; the pooling results are concatenated along the channel direction and input into a 7×7 convolutional layer; a sigmoid activation is then applied to obtain the spatial attention map. Finally, element-wise multiplication is used to multiply the features sequentially with the channel attention map and the spatial attention map to obtain the enhanced features.

[0031] The network architecture described above was trained using the dataset created in this study. Model development was performed on a workstation equipped with an NVIDIA RTX3090 GPU using Python 3.8 and CUDA 11.3. Training was conducted using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and a total of 100 epochs. Performance evaluation was based on standard metrics, including accuracy, recall, mean average precision (mAP), inference time, and preprocessing time. The trained model achieved an mAP50 of 0.952, representing an improvement of approximately 6.7% over the baseline model without an attention module.

[0032] Detection and classification: The trained image recognition network is used to detect and classify particles in the image frames during the sorting process of the microfluidic chip to obtain the particle detection set Dk.

[0033] S2. Particle Tracking and Counting: Based on the detection results of step S1, a tracker is established for the particles and a counting process for lifecycle management is executed.

[0034] This step specifically includes: S2-1. Use Kalman filtering to predict the position of the tracker that already exists in the previous frame.

[0035] Specific particle motion prediction based on Kalman filtering: For the set Tk-1 of all particle trajectories established in the previous frame, this embodiment configures a state dimension of [ , , , A Kalman filter is used, where position and velocity constitute a complete motion state. The position of the particle in the current frame is predicted by the state prediction model, resulting in a prediction set Pk.

[0036] The state prediction model is based on the assumption of quasi-uniform motion in a two-dimensional plane. Within the microfluidic channel, particles are primarily driven by fluid drag forces over short timescales, and their trajectories can be approximated as uniform linear motion between adjacent frames. Therefore, position and velocity are used together to describe the particle's motion state. Under this assumption, the particle's motion in the x and y directions can be expressed as:

[0037] The formula for predicting the particle motion state between two adjacent frames can be expressed as follows:

[0038] in, , Let be the position coordinates of the particle in the k-th frame, and Δt be the time interval between two adjacent frames; define Let be the particle's state vector, representing the particle's position and velocity in the k-th frame. , This represents the velocity components of the particle in the x and y directions.

[0039] Uncertainty propagation is performed using the covariance matrix of the previous frame.

[0040] In the Kalman filter framework, in addition to estimating the state, it is also necessary to estimate the uncertainty of that state, which is determined by the covariance matrix. The covariance matrix reflects the magnitude of the current state estimation error and the correlation between various state variables. Uncertainty propagation is represented in the Kalman filter framework as the propagation form of the covariance matrix between two adjacent frames, where the covariance matrix of the previous frame is... The state transition matrix will be used to propagate to the current time step, specifically in the following form:

[0041] in Let Q be the state covariance matrix during the prediction phase, and let Q be the process noise covariance matrix. The process noise covariance matrix is ​​used to model the uncertainty of the motion model. The propagation of the covariance matrix does not directly change the numerical result of the predicted state vector, but rather characterizes the degree of uncertainty in the prediction result. This uncertainty will indirectly affect the state estimation result through the calculation of the Kalman gain during subsequent observation updates. When the prediction covariance is large, the filter will rely more on observation information to correct the state; conversely, when the prediction covariance is small, the filter tends to keep the model prediction result unchanged. Therefore, covariance propagation plays a crucial role in weight adjustment during the prediction and observation fusion process, and also provides a reliable uncertainty measure for data association and trajectory management.

[0042] The prediction steps described above yield the predicted position set Pk for the current frame. Since high-speed particles exhibit significant displacement between adjacent frames, relying solely on the positional similarity between detection boxes can easily lead to mismatches. The prediction steps narrow down the true matching range, effectively reducing association errors.

[0043] S2-2. Construct a matching cost matrix by combining the prediction results with the detection results of the current frame, and perform global optimal matching using the Hungarian algorithm.

[0044] Cross-frame association matching mechanism: In this embodiment, a Euclidean distance cost matrix is ​​constructed based on the prediction set Pk and the current detection set Dk.

[0045] Where pi is the position vector of the i-th predicted particle in the image coordinate system, dj is the position vector of the j-th detected particle in the image coordinate system, C represents the Euclidean distance cost matrix between the predicted particle and the detected particle, and C(i,j) are the elements of the cost matrix.

[0046] The Hungarian algorithm is used to find the globally optimal match with the minimum substitution cost for C(i,j). To further improve the consistency of trajectory IDs, for matches with a distance exceeding a preset matching threshold θ... d The candidate matches are assigned invalid values ​​so that they are automatically filtered out during the solution process, thereby avoiding erroneous associations caused by abnormal displacement, transient occlusion or detection noise.

[0047] S2-3. Update the tracker state based on the matching results; generate a new tracker for unmatched detection results.

[0048] Lifecycle management and counting rules: For successfully matched trajectories, this embodiment updates their position and lifecycle; creates new trackers for unmatched detection results; and accumulates unmatched counts for consecutive unmatched trajectories.

[0049] S2-4. Record the number of consecutive survival frames for each tracker. When the preset lifespan threshold is reached, perform a count and remove the tracker.

[0050] The trajectory lifecycle is represented by the number of frames L, the cumulative unmatch count of the trajectory is represented by M, τ_live is the lifecycle threshold, and τ_miss is the cumulative unmatch count threshold. When L≥τ_live, it is considered that the particle exists stably in multiple consecutive frames. In this embodiment, the counter of its corresponding category is incremented by one, and it is marked as a counted trajectory to avoid duplicate counting.

[0051] Furthermore, if the number of unmatched trajectories satisfies M ≥ τ_miss, the trajectory is deleted to avoid low-speed particles lingering in the monitoring area for extended periods, thus preventing duplicate counting.

[0052] S3. Real-time monitoring system: Steps S1 and S2 are executed in the inlet area, target outlet area and waste liquid outlet area of ​​the microfluidic chip, respectively. The counting algorithm based on life cycle management is used to continuously count the particles in each area, and real-time monitoring data of the sorting process is generated based on the counting results.

[0053] In the monitoring system of this embodiment, the inlet, target outlet and waste liquid outlet can be independently detected, correlated and counted, thereby realizing real-time event identification and sorting status judgment of the surface acoustic wave microfluidic sorting process.

[0054] Monitoring area division and application process: such as Figure 4 As shown, three monitoring regions are manually calibrated on the microscopic image of the surface acoustic wave microfluidic chip: the inlet region (ROI_in), the target outlet region (ROI_tar), and the waste liquid outlet region (ROI_waste). The real-time processing flow is as follows: each frame of image is input into the image recognition network implemented by S1, and the counting algorithm implemented by S2 is applied to filter the detection results of the corresponding region according to ROI. Tracking and counting are performed independently for each ROI, and the time series of various particles in the three regions are continuously updated. By comparing and analyzing the counting relationship between ROI_in and ROI_tar and ROI_waste in real time, missed sorting and missorting events are identified. For example,

[0055] This indicates that 75 target particles were detected in the target outlet area, resulting in 5 missed sorting events, while 16 non-target particles were detected in the waste liquid outlet area, resulting in 4 incorrect sorting events. Through the above-described monitoring area division and application process, real-time monitoring of the surface acoustic wave microfluidic particle sorting process was achieved, enabling visualization, quantification, and traceability, significantly improving the intelligence level and reliability of the sorting system.

[0056] Experimental results test: Following the specific implementation process of S1-S3, a real-time monitoring method for surface acoustic wave microfluidic single-particle sorting scenarios was obtained, which can evaluate the entire sorting process of the sorting system with high accuracy and high robustness.

[0057] To further illustrate the experimental effects of the present invention, in this specific embodiment, the method of the present invention was experimentally tested. The particle recognition performance evaluation was based on standard indicators, including accuracy, recall, average precision (mAP), and inference time. The real-time performance evaluation indicators for the particle counting task included the processing time of the four processes: preprocessing, inference, postprocessing, and counting, as well as the counting accuracy. The evaluation results are shown in the table below (the following experimental data are only the results of one embodiment of the present invention, used to illustrate that the present invention can achieve the above technical effects, and do not constitute a limitation on the scope of protection of the present invention): Table 1. Performance evaluation of the target recognition model of the present invention and several existing models in particle recognition tasks.

[0058] Table 2. Performance evaluation of the attention mechanism of this invention and several existing attention mechanisms in particle counting tasks (unit: ms)

[0059] The experimental results show that the particle recognition and counting performance of the present invention is superior to that of existing models and existing attention mechanisms, verifying the feasibility and robustness of the present invention in monitoring sorting events in a surface acoustic wave microfluidic particle sorting system.

[0060] Example 2: A real-time monitoring system for surface acoustic wave microfluidic particle sorting, characterized in that it comprises: High-speed imaging device for acquiring continuous image frames of microfluidic chips; The image recognition module is used to detect and classify particles in image frames; The counting module is used to establish trackers for particles and perform the counting process for lifecycle management; The data generation module is used to generate real-time monitoring data for the sorting process based on the counting results of the inlet area, target outlet area, and waste liquid outlet area.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method and system for real-time monitoring of microfluidic particle composition using surface acoustic waves, characterized in that, Includes the following steps: S1. Image recognition: Acquire continuous image frames of the microfluidic chip sorting process, and use a deep learning image recognition network containing a normalized channel-spatial attention module to detect and classify particles in the image frames. S2. Particle Tracking and Counting: Based on the detection results of step S1, a tracker is established for the particles, and a lifecycle management counting process is executed, including: S2-1. Use Kalman filtering to predict the position of the tracker that already exists in the previous frame; S2-2. Construct a matching cost matrix by combining the prediction results with the detection results of the current frame, and perform global optimal matching using the Hungarian algorithm; S2-3. Update the tracker state based on the matching results; generate a new tracker for unmatched detection results. S2-4. Record the number of consecutive survival frames for each tracker. When the preset lifespan threshold is reached, perform a count and remove the tracker. S3. Real-time monitoring system: Steps S1 and S2 are executed in the inlet area, target outlet area and waste liquid outlet area of ​​the microfluidic chip, respectively, to continuously count the particles in each area and generate real-time monitoring data of the sorting process based on the counting results.

2. The method according to claim 1, wherein the normalized channel-spatial attention module comprises: S1-1, Channel Attention: Channel weights are generated using the γ parameter of the batch normalization layer, and the input features are weighted according to the channel direction. S1-2, Spatial Attention: Perform average pooling and max pooling operations on the input feature map along the channel dimension, and generate a spatial attention map by convolution and sigmoid activation of the pooling results; S1-3, Fusion Weighting: Perform element-wise multiplication of the channel attention map and spatial attention map with the input features respectively to obtain weighted features.

3. The method according to claim 1, wherein the lifecycle management process includes: For each tracker, record the number of consecutive frames it appears in. When its survival frame count reaches a preset lifespan threshold, count the particles corresponding to that tracker and delete the tracker.

4. The method according to claim 1, wherein the Hungarian matching uses a distance cost matrix and sets a maximum matching distance threshold to filter out non-true matches caused by abnormal displacement.

5. The method according to claim 1, wherein the inlet region, the target outlet region, and the waste liquid outlet region independently perform particle detection, tracking, and counting, and form particle time series data for each region.

6. A real-time monitoring system for surface acoustic wave microfluidic particle sorting, characterized in that, include: High-speed imaging device for acquiring continuous image frames of microfluidic chips; The image recognition module is used to detect and classify particles in image frames; The counting module is used to establish trackers for particles and perform counting processes for lifecycle management. The data generation module is used to generate real-time monitoring data for the sorting process based on the counting results of the inlet area, target outlet area, and waste liquid outlet area.

Images