Blood cell detection system and detection method

Abstract

The invention discloses a blood cell detection system which comprises a data input and processing layer, a training data preparation layer, a control and optimization layer and a neural network model based on a YOLOv5-ALT architecture. Through organic fusion and synergistic effect of three core technologies of entropy screening, consistency regularization and Mixup mixing, the detection accuracy of the blood cells is improved; a set of complete efficient semi-supervised learning framework is formed, and high-precision and high-robustness blood cell detection under an extremely low labeling proportion is finally realized.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and medical image processing, specifically to a blood cell detection system and detection method. Background Technology

[0002] Blood cell testing is a fundamental and crucial task in medical diagnosis, and its accuracy directly impacts the diagnosis of various diseases such as anemia, infection, and leukemia. Traditional blood cell testing relies on manual observation and counting by professionals using microscopes. This method is inefficient, highly subjective, and prone to errors due to fatigue. With the development of computer vision and deep learning technologies, deep learning-based object detection algorithms (such as Faster R-CNN, SSD, and YOLO series) have been widely applied to automated blood cell testing. While these fully supervised learning methods have achieved high accuracy rates, their performance heavily depends on a large amount of high-quality manually labeled data. However, in the field of medical imaging, obtaining such labeled data presents significant challenges: (1) Poor quality of pseudo-labels leads to error accumulation and propagation: Existing methods often simply use the confidence score of the model's prediction as the criterion for filtering pseudo-labels. However, the initial model is trained with limited labeled data, and its predictions may be inaccurate, resulting in a large number of errors or low-quality pseudo-labels. Adding these noisy labels to the training set will cause the errors to be continuously amplified in subsequent training iterations, ultimately damaging the model's performance; (2) Insufficient robustness and generalization ability of the model: Many existing semi-supervised methods do not fully consider the consistency of the model with input data (such as image transformation and perturbation). This means that for different enhanced versions of the same image, the model may give very different predictions, indicating that the model is not stable and robust enough, and its ability to generalize to unseen data is also limited; (3) Low utilization efficiency of unlabeled data: Traditional methods usually use these data directly after generating pseudo-labels and performing consistency training.

[0003] Therefore, this invention aims to address the problem of providing a blood cell detection system and method that organically integrates and synergistically utilizes three core technologies—entropy screening, consistency regularization, and Mixup—to form a complete and efficient semi-supervised learning framework, ultimately achieving high-precision and robust blood cell detection with extremely low annotation ratios. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this invention is to overcome the issues of poor pseudo-label quality leading to error accumulation and propagation, insufficient model robustness and generalization ability, and low utilization efficiency of unlabeled data in existing technologies. This invention provides a blood cell detection system and method that organically integrates and synergizes three core technologies—entropy filtering, consistency regularization, and Mixup—to form a complete and efficient semi-supervised learning framework. Ultimately, this achieves high-precision and highly robust blood cell detection with extremely low labeling ratios.

[0005] To achieve the above objectives, the present invention provides a blood cell detection system, the system comprising: a data input and processing layer, a training data preparation layer, a control and optimization layer, and a neural network model based on the YOLOv5-ALT architecture; wherein, The data input and processing layer includes a storage module for classifying, storing, and managing labeled and unlabeled medical image data, and a preprocessing module for processing unlabeled data to obtain a high-quality pseudo-labeled dataset. The training data preparation layer includes: a pseudo-label management module responsible for storing, versioning, dynamically updating and monitoring the quality of pseudo-labels; and a data mixing module that uses adaptive Mixup technology to mix high-quality pseudo-labels and labeled or unlabeled data in the image and label space to obtain diverse training samples. The core layer of model optimization includes: a loss calculation module for calculating the total loss of the neural network model to integrate supervision signals from different data sources and different learning strategies into a unified optimization objective; and a model training and optimization module for performing total loss calculation and model parameter updates.

[0006] Preferably, the loss calculation module includes: an entropy calculation unit, a filtering unit, a supervised loss unit, and a consistency loss unit; wherein, The entropy calculation unit is used to calculate the predicted entropy value H of the predicted target; The filtering unit is used to filter out those high-deterministic prediction results whose prediction entropy H is lower than the threshold T, and use their categories and bounding boxes as high-quality pseudo-labels to form a pseudo-label dataset R. The supervision loss unit is used to calculate the supervision loss of labeled data; The consistency loss unit is used to calculate the consistency loss between the model's prediction results for different views of the same image.

[0007] A method for detecting blood cells, the method comprising the following steps: Step S1: Supervised pre-training with labeled data The YOLOv5-ALT model was trained in a fully supervised manner using a small amount of labeled blood cell dataset L to obtain the initial weights W. Step S2: Pseudo-label generation and filtering based on prediction entropy Step S201, Model Prediction: Use the initial model trained in step S1 to predict the images in a large number of unlabeled datasets U to obtain the class probability distribution and bounding box of each predicted target; Step S202: Calculate the prediction entropy: For each prediction target, calculate the prediction entropy value H based on its category probability distribution; Step S203, Threshold Filtering: Set a prediction entropy threshold T, and retain only those prediction results whose prediction entropy H is lower than the threshold T. Use their categories and bounding boxes as high-quality pseudo-labels to form a pseudo-label dataset R. Step S3: Consistency Learning Regularization Step S301, Data Augmentation: Apply various random data augmentation transformations to the labeled data L and the unlabeled data U respectively to generate multiple augmented views; Step S302, Consistency Constraint: Input the original unlabeled image and its enhanced view into the current model, and calculate the consistency loss between the model's prediction results for different views of the same image; Step S4: Adaptive Mixup Data Mixing Operation Step S401, Data Mixing: For labeled data, pseudo-labeled data, and data that has undergone consistency learning, an improved Mixup strategy is used to perform linear interpolation mixing; Step S402: Construct the training set: Combine the new mixed data with the original labeled data and pseudo-labeled data to form the final training set S; Step S5: Joint Training and Iterative Optimization The hybrid training set S constructed in step S4 is used to train the YOLOv5-ALT model.

[0008] Preferably, in step S202, the formula for calculating the entropy value H is: in, This represents each element of the predicted probability distribution, where N is the number of categories.

[0009] Preferably, in step S302, the function of the consistency loss L -consistency Using cross-entropy, its calculation formula is as follows: in, P pseudoPredict probabilities for pseudo-labels in the original view. Let be the predicted probability of the i-th augmented view.

[0010] Preferably, in step S401, the adaptive Mixup data blending operation further includes: blending the image and its corresponding labels and bounding boxes, the calculation formula of which is: in, and Two input images, and These are the coordinates of their corresponding bounding boxes. and These are their respective category labels.

[0011] Preferably, in step S5, the total loss function of the YOLOv5-ALT model is... Loss It consists of three weighted parts: in, l The supervised loss is for labeled data; To address the supervisory loss in high-quality pseudo-label data; The consistency loss mentioned in step S3; α, β and λ are the weighting coefficients of each loss term.

[0012] Preferably, steps S2 to S5 can be iterated multiple times to continuously optimize the model weights until the model converges.

[0013] Preferably, step S301, data augmentation, involves applying various random data augmentation transformations, including rotation, scaling, cropping, and color jitter transformations, to the labeled data L and the unlabeled data U respectively, to generate multiple augmented samples.

[0014] Preferably, the system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the blood cell detection method.

[0015] According to the above technical solution, the beneficial effects of the blood cell detection system and method provided by the present invention in use are as follows: (1) This invention significantly improves the performance of the model on low-proportion labeled data by integrating pseudo-label selection based on prediction entropy, consistency learning, and the Mixup hybrid strategy. Experiments show that on the BCCD dataset using only 5% labeled data, the average accuracy (mAP) of this invention reaches 91.02%, which is significantly better than other mainstream semi-supervised methods (such as STAC, DSL, etc.) and even approaches the performance of some fully supervised methods, effectively solving the core pain point of poor model performance in low-label scenarios.

[0016] (2) Significantly reduces data annotation costs and dependence: This invention can efficiently utilize massive amounts of unannotated medical image data. Only a very small amount of labeled data is needed to achieve performance close to that of a fully supervised model. This means that more than 90% of expensive professional annotation work can be reduced, breaking down a major barrier to the application of deep learning in the medical field, and has extremely high economic and practical value.

[0017] (3) The generated pseudo-labels are of high quality and effectively suppress error propagation: Traditional pseudo-labeling methods based on simple confidence thresholds introduce a lot of noise. The innovative prediction entropy-based screening mechanism of this invention can more accurately measure the uncertainty of model predictions, thereby filtering out low-quality, uncertain predictions and retaining only highly certain predictions as pseudo-labels. This ensures the reliability of pseudo-labels from the source, cuts off the chain of error propagation, and lays the foundation for stable optimization of the model.

[0018] (4) The model has excellent robustness and generalization ability: This invention uses dual consistency learning regularization (data augmentation consistency + Mixup implicit consistency) to force the model to learn robust features that are insensitive to image transformations (such as changes in illumination, color, and angle). This makes the trained model more adaptable and stable to image quality fluctuations and differences caused by different acquisition devices encountered in actual clinical practice, greatly improving the reliability of deployment in actual medical environments.

[0019] (5) Effectively prevents overfitting and enhances model generalization performance: The adaptive Mixup strategy introduced in this invention constructs a near-infinite number of virtual training samples by performing linear interpolation between the image and label spaces, greatly enriching the distribution of training data. This not only expands the dataset but also plays a powerful regularization role, smoothing the decision boundary of the model and making it less prone to overfitting the limited labeled data, thereby achieving superior generalization performance.

[0020] (6) Strong compatibility, easy integration and deployment: The framework of this invention is based on the mature YOLOv5 architecture and is improved without introducing extremely complex or difficult-to-implement modules. Therefore, the method is easy to implement, stable in training, and can be easily integrated into existing medical image analysis systems or automated diagnostic equipment without replacing expensive hardware. It has good backward compatibility and engineering prospects.

[0021] (7) Although the present invention takes blood cell detection as an example, its core semi-supervised learning framework has universality and can be widely applied to other medical image detection tasks with scarce labeled data, such as pathological cell detection, lesion detection in medical images (lung nodules, tumors, etc.), abnormal detection in X-ray images, etc., and has strong technical extensibility and promotion value.

[0022] In summary, this invention, through the organic integration and synergistic effect of three core technologies—entropy screening, consistency regularization, and Mixup—constitutes a complete and efficient semi-supervised learning framework, ultimately achieving high-precision and robust blood cell detection with extremely low annotation ratios. This not only represents a breakthrough in technical indicators but also brings multiple beneficial effects in terms of practicality, economy, and reliability, providing an efficient and reliable solution to the "annotation shortage" problem in the field of medical artificial intelligence.

[0023] Other features and advantages of the present invention will be described in detail in the following detailed description section; and all parts not covered in the present invention are the same as or can be implemented using the prior art. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a blood cell detection method provided in a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of the working principle of a blood cell detection system provided in a preferred embodiment of the present invention.

[0025] Explanation of reference numerals in the attached figures 1. Data Input and Processing Layer; 101. Storage Module; 102. Preprocessing Module; 2. Training Data Preparation Layer; 201. Pseudo-Label Management Module; 202. Data Mixing Module; 3. Model Optimization Core Layer; 301. Loss Calculation Module; 302. Model Training and Optimization Module; 4. YOLOv5-ALT Model. Detailed Implementation

[0026] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0027] like Figure 1-2 As shown, a blood cell detection system includes: a data input and processing layer 1, a training data preparation layer 2, a control and optimization layer 3, and a YOLOv5-ALT model 4; wherein, The data input and processing layer 1 includes a storage module 101 for classifying, storing and managing labeled and unlabeled medical image data, and a preprocessing module 102 for processing unlabeled data to obtain a high-quality pseudo-labeled dataset. The training data preparation layer 2 includes: a pseudo-label management module 201 responsible for storing, versioning, dynamically updating and monitoring the quality of pseudo-labels; and a data mixing module 202 that uses adaptive Mixup technology to mix high-quality pseudo-labels and labeled or unlabeled data in the image and label space to obtain diverse training samples. The model optimization core layer 3 includes: a loss calculation module 301 for calculating the total loss of the neural network model 4 to integrate supervision signals from different data sources and different learning strategies into a unified optimization objective, and a model training and optimization module 302 for performing total loss calculation and model parameter updates.

[0028] In the above scheme, the storage module 101 functions as a centralized data management center to classify, store, and manage the labeled and unlabeled medical image data required by the system, providing stable and reliable data support for all subsequent processing and analysis steps, and is the foundation for the operation of the entire blood cell detection system.

[0029] The preprocessing module 102 uses information entropy as a metric to evaluate the uncertainty and quality of the raw predictions generated by the initial model on unlabeled data, thereby generating a highly deterministic, low-noise pseudo-label dataset, laying a solid data foundation for the stability and efficient optimization of the entire semi-supervised learning framework.

[0030] The pseudo-label management module 201 serves as the core management center for pseudo-label data. It is responsible for the storage, version control, dynamic updates, quality monitoring, and collaborative scheduling with downstream training processes of pseudo-labels, ensuring that the system can continuously utilize increasingly high-quality self-generated data to drive the continuous evolution of model performance.

[0031] The data mixing module 202 uses an adaptive mixup technique for object detection to perform linear interpolation on data from different sources (labeled data, high-quality pseudo-labeled data, etc.) in the image and label spaces, thereby generating a large number of diverse virtual training samples. This expands the dataset, smooths the decision boundary, and serves as an effective regularization method, ultimately training a robust model with stronger generalization ability and less prone to overfitting.

[0032] The function of the loss calculation module 301 is to integrate supervision signals from different data sources and different learning strategies into a unified optimization objective. It guides the model not only to fit the limited labeled data, but also to make full use of the patterns and consistency in the unlabeled data, ultimately training a high-precision and highly robust blood cell detection model.

[0033] The model training and optimization module 302 serves as the control core and execution engine of the entire system. It calculates the overall objective function by integrating multi-source loss and iteratively updates the model parameters using backpropagation and optimization algorithms. At the same time, it drives a collaborative working loop of pseudo-label generation, consistency learning, and data mixing until the model converges, ultimately producing a high-precision blood cell detection model that can be used for practical deployment.

[0034] The relationship between these modules is as follows: Data-driven: The output of the upstream module is the input of the downstream module, and data (images, labels, loss) is the link connecting the various modules.

[0035] Iterative Loop: The entire system forms a closed loop driven by the model training and optimization modules. Updates to model parameters improve the quality of pseudo-labels, while better data further optimizes the model, creating a self-reinforcing virtuous cycle.

[0036] Functional synergy: Each module performs its own function, yet they work together to serve the same ultimate goal—training a high-precision, highly robust blood cell detection model when labeled data is scarce. Together, they achieve the synergistic effect of "1+1+1>3" described in this invention.

[0037] In a preferred embodiment of the present invention, the loss calculation module 203 includes: an entropy calculation unit, a filtering unit, a supervised loss unit, and a consistency loss unit; wherein, The entropy calculation unit is used to calculate the predicted entropy value H of the predicted target; The filtering unit is used to filter out those high-deterministic prediction results whose prediction entropy H is lower than the threshold T, and use their categories and bounding boxes as high-quality pseudo-labels to form a pseudo-label dataset R. The supervision loss unit is used to calculate the supervision loss of labeled data; The consistency loss unit is used to calculate the consistency loss between the model's prediction results for different views of the same image.

[0038] In the above scheme, the entropy calculation unit is used to calculate the prediction entropy value H of the prediction target. The higher the prediction entropy H, the lower the certainty of the model for the prediction result.

[0039] The filtering unit is used to filter out high-deterministic prediction results whose prediction entropy H is lower than the threshold T, and their categories and bounding boxes are used as high-quality pseudo-labels to form a pseudo-label dataset R. Its core purpose is to filter out low-confidence noise predictions and improve the quality of false labels from the source.

[0040] The supervision loss unit is used to calculate the supervision loss of labeled data. Supervision loss of high-quality pseudo-label data .

[0041] The consistency loss of the model is calculated using the consistency loss unit, which forces the model to maintain consistent predictions for various changes in the input, thereby greatly enhancing the robustness and generalization ability of the model.

[0042] In a preferred embodiment of the present invention, the system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the blood cell detection method based on YOLOv5-ALT and semi-supervised learning.

[0043] This invention provides a method for detecting blood cells, the method comprising the following steps: Step S1: Supervised pre-training with labeled data The YOLOv5-ALT model was trained in a fully supervised manner using a small amount of labeled blood cell dataset L to obtain the initial weights W. This step provides a relatively reliable initial model for generating high-quality pseudo-labels in the future. Step S2: Pseudo-label generation and filtering based on prediction entropy Step S201, Model Prediction: Use the initial model trained in step S1 to predict the images in a large number of unlabeled datasets U to obtain the class probability distribution and bounding box of each predicted target; Step S202: Calculate the prediction entropy: For each predicted target, calculate the prediction entropy value H based on its category probability distribution. The formula for calculating the entropy value H is: in, P iH represents each element of the predicted probability distribution, where N is the number of classes. A higher prediction entropy H indicates a lower certainty in the model's prediction outcome. Step S203, Threshold Filtering: Set a prediction entropy threshold T, and retain only those prediction results whose prediction entropy H is lower than the threshold T. Use their categories and bounding boxes as high-quality pseudo-labels to form a pseudo-label dataset R. The core purpose is to filter out noisy predictions with low confidence and improve the quality of pseudo-labels from the source.

[0044] Step S3: Consistency Learning Regularization Step S301, Data Augmentation: Apply various random data augmentation transformations (such as rotation, scaling, cropping, color jitter, etc.) to the labeled data L and the unlabeled data U respectively to generate multiple augmented views; Step S302, Consistency Constraint: Input the original unlabeled image and its enhanced view into the current model, and calculate the consistency loss between the model's prediction results for different views of the same image; the consistency loss function... L -consistency Using cross-entropy, its calculation formula is as follows: in, Predict probabilities for pseudo-labels in the original view (as a supervision signal). Let be the predicted probability of the i-th augmented view. By minimizing this loss, the model is forced to maintain consistent predictions across various transformations of the input, thereby greatly enhancing the model's robustness and generalization ability.

[0045] Step S4: Adaptive Mixup Data Mixing Operation Step S401, Data Mixing: For labeled data, pseudo-labeled data, and data processed through consistency learning, an improved Mixup strategy is used for linear interpolation mixing. The Mixup operation not only mixes the images but also their corresponding labels and bounding boxes. The calculation formula is as follows: in, and Two input images, and These are the coordinates of their corresponding bounding boxes. and These are their corresponding category labels (for pseudo-label data, these are pseudo-labels). Step S402: Construct the training set: Combine the new mixed data with the original labeled data and pseudo-labeled data to form the final training set S; This step, by constructing training samples in the feature space, can significantly increase the diversity of data, smooth the decision boundary, thereby effectively preventing model overfitting and further improving generalization performance. Step S5: Joint Training and Iterative Optimization The hybrid training set S constructed in step S4 is used to train the YOLOv5-ALT model, and the total loss function of the YOLOv5-ALT model is... Loss It consists of three weighted parts: in, The supervised loss for labeled data (such as cross-entropy loss and bounding box regression loss). To address the supervisory loss associated with high-quality pseudo-label data, This refers to the consistency loss described in step three. α, β, and λ are the weighting coefficients for each loss term.

[0046] Steps S2 to S5 can be iterated multiple times to continuously optimize the model weights until the model converges.

[0047] In this embodiment, a blood cell detection method based on YOLOv5-ALT and semi-supervised learning is used as described above, and a comparative experiment is conducted with existing well-known semi-supervised or fully supervised algorithms. The experimental comparison data is shown in the table below.

[0048] in: Table 1 shows a comparison of experimental results between this method and the semi-supervised algorithm; Table 2 compares the experimental results of this method with those of the fully supervised algorithm.

[0049] Table 1 compares the experimental results of this method with those of the semi-supervised algorithm. Table 2 shows a comparison of the experimental results of this method and the fully supervised algorithm. The above experimental data proves that: (1) On the BCCD dataset with only 5% labeled data, the average accuracy (mAP) of this invention reached 91.02%, which is significantly better than other mainstream semi-supervised methods (such as STAC, DSL, etc.) and even close to the performance of some fully supervised methods. It effectively solves the core pain point of poor model performance in low-label scenarios.

[0050] (2) Significantly reduces data annotation costs and dependence: This invention can efficiently utilize massive amounts of unannotated medical image data. Only a very small amount of annotated data (such as 1%, 5%, 10%) is needed to achieve performance close to that of a fully supervised model. This means that more than 90% of expensive professional annotation work can be reduced, breaking down a major barrier to the application of deep learning in the medical field, and has extremely high economic and practical value.

[0051] Therefore, the above method has the following advantages: (1) This invention significantly improves the performance of the model on low-proportion labeled data by integrating pseudo-label selection based on prediction entropy, consistency learning, and the Mixup hybrid strategy. Experiments show that on the BCCD dataset using only 5% labeled data, the average accuracy (mAP) of this invention reaches 91.02%, which is significantly better than other mainstream semi-supervised methods (such as STAC, DSL, etc.) and even approaches the performance of some fully supervised methods, effectively solving the core pain point of poor model performance in low-label scenarios.

[0052] (2) Significantly reduces data annotation costs and dependence: This invention can efficiently utilize massive amounts of unannotated medical image data. Only a very small amount of labeled data is needed to achieve performance close to that of a fully supervised model. This means that more than 90% of expensive professional annotation work can be reduced, breaking down a major barrier to the application of deep learning in the medical field, and has extremely high economic and practical value.

[0053] (3) The generated pseudo-labels are of high quality and effectively suppress error propagation: Traditional pseudo-labeling methods based on simple confidence thresholds introduce a lot of noise. The innovative prediction entropy-based screening mechanism of this invention can more accurately measure the uncertainty of model predictions, thereby filtering out low-quality, uncertain predictions and retaining only highly certain predictions as pseudo-labels. This ensures the reliability of pseudo-labels from the source, cuts off the chain of error propagation, and lays the foundation for stable optimization of the model.

[0054] (4) The model has excellent robustness and generalization ability: This invention uses dual consistency learning regularization (data augmentation consistency + Mixup implicit consistency) to force the model to learn robust features that are insensitive to image transformations (such as changes in illumination, color, and angle). This makes the trained model more adaptable and stable to image quality fluctuations and differences caused by different acquisition devices encountered in actual clinical practice, greatly improving the reliability of deployment in actual medical environments.

[0055] (5) Effectively prevents overfitting and enhances model generalization performance: The adaptive Mixup strategy introduced in this invention constructs a near-infinite number of virtual training samples by performing linear interpolation between the image and label spaces, greatly enriching the distribution of training data. This not only expands the dataset but also plays a powerful regularization role, smoothing the decision boundary of the model and making it less prone to overfitting the limited labeled data, thereby achieving superior generalization performance.

[0056] (6) Strong compatibility, easy integration and deployment: The framework of this invention is based on the mature YOLOv5 architecture and is improved without introducing extremely complex or difficult-to-implement modules. Therefore, the method is easy to implement, stable in training, and can be easily integrated into existing medical image analysis systems or automated diagnostic equipment without replacing expensive hardware. It has good backward compatibility and engineering prospects.

[0057] (7) Although the present invention takes blood cell detection as an example, its core semi-supervised learning framework has universality and can be widely applied to other medical image detection tasks with scarce labeled data, such as pathological cell detection, lesion detection in medical images (lung nodules, tumors, etc.), abnormal detection in X-ray images, etc., and has strong technical extensibility and promotion value.

[0058] In summary, this invention not only achieves breakthroughs in technical indicators, but also brings multiple beneficial effects in terms of practicality, economy, and reliability, providing an efficient and reliable solution to the problem of "labeling shortage" in the field of medical artificial intelligence.

[0059] In summary, the blood cell detection system and method provided by this invention overcome the problems of poor quality of pseudo-labels, which leads to error accumulation and propagation, insufficient model robustness and generalization ability, and low utilization efficiency of unlabeled data in the prior art.

[0060] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

[0061] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0062] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. A blood cell detection system, characterized in that, The system includes: a data input and processing layer (1), a training data preparation layer (2), a control and optimization layer (3), and a YOLOv5-ALT model-based layer (4); wherein, The data input and processing layer (1) includes: a storage module (101) for classifying, storing and managing labeled and unlabeled medical image data, and a preprocessing module (102) for processing unlabeled data to obtain a high-quality pseudo-labeled dataset. The training data preparation layer (2) includes: a pseudo-label management module (201) responsible for storing, versioning, dynamically updating and monitoring the quality of pseudo-labels, and a data mixing module (202) that uses adaptive Mixup technology to mix high-quality pseudo-labels and labeled or unlabeled data in the image and label space to obtain diverse training samples. The model optimization core layer (3) includes: a loss calculation module (301) for calculating the total loss of the YOLOv5-ALT model (4) to integrate supervision signals from different data sources and different learning strategies into a unified optimization objective, and a model training and optimization module (302) for performing total loss calculation and model parameter update.

2. The blood cell detection system according to claim 1, characterized in that, The loss calculation module (301) includes: an entropy calculation unit, a filtering unit, a supervised loss unit, and a consistency loss unit; wherein, The entropy calculation unit is used to calculate the predicted entropy value H of the predicted target; The filtering unit is used to filter out those high-deterministic prediction results whose prediction entropy H is lower than the threshold T, and use their categories and bounding boxes as high-quality pseudo-labels to form a pseudo-label dataset R. The supervision loss unit is used to calculate the supervision loss of labeled data; The consistency loss unit is used to calculate the consistency loss between the model's prediction results for different views of the same image.

3. The blood cell detection method according to claim 1 or 2, characterized in that, The method includes the following steps: Step S1: Supervised pre-training with labeled data The YOLOv5-ALT model (4) was trained in a fully supervised manner using a small amount of labeled blood cell dataset L to obtain the initial weights W; Step S2: Pseudo-label generation and filtering based on prediction entropy Step S201, Model Prediction: Use the initial model trained in step S1 to predict the images in a large number of unlabeled datasets U to obtain the class probability distribution and bounding box of each predicted target; Step S202: Calculate the prediction entropy: For each prediction target, calculate the prediction entropy value H based on its category probability distribution; Step S203, Threshold Filtering: Set a prediction entropy threshold T, and retain only those prediction results whose prediction entropy H is lower than the threshold T. Use their categories and bounding boxes as high-quality pseudo-labels to form a pseudo-label dataset R. Step S3: Consistency Learning Regularization Step S301, Data Augmentation: Apply various random data augmentation transformations to the labeled data L and the unlabeled data U respectively to generate multiple augmented views; Step S302, Consistency Constraint: Input the original unlabeled image and its enhanced view into the current model, and calculate the consistency loss between the model's prediction results for different views of the same image; Step S4: Adaptive Mixup Data Mixing Operation Step S401, Data Mixing: For labeled data, pseudo-labeled data, and data that has undergone consistency learning, an improved Mixup strategy is used to perform linear interpolation mixing; Step S402: Construct the training set: Combine the new mixed data with the original labeled data and pseudo-labeled data to form the final training set S; Step S5: Joint Training and Iterative Optimization The hybrid training set S constructed in step S4 is used to train the YOLOv5-ALT model.

4. The blood cell detection method according to claim 3, characterized in that, In step S202, the formula for calculating the entropy value H is: in, This represents each element of the predicted probability distribution, where N is the number of categories.

5. The blood cell detection method according to claim 4, characterized in that, In step S302, the function of the consistency loss Using cross-entropy, its calculation formula is as follows: in, Predict probabilities for pseudo-labels in the original view. Let be the predicted probability of the i-th augmented view.

6. The blood cell detection method according to claim 5, characterized in that, In step S401, the adaptive Mixup data blending operation further includes: blending the image and its corresponding labels and bounding boxes, the calculation formula of which is: in, and Two input images, and These are the coordinates of their corresponding bounding boxes. and These are their respective category labels.

7. The blood cell detection method according to claim 6, characterized in that, In step S5, the total loss function of the YOLOv5-ALT model Loss Weighted by three parts composition: in, The supervised loss is for labeled data; To address the supervisory loss in high-quality pseudo-label data; y The consistency loss mentioned in step S3; α, β and λ are the weighting coefficients of each loss term.

8. The blood cell detection method according to claim 7, characterized in that, Steps S2 to S5 can be iterated multiple times to continuously optimize the model weights until the model converges.

9. The blood cell detection method according to claim 8, characterized in that, Step S301, data augmentation: Apply various random data augmentation transformations, including rotation, scaling, cropping, and color jitter transformations, to the labeled data L and the unlabeled data U respectively to generate multiple augmented samples.

10. The blood cell detection system according to claim 1 or 2, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the blood cell detection method as described in any one of claims 3 to 9.

Images