Blood cell analysis system and method based on computer vision

Abstract

The invention relates to the technical field of blood cell analysis, and particularly discloses a blood cell analysis system and method based on computer vision, which uses an image recognition and generation technology based on computer vision to carry out grid division and semantic feature extraction on blood image data. Performing global text generation and semantic coding on each blood image data grid block feature after semantic extraction so as to perform adjustment response aggregation on attention of different blood local areas in each blood image data grid block feature according to blood image text description semantic features; therefore, intelligent optimization of the blood image data is realized. In this way, the complex deep-layer mutual influence association relationship between cells can be captured, so that a more accurate diagnosis basis is provided. Meanwhile, the processing strategy can be dynamically adjusted according to the specific description content of the blood image so as to better adapt to diversified samples, and the method does not only depend on fixed preprocessing parameters.

Description

Technical Field

[0001] The present application relates to the technical field of blood cell analysis, and more specifically, to a blood cell analysis system and method based on computer vision. Background Art

[0002] Blood cell analysis plays an indispensable role in clinical diagnosis and pathological research, and is crucial for the early detection, accurate diagnosis, and effective treatment of various blood diseases, such as anemia, leukemia, and lymphoma. By performing detailed morphological, quantitative, and functional analyses of red blood cells, white blood cells, and platelets in blood samples, physicians can assess patient health, monitor disease progression, and adjust treatment plans.

[0003] Patent application number CN118397620A discloses a vision-based blood cell analysis method. The method first captures a blood sample image and performs image preprocessing. Next, image segmentation techniques are used to separate blood cells and extract features such as shape, size, and color. Based on these features, a machine learning algorithm is used to train an analysis model to classify and count blood cells. Finally, the trained model is applied to new images in real time, outputting a report on the number and proportion of different blood cell types.

[0004] In the above invention, the collected images are preprocessed by eliminating noise, enhancing contrast and smoothing the images. Specifically, the preprocessing method focuses on pixel-level operations, such as eliminating noise, enhancing contrast and smoothing images, but they mainly target low-level visual features (such as edges and textures). This processing method cannot capture high-level information or contextual semantics in the image, such as the specific morphology of cells, internal structures and their relationships with each other, which is particularly important in medical image analysis. In addition, traditional preprocessing methods are usually fixed, that is, the same algorithm parameters are applied to all input images. However, in reality, blood samples not only have differences between individuals, but samples collected from the same patient at different time points may also be different. A fixed preprocessing process cannot be adjusted according to the unique properties of the sample, which may lead to poor preprocessing effects in some cases.

[0005] Therefore, an optimized computer vision-based blood cell analysis solution is desired. Summary of the Invention

[0006] This application provides a computer vision-based blood cell analysis system and method that can capture the complex and deep interplay between cells, thereby providing a more accurate basis for diagnosis. Furthermore, the system can dynamically adjust processing strategies based on the specific description of blood images to better adapt to diverse samples, rather than relying solely on fixed preprocessing parameters.

[0007] In a first aspect, a computer vision-based blood cell analysis system is provided, comprising: a blood image acquisition module for acquiring an image of a blood sample using an image acquisition device to obtain blood image data; a blood image preprocessing module for preprocessing the blood image data based on text semantic description to obtain preprocessed blood image data; a blood cell segmentation module for separating blood cells and background in the preprocessed blood image data using image segmentation technology to obtain a blood cell image; a blood cell feature extraction module for extracting blood cell features from the blood cell image; a blood cell classification module for classifying and counting the blood cell features to obtain classification and counting results; and a result generation module for generating a blood cell report or chart based on the classification and counting results. The blood image preprocessing module comprises: a blood image segmentation and semantic analysis unit, configured to perform grid segmentation and semantic feature extraction on the blood image data to obtain a sequence of grid block features of the blood image data; A blood image global semantic representation unit, configured to perform text generation and semantic encoding on the sequence of grid block features of the blood image data based on the global blood image to obtain semantic features of the blood image text description; A blood image text scanning query unit is used to perform efficient semantic scanning query on a sequence of semantic features of the blood image text description and features of the blood image data grid block to obtain an aggregated representation of semantic query responses of the blood image data grid block, and obtain the preprocessed blood image data based on the aggregated representation of semantic query responses of the blood image data grid block.

[0008] Furthermore, the blood image segmentation semantic analysis unit includes: a blood image grid division subunit, configured to grid the blood image data using a grid divider to obtain a sequence of blood image data grid blocks; The grid block semantic feature extraction subunit is used to pass the sequence of blood image data grid blocks through a blood image grid block semantic feature extractor based on a CNN-BiLSTM-SE network to obtain a sequence of blood image data grid block feature vectors as the sequence of blood image data grid block features.

[0009] Furthermore, the grid block semantic feature extraction subunit is used to: Aggregating the sequence of blood image data grid blocks along the channel dimension and inputting them into SEBlock and CNN models respectively to obtain a blood image data grid weight feature map and a blood image data grid aggregation feature map; Multiplying each feature matrix along the channel dimension in the blood image data grid weight feature map by position with each feature matrix along the channel dimension in the blood image data grid aggregate feature map to obtain a blood image data grid aggregate weighted feature map; Performing feature expansion along the channel dimension on the blood image data grid aggregated weighted feature map to obtain a sequence of blood image data grid feature matrices; After flattening the sequence of blood image data grid feature matrices to obtain a sequence of blood image data grid flattened feature vectors, the sequence of blood image data grid flattened feature vectors is sequentially input into a BiLSTM model and a fully connected layer to obtain a sequence of blood image data grid block feature vectors.

[0010] Furthermore, the blood image global semantic representation unit includes: a blood image data cascading subunit, configured to cascade the sequence of blood image data grid block feature vectors to obtain a global semantic representation vector of the blood image data; a blood image text description generating subunit, configured to pass the global semantic representation vector of the blood image data through a blood image text generator based on a large language model to obtain a key semantic text description of the blood image; The blood image text description semantic encoding subunit is used to perform semantic encoding on the key semantic text description of the blood image to obtain a blood image text description semantic feature vector as the blood image text description semantic feature.

[0011] Furthermore, the blood image text scanning query unit includes: a blood image text efficient semantic scanning query subunit, configured to perform an efficient semantic scanning query on a sequence of the blood image text description semantic feature vector and the blood image data grid block feature vector to obtain a blood image data grid block semantic query response aggregate representation vector as the blood image data grid block semantic query response aggregate representation; The pre-processed blood image data generating subunit is configured to obtain the pre-processed blood image data by aggregating representation vectors according to the semantic query responses of the blood image data grid blocks.

[0012] Furthermore, the blood image text efficient semantic scanning query subunit includes: a blood image data fast matching secondary subunit, configured to determine a fast matching subsequence of blood image data grid block feature vectors from the sequence of blood image data grid block feature vectors based on matching information between the blood image text description semantic feature vector and the sequence of blood image data grid block feature vectors; The blood image data query response aggregation secondary sub-unit is used to perform a linear transformation on the blood image text description semantic feature vector to obtain a blood image text description query vector and a blood image text description value vector, and use a subsequence of the blood image data grid block feature vector as a subsequence of the key vector to perform Transformer-based query response aggregation on the blood image text description query vector, the blood image text description value vector and the subsequence of the key vector to obtain the blood image data grid block semantic query response aggregation representation vector.

[0013] Furthermore, the blood image data fast matching secondary subunit is used to: calculating the mutual information between the blood image text description semantic feature vector and each blood image data grid block feature vector in the sequence of blood image data grid block feature vectors to obtain a sequence of blood image text-grid fast semantic matching factors as the matching information; identifying a first blood image text-grid fast semantic maximum value and a second blood image text-grid fast semantic maximum value from the sequence of the blood image text-grid fast semantic matching factors; Based on the positions of the first blood image text-grid fast semantic maximum and the second blood image text-grid fast semantic maximum in the sequence of blood image text-grid fast semantic matching factors, a subsequence of the blood image data grid block feature vectors that are fast matched is determined from the sequence of the blood image data grid block feature vectors.

[0014] Furthermore, the preprocessed blood image data generation subunit is used to: add generation prompt information to the end of the blood image data grid block semantic query response aggregate representation vector, and then input it together with the blood image data into the AIGC-based optimized blood image generator to obtain the preprocessed blood image data.

[0015] In a second aspect, a computer vision-based blood cell analysis method is provided, comprising: using an image acquisition device to acquire an image of a blood sample to obtain blood image data; preprocessing the blood image data based on text semantic description to obtain preprocessed blood image data; separating blood cells and background in the preprocessed blood image data using image segmentation technology to obtain a blood cell image; extracting blood cell features from the blood cell image; classifying and counting the blood cell features to obtain classification and counting results; and generating a blood cell report or chart based on the classification and counting results. The preprocessing of the blood image data based on text semantic description to obtain the preprocessed blood image data comprises: Performing grid division and semantic feature extraction on the blood image data to obtain a sequence of grid block features of the blood image data; Performing text generation and semantic encoding based on the global blood image on the sequence of grid block features of the blood image data to obtain semantic features of blood image text description; An efficient semantic scanning query is performed on a sequence of the semantic features of the blood image text description and the features of the blood image data grid block to obtain an aggregated representation of the semantic query response of the blood image data grid block, and the preprocessed blood image data is obtained based on the aggregated representation of the semantic query response of the blood image data grid block.

[0016] Furthermore, grid division and semantic feature extraction are performed on the blood image data to obtain a sequence of grid block features of the blood image data, including: gridding the blood image data using a gridder to obtain a sequence of blood image data grid blocks; The sequence of blood image data grid blocks is passed through a blood image grid block semantic feature extractor based on a CNN-BiLSTM-SE network to obtain a sequence of blood image data grid block feature vectors as the sequence of blood image data grid block features.

[0017] The present application provides a computer vision-based blood cell analysis system and method, which uses computer vision-based image recognition and generation technology to grid the blood image data and extract semantic features, and then performs global text generation and semantic encoding on the semantically extracted grid block features of each blood image data, so as to adjust the attention of different blood local areas in each blood image data grid block feature according to the semantic features of the blood image text description, and respond to aggregation, thereby realizing intelligent optimization of the blood image data. Through the method of the present application, the complex and deep mutual influence and correlation relationship between cells can be captured, thereby providing a more accurate basis for diagnosis. At the same time, the processing strategy can be dynamically adjusted according to the specific description content of the blood image to better adapt to diverse samples, rather than just relying on fixed preprocessing parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings of the embodiments of the present application. Obviously, the drawings described below only relate to some embodiments of the present application, and are not intended to limit the present application.

[0019] Figure 1 This is a schematic block diagram of a computer vision-based blood cell analysis system according to an embodiment of the present application.

[0020] Figure 2Schematic diagram of data flow of a computer vision-based blood cell analysis system according to an embodiment of the present application.

[0021] Figure 3 This is a schematic block diagram of a blood image preprocessing module in a computer vision-based blood cell analysis system according to an embodiment of the present application.

[0022] Figure 4 This is a schematic block diagram of a blood image segmentation semantic analysis unit in a computer vision-based blood cell analysis system according to an embodiment of the present application.

[0023] Figure 5 This is a schematic block diagram of a blood image global semantic representation unit in a computer vision-based blood cell analysis system according to an embodiment of the present application.

[0024] Figure 6 This is a schematic flowchart of a blood cell analysis method based on computer vision according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without making creative efforts also fall within the scope of protection of this application.

[0026] Based on this, this application proposes a blood cell analysis system based on computer vision, such as Figure 1 and Figure 2 As shown, the computer vision-based blood cell analysis system includes: a blood image acquisition module 10 for using an image acquisition device to acquire an image of a blood sample to obtain blood image data; a blood image preprocessing module 20 for preprocessing the blood image data based on text semantic description to obtain preprocessed blood image data; a blood cell segmentation module 30 for separating blood cells and background in the preprocessed blood image data using image segmentation technology to obtain a blood cell image; a blood cell feature extraction module 40 for extracting blood cell features from the blood cell image; a blood cell classification module 50 for classifying and counting the blood cell features to obtain classification and counting results; and a result generation module 60 for generating a blood cell report or chart based on the classification and counting results. In particular, the blood cell features include the shape, size, and color features of the blood cells.

[0027] For example, in the blood image acquisition module 10, an image acquisition device is used to capture images of blood samples to obtain blood image data. It should be understood that analyzing the status and quantity of blood components such as red blood cells, white blood cells, and platelets can provide doctors with important health assessment information, assisting in early disease detection, accurate diagnosis, and the development of treatment plans. To achieve these goals, high-quality blood images are first required as basic data.

[0028] In one embodiment, when a detailed analysis of blood cells in a blood sample is required, the diluted and mixed blood sample is first placed on a special glass slide or blood processing dish. Next, a high-resolution camera, such as a microscope camera or a specially designed medical-grade imaging system, is used to capture the microstructure of the blood sample. To ensure that the captured image can truly reflect the internal conditions of the blood sample and meet the requirements of subsequent image processing and analysis, the acquisition process must consider multiple factors: on the one hand, the consistency and stability of the acquisition environment must be ensured, including but not limited to the light source type (for example, using an LED cold light source to reduce the impact of heat on the sample), light intensity and uniformity, focus adjustment, magnification selection, etc. This helps to improve the quality of the image, making the blood cell features more clearly visible, thereby facilitating subsequent feature extraction and classification counting.

[0029] For example, in the blood image preprocessing module 20, the blood image data is preprocessed based on textual semantic descriptions to obtain preprocessed blood image data. Traditional preprocessing methods typically focus on pixel-level operations, such as noise removal, contrast enhancement, and image smoothing. However, these operations primarily target low-level visual features (such as edges and textures) and fail to fully capture high-level information or contextual semantics within the image, such as the specific morphology, internal structure, and interrelationships of cells. This is particularly important in medical image analysis, as doctors often need to understand the complex interactions between blood cells to make correct diagnoses. Furthermore, in practice, blood samples vary from individual to individual, and even samples collected from the same patient at different time points may differ. A fixed preprocessing process struggles to cope with this diversity, potentially resulting in poor preprocessing results in some cases. Therefore, introducing a preprocessing method based on textual semantic descriptions can more flexibly adapt to various sample characteristics, improving the reliability and accuracy of analysis results.

[0030] Accordingly, in the blood image preprocessing module, the technical concept of the present application is to use computer vision-based image recognition and generation technology to perform grid division and semantic feature extraction on the blood image data, and then perform global text generation and semantic encoding on the semantically extracted grid block features of each blood image data, so as to adjust the attention of different blood local areas in each blood image data grid block feature according to the semantic features of the blood image text description, and respond to aggregation, thereby realizing intelligent optimization of the blood image data. Through the method of the present application, the complex and deep mutual influence and correlation relationship between cells can be captured, thereby providing a more accurate basis for diagnosis. At the same time, the processing strategy can be dynamically adjusted according to the specific description content of the blood image to better adapt to diverse samples, rather than just relying on fixed preprocessing parameters.

[0031] like Figure 3 As shown, the blood image preprocessing module 20 includes: a blood image division semantic analysis unit 21, which is used to perform grid division and semantic feature extraction on the blood image data to obtain a sequence of blood image data grid block features; a blood image global semantic representation unit 22, which is used to perform text generation and semantic encoding based on the global blood image on the sequence of blood image data grid block features to obtain blood image text description semantic features; a blood image text scanning query unit 23, which is used to perform efficient semantic scanning query on the blood image text description semantic features and the sequence of blood image data grid block features to obtain a blood image data grid block semantic query response aggregate representation, and obtain the preprocessed blood image data based on the blood image data grid block semantic query response aggregate representation.

[0032] In one embodiment, Figure 4 As shown, the blood image division semantic analysis unit 21 includes: a blood image grid division subunit 211, used to use a grid divider to grid divide the blood image data to obtain a sequence of blood image data grid blocks; a grid block semantic feature extraction subunit 212, used to pass the sequence of blood image data grid blocks through a blood image grid block semantic feature extractor based on a CNN-BiLSTM-SE network to obtain a sequence of blood image data grid block feature vectors as the sequence of blood image data grid block features.

[0033] Exemplarily, in the blood image gridding subunit 211, a gridder is used to grid the blood image data to obtain a sequence of blood image data grid blocks. Considering that different blood cell types in the blood image data may have specific morphological characteristics, these specific characteristics are important for subsequent blood cell classification and counting. Therefore, in order to accurately characterize the local features within each region and improve the quality of subsequent feature analysis, the technical solution of this application uses a gridder to grid the blood image data, dividing the image into multiple small regions to obtain a sequence of small, detail-rich blood image data grid blocks.

[0034] Exemplarily, in the grid block semantic feature extraction subunit 212, the sequence of blood image data grid blocks is passed through a blood image grid block semantic feature extractor based on a CNN-BiLSTM-SE network to obtain a sequence of blood image data grid block feature vectors as the sequence of blood image data grid block features. It should be understood that in order to effectively capture and utilize spatial and temporal sequence information in blood images, thereby improving the quality of blood images, in the technical solution of the present application, the sequence of blood image data grid blocks is passed through a blood image grid block semantic feature extractor based on a CNN-BiLSTM-SE network to obtain a sequence of blood image data grid block feature vectors. It should be understood that in the CNN-BiLSTM-SE network, CNNs (Convolutional Neural Networks) are used to extract spatial features from blood images. The convolutional layers can capture local patterns in the image. Each convolutional layer contains multiple convolution kernels to capture key local information in the input features, such as cell shape and texture, which are crucial for recognition and classification tasks. SEBlocks (Squeeze-and-Excitation Blocks) are used to enhance or suppress features in different channels. The SE module can improve the quality of feature representation by adaptively reweighting the importance of each channel, allowing the model to focus more on features that are helpful for the task. Specifically, the internal operation process of SEBlock includes Squeeze, Excitation, and Reweight. Among them, the Squeeze operation compresses the features of the spatial dimension, the Excitation operation reduces the dimension of the features first and then increases the dimension, and is implemented through a fully connected layer and a Sigmoid function. The Reweight operation redistributes the original features across the channels according to the weights obtained in the Excitation stage; BiLSTM (Bidirectional LongShort-Term Memory) is used to process time series data composed of spatial features extracted by CNN. Learning sequence information from both the forward and backward directions helps capture the spatiotemporal dependencies between blood image grid blocks, thereby better understanding the association between the contexts of each blood image grid block.

[0035] In one embodiment, the grid block semantic feature extraction subunit 212 is used to: aggregate the sequence of blood image data grid blocks along the channel dimension and input them into SEBlock and CNN models respectively to obtain a blood image data grid weight feature map and a blood image data grid aggregate feature map; multiply each feature matrix along the channel dimension in the blood image data grid weight feature map with each feature matrix along the channel dimension in the blood image data grid aggregate feature map by position to obtain a blood image data grid aggregate weighted feature map; perform feature expansion along the channel dimension on the blood image data grid aggregate weighted feature map to obtain a sequence of blood image data grid feature matrices; flatten the sequence of blood image data grid feature matrices to obtain a sequence of blood image data grid flattened feature vectors, and then input the sequence of blood image data grid flattened feature vectors into the BiLSTM model and the fully connected layer in sequence to obtain a sequence of blood image data grid block feature vectors.

[0036] In one embodiment, Figure 5 As shown, the blood image global semantic representation unit 22 includes: a blood image data cascade subunit 221, used to cascade the sequence of the blood image data grid block feature vectors to obtain a blood image data global semantic representation vector; a blood image text description generation subunit 222, used to pass the blood image data global semantic representation vector through a blood image text generator based on a large language model to obtain a blood image key semantic text description; a blood image text description semantic encoding subunit 223, used to semantically encode the blood image key semantic text description to obtain a blood image text description semantic feature vector as the blood image text description semantic feature.

[0037] Exemplarily, in the blood image data concatenation subunit 221, the sequence of blood image data grid block feature vectors is concatenated to obtain a global semantic representation vector for the blood image data. It should be understood that in order to integrate all local features to form a comprehensive global representation, thereby more comprehensively capturing and expressing the information of the entire blood image, in the technical solution of this application, the sequence of blood image data grid block feature vectors is concatenated to connect the local feature vectors of each grid block together, thereby obtaining a global semantic representation vector for the blood image data that contains all local information of the blood image.

[0038] Exemplarily, in the blood image text description generation subunit 222, the global semantic representation vector of the blood image data is passed through a blood image text generator based on a large language model to generate a key semantic text description of the blood image. It should be understood that, given that blood image data often contains rich semantic information, in order to more intuitively and clearly understand the image, in the technical solution of the present application, the global semantic representation vector of the blood image data is passed through a blood image text generator based on a large language model to convert the complex image data information into an easily understandable text description, thereby generating a key semantic text description of the blood image. This enables a better understanding of the specific content information in the image. In particular, in a specific embodiment of the present application, the global semantic representation vector of the blood image data can be processed using GPT-4 or a decoder structure based on a recurrent neural network (RNN), a long short-term memory network (LSTM), or a gated recurrent unit to generate a key semantic text description of the blood image containing rich semantic information and content.

[0039] Exemplarily, in the blood image text description semantic encoding subunit 223, the key semantic text description of the blood image is semantically encoded to obtain a blood image text description semantic feature vector as the blood image text description semantic feature. It should be understood that in order to capture and extract the semantic information between each image area in the blood image and the semantic association between the context from the key semantic text description of the blood image, in the technical solution of the present application, the key semantic text description of the blood image is semantically encoded to capture and mine the deep-level blood image semantic information in the text to obtain a blood image text description semantic feature vector. In particular, in a specific embodiment of the present application, the key semantic text description of the blood image can be semantically encoded by BERT to utilize a bidirectional Transformer architecture to capture complex and professional terms and deep descriptions in the medical field to obtain a blood image text description semantic feature vector.

[0040] In one embodiment, the blood image text scanning query unit 23 includes: a blood image text efficient semantic scanning query subunit 231, which is used to perform an efficient semantic scanning query on a sequence of the blood image text description semantic feature vector and the blood image data grid block feature vector to obtain a blood image data grid block semantic query response aggregate representation vector as the blood image data grid block semantic query response aggregate representation; and a preprocessed blood image data generation subunit 232, which is used to obtain the preprocessed blood image data based on the blood image data grid block semantic query response aggregate representation vector.

[0041] Exemplarily, in the blood image text efficient semantic scanning query subunit 231, an efficient semantic scanning query is performed on the sequence of the blood image text description semantic feature vectors and the blood image data grid block feature vectors to obtain a blood image data grid block semantic query response aggregate representation vector as the blood image data grid block semantic query response aggregate representation. In particular, considering that the blood image text description semantic features capture the textual description of the image content, including cell type, morphology, density, distribution pattern, etc., the sequence of blood image data grid block features represents visual features within different regions, such as cell shape, texture, color, etc., as well as the dependencies between different grid blocks. Therefore, in order to dynamically adjust the attention of each blood image data grid block feature based on the blood image text description semantic features and aggregate them to more comprehensively understand the content of the blood image, the technical solution of this application introduces an efficient semantic scanning query mechanism to quickly query and respond to the sequence of the blood image text description semantic features and the blood image data grid block features to obtain a blood image data grid block semantic query response aggregate representation.

[0042] In one embodiment, the blood image text efficient semantic scanning query subunit 231 includes: a blood image data fast matching secondary subunit, used to determine a fast matching subsequence of blood image data grid block feature vectors from the sequence of blood image data grid block feature vectors based on the matching information between the blood image text description semantic feature vector and the sequence of blood image data grid block feature vectors; a blood image data query response aggregation secondary subunit, used to perform a linear transformation on the blood image text description semantic feature vector to obtain a blood image text description query vector and a blood image text description value vector, and use the subsequence of the blood image data grid block feature vector as a subsequence of the key vector, and perform Transformer-based query response aggregation on the blood image text description query vector, the blood image text description value vector and the subsequence of the key vector to obtain the blood image data grid block semantic query response aggregation representation vector.

[0043] In one embodiment, the blood image data fast matching secondary sub-unit is used to: calculate the mutual information between the blood image text description semantic feature vector and each blood image data grid block feature vector in the sequence of blood image data grid block feature vectors to obtain a sequence of blood image text-grid fast semantic matching factors as the matching information; identify a first blood image text-grid fast semantic maximum and a second blood image text-grid fast semantic maximum from the sequence of blood image text-grid fast semantic matching factors; and determine a subsequence of the blood image data grid block feature vectors for fast matching from the sequence of blood image data grid block feature vectors based on the positions of the first blood image text-grid fast semantic maximum and the second blood image text-grid fast semantic maximum in the sequence of blood image text-grid fast semantic matching factors.

[0044] Specifically, the mutual information between the blood image text description semantic feature vector and each blood image data grid block feature vector in the sequence of blood image data grid block feature vectors is calculated to obtain a sequence of blood image text-grid fast semantic matching factors as the matching information. This process can be expressed as follows: ; ; ; ; in, is a sequence of grid block feature vectors of the blood image data, , , and are the first, second, and third in the sequence of the blood image data grid block feature vectors. and blood image data grid block feature vectors, is the semantic feature vector of the blood image text description, yes function, is the semantic transformation feature vector of the blood image text description, yes The corresponding blood image data grid block transformation feature vector, express The information entropy of express The information entropy of express and The joint entropy between express and Fast semantic matching factor between blood images,text-grids.

[0045] As you can understand, mutual information (MI), a statistic that measures the dependence between two random variables, is used here to assess the degree of correlation between the semantic feature vectors of the blood image text description and the feature vectors of each blood image data grid block. A larger mutual information value indicates greater information overlap between the two, meaning that the concepts or information they represent are highly correlated. In the context of blood cell analysis, this means that the image features of certain local regions have a strong semantic connection with the text description of the entire image. For example, when the text description mentions "clustered white blood cells," mutual information can help identify grid blocks that actually contain a large number of white blood cell clusters, rather than simply areas with similar morphology but different content. Next, based on the calculated mutual information values, the system constructs a sequence of blood image text-grid fast semantic matching factors. This sequence not only reflects the strength of the relationship between each grid block feature vector and the global text description, but also provides an objective criterion for subsequent fast matching. In practical applications, this step helps screen the grid blocks most likely to be related to the text description, thereby reducing unnecessary computational burden and improving processing speed. For example, in a complex blood sample image, there may be multiple regions containing different types of blood cells; by calculating mutual information, we can prioritize those parts that best match the text description and concentrate resources for more in-depth analysis.

[0046] Specifically, the first blood image text-grid fast semantic maximum and the second blood image text-grid fast semantic maximum are identified from the sequence of the blood image text-grid fast semantic matching factors. This process can be expressed as follows: ; ; in, is a sequence of blood image text-grid fast semantic matching factors, is the number of blood image text-grid fast semantic matching factors in the sequence of blood image text-grid fast semantic matching factors, To get the maximum value in the sequence, is the first blood image text-grid fast semantic maximum, It is the second blood image text-grid fast semantic maximum.

[0047] It should be understood that the sequence of blood image text-grid fast semantic matching factors, obtained by calculating mutual information, reflects the semantic similarity between each grid block feature vector and the overall text description. Higher values in this sequence indicate local regions where the image features have a strong correlation with the text description. For example, a complex blood sample image may contain multiple regions containing different types of blood cells. Mutual information calculation can quantify the strength of the relationship between these regions and the text description (e.g., "clustered white blood cells"). Identifying the maximum values of the text-grid fast semantic matching factors for the first and second blood images is essentially finding the two grid block feature vectors with the highest semantic relevance to the text description. These two maxima not only represent the highest mutual information scores but also provide key clues as to which local regions best reflect the content of the text description. For example, if the text description specifically emphasizes the presence of a certain type of blood cell or a particular morphology, the corresponding maxima will point to the grid blocks that actually contain that type of blood cell or exhibit that morphology. This precise matching helps ensure that subsequent processing focuses on the most representative and relevant regions, avoiding distraction by irrelevant data.

[0048] Specifically, based on the positions of the first blood image text-grid fast semantic maximum and the second blood image text-grid fast semantic maximum in the sequence of blood image text-grid fast semantic matching factors, a subsequence of the blood image data grid block feature vectors for fast matching is determined from the sequence of blood image data grid block feature vectors. This process can be expressed as follows: ; ; ; in, To find the position of the maximum value in the sequence, To first determine the quick match position, Determine the quick match position for the second, is a subsequence of the feature vectors of the blood image data grid block.

[0049] As you can understand, based on the positions of the two maxima, a fast-matching subsequence can be identified from the original grid block feature vector sequence. This step acts as an intelligent feature selection or filtering mechanism, enabling the system to significantly improve efficiency while maintaining high accuracy. Specifically, it allows the system to focus on local regions most likely to provide valuable information, reducing noise interference and improving the model's overall performance and generalization. For example, when rapidly classifying a large number of blood samples, this approach can effectively narrow the search scope and focus on the most promising candidates, thereby speeding up processing without sacrificing accuracy. In this way, the system can more efficiently locate local regions that are highly relevant to the text description, ensuring that subsequent feature extraction, classification and counting operations are more focused on the actual target cells, thereby improving the accuracy of the overall analysis. Furthermore, this semantic relevance-based filtering method enhances the system's robustness by avoiding false positives caused by background noise or other irrelevant factors. For example, in some cases, even if there are regions in an image that are morphologically similar but differ in content, mutual information calculation can help distinguish the truly relevant parts, ensuring the reliability of the results.

[0050] Specifically, in the blood image data query response aggregation secondary subunit, a linear transformation is performed on the blood image text description semantic feature vector to obtain a blood image text description query vector and a blood image text description value vector. Furthermore, using a subsequence of the blood image data grid block feature vector as a subsequence of the key vector, a Transformer-based query response aggregation is performed on the blood image text description query vector, the blood image text description value vector, and the subsequence of the key vector to obtain the blood image data grid block semantic query response aggregation representation vector. This process can be expressed as follows: ; ; ; in, and are the query embedding matrix and the query bias vector respectively, is the blood image text description query vector, and are the value embedding matrix and value bias vector respectively, is the blood image text description value vector, is the first in the sequence of the blood image data grid block feature vectors blood image data grid block feature vectors, yes The transposed vector of yes length, yes function, is matrix multiplication, is an aggregated representation vector of the blood image data grid block semantic query response.

[0051] Specifically, a linear transformation is first performed on the semantic feature vector of the blood image text description, resulting in two important outputs: a query vector and a value vector. The query vector is used to identify the most relevant key vectors, while the value vector carries the specific information about these relevant parts. This approach ensures that the query process dynamically focuses on the most relevant local regions, avoiding the limitation of traditional methods that process each element sequentially. Next, a subsequence of the feature vectors of the blood image data grid blocks identified in the fast matching step is used as a subsequence of the key vectors. This step effectively utilizes the most semantically relevant grid blocks previously selected for further detailed analysis. This approach significantly improves efficiency while maintaining high accuracy, by focusing resources on the local regions most likely to provide valuable information and reducing noise interference. The query vector, value vector, and key vector subsequence are then input into a cross-domain query encoding module based on the Transformer architecture. The core of the Transformer is the self-attention mechanism, which allows the model to dynamically focus on different parts when processing sequential data. In this scenario, the query vector is responsible for finding the key vector that is most relevant to it, while the value vector carries the specific information of these relevant parts. In this way, the converter model can effectively capture and utilize the long-distance dependencies between the semantic feature vector of the blood image text description and the sequence of the blood image data grid block feature vector to generate an accurate query response semantic encoding vector. For example, in blood cell analysis, this encoding can more accurately describe the complex interactions between blood cells and their internal structure. Assuming that the text description particularly emphasizes "aggregated white blood cells", then through the above process, the system can accurately locate those grid blocks that do contain a large number of white blood cell aggregates and further extract their detailed morphological features. This not only improves the accuracy of classification and counting, but also helps identify abnormalities or special patterns, providing doctors with a more reliable basis for diagnosis.

[0052] Exemplarily, the preprocessed blood image data generation subunit 232 generates the preprocessed blood image data based on the aggregated representation vector of the semantic query response of the blood image data grid block. In one embodiment, the preprocessed blood image data generation subunit 232 is configured to: append generation prompt information to the end of the aggregated representation vector of the semantic query response of the blood image data grid block, and then input the generated information into an AIGC-based optimized blood image generator along with the blood image data to generate the preprocessed blood image data. The generation prompt information is "optimize the blood image data based on the input corpus." Specifically, the blood image text description semantic feature vector is used to generate and process the blood image text description semantic features obtained by performing a semantic adaptive query on each local region in the sequence of blood image data grid block feature vectors, thereby achieving intelligent optimization of the blood image data. This allows the capture of complex and deep-seated interactions and correlations between cells, thereby providing more accurate diagnostic evidence. Furthermore, the processing strategy can be dynamically adjusted based on the specific description content of the blood image to better adapt to diverse samples, rather than relying solely on fixed preprocessing parameters.

[0053] In one specific embodiment, a variational autoencoder (VAE) serves as the core model for an AIGC-based optimized blood image generator. The VAE's processing is as follows: First, preprocessed blood image data is combined with semantic features from blood image text descriptions obtained from efficient semantic scanning queries. A generation prompt, such as "Based on the input corpus, optimize the blood image data," is appended to the end of the semantic feature vector. This step provides additional guidance to the VAE, allowing it to better understand the desired optimization direction. Next, the prepared data is fed into the VAE-based optimized blood image generator. The VAE consists of an encoder and a decoder. The encoder maps the input blood image into a low-dimensional latent space, where each point represents a possible representation of the input image. To ensure reasonable diversity in the generated images, the VAE introduces a KL divergence loss term, ensuring that the latent variables follow a standard normal distribution. This allows for random sampling from the latent space during the decoding phase, generating new images that still conform to the original data distribution. During training, the encoder learns how to transform the input image into coordinates in the latent space, while the decoder learns to reconstruct an image from these coordinates that is as close to the original input as possible. For the blood image optimization task, this means that the decoder must not only recover the basic structure of the image but also adjust details based on additional semantic information (such as the focus on areas with concentrated red blood cells), such as reducing noise and enhancing contrast. Once the VAE model is trained, it can be used for actual image optimization. During inference, given a preprocessed blood image and its corresponding semantic description, the encoder transforms it into one or more points in the latent space. The decoder then uses these latent points to generate an optimized blood image. Because the VAE has learned to capture the key features of the input image and mutate them appropriately during training, the generated image not only retains important details from the original image but also exhibits greater clarity and smoothness.

[0054] In a specific example, consider a blood sample image containing a large number of RBC aggregates, but the image suffers from noise interference and insufficient contrast. Using the VAE-based method described above, the system specifically focuses on these RBC aggregates, automatically adjusting parameters during the generation process to reduce noise and enhance contrast. Ultimately, the system outputs an optimized, high-quality blood image in which the RBC aggregates are more clearly visible and background noise is effectively suppressed. This optimized image helps improve the accuracy of subsequent blood cell classification and counting, thereby supporting more precise medical diagnoses.

[0055] Preferably, considering that the sequence of the blood image text description semantic feature vector and the blood image data grid block feature vector respectively represent the key semantic text semantic features of the blood image and the grid-based local-global correlation image semantic features of the blood image data, when performing semantic query encoding based on a fast scanning mechanism, the insufficient correspondence of the scanning mechanism caused by the differences in the semantic representation of different modal data will lead to the lack of query encoding feature instance determination of the blood image data grid block semantic query response aggregate representation vector, thereby affecting the data quality of the preprocessed blood image data obtained by the optimized blood image generator based on AIGC.

[0056] Therefore, after adding generation prompt information to the tail of the blood image data grid block semantic query response aggregate representation vector, and then inputting it together with the blood image data into the AIGC-based optimized blood image generator to obtain pre-processed blood image data, the blood image data grid block semantic query response aggregate representation vector is first optimized, specifically including the following steps: Calculate the blood image data grid block semantic query response aggregation representation vector, for example, represented as Each pair of eigenvalues and The square root of the product and the quotient between them are used to obtain the first blood image data grid block semantic query response aggregate topological association matrix and the second blood image data grid block semantic query response aggregate topological association matrix ,Right now ; ; in, A semantic query response aggregate representation vector representing the blood image data grid block, and Respectively represent each pair of feature values of the blood image data grid block semantic query response aggregation representation vector, that is, the first and The eigenvalues at the positions, Representing the aggregated topological correlation matrix of the semantic query response of the first blood image data grid block The value of the position, Representing the second blood image data grid block semantic query response aggregate topological association matrix The value of the position.

[0057] Aggregate the semantic query response of the first blood image data grid block into a topological correlation matrix Aggregate topological association matrix with semantic query response of the second blood image data grid block Perform matrix multiplication on the transpose of to obtain the first blood image data grid block semantic query response aggregation distribution level matrix , represents matrix multiplication, Represents the transpose of a matrix; Aggregate the topological correlation matrix of the second blood image data grid block semantic query response Aggregate topological association matrix with semantic query response of the first blood image data grid block Perform matrix multiplication on the transpose of to obtain the second blood image data grid block semantic query response aggregation distribution level matrix ; Aggregate the semantic query response of the first blood image data grid block into a distribution hierarchy matrix Aggregate distribution hierarchical matrix of semantic query response with the second blood image data grid block Perform weighted points to obtain blood image data grid block semantic query response aggregation cross compensation matrix ,in, and represents different weight hyperparameters, represents dot product, Indicates point addition; Aggregate the blood image data grid block semantic query response cross compensation matrix Aggregate representation vectors for semantic query responses with the blood image data grid blocks Matrix multiplication is performed to obtain the optimized blood image data grid block semantic query response aggregate representation vector.

[0058] That is, by using the nonlinear feature encoding representation of the differential topological association matrix of the blood image data grid block semantic query response aggregate representation vector, the self-correlation complete similarity instantiation of the blood image data grid block semantic query response aggregate representation vector is subjected to a bidirectional target conversion representation of a distributed hierarchical architecture, and the cross-validation noise is compensated by dynamic fusion weight adjustment to improve the probability generation optimization judgment degree of the blood image data grid block semantic query response aggregate representation vector under the discriminant feature significance evaluation, that is, the effectiveness degree of feature significance as an instance for the probability generation optimization judgment, thereby improving the data quality of the pre-processed blood image data obtained by the optimized blood image grid block semantic query response aggregate representation vector through the AIGC-based optimized blood image generator.

[0059] For example, in the blood cell segmentation module 30, image segmentation techniques are used to separate blood cells from the background in the preprocessed blood image data to obtain a blood cell image. It should be understood that, given that blood cells are often mixed with the background and other non-target elements in preprocessed blood images, directly extracting features or performing classification from such images can introduce significant noise, reducing the quality of diagnostic results. Effective segmentation can accurately extract blood cells of interest from the complex background, allowing subsequent feature extraction, classification, and counting operations to better focus on the actual target cells, thereby improving the accuracy and efficiency of the entire analysis process.

[0060] In one embodiment, threshold segmentation can be used to distinguish blood cells from the background based on differences in grayscale or color intensity. Specifically, by performing a histogram analysis on the image, an appropriate threshold is selected, and pixels above the threshold are considered to be foreground (i.e., blood cells), while pixels below the threshold are considered to be background. This method is suitable for situations where there is a clear difference between blood cells and the background, such as when blood cells appear dark and the background is lighter. However, the selection of the threshold often relies on empirical values and may not be robust enough for complex samples (such as cells that overlap or background noise is large). To overcome these problems, adaptive threshold segmentation technology can be used to dynamically adjust the threshold based on the statistical characteristics of the local area. In addition, morphological operations can be combined to remove isolated small objects and smooth segmentation boundaries to improve the quality of the segmentation results.

[0061] In another embodiment, the U-Net architecture can be used. U-Net is a convolutional neural network (CNN) specifically designed for medical image segmentation. It can directly predict the probability distribution of whether each pixel belongs to a blood cell or background from the original image. The U-Net encoder captures the high-level features of the input image, while the decoder gradually restores the spatial resolution, ultimately generating a segmentation mask of the same size as the input image. This end-to-end learning approach not only avoids the problem of manually setting thresholds but also automatically learns the optimal segmentation boundaries, achieving good results even in the presence of irregular cell shapes and sizes and high background noise. The advantage of U-Net is that it can be trained with a large amount of labeled data, allowing it to learn to recognize different types of blood cells (such as red blood cells, white blood cells, and platelets). The trained model can be directly applied to newly acquired blood images, outputting high-quality segmentation results in real time. Furthermore, U-Net has excellent generalization capabilities, maintaining high segmentation accuracy across different blood sample types. Therefore, it has become one of the preferred methods for blood cell segmentation tasks.

[0062] For example, the blood cell feature extraction module 40 extracts blood cell features from the blood cell image. It should be understood that when extracting features from a blood cell image, the unique morphological characteristics of different blood cell types (such as red blood cells, white blood cells, and platelets) must be taken into account. For example, red blood cells are typically biconcave discs, while white blood cells have diverse shapes and complex internal structures. To distinguish these differences, the feature extraction module uses various shape descriptors, such as circularity and major-minor axis ratio, to quantify the geometric properties of the cells. This not only helps identify specific blood cell types but also detects the presence of abnormal cells, which is particularly important for disease diagnosis. Furthermore, in addition to shape characteristics, blood cell size is also an important basis for classification. The feature extraction module can describe cell size by measuring parameters such as area, perimeter, or distance from the center of mass. This information can help distinguish blood cells at different developmental stages. In some cases, changes in cell size can be an indicator of a pathological condition. Therefore, the extraction of size features is equally important. Furthermore, the color characteristics of blood cells should not be underestimated. Due to the staining of blood samples, different types of blood cells will appear different colors under a microscope. The feature extraction module extracts values from the RGB color channels and calculates metrics such as color distribution and consistency to describe the color properties of cells. This approach not only enhances the ability to identify cell types but also helps determine whether cells are healthy or diseased.

[0063] For example, in the blood cell classification module 50, the blood cell features are classified and counted to obtain classification and counting results. It should be understood that by classifying blood cell features (such as shape, size, and color), different blood cell types, such as red blood cells, white blood cells, and platelets, can be distinguished and their numbers counted. The results of this classification and counting are crucial for assessing patient health, monitoring disease progression, and adjusting treatment plans. For example, an abnormally increased white blood cell count may indicate infection or inflammation, while a decreased red blood cell count may be a sign of anemia. In one embodiment, a support vector machine (SVM) is used as a classifier to distinguish different blood cell types. A multi-class classifier is trained with the various features mentioned above as input and outputs specific blood cell class labels (red blood cells, white blood cells, platelets, etc.). The trained model can then perform classification predictions on newly input unlabeled blood cell images. Once training is complete, new unlabeled blood cell images can be fed into the trained model for classification predictions. For each new image, the model outputs a probability distribution for each pixel belonging to a specific blood cell type, thereby generating a complete classification mask. The number of each type of blood cell can then be counted using simple connected component analysis or other counting algorithms. For example, morphological operations can be used to remove isolated small objects and smooth segmentation boundaries, followed by connected component analysis to ensure that the final blood cell image is complete and free of redundant background information.

[0064] Illustratively, in the result generation module 60, a blood cell report or chart is generated based on the classification and counting results. In particular, the blood cell characteristics include the shape, size, and color characteristics of the blood cells. It should be understood that the results of blood cell analysis usually contain a large number of numerical values and technical terms, which are difficult for non-professionals to interpret directly. Therefore, by converting these data into structured reports or visual charts, it can make it easier for doctors to understand the specific changes in patients' blood samples, assisting them in making more accurate diagnoses and treatment recommendations. For example, a detailed blood cell report can list the number and proportion of different types of blood cells (such as red blood cells, white blood cells, and platelets), helping doctors identify potential health problems such as anemia, infection, or other blood diseases. In addition, clear chart displays allow doctors to see data trends at a glance, which is particularly useful for monitoring the progression of the disease.

[0065] In summary, according to the embodiment of the present application, a blood cell analysis system based on computer vision is illustrated, which uses computer vision-based image recognition and generation technology to perform grid division and semantic feature extraction on the blood image data, and then performs global text generation and semantic encoding on the grid block features of each blood image data after semantic extraction, so as to adjust the attention of different blood local areas in each blood image data grid block feature according to the semantic features of the blood image text description, and respond to aggregation, thereby realizing intelligent optimization of the blood image data. Through the method of the present application, the complex and deep mutual influence and correlation relationship between cells can be captured, thereby providing a more accurate basis for diagnosis. At the same time, the processing strategy can be dynamically adjusted according to the specific description content of the blood image to better adapt to diverse samples, rather than just relying on fixed preprocessing parameters.

[0066] Figure 6 Schematic flow chart of the blood cell analysis method based on computer vision according to an embodiment of the present application. Figure 6 As shown, the blood cell analysis method based on computer vision includes: S1, using an image acquisition device to acquire an image of a blood sample to obtain blood image data; S2, preprocessing the blood image data based on text semantic description to obtain preprocessed blood image data; S3, separating blood cells and background in the preprocessed blood image data by image segmentation technology to obtain a blood cell image; S4, extracting blood cell features from the blood cell image; S5, classifying and counting the blood cell features to obtain classification and counting results; S6, generating a blood cell report or chart based on the classification and counting results.

[0067] In one embodiment, the blood image data is preprocessed based on text semantic description to obtain preprocessed blood image data, including: grid division and semantic feature extraction of the blood image data to obtain a sequence of blood image data grid block features; text generation and semantic encoding based on the global blood image on the sequence of blood image data grid block features to obtain blood image text description semantic features; efficient semantic scanning query on the blood image text description semantic features and the sequence of blood image data grid block features to obtain a blood image data grid block semantic query response aggregate representation, and obtaining the preprocessed blood image data based on the blood image data grid block semantic query response aggregate representation.

[0068] In one embodiment, the blood image data is gridded and semantic features are extracted to obtain a sequence of blood image data grid block features, including: using a grid divider to grid divide the blood image data to obtain a sequence of blood image data grid blocks; passing the sequence of blood image data grid blocks through a blood image grid block semantic feature extractor based on a CNN-BiLSTM-SE network to obtain a sequence of blood image data grid block feature vectors as the sequence of blood image data grid block features.

[0069] Here, those skilled in the art will appreciate that the specific operations of each step in the above-mentioned blood cell analysis method based on computer vision have been described in detail above. Figures 1 to 5 The computer vision-based blood cell analysis system has been described in detail, and therefore, its repeated description will be omitted.

[0070] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0071] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A blood cell analysis system based on computer vision, characterized in that: include: A blood image acquisition module, configured to acquire an image of a blood sample using an image acquisition device to obtain blood image data; a blood image preprocessing module, configured to perform preprocessing on the blood image data based on text semantic description to obtain preprocessed blood image data; a blood cell segmentation module, configured to separate the blood cells and the background in the pre-processed blood image data by using an image segmentation technique to obtain a blood cell image; a blood cell feature extraction module, configured to extract blood cell features from the blood cell image; a blood cell classification module, configured to classify and count the blood cell characteristics to obtain classification and counting results; A result generation module is configured to generate a blood cell report or chart based on the classification and counting results; wherein the blood image preprocessing module includes: a blood image segmentation and semantic analysis unit, configured to perform grid segmentation and semantic feature extraction on the blood image data to obtain a sequence of grid block features of the blood image data; A blood image global semantic representation unit, configured to perform text generation and semantic encoding on the sequence of grid block features of the blood image data based on the global blood image to obtain semantic features of the blood image text description; A blood image text scanning query unit is used to perform efficient semantic scanning query on a sequence of semantic features of the blood image text description and features of the blood image data grid block to obtain an aggregated representation of semantic query responses of the blood image data grid block, and obtain the preprocessed blood image data based on the aggregated representation of semantic query responses of the blood image data grid block.

2. The computer vision-based blood cell analysis system according to claim 1, characterized in that: The blood image segmentation semantic analysis unit includes: a blood image grid division subunit, configured to grid the blood image data using a grid divider to obtain a sequence of blood image data grid blocks; The grid block semantic feature extraction subunit is used to pass the sequence of blood image data grid blocks through a blood image grid block semantic feature extractor based on a CNN-BiLSTM-SE network to obtain a sequence of blood image data grid block feature vectors as the sequence of blood image data grid block features.

3. The computer vision-based blood cell analysis system according to claim 2, characterized in that: The grid block semantic feature extraction subunit is used to: Aggregating the sequence of blood image data grid blocks along the channel dimension and inputting them into SEBlock and CNN models respectively to obtain a blood image data grid weight feature map and a blood image data grid aggregation feature map; Multiplying each feature matrix along the channel dimension in the blood image data grid weight feature map by position with each feature matrix along the channel dimension in the blood image data grid aggregate feature map to obtain a blood image data grid aggregate weighted feature map; Performing feature expansion along the channel dimension on the blood image data grid aggregated weighted feature map to obtain a sequence of blood image data grid feature matrices; After flattening the sequence of blood image data grid feature matrices to obtain a sequence of blood image data grid flattened feature vectors, the sequence of blood image data grid flattened feature vectors is sequentially input into a BiLSTM model and a fully connected layer to obtain a sequence of blood image data grid block feature vectors.

4. The computer vision-based blood cell analysis system according to claim 3, characterized in that: The blood image global semantic representation unit includes: a blood image data cascading subunit, configured to cascade the sequence of blood image data grid block feature vectors to obtain a global semantic representation vector of the blood image data; a blood image text description generating subunit, configured to pass the global semantic representation vector of the blood image data through a blood image text generator based on a large language model to obtain a key semantic text description of the blood image; The blood image text description semantic encoding subunit is used to perform semantic encoding on the key semantic text description of the blood image to obtain a blood image text description semantic feature vector as the blood image text description semantic feature.

5. The computer vision-based blood cell analysis system according to claim 4, characterized in that: The blood image text scanning query unit includes: a blood image text efficient semantic scanning query subunit, configured to perform an efficient semantic scanning query on a sequence of the blood image text description semantic feature vector and the blood image data grid block feature vector to obtain a blood image data grid block semantic query response aggregate representation vector as the blood image data grid block semantic query response aggregate representation; The pre-processed blood image data generating subunit is configured to obtain the pre-processed blood image data by aggregating representation vectors according to the semantic query responses of the blood image data grid blocks.

6. The computer vision-based blood cell analysis system according to claim 5, characterized in that: The blood image text efficient semantic scanning query subunit includes: a blood image data fast matching secondary subunit, configured to determine a fast matching subsequence of blood image data grid block feature vectors from the sequence of blood image data grid block feature vectors based on matching information between the blood image text description semantic feature vector and the sequence of blood image data grid block feature vectors; The blood image data query response aggregation secondary sub-unit is used to perform a linear transformation on the blood image text description semantic feature vector to obtain a blood image text description query vector and a blood image text description value vector, and use a subsequence of the blood image data grid block feature vector as a subsequence of the key vector to perform Transformer-based query response aggregation on the blood image text description query vector, the blood image text description value vector and the subsequence of the key vector to obtain the blood image data grid block semantic query response aggregation representation vector.

7. The computer vision-based blood cell analysis system according to claim 6, characterized in that: The blood image data fast matching secondary subunit is used to: calculating the mutual information between the blood image text description semantic feature vector and each blood image data grid block feature vector in the sequence of blood image data grid block feature vectors to obtain a sequence of blood image text-grid fast semantic matching factors as the matching information; identifying a first blood image text-grid fast semantic maximum value and a second blood image text-grid fast semantic maximum value from the sequence of the blood image text-grid fast semantic matching factors; Based on the positions of the first blood image text-grid fast semantic maximum and the second blood image text-grid fast semantic maximum in the sequence of blood image text-grid fast semantic matching factors, a subsequence of the blood image data grid block feature vectors that are fast matched is determined from the sequence of the blood image data grid block feature vectors.

8. The computer vision-based blood cell analysis system according to claim 7, characterized in that: The preprocessed blood image data generating subunit is used to: add generation prompt information to the end of the blood image data grid block semantic query response aggregate representation vector, and then input it together with the blood image data into the AIGC-based optimized blood image generator to obtain the preprocessed blood image data.

9. A blood cell analysis method based on computer vision, characterized in that: include: Using an image acquisition device to acquire an image of the blood sample to obtain blood image data; Preprocessing the blood image data based on text semantic description to obtain preprocessed blood image data; separating blood cells and background in the preprocessed blood image data using image segmentation technology to obtain a blood cell image; and extracting blood cell features from the blood cell image; Classifying and counting the blood cell features to obtain classification and counting results; generating a blood cell report or chart based on the classification and counting results; wherein preprocessing the blood image data based on text semantic description to obtain preprocessed blood image data includes: Performing grid division and semantic feature extraction on the blood image data to obtain a sequence of grid block features of the blood image data; Performing text generation and semantic encoding based on the global blood image on the sequence of grid block features of the blood image data to obtain semantic features of blood image text description; An efficient semantic scanning query is performed on a sequence of the semantic features of the blood image text description and the features of the blood image data grid block to obtain an aggregated representation of the semantic query response of the blood image data grid block, and the preprocessed blood image data is obtained based on the aggregated representation of the semantic query response of the blood image data grid block.

10. The computer vision-based blood cell analysis method according to claim 9, characterized in that: Performing grid division and semantic feature extraction on the blood image data to obtain a sequence of grid block features of the blood image data, comprising: gridding the blood image data using a gridder to obtain a sequence of blood image data grid blocks; The sequence of blood image data grid blocks is passed through a blood image grid block semantic feature extractor based on a CNN-BiLSTM-SE network to obtain a sequence of blood image data grid block feature vectors as the sequence of blood image data grid block features.

Images