Blood cell image detection method and system based on convolutional neural network

Abstract

The invention relates to the field of data processing, in particular to a blood cell image detection method and system based on a convolutional neural network. The method comprises the following steps: acquiring a blood cell image to be processed; extracting multi-size image features corresponding to each local image area in the blood cell image, and compressing the multi-size image features to obtain corresponding multi-size blood cell local feature information; performing cell type prediction on the local feature information of the multi-size blood cells, and identifying corresponding cell type prediction information; generating a cell space distribution network corresponding to each local image area based on the cell type prediction information, and combining to obtain a corresponding global cell space distribution network; and based on the cell type prediction information and the global cell space distribution network, generating visual prediction information corresponding to the blood cell image. Through cell type prediction and construction of a global cell space distribution network, the accuracy, reliability and efficiency of blood cell detection are improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a blood cell image detection method and system based on a convolutional neural network. Background Art

[0002] Blood cell detection is one of the most commonly used and important detection items in clinical diagnosis. By mainly analyzing indicators such as the morphology, quantity, and proportion of various cells such as red blood cells, white blood cells, and platelets in the blood, doctors can diagnose, monitor the treatment, and evaluate the prognosis of various diseases such as anemia, infection, and leukemia.

[0003] Currently, laboratory technicians observe blood smears under a microscope to classify and count cells. It relies on the experience and professional level of laboratory technicians, is highly subjective, has low efficiency, and long-term observation easily leads to fatigue and errors. In related technologies, principles such as impedance method and optical method are also used to count and preliminarily classify blood cells to quickly give some basic parameters, but the ability to identify subtle changes in cell morphology and abnormal cells is limited, the detection accuracy is not high, and the workload of manual recheck is relatively large.

[0004] In related technologies, different collection devices, lighting conditions, staining methods, etc. will cause differences in the clarity, contrast, color, etc. of blood cell images, bringing difficulties to subsequent analysis and processing. In addition, during the blood smear process, the cell distribution in the blood sample may be uneven, with some areas having overly dense cells and some areas having sparse cells, affecting the accuracy of cell recognition and counting.

[0005] Therefore, there is an urgent need for a brand-new solution to improve the detection efficiency and assist in improving the accuracy of cell recognition and counting. Summary of the Invention

[0006] The present invention aims at the technical problems existing in the prior art and provides a blood cell image detection method and system based on a convolutional neural network for improving the accuracy, reliability, and efficiency of blood cell detection.

[0007] In a first aspect, an embodiment of the present application provides a blood cell image detection method based on a convolutional neural network, including: Obtaining a blood cell image to be processed; the blood cell image is from a blood image acquisition device; Extracting multi-size image features corresponding to each local image region in the blood cell image, and compressing the multi-size image features to obtain multi-size blood cell local feature information corresponding to each local image region; Performing cell type prediction on the multi-size blood cell local feature information to identify cell type prediction information corresponding to each local image region; Generate a cell spatial distribution network corresponding to each local image region based on the cell type prediction information, so as to obtain a global cell spatial distribution network corresponding to the blood cell image by combination; Generate visualization prediction information corresponding to the blood cell image based on the cell type prediction information and the global cell spatial distribution network.

[0008] In a second aspect, an embodiment of the present application provides a blood cell image detection system based on a convolutional neural network. The system includes the following units: An acquisition unit, configured to acquire a blood cell image to be processed; the blood cell image is from a blood image acquisition device; An extraction unit, configured to extract multi-scale image features corresponding to each local image region in the blood cell image, and compress the multi-scale image features to obtain multi-scale blood cell local feature information corresponding to each local image region; A prediction unit, configured to perform cell type prediction on the multi-scale blood cell local feature information to identify cell type prediction information corresponding to each local image region; A generation unit, configured to generate a cell spatial distribution network corresponding to each local image region based on the cell type prediction information, so as to obtain a global cell spatial distribution network corresponding to the blood cell image by combination; A reporting unit, configured to generate visualization prediction information corresponding to the blood cell image based on the cell type prediction information and the global cell spatial distribution network.

[0009] In a third aspect, an embodiment of the present application provides an electronic device, and the electronic device includes: At least one processor, a memory, and an input-output unit; Wherein, the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the blood cell image detection method based on a convolutional neural network in the first aspect.

[0010] In a fourth aspect, a computer-readable storage medium is provided, which includes instructions. When the instructions are run on a computer, the computer is made to execute the blood cell image detection method based on a convolutional neural network in the first aspect.

[0011] The beneficial effects of the present invention are as follows: A method and system for detecting blood cell images based on a convolutional neural network are provided. In this technical solution, a blood cell image to be processed is acquired; the blood cell image is from a blood image acquisition device. Furthermore, multi-size image features corresponding to each local image region in the blood cell image are extracted, and the multi-size image features are compressed to obtain multi-size blood cell local feature information corresponding to each local image region. Then, cell type prediction is performed on the multi-size blood cell local feature information to identify cell type prediction information corresponding to each local image region. Next, a cell spatial distribution network corresponding to each local image region is generated based on the cell type prediction information, and the global cell spatial distribution network corresponding to the blood cell image is obtained by combination. Finally, visualization prediction information corresponding to the blood cell image is generated based on the cell type prediction information and the global cell spatial distribution network. In the embodiments of the present application, through cell type prediction and the construction of the global cell spatial distribution network, the technical problems existing in the related art can be effectively overcome, thereby improving the accuracy, reliability, and efficiency of blood cell detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a flowchart of a method for detecting blood cell images based on a convolutional neural network according to an embodiment of the present application; Figure 2 is a structural schematic diagram of a system for detecting blood cell images based on a convolutional neural network according to an embodiment of the present application; Figure 3 is a structural schematic diagram of an electronic device according to an embodiment of the present application; Figure 4 is a structural schematic diagram of a medium device according to an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0013] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present application.

[0014] In the description of the present application, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of the described features. In the description of the present application, "a plurality" means two or more, unless otherwise specifically defined.

[0015] In the description of the present application, the term "for example" is used to mean "serving as an example, illustration, or explanation". Any embodiment described as "for example" in the present application is not necessarily construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to implement and use the present invention. In the following description, details are set forth for purposes of explanation. It should be understood that those of ordinary skill in the art can recognize that the present invention can be implemented without the use of these specific details. In other instances, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but rather to be in line with the broadest scope consistent with the principles and features disclosed in the present application.

[0016] An embodiment of the present application provides a method and system for detecting blood cell images based on a convolutional neural network. In this technical solution, a blood cell image to be processed is obtained; the blood cell image is from a blood image acquisition device; multi-size image features corresponding to each local image region in the blood cell image are extracted, and the multi-size image features are compressed to obtain multi-size blood cell local feature information corresponding to each local image region; cell type prediction is performed on the multi-size blood cell local feature information to identify cell type prediction information corresponding to each local image region; a cell spatial distribution network corresponding to each local image region is generated based on the cell type prediction information to obtain a global cell spatial distribution network corresponding to the blood cell image in combination; and visualization prediction information corresponding to the blood cell image is generated based on the cell type prediction information and the global cell spatial distribution network.

[0017] In one aspect of the embodiments of the present application, a cell spatial distribution network is generated based on cell type prediction information, which can not only determine the cell type, but also understand the distribution position and mutual relationship of cells in the image. This is very important for analyzing the aggregation, dispersion, etc. of cells in a blood sample, helps to discover abnormal cell distributions related to diseases, provides more comprehensive information for disease diagnosis, and further improves the accuracy of detection. On the other hand, through targeted model training and optimization of the local feature information of blood cells of multiple sizes, the model can quickly give accurate results when predicting cell types. At the same time, when generating the cell spatial distribution network and visualizing the prediction information, efficient data processing algorithms and preset templates are used to quickly complete data integration and report generation, meeting the needs of rapid clinical detection. In addition, the finally generated visual prediction information is based on cell type prediction information and the global cell spatial distribution network, covering multiple aspects of information such as cell type, quantity, and distribution position, which helps to assist doctors in comprehensively understanding the blood sample situation through a single report, providing great convenience for clinical diagnosis. At the same time, the cell spatial distribution network can reveal the mutual relationship and distribution pattern between cells, and this information may contain some potential disease clues. For example, certain diseases may cause specific types of cells to aggregate in specific areas. By analyzing the cell spatial distribution network, it helps to assist doctors in discovering these hidden disease characteristics early, providing assistance for the early diagnosis and treatment of diseases. In summary, the technical solution of the present application can effectively overcome the technical problems existing in the related art through cell type prediction and the construction of the global cell spatial distribution network, thereby improving the accuracy, reliability, and efficiency of blood cell detection.

[0018] The blood cell image detection solution based on a convolutional neural network provided by the embodiments of the present application can also be executed by an electronic device, and the electronic device can be a server, a server cluster, or a cloud server. The electronic device can also be a terminal device such as a mobile phone, a computer, a tablet computer, a wearable device, or a dedicated device (such as a dedicated terminal device with a blood cell image detection system based on a convolutional neural network, etc.). The above-mentioned chips introduced in the above embodiments can also be installed in these electronic devices. Or, these electronic devices can also install a service program for executing the blood cell image detection solution based on a convolutional neural network.

[0019] Figure 1 Schematic diagram of a blood cell image detection method based on a convolutional neural network provided by the embodiments of the present application, as Figure 1 shown, the method includes the following steps: 101. Obtain a blood cell image to be processed; 102. Extract the multi - scale image features corresponding to each local image region in the blood cell image, and compress the multi - scale image features to obtain the multi - scale blood cell local feature information corresponding to each local image region; 103. Perform cell type prediction on the multi - scale blood cell local feature information to identify the cell type prediction information corresponding to each local image region; 104. Generate a cell spatial distribution network corresponding to each local image region based on the cell type prediction information, and combine them to obtain the global cell spatial distribution network corresponding to the blood cell image; 105. Generate the visualization prediction information corresponding to the blood cell image based on the cell type prediction information and the global cell spatial distribution network.

[0020] In the embodiments of the present application, the blood cell image refers to a visual image obtained by photographing or scanning a blood sample using a blood image acquisition device, such as a microscope paired with an image acquisition device, a specific blood cell analyzer, etc. In these images, the morphology, size, color, texture and other characteristics of various cells in the blood can be presented. For example, red blood cells usually appear as biconcave discs without a nucleus and are generally red or dark red in the image; white blood cells are relatively large in volume, with diverse nuclear morphologies, such as lobed, round, etc., and the cytoplasm will show different colors after staining; platelets are small in volume and irregular in shape.

[0021] Blood cell images are crucial for clinical diagnosis. Doctors and researchers can assist in diagnosing various diseases such as anemia, infection, leukemia, etc. by analyzing information such as the quantity, proportion, morphological changes and spatial distribution of various cells in blood cell images, and can also be used for disease treatment monitoring and prognosis assessment. In the above claims, the blood cell image, as the basic data, has achieved efficient and accurate cell detection and classification through a series of steps such as pre - processing, feature extraction, model construction and training, providing strong support for clinical diagnosis.

[0022] The blood cell image comes from a blood image acquisition device. For example, an optical microscope paired with an image acquisition device is the most basic and commonly used device combination. The optical microscope uses optical principles to magnify the blood sample, allowing the examiner to directly observe the morphology of blood cells. By connecting with devices such as a camera and an image acquisition card, the image observed under the microscope can be converted into a digital signal, and then collected and stored as a blood cell image. Its advantages are relatively simple operation, low cost, and the ability to clearly present the basic morphology of cells, such as the biconcave disc shape of red blood cells and the nuclear morphology of white blood cells, and it is widely used in clinical examinations.

[0023] For example, in blood cell detection, a fluorescence microscope can be used to observe specific cell components or markers after fluorescence labeling. By exciting the fluorescent substance to emit fluorescence of a specific wavelength, the target cell structure or component presents a bright fluorescent image against a dark background, enabling more clear resolution of the fine structure and specific components of the cell, which is of great significance for studying the physiological functions and pathological changes of cells, especially suitable for the detection and analysis of some special cells or intracellular substances.

[0024] For example, a flow cytometer is a relatively advanced blood cell analysis device that can perform rapid and accurate multi-parameter analysis on flowing single cells or biological particles. When collecting blood cell images, the flow cytometer makes the cells pass through the detection area one by one, irradiates the cells with a laser, and obtains various information of the cells based on the scattering and fluorescence signals of the cells to the laser, and can convert this information into cell images. It can not only obtain the morphological information of the cells, but also measure multiple physical and chemical characteristics of the cells simultaneously, such as cell size, internal structure, surface marker expression, etc., to achieve the classification and quantitative analysis of blood cells, with a fast detection speed and high accuracy.

[0025] As an alternative embodiment, in 102, extracting the multi-scale image features corresponding to each local image region in the blood cell image can be implemented as follows: Perform adaptive normalization processing on the blood cell image according to the type of the blood image acquisition device to obtain a first preprocessed blood cell image; perform multi-edge filtering and noise reduction on the first preprocessed blood cell image to obtain a second preprocessed blood cell image; perform region segmentation on the second preprocessed blood cell image to segment out each local image region; perform adaptive enhancement processing on each local image region, and extract the corresponding multi-scale image features from each enhanced local image region.

[0026] In this alternative embodiment, the process of extracting multi-scale image features from a blood cell image involves multiple closely connected image processing steps, each step having its unique role and jointly serving for accurate extraction of image features. The implementation methods of each step will be introduced in detail below: Due to different types of blood image acquisition devices, such as optical microscopes paired with image acquisition devices, fluorescence microscopes, flow cytometers, etc., their imaging principles and parameters are different, which will lead to differences in the brightness, contrast, color, etc. of the collected blood cell images. Therefore, adaptive normalization processing needs to be carried out according to the device type. For example, for images collected by an optical microscope, specific normalization parameters are set according to its common brightness range and noise characteristics; for images converted from data collected by a flow cytometer, corresponding normalization methods are adopted based on its data characteristics and output format. For instance, the image pixel values are mapped to the [0, 1] interval to eliminate the influence of device differences on the image, obtaining the first preprocessed blood cell image, which provides a unified standard data basis for subsequent processing.

[0027] After obtaining the first preprocessed blood cell image, a bilateral filtering algorithm is used for noise reduction. Bilateral filtering not only considers the spatial distance of pixels but also the similarity of pixel values, and can better retain the edges and detailed information of cells while removing noise. According to the noise situation and cell detail characteristics of the image, the standard deviation of the spatial Gaussian kernel and the standard deviation of the range Gaussian kernel are dynamically adjusted. For example, for an image with more noise and rich cell details, the standard deviation of the range Gaussian kernel is appropriately increased to better retain the details; for an image with less noise, the standard deviation of the spatial Gaussian kernel can be appropriately reduced to improve the filtering efficiency. After bilateral filtering for noise reduction, the second preprocessed blood cell image is obtained, reducing the interference of noise on subsequent processing.

[0028] Furthermore, region segmentation is performed on the second preprocessed blood cell image, aiming to separate different cell regions and the background in the image, and each local image region is segmented. Methods such as threshold segmentation, edge detection, and clustering analysis can be used. For example, using the threshold segmentation method, according to the gray-scale characteristics of the blood cell image, an appropriate threshold is set to divide the image into cell regions and background regions; or an edge detection algorithm such as the Canny algorithm is used to detect the edges of cells, thereby determining the contours of cells and achieving region segmentation. Each segmented local image region provides a basis for subsequent feature extraction for each region.

[0029] Finally, perform adaptive enhancement processing on each of the segmented local image regions. According to the characteristics of each local image region, such as contrast, brightness, texture, etc., different enhancement methods are adopted. For example, for local regions with low contrast, histogram equalization or adaptive histogram equalization (CLAHE) techniques are used to enhance their contrast; for regions with uneven brightness, local brightness adjustment is performed. After the enhancement processing, corresponding multi-scale image features are extracted from each local image region. Convolution operations are performed on the local image regions using convolution kernels of different sizes, such as 3×3, 5×5, 7×7, etc., to obtain image features at different scales. Small-sized convolution kernels focus on local fine features, and large-sized convolution kernels obtain more macroscopic structural information, thereby comprehensively capturing features such as the morphology and texture of cells, providing rich data support for subsequent cell type prediction and analysis.

[0030] Further, in 102, multi-lateral filtering noise reduction is performed on the first preprocessed blood cell image to obtain a second preprocessed blood cell image, which can be implemented as follows: obtaining the similarity between each pixel value in the first preprocessed blood cell image and other surrounding pixel values; based on the similarity, performing multi-lateral filtering processing on each pixel value in the first preprocessed blood cell image to obtain an initial filtered image; identifying the nuclear morphology type in the initial filtered image; pre-entering each nuclear morphology type of blood cells in a pre-constructed expert knowledge base; based on the nuclear morphology type, performing optimized filling processing on each pixel value in the initial filtered image to strengthen the contours of each nuclear morphology in the initial filtered image, and obtaining the second preprocessed blood cell image.

[0031] In this image noise reduction and optimization processing flow, each step is closely connected and plays a key role in improving the quality of blood cell images. Specific examples and technical effects of each step will be elaborated in detail in combination with the characteristics of actual blood cell images.

[0032] Exemplarily, assume there is a white blood cell image collected by an optical microscope. In the first preprocessed blood cell image, a pixel point at the edge of a red blood cell is selected. By using methods such as calculating the Euclidean distance, the similarity between this pixel point and the pixel values within a certain surrounding neighborhood is obtained. For example, with this pixel as the center, the other 8 pixels within a 3×3 neighborhood are taken, and the Euclidean distances between them and the central pixel in the RGB color space are calculated. The smaller the distance, the higher the similarity. Furthermore, based on the similarity obtained in the previous step, the image is subjected to bilateral filtering. In the above-mentioned white blood cell image, for each pixel, different weights are assigned according to its similarity with the neighboring pixels. Pixels with high similarity have larger weights and have a greater impact on the current pixel value during the filtering calculation. After bilateral filtering, the noise points (with low similarity to the surrounding pixels) in the image are effectively suppressed, while the features such as the edges and internal textures of the cells (regions with high similarity) are well preserved, obtaining an initial filtered image. Then, in the obtained initial filtered image, for the nucleus of the white blood cell, through morphological analysis algorithms such as contour detection and geometric feature calculation, its morphological type is identified. For example, for a lobed nucleus, by detecting features such as the concavity and convexity of its contour and the number of lobes, it is determined that it belongs to a segmented neutrophil. Finally, assume that various nuclear morphological types and their corresponding standard contour features have been entered into the expert knowledge base. For the identified segmented neutrophil, in the initial filtered image, according to the standard contour of this nuclear morphology in the knowledge base, the pixel values inside and at the edge of the nucleus are optimized and filled. For example, for the blurred part of the nucleus edge, referring to the standard contour, the pixel values are adjusted to make its contour clearer and more complete, thereby obtaining the second preprocessed blood cell image.

[0033] In this way, by obtaining pixel similarity and performing bilateral filtering, the noise in the image can be accurately removed while retaining the key features of the cells, providing a clean image basis for subsequent cell analysis and reducing the interference of noise on cell feature extraction and recognition. Identifying the nuclear morphological type helps doctors and researchers quickly understand the type of cells and possible existing pathological conditions, because different nuclear morphologies are often related to different blood diseases. The optimization and filling process based on the expert knowledge base can strengthen the contour of the nuclear morphology, making the morphological features of the cells more prominent, facilitating subsequent cell classification and counting, and improving the accuracy and reliability of detection.

[0034] It can be understood that by using the bilateral filtering algorithm, while considering the pixel spatial distance, the filtering operation is carried out based on the similarity of pixel values, effectively removing image noise and at the same time maximizing the retention of the edges and detailed features of the cells, avoiding the destruction of cell morphological information. The nuclei of white blood cells in the image have diverse morphologies, such as lobed and round, and their detailed features are crucial for classification.

[0035] Further optionally, in the foregoing steps, according to the image noise intensity distribution and the requirement for retaining cell detail features, dynamically adjust the standard deviation of the spatial Gaussian kernel and the standard deviation of the range Gaussian kernel to ensure that while reducing noise, the key morphological information of cells is retained to the greatest extent, providing a basis for subsequent accurate feature extraction and classification, and thus improving the detection accuracy. For example, to analyze the image noise intensity distribution, it can be evaluated by calculating the noise power spectrum of the image. For the white blood cell region, since the nuclear morphology is diverse and the detail features are crucial for classification, adjust the standard deviation according to the requirement for retaining cell detail features. If it is found that the noise is mainly concentrated in the high-frequency part and the edge details of the white blood cell nucleus are rich, appropriately increase it, for example, increase it to 0.2 to better retain the details. If the noise is evenly distributed, adjust it appropriately, for example, adjust it to 2.5 to more effectively remove the noise.

[0036] In this way, through the above-mentioned bilateral filtering algorithm, the noise in the blood cell image can be effectively removed, making the image clearer, providing a clean data basis for subsequent feature extraction and classification, and avoiding misclassification caused by noise interference. Filtering based on considering the pixel spatial distance and pixel value similarity can retain the edge and detail features of cells to the greatest extent. For the diverse nuclear morphologies of white blood cells, such as lobed and round shapes, the key morphological information such as the contour and internal texture is well preserved, providing a guarantee for accurately identifying the types of white blood cells. Dynamically adjusting the standard deviation of the spatial Gaussian kernel and the standard deviation of the range Gaussian kernel enables the filtering process to be optimized according to the actual situation of the image and the cell feature requirements, retaining the key morphological information of cells to the greatest extent, thus laying a solid foundation for subsequent accurate feature extraction and classification and significantly improving the accuracy of blood cell detection.

[0037] Further, in 102, the adaptive enhancement processing of each local image region can be realized as follows: obtain the resolution, cell distribution density, density, and image texture complexity of each local image region; identify the nuclear morphology types in each local image region; use the adaptive image enhancement branch corresponding to the nuclear morphology type to determine the regions to be enhanced in each local image region; the adaptive image enhancement branch is trained using the historical image data corresponding to the nuclear morphology type; according to the resolution, cell distribution density, density, and image texture complexity of each local image region, perform histogram equalization processing on the regions to be enhanced in each local image region to enhance the visual features of cells in each local image region.

[0038] Next, the implementation method and technical effects of adaptive enhancement processing will be elaborated in detail around the blood cell image detection method. Specifically, for each local image region obtained by segmentation, information such as its resolution, cell distribution density, density degree, and image texture complexity is first obtained through image analysis algorithms. For example, the resolution is obtained using the image size parameter. The cell distribution density is determined by counting the number of cells per unit area. The density degree is judged according to the degree of cell aggregation or dispersion. The texture analysis algorithm, such as the gray-level co-occurrence matrix, is used to calculate the image texture complexity. At the same time, through the morphological recognition algorithm, such as contour detection combined with geometric feature analysis, the nuclear morphology type in the local image region is recognized, and it is judged whether it is circular, lobed or other morphologies.

[0039] Furthermore, according to the recognized nuclear morphology type, the corresponding branch is selected from multiple pre-trained adaptive image enhancement branches. These adaptive image enhancement branches are trained using the historical image data corresponding to different nuclear morphology types. For example, for the image data of circular nuclei, during training, the features of this type of nuclear image and the appropriate enhancement method are focused on learning, and an adaptive image enhancement branch for circular nuclei is trained.

[0040] Then, the selected adaptive image enhancement branch analyzes the local image region to determine the region to be enhanced. For example, for some nuclear morphologies, the edge part may be the key enhancement region, and the branch uses specific algorithms, such as methods based on edge detection and feature extraction, to identify these regions to be enhanced.

[0041] Finally, according to the information such as the resolution, cell distribution density, density degree, and image texture complexity of the local image region obtained previously, histogram equalization processing is performed on the region to be enhanced. If the cell distribution is dense and the texture is complex, when performing histogram equalization, the parameters will be adjusted to enhance the contrast while avoiding detail loss caused by over-enhancement. For regions with low resolution, the equalization range will be appropriately adjusted to achieve the best enhancement effect and improve the visual features of the cells.

[0042] In this way, by identifying the nuclear morphological types and adopting the corresponding adaptive image enhancement branches, targeted enhancement can be performed on different types of cell images, avoiding the inadaptability of the unified enhancement method to different cell features and better highlighting the features of various cells. Histogram equalization processing is carried out by combining multiple features of the image, which can effectively enhance the visual features of the cells, make the contours, textures, etc. of the cells clearer, facilitate subsequent feature extraction and classification, and improve the detection accuracy. The adaptive image enhancement branches are trained using the historical image data corresponding to different nuclear morphological types, fully mining the information in the historical data, making the enhancement process more scientific and effective, and being able to continuously optimize the enhancement effect. As the historical data accumulates and updates, the performance of the adaptive image enhancement branches can also be continuously improved.

[0043] It can be understood that by using the adaptive histogram equalization (CLAHE) technology, the size of the image patches and the overlapping regions are adaptively adjusted according to the resolution of the image, the cell distribution density, the degree of density and sparseness, and the texture complexity of the image. Histogram equalization is performed on each patch separately to enhance the image contrast, highlight the cell morphological features, and improve the visual effect and feature recognizability of the image. Platelets are small and irregular in shape in the image. Through targeted enhancement processing, these cell features are more easily recognized by the model, thereby improving the detection accuracy.

[0044] That is to say, in the adaptive histogram equalization (CLAHE) processing, the size of the patches and the overlapping regions are adaptively adjusted according to the resolution of the image, the degree of density and sparseness of the cell distribution, and the texture complexity of the image to achieve the optimal contrast enhancement effect for images with different characteristics, make the cell features clearer, facilitate accurate recognition by the model, and improve the detection accuracy. For example, in the adaptive histogram equalization (CLAHE) technology, the size of the patches and the overlapping regions are adaptively adjusted according to the resolution of the image, the degree of density and sparseness of the cell distribution, and the texture complexity of the image. For example, when the resolution of the red blood cell image is high, the distribution is uniform, and the texture is simple, the patch is set to 32×32 pixels, and the overlapping region is 8 pixels. When the white blood cells are densely distributed and the texture is complex, the patch is set to 16×16 pixels, and the overlapping region is 12 pixels. When the platelets are sparsely distributed, the patch is set to 8×8 pixels, and the overlapping region is 6 pixels to achieve the optimal contrast enhancement effect for images with different characteristics, make the cell features clearer, facilitate accurate recognition by the model, and improve the detection accuracy.

[0045] For example, in blood cell image detection, the Contrast Limited Adaptive Histogram Equalization (CLAHE) technique plays a crucial role. Taking the red blood cell image as an example, when the image resolution is high, the cell distribution is relatively uniform, and the texture is relatively simple, the image patches can be set to a larger size, such as 32×32 pixels, and the overlapping area can be set to 8 pixels. Since the red blood cells have regular shapes and uniform distributions, larger patches can perform histogram equalization more efficiently, enhancing the overall contrast and highlighting the biconcave disc shape characteristics of red blood cells.

[0046] For white blood cell images, their cell nuclei have diverse shapes and high texture complexity. If white blood cells are densely distributed, to better capture the detailed features of the cell nuclei, the patch size can be reduced to 16×16 pixels, and at the same time, the overlapping area can be increased to 12 pixels. In this way, when performing histogram equalization, each local area can be processed more precisely, avoiding the loss of detailed information of the cell nuclei due to large-scale processing, making the morphological features of the white blood cell nuclei, such as lobed and round shapes, more distinct, and facilitating accurate identification by the model.

[0047] For platelet images, due to their small size and irregular shapes, in the case of sparse cell distribution, the patch size is set to 8×8 pixels, and the overlapping area is set to 6 pixels. Through this setting of small-sized patches and larger overlapping areas, the local area where platelets are located is enhanced specifically, highlighting their irregular morphological features, improving the recognition rate of platelets in the image, making it easier for the model to identify, and thus improving the accuracy of blood cell detection.

[0048] As an optional embodiment, in 103, cell type prediction is performed on the multi-size local feature information of blood cells to obtain cell type prediction information corresponding to each local image area, including at least the following steps: Perform nuclear morphological feature recognition on the multi-size local feature information of blood cells to obtain initial nuclear morphological prediction information corresponding to each local image area; according to the initial nuclear morphological prediction information, dynamically configure the cell type prediction branch parameters in the multi-layer cell type prediction network, and configure the cell type prediction branch in the multi-layer cell type prediction network based on the branch parameters; input the multi-size local feature information of blood cells into the configured cell type prediction branch to obtain multiple local classification information of blood cells; construct the multiple local classification information of blood cells into a multi-layer virtual cell distribution state; the multi-layer virtual cell distribution state at least includes: the local area cell distribution state predicted by each cell type prediction branch; perform multi-dimensional fusion processing on the multi-layer virtual cell distribution state to obtain cell type prediction information corresponding to each local image area.

[0049] It can be understood that the local feature information of blood cells of multiple sizes is analyzed using image recognition algorithms, such as feature extraction methods based on convolutional neural networks. By training the model to learn the characteristics of different cell nucleus morphologies, such as round, lobed, etc., the initial predicted information of the cell nucleus morphology corresponding to each local image region is obtained. For example, the local feature information of blood cells of multiple sizes is input into a pre-trained cell nucleus morphology recognition model, and the model outputs the predicted results of the cell nucleus morphology in each local region. Furthermore, based on the initial predicted information of the cell nucleus morphology, the possible range of cell types is judged. For example, if the predicted cell nucleus is lobed, it may be a neutrophil, etc. Then, according to different cell types, the cell type prediction branch parameters in the multi-layer cell type prediction network are dynamically configured. For example, targeted adjustments are made to the feature weight allocation, number of network layers, and number of neurons for different types of cells. If a certain type of cell is more dependent on a specific feature dimension in terms of morphology and characteristics, the weight of that dimension feature in the prediction branch is correspondingly increased. Next, the local feature information of blood cells of multiple sizes is input into the configured cell type prediction branch to obtain multiple local classification information of blood cells. These classification information include the possible types of cells in each local region. Then, these local classification information are constructed into a multi-layer virtual cell distribution state, with each layer corresponding to the local region cell distribution state predicted by a cell type prediction branch. For example, the first layer represents the cell distribution predicted based on a certain feature dimension, and the second layer represents the cell distribution predicted based on another feature dimension. Finally, multi-dimensional fusion processing is performed on the multi-layer virtual cell distribution state. A weighted fusion method can be adopted, and different weights are assigned according to the reliability and importance of each cell type prediction branch. For example, a higher weight is given to a prediction branch that has been verified by a large amount of data and has a high accuracy. Through fusion, information from multiple dimensions is comprehensively considered to obtain the cell type prediction information corresponding to each local image region.

[0050] In this way, by identifying the cell nucleus morphology characteristics, the cell types can be initially screened and judged, providing a basis for subsequent predictions. Dynamically configuring the cell type prediction branch parameters makes the network more suitable for the characteristics of different cell types and improves the prediction accuracy. Multi-dimensional fusion processing comprehensively considers information from multiple dimensions, avoids the limitations of single-dimensional information, and further improves the accuracy of cell type prediction. Dynamically configuring parameters according to the initial predicted information of the cell nucleus morphology enables the model to adapt to the characteristics of different types of cells, enhances the adaptability of the model to complex and diverse blood cell images, and can better process blood cell images with various different characteristics. Constructing a multi-layer virtual cell distribution state can display the distribution of cells in the local region from multiple perspectives, providing more comprehensive and rich information for subsequent analysis and diagnosis, and helping doctors to more accurately understand the state of cells in the blood sample.

[0051] Exemplarily, in 103, the SENet (Squeeze-and-Excitation Networks) neural network is adopted. In the feature extraction layer, first, the convolutional layer is used to perform preliminary feature extraction on the input image, and then the Squeeze-and-Excitation module is introduced. Through the Squeeze operation, this module compresses each feature map into a single value using global average pooling to obtain the global information of the feature channels; then through the Excitation operation, after passing through two fully connected layers, according to the number of channels of the feature map, the spatial dimension of the image, and the expected abstraction level of the features to be extracted, the number of neurons is dynamically set, the importance weights of the feature channels are learned and weight coefficients are generated. Finally, the weight coefficients are multiplied by the original feature map to recalibrate the features and enhance the expression of key features. In the classification layer, after extracting features through multiple SENet modules, a fully connected layer is connected, and the class probabilities are calculated through the Softmax function to achieve accurate judgment of cell types. For example, for the unique biconcave disc shape feature of red blood cells and the complex nuclear morphology feature of white blood cells, by strengthening the expression of key features, the ability of the model to distinguish different types of blood cells is improved, thereby enhancing the detection accuracy.

[0052] In another example, DenseNet (Dense Connected Convolutional Network) is adopted. In the model, each layer is directly connected to all subsequent layers, enabling efficient transfer and reuse of features in the network; during the feature extraction process, different levels of features are continuously fused through dense connection blocks, reducing the problem of gradient vanishing; the classification layer also completes the determination of cell types through a fully connected layer and the Softmax function.

[0053] Further optionally, in the DenseNet, the number of layers of the dense connection block and the number of convolutional kernels in each layer are set according to the morphological diversity of cell images, the computational resource limitations of the model, and the expected depth of feature extraction, ensuring that the model can fully learn the key features of blood cells under limited computational resources, enhancing the model's ability to recognize various cells, and improving the detection accuracy.

[0054] Or, in other examples, EfficientNet is adopted. Through the compound scaling method, the depth, width, and resolution of the network are jointly optimized; when constructing the model, according to the hardware computational resources and the expected detection accuracy, the network parameters are reasonably adjusted, while ensuring a low computational cost, enhancing the model's ability to extract and classify the features of blood cell images.

[0055] Further optionally, in the construction of EfficientNet, the coefficients of compound scaling are dynamically adjusted according to the memory capacity of the hardware resources, the performance of the computing cores, and the actual requirements for model accuracy and inference speed, so as to achieve the optimal matching of model performance and computing resources, and improve the detection accuracy of the model for blood cell images while ensuring the computing efficiency.

[0056] Alternatively, in other examples, Vision Transformer (ViT) can also be adopted. The blood cell image is divided into multiple small patches and regarded as a sequence input. The multi-head self-attention mechanism is used to model the global information between the image patches to capture the long-range dependencies between cells. In the classification stage, the classification head combines the positional encoding information to complete the prediction of cell categories.

[0057] Further optionally, in the Vision Transformer (ViT), the number of heads of the multi-head self-attention mechanism and the dimension of the positional encoding are set according to the size of the image, the spatial structure of cell distribution, and the richness of cell features, so as to effectively capture the global information in the image and the complex relationships between cells, improve the model's understanding and classification ability for complex cell images, and enhance the detection accuracy.

[0058] Further optionally, in the model training step, the number of training epochs and the batch size are dynamically optimized and adjusted according to the scale of the dataset, the diversity of samples, the complexity of the model, and the real-time load of the hardware computing resources, so as to ensure the stability and convergence of the model during training, enable the model to fully learn the data features, and improve the detection accuracy.

[0059] Further optionally, during the model training process, the model is periodically evaluated using the validation set, and the model parameters and training strategies are adjusted according to the evaluation results to ensure that the accuracy of the model on the validation set continues to improve until the preset accuracy threshold is reached, so as to ensure the accuracy of the model in actual detection.

[0060] Further optionally, in the model testing and application step, the detection results are manually spot-checked and verified, and the verification results are fed back to the model training link to optimize the model in a targeted manner to further improve the detection accuracy.

[0061] As an alternative embodiment, in 104, based on the cell type prediction information, a cell spatial distribution network corresponding to each local image region is generated to combine and obtain the global cell spatial distribution network corresponding to the blood cell image, which at least includes the following steps: Based on the cell type prediction information, obtain the predicted cell count, predicted cell type, and predicted cell position within each local image region; based on the predicted cell count, predicted cell type, and predicted cell position within each local image region, construct an initial cell model for each local image region; perform spatial mapping on the initial cell model according to the predicted cell position to obtain a three-dimensional cell spatial image corresponding to each local image region; dynamically adjust the relative position relationship between the initial cell models in the three-dimensional cell spatial image to obtain a cell spatial distribution network corresponding to each local image region; based on the adjacency relationship between the initial cell models, perform spatial fusion on the cell spatial distribution network to obtain the global cell spatial distribution network.

[0062] It can be understood that through the analysis of the cell type prediction information, the predicted cell count, predicted cell type, and predicted cell position within each local image region are parsed using algorithms. For example, the predicted cell count is obtained by counting the number of markers of different cell types in each local region; according to the probability distribution output by the cell type prediction branch, the predicted cell type value for each cell is determined; the predicted cell position is extracted from the image coordinate information. Furthermore, based on the obtained predicted cell count, predicted cell type, and predicted cell position, an initial cell model is constructed for each local image region. Simple geometric shapes can be used to represent cells, such as using spheres to represent red blood cells and irregular polygons to represent white blood cells, etc. Each model contains attribute information such as the type, position, and number of cells. Then, the constructed initial cell model is spatially mapped according to the predicted cell position to generate a three-dimensional cell spatial image corresponding to each local image region. In this process, the cell position information in the two-dimensional image is extended to three-dimensional space, providing a basis for subsequent analysis of the spatial relationship between cells. For example, the two-dimensional coordinates (x, y) of the cell are extended to three-dimensional coordinates (x, y, z), where z can be assigned according to the level of the image or other relevant information. In the three-dimensional cell spatial image, considering the interaction between cells and actual biological characteristics, the relative position relationship between the initial cell models is dynamically adjusted. For example, according to the attraction or repulsion force model between cells, the distance and angle between cells are adjusted to make the cell spatial distribution more in line with the actual situation, thereby obtaining a cell spatial distribution network corresponding to each local image region. Finally, based on the adjacency relationship between the initial cell models, spatial fusion is performed on the cell spatial distribution network corresponding to each local image region. By splicing and integrating the cell spatial distribution networks of adjacent local regions and considering the continuity and consistency of cells at the boundary, the global cell spatial distribution network corresponding to the blood cell image is finally obtained.

[0063] In this way, by constructing a global network of cell spatial distribution, the spatial distribution of blood cells in the image can be comprehensively displayed, including not only the types and quantities of cells, but also the relative positional relationships between cells, providing richer information for doctors and researchers. The spatial distribution of cells is closely related to certain diseases. For example, the blood cell distribution in leukemia patients may show abnormal aggregation or dispersion. By analyzing the global network of cell spatial distribution, these abnormalities can be detected, providing a strong basis for the diagnosis and treatment of diseases. Considering the relative positional relationships and adjacency relationships between cells can further verify and optimize the cell type prediction results, improving the accuracy of blood cell detection. For example, if the predicted cell type does not match the spatial relationship with surrounding cells, the predicted cell type can be re-evaluated. The generated three-dimensional cell spatial image and the network of cell spatial distribution can be visually displayed, enabling doctors and researchers to more intuitively observe the distribution of blood cells and helping to more deeply understand the characteristics of blood samples and the pathogenesis of diseases.

[0064] As an optional embodiment, in 105, based on the cell type prediction information and the global network of cell spatial distribution, generating the visualization prediction information corresponding to the blood cell image includes at least the following steps: Generating a three-dimensional visualization model corresponding to the global network of cell spatial distribution; generating multi-dimensional annotation information in the three-dimensional visualization model based on the cell type prediction information; and loading the multi-dimensional annotation information into the three-dimensional visualization model as the visualization interaction description information of the three-dimensional visualization model.

[0065] In the embodiment of the present application, the multi-dimensional annotation information at least includes: cell type prediction information, cell distribution prediction information, and cell state prediction information.

[0066] Exemplarily, using three-dimensional modeling technology, information such as the positions, quantities, and mutual relationships of cells in the global cell spatial distribution network is transformed into a three-dimensional model. For example, professional three-dimensional modeling software such as Blender, 3ds Max, or open-source three-dimensional visualization libraries such as Three.js are used. Based on the three-dimensional coordinate information of the cells, geometric shapes representing the cells are constructed. For example, spheres are used to represent red blood cells, and irregular polyhedrons are used to represent white blood cells. These geometric shapes are arranged in three-dimensional space according to the cell distribution to generate a preliminary three-dimensional visualization model. The predicted type of each cell, such as red blood cells, white blood cells, platelets, etc., is obtained from the cell type prediction results and organized into an easily displayable format, such as text labels. According to the global cell spatial distribution network, information such as the distribution density and aggregation regions of cells in the image is extracted and presented in the form of charts or text descriptions. For example, a cell distribution heatmap is generated, where the darker the color, the denser the cell distribution. If the model predicts the state of the cells (such as whether they are diseased, high or low activity, etc.), this information is presented in the form of specific symbols or text annotations. For example, red borders are used to represent diseased cells, and green represents normal cells. Furthermore, the generated multi-dimensional annotation information is associated with the three-dimensional visualization model. In the three-dimensional modeling software or visualization library, the annotation information is loaded onto the corresponding cells or regions by setting attributes or adding plugins. For example, in Three.js, custom attributes can be added to each geometry representing a cell, and annotation information such as cell type, distribution, and state is stored in these attributes. When the model is rendered, the corresponding annotation information is displayed according to the user's interaction operations (such as mouse hovering).

[0067] In this way, the three-dimensional visualization model can intuitively present the spatial distribution of blood cells. Combining multi-dimensional annotation information, doctors and researchers can clearly understand the cell type, distribution, and state at a glance, avoiding the difficulty of interpreting information from complex data. Through the visual interactive explanatory information, it helps doctors obtain key information more quickly and accurately, assisting in disease diagnosis and treatment decision-making.

[0068] For example, by observing the abnormal regions of cell distribution and the positions of diseased cells, the development stage and severity of the disease can be judged. For researchers, the visual prediction information helps to deeply analyze the characteristics and mutual relationships of blood cells, discover potential laws and abnormalities, and provide strong support for further medical research. When communicating between medical teams or with patients, the visual information is easier to understand and communicate, helping to improve communication efficiency and reduce misunderstandings.

[0069] In an embodiment of the present application, a blood cell image to be processed is obtained; the blood cell image is from a blood image acquisition device. Furthermore, multi-size image features corresponding to each local image region in the blood cell image are extracted, and the multi-size image features are compressed to obtain multi-size blood cell local feature information corresponding to each local image region. Then, cell type prediction is performed on the multi-size blood cell local feature information to identify cell type prediction information corresponding to each local image region. Next, a cell spatial distribution network corresponding to each local image region is generated based on the cell type prediction information, and the global cell spatial distribution network corresponding to the blood cell image is obtained by combination. Finally, visualization prediction information corresponding to the blood cell image is generated based on the cell type prediction information and the global cell spatial distribution network. In the embodiment of the present application, through cell type prediction and the construction of the global cell spatial distribution network, the technical problems existing in the related art can be effectively overcome, thereby improving the accuracy, reliability, and efficiency of blood cell detection.

[0070] Figure 2 FIG. is a schematic structural diagram of a blood cell image detection system based on a convolutional neural network provided by an embodiment of the present application, as Figure 2 shown, the system includes the following steps: An acquisition unit, configured to acquire a blood cell image to be processed; the blood cell image is from a blood image acquisition device; An extraction unit, configured to extract multi-size image features corresponding to each local image region in the blood cell image, and compress the multi-size image features to obtain multi-size blood cell local feature information corresponding to each local image region; A prediction unit, configured to perform cell type prediction on the multi-size blood cell local feature information to identify cell type prediction information corresponding to each local image region; A generation unit, configured to generate a cell spatial distribution network corresponding to each local image region based on the cell type prediction information, and obtain the global cell spatial distribution network corresponding to the blood cell image by combination; A reporting unit, configured to generate visualization prediction information corresponding to the blood cell image based on the cell type prediction information and the global cell spatial distribution network.

[0071] Further optionally, the extraction unit extracts multi-size image features corresponding to each local image region in the blood cell image, and is configured to perform adaptive normalization processing on the blood cell image according to the type of the blood image acquisition device to obtain a first preprocessed blood cell image; Perform bilateral filtering denoising on the first preprocessed blood cell image to obtain a second preprocessed blood cell image; Perform region segmentation on the second preprocessed blood cell image to obtain each local image region; Perform adaptive enhancement processing on each local image region, and extract corresponding multi-scale image features from each enhanced local image region.

[0072] Further optionally, the extraction unit performs multi-edge filtering and noise reduction on the first preprocessed blood cell image to obtain a second preprocessed blood cell image, specifically for: Obtain the similarity between each pixel value in the first preprocessed blood cell image and other surrounding pixel values; Based on the similarity, perform multi-edge filtering on each pixel value in the first preprocessed blood cell image to obtain an initial filtered image; Identify the nuclear morphology type in the initial filtered image; each nuclear morphology type of blood cells is pre-entered in a pre-constructed expert knowledge base; Based on the nuclear morphology type, perform optimized filling on each pixel value in the initial filtered image to strengthen the contours of each nuclear morphology in the initial filtered image, and obtain the second preprocessed blood cell image.

[0073] Further optionally, the extraction unit performs adaptive enhancement processing on each local image region, specifically for: Obtain the resolution, cell distribution density, density, and image texture complexity of each local image region; Identify the nuclear morphology type in each local image region; Use the adaptive image enhancement branch corresponding to the nuclear morphology type to determine the regions to be enhanced in each local image region; the adaptive image enhancement branch is trained using historical image data corresponding to the nuclear morphology type; According to the resolution, cell distribution density, density, and image texture complexity of each local image region, perform histogram equalization on the regions to be enhanced in each local image region to enhance the cell visual features in each local image region.

[0074] Further optionally, the prediction unit performs cell type prediction on the multi-scale blood cell local feature information to identify the cell type prediction information corresponding to each local image region, for: Perform nuclear morphology feature recognition on the multi-scale blood cell local feature information to obtain the initial nuclear morphology prediction information corresponding to each local image region; According to the initial prediction information of the cell nucleus morphology, dynamically configure the cell type prediction branch parameters in the multi-layer cell type prediction network, and configure the cell type prediction branch in the multi-layer cell type prediction network based on the branch parameters; Input the multi-size blood cell local feature information into the configured cell type prediction branch to obtain multiple blood cell local classification information; Construct the multiple blood cell local classification information into a multi-layer virtual cell distribution state; the multi-layer virtual cell distribution state at least includes: the local area cell distribution state predicted by each cell type prediction branch; Perform multi-dimensional fusion processing on the multi-layer virtual cell distribution state to obtain the cell type prediction information corresponding to each local image area.

[0075] Further optionally, a generation unit generates a cell spatial distribution network corresponding to each local image area based on the cell type prediction information, and combines them to obtain the global cell spatial distribution network corresponding to the blood cell image, specifically for: Based on the cell type prediction information, obtain the cell number prediction value, cell type prediction value, and cell position prediction value in each local image area; Based on the cell number prediction value, cell type prediction value, and cell position prediction value in each local image area, construct an initial cell model in each local image area; Perform spatial mapping on the initial cell model according to the cell position prediction value to obtain a three-dimensional cell spatial image corresponding to each local image area; Dynamically adjust the relative position relationship between the initial cell models in the three-dimensional cell spatial image to obtain a cell spatial distribution network corresponding to each local image area; Based on the adjacency relationship between the initial cell models, perform spatial fusion on the cell spatial distribution network to obtain the global cell spatial distribution network.

[0076] Further optionally, a reporting unit generates visual prediction information corresponding to the blood cell image based on the cell type prediction information and the global cell spatial distribution network, specifically for: Generate a three-dimensional visualization model corresponding to the global cell spatial distribution network; Based on the cell type prediction information, generate multi-dimensional annotation information in the three-dimensional visualization model; the multi-dimensional annotation information at least includes: cell type prediction information, cell distribution prediction information, and cell state prediction information; Load the multi-dimensional annotation information into the three-dimensional visualization model as the visual interaction description information of the three-dimensional visualization model.

[0077] Please refer to Figure 3 , Figure 3 , which is a schematic diagram of an embodiment of an electronic device provided by an embodiment of the present application. As Figure 3 shown, an embodiment of the present application provides an electronic device 500, including a memory 510, a processor 520, and a computer program 511 stored on the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 511, the foregoing embodiments are implemented.

[0078] Please refer to Figure 4 , Figure 4 , which is a schematic diagram of an embodiment of a computer-readable storage medium provided by an embodiment of the present application. As Figure 4 shown, this embodiment provides a computer-readable storage medium 600, on which a computer program 611 is stored. When the computer program 611 is executed by a processor, the foregoing embodiments are implemented.

[0079] It should be noted that in the above embodiments, the descriptions of the various embodiments have their own emphases. For parts not described in detail in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0080] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one or more flows or multiple flows and / or blocks Figure 1 one or more blocks or multiple blocks.

[0082] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction means that implements the functions specified in one or more of the processes and / or blocks Figure 1 in one or more of the processes and / or blocks Figure 1 specified in the one or more of the processes and / or blocks.

[0083] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the processes and / or blocks Figure 1 in one or more of the processes and / or blocks Figure 1 specified in the one or more of the processes and / or blocks.

[0084] Although the preferred embodiments of the present invention have been described, additional changes and modifications can be made by those skilled in the art once they learn of the basic inventive concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications that fall within the scope of the present invention.

[0085] It is obvious that those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A blood cell image detection method based on convolutional neural network, characterized in that: The method at least comprises: Acquire a blood cell image to be processed; the blood cell image comes from a blood image acquisition device; Extracting multi-size image features corresponding to each local image region in the blood cell image, and compressing the multi-size image features to obtain multi-size blood cell local feature information corresponding to each local image region; Performing cell type prediction on the multi-size blood cell local feature information to identify and obtain cell type prediction information corresponding to each local image region; Generate a cell space distribution network corresponding to each local image region based on the cell type prediction information, so as to combine and obtain a global cell space distribution network corresponding to the blood cell image; Based on the cell type prediction information and the global cell spatial distribution network, visualization prediction information corresponding to the blood cell image is generated.

2. The blood cell image detection method based on convolutional neural network according to claim 1, characterized in that: The extracting of multi-size image features corresponding to each local image region in the blood cell image includes: According to the type of the blood image acquisition device, the blood cell image is adaptively normalized to obtain a first blood cell preprocessed image; performing multilateral filtering and noise reduction on the first blood cell preprocessed image to obtain a second blood cell preprocessed image; Performing region segmentation on the second blood cell preprocessed image to obtain various local image regions; Adaptive enhancement processing is performed on each local image region, and corresponding multi-scale image features are extracted from each local image region after the enhancement processing.

3. The blood cell image detection method based on convolutional neural network according to claim 2 is characterized in that: The performing multilateral filtering and noise reduction on the first blood cell preprocessed image to obtain a second blood cell preprocessed image includes: Acquire the similarity between each pixel value in the first blood cell preprocessed image and other surrounding pixel values; Based on the similarity, performing a multilateral filtering process on each pixel value in the first blood cell preprocessed image to obtain an initial filtered image; Identifying the cell nuclear morphology type in the initial filtered image; pre-entering each cell nuclear morphology type of blood cells in a pre-built expert knowledge base; Based on the cell nuclear morphology type, each pixel value in the initial filtered image is optimized and filled to strengthen the contour of each cell nuclear morphology in the initial filtered image to obtain the second blood cell preprocessing image.

4. The blood cell image detection method based on convolutional neural network according to claim 2, characterized in that: The adaptive enhancement processing of each local image area includes: Obtain the resolution, cell distribution density, density, and image texture complexity of each local image area; Identify the cell nuclear morphology type in each local image region; Adopting the adaptive image enhancement branch corresponding to the cell nucleus morphology type to determine the area to be enhanced in each local image area; the adaptive image enhancement branch is trained using the historical image data corresponding to the cell nucleus morphology type; According to the resolution, cell distribution density, density and image texture complexity of each local image area, the area to be enhanced in each local image area is subjected to histogram equalization processing to enhance the cell visual characteristics in each local image area.

5. The blood cell image detection method based on convolutional neural network according to claim 1, characterized in that: The performing cell type prediction on the multi-size blood cell local feature information to identify and obtain cell type prediction information corresponding to each local image region includes: Performing cell nucleus morphology feature recognition on the multi-size blood cell local feature information to obtain initial prediction information of cell nucleus morphology corresponding to each local image region; According to the initial prediction information of the cell nuclear morphology, dynamically configuring the cell type prediction branch parameters in the multi-layer cell type prediction network, and configuring the cell type prediction branches in the multi-layer cell type prediction network based on the branch parameters; Inputting the multi-size blood cell local feature information into the configured cell type prediction branch to obtain a plurality of blood cell local classification information; The plurality of local classification information of blood cells are constructed into a multi-layer virtual cell distribution state; the multi-layer virtual cell distribution state at least includes: a local area cell distribution state predicted by each cell type prediction branch; The multi-layer virtual cell distribution states are subjected to multi-dimensional fusion processing to obtain cell type prediction information corresponding to each local image region.

6. The blood cell image detection method based on convolutional neural network according to claim 1, characterized in that: The generating of the cell space distribution network corresponding to each local image region based on the cell type prediction information to obtain the global cell space distribution network corresponding to the blood cell image by combination includes: Based on the cell type prediction information, obtaining a cell quantity prediction value, a cell type prediction value, and a cell position prediction value in each local image area; Based on the predicted values ​​of cell number, cell type and cell position in each local image area, an initial cell model in each local image area is constructed; The initial cell model is spatially mapped according to the cell position prediction value to obtain a three-dimensional cell space image corresponding to each local image area; Dynamically adjusting the relative position relationship between each initial cell model in the three-dimensional cell space image to obtain a cell space distribution network corresponding to each local image area; Based on the adjacency relationship between each initial cell model, the cell spatial distribution network is spatially fused to obtain the global cell spatial distribution network.

7. The blood cell image detection method based on convolutional neural network according to claim 1, characterized in that: The generating of visualization prediction information corresponding to the blood cell image based on the cell type prediction information and the global cell spatial distribution network includes: Generating a three-dimensional visualization model corresponding to the global cell spatial distribution network; Based on the cell type prediction information, generating multidimensional annotation information in the three-dimensional visualization model; the multidimensional annotation information at least includes: cell type prediction information, cell distribution prediction information, and cell state prediction information; The multi-dimensional annotation information is loaded into the three-dimensional visualization model as visualization interaction description information of the three-dimensional visualization model.

8. A blood cell image detection system based on convolutional neural network, characterized in that: The system comprises at least the following units: An acquisition unit, used for acquiring a blood cell image to be processed; The blood cell image comes from a blood image acquisition device; An extraction unit, used to extract multi-size image features corresponding to each local image region in the blood cell image, and compress the multi-size image features to obtain multi-size blood cell local feature information corresponding to each local image region; A prediction unit, configured to perform cell type prediction on the multi-size blood cell local feature information to identify and obtain cell type prediction information corresponding to each local image region; A generating unit, configured to generate a cell space distribution network corresponding to each local image region based on the cell type prediction information, so as to combine and obtain a global cell space distribution network corresponding to the blood cell image; A reporting unit is used to generate visual prediction information corresponding to the blood cell image based on the cell type prediction information and the global cell spatial distribution network.

9. An electronic device, characterized in that: including a memory for storing a computer software program; A processor is used to read and execute the computer software program, thereby realizing the functions of each component part in the blood cell image detection system based on convolutional neural network as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that: The storage medium stores a computer software program, which, when executed by a processor, implements the functions of each component part in the blood cell image detection system based on a convolutional neural network as described in any one of claims 1 to 7.

Images