Deep learning-based blood smear abnormal cell intelligent identification system and method thereof

Abstract

The invention relates to the technical field of medical information systems, in particular to a blood smear abnormal cell intelligent recognition system and method based on deep learning. The multi-stage convolutional neural network module is in data connection with the image acquisition module, and the abnormal cell bank module is in data connection with the multi-stage convolutional neural network module; the knowledge graph module is in data connection with the abnormal cell bank module, and the blood index calculation module is in data connection with the multistage convolutional neural network module and the abnormal cell bank module and is used for calculating various blood indexes; generating a graphical display result of the blood indexes; the result output module is in data connection with the knowledge graph module and the blood index calculation module; according to the blood disease diagnosis method and system, the whole process from image acquisition to diagnosis report generation can be automatically completed, detailed analysis results and diagnosis suggestions can be provided for doctors, and the efficiency and accuracy of blood disease diagnosis are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information systems, and in particular to a deep learning-based intelligent recognition system and method for abnormal cells in blood smears. Background Art

[0002] In recent years, with the continuous advancement of medical technology, the diagnosis and treatment of blood diseases have made significant progress. In the field of hematopathology, blood smear examination has always been one of the important means of diagnosing blood diseases. Traditional blood smear analysis mainly relies on the experience and judgment of professional physicians. Although this method is reliable, it has problems such as low efficiency, strong subjectivity, and inconsistent results. With the development of artificial intelligence technology, especially the breakthrough of deep learning algorithms in the field of image recognition, the application of AI technology to blood smear analysis has become a hot research direction.

[0003] At present, some blood smear analysis systems based on machine learning have been proposed and applied. These systems usually use a single convolutional neural network model to analyze blood smear images, which can identify some common abnormal cells to a certain extent. However, these systems still have some obvious limitations. First, the recognition ability of a single model is limited, and it is difficult to cope with complex and diverse abnormal cell types. Secondly, these systems often only focus on the classification of cells and ignore the importance of cell localization, which makes it difficult to provide accurate abnormal cell location information in practical applications. In addition, most of the existing systems lack effective integration with medical knowledge and cannot provide doctors with comprehensive diagnostic advice.

[0004] On the other hand, some studies have tried to introduce knowledge graphs to assist in the diagnosis of blood diseases, but these methods often separate image analysis and knowledge reasoning and fail to achieve an organic integration of the two. At the same time, the existing system also faces challenges in balancing computational efficiency and accuracy when processing a large number of blood smear images. In addition, patient privacy protection is often overlooked in existing systems, which is particularly important today as medical data security is increasingly valued.

[0005] In view of the above problems, there is an urgent need for an intelligent system that can comprehensively, accurately and efficiently identify abnormal cells in blood smears. The system should be able to accurately classify and locate cells at the same time, have the ability to handle complex and diverse abnormal cell types, and effectively integrate medical knowledge to provide reliable assistance for clinical diagnosis. At the same time, the system should also have good scalability and privacy protection capabilities to meet the needs of actual clinical applications. Summary of the invention

[0006] The present invention aims at the above technical problems and proposes a deep learning-based intelligent recognition system and method for abnormal cells in blood smears. The system realizes efficient and accurate recognition of abnormal cells in blood smears by innovatively designing a multi-level convolutional neural network module and combining an abnormal cell library and a knowledge graph.

[0007] The present invention proposes an intelligent clinical nursing skills training and evaluation system, comprising:

[0008] Image acquisition module for:

[0009] Collect microscopic images of blood smears;

[0010] Preprocess the collected images, including denoising and morphological processing;

[0011] A multi-level convolutional neural network module is data-connected to the image acquisition module and is used to:

[0012] Receiving the preprocessed image sent by the image acquisition module;

[0013] Performing multi-task analysis on the received image through at least two-level convolutional neural network models, and outputting classification information of cells in the image, wherein the classification information includes cell category and cell location;

[0014] Among them, the models at different levels receive the output results of the previous level model and compare them with their own results to obtain the difference areas, merge and optimize the difference areas before proceeding to the next level of recognition;

[0015] The abnormal cell library module is connected to the multi-level convolutional neural network module data and is used to:

[0016] Store images and related information of known abnormal cells;

[0017] Receiving the output result of the multi-level convolutional neural network module and matching it with the stored abnormal cell information;

[0018] The knowledge graph module is connected with the abnormal cell library module data and is used to:

[0019] Based on the matching results of the abnormal cell bank module, provide relevant disease knowledge and diagnostic suggestions;

[0020] The blood index calculation module is connected to the multi-level convolutional neural network module and the abnormal cell library module data, and is used to:

[0021] Calculating a plurality of blood indicators based on the output results of the multi-level convolutional neural network module and the matching results of the abnormal cell library module;

[0022] Generate graphical display results of blood indicators;

[0023] The result output module is data-connected with the knowledge graph module and the blood index calculation module, and is used to:

[0024] Integrate the diagnostic suggestions provided by the knowledge graph module and the graphical display results generated by the blood index calculation module;

[0025] Generate a final diagnostic report.

[0026] Preferably, the multi-level convolutional neural network module comprises:

[0027] The first level of convolutional neural network units is used to:

[0028] Perform preliminary analysis on the input image and output preliminary cell classification and location information;

[0029] The second-level convolutional neural network unit is data-connected to the first-level convolutional neural network unit, and is used to:

[0030] Receiving the output result of the first-level convolutional neural network unit;

[0031] Further analysis of the received results to identify areas of difference;

[0032] Optimize the identified difference areas;

[0033] The third-level convolutional neural network unit is data-connected to the second-level convolutional neural network unit, and is used to:

[0034] Receiving the optimization result of the second-level convolutional neural network unit;

[0035] Conduct a final analysis on the optimized results and output the final cell classification and localization information;

[0036] Among them, the first-level, second-level and third-level convolutional neural network units all adopt a network structure based on Inception-Resnet, with a network width of 38, and the input and output of each layer of the network have the same resolution.

[0037] Preferably, the multi-level convolutional neural network module further includes:

[0038] The attention mechanism unit is connected to the first, second and third level convolutional neural network unit data to:

[0039] Introducing a class-attention mechanism into each level of convolutional neural network to integrate multi-category image classification tasks;

[0040] Set weights for each channel in the loss function of the cell class and train the weights;

[0041] The spatial attention mechanism is introduced to enhance the model's attention to key areas.

[0042] Preferably, the abnormal cell bank module comprises:

[0043] Abnormal cell image submodule, used for:

[0044] Store standard images of various known abnormal cells;

[0045] The abnormal cell feature submodule is data-connected to the abnormal cell image submodule and is used to:

[0046] storing feature information corresponding to the abnormal cell image;

[0047] The matching analysis submodule is data-connected with the abnormal cell image submodule and the abnormal cell feature submodule, and is used to:

[0048] Receiving the output result of the multi-level convolutional neural network module;

[0049] matching and analyzing the received results with the stored abnormal cell images and features;

[0050] Output the matching results.

[0051] Preferably, the knowledge graph module includes:

[0052] Disease knowledge base submodule, used for:

[0053] Store disease knowledge related to blood abnormalities;

[0054] Diagnostic rules submodule, used to:

[0055] storing diagnostic rules based on abnormal cell types;

[0056] The inference engine submodule is data-connected with the disease knowledge base submodule and the diagnosis rule submodule, and is used to:

[0057] Based on the matching results of the abnormal cell bank module, combined with the stored disease knowledge and diagnosis rules;

[0058] Generate diagnostic recommendations.

[0059] Preferably, the blood index calculation module comprises:

[0060] Cell counting submodule for:

[0061] Calculating the number of each type of cells based on the output results of the multi-level convolutional neural network module;

[0062] The ratio calculation submodule is data-connected with the cell counting submodule and is used for:

[0063] Calculate the proportion of each type of cell;

[0064] Morphological analysis submodule, used to:

[0065] Analyze the morphological characteristics of cells, such as the degree of deformity of red blood cells;

[0066] The index generation submodule is data-connected with the cell counting submodule, the ratio calculation submodule and the morphology analysis submodule, and is used to:

[0067] The calculation results of each submodule are integrated to generate a complete blood index report.

[0068] As a preference, it also includes:

[0069] Biochemical test module for:

[0070] Conduct biochemical tests on blood samples to obtain biochemical index data;

[0071] A data fusion module is data-connected to the multi-level convolutional neural network module and the biochemical testing module, and is used to:

[0072] Receiving the image analysis results of the multi-level convolutional neural network module and the biochemical index data of the biochemical testing module;

[0073] The received data is fused and analyzed to improve the accuracy of abnormal cell detection.

[0074] Preferably, the multi-level convolutional neural network module adopts a multi-task learning strategy to simultaneously perform cell classification and cell localization tasks, wherein:

[0075] The cell classification task is used to determine the category of cells;

[0076] The cell localization task is used to determine the position and boundaries of cells in the image;

[0077] By sharing the feature extraction layer, the collaborative optimization of the two tasks is achieved.

[0078] As a preference, it also includes:

[0079] A privacy protection module, which is data-connected to the image acquisition module and the result output module, is used to:

[0080] Desensitize the collected blood smear microscopic images to remove any personal identifying information that may be contained;

[0081] Encrypt and store the generated diagnostic report;

[0082] Controls access to diagnostic reports.

[0083] The deep learning-based intelligent recognition method for abnormal cells in blood smears based on the system comprises the following steps:

[0084] S1: Collect blood smear microscopic images and preprocess the collected images;

[0085] S2: inputting the preprocessed image into a multi-level convolutional neural network module, which includes at least two levels of convolutional neural network models, analyzes the image in a multi-task manner, and outputs classification information of cells in the image, wherein the classification information includes cell category and cell location;

[0086] S3: In the multi-level convolutional neural network analysis process, the models at different levels receive the output results of the previous level model and compare them with their own results to obtain the difference areas, merge and optimize the difference areas before proceeding to the next level of recognition;

[0087] S4: matching the output result of the multi-level convolutional neural network module with the abnormal cell information stored in the abnormal cell library;

[0088] S5: Based on the matching results of the abnormal cell library, the knowledge graph is used to provide relevant disease knowledge and diagnostic suggestions;

[0089] S6: Based on the output results of the multi-level convolutional neural network module and the matching results of the abnormal cell library, multiple blood indicators are calculated, and graphical display results of the blood indicators are generated;

[0090] S7: Integrate the diagnostic suggestions provided by the knowledge graph and the graphical display results of blood indicators to generate the final diagnostic report;

[0091] Among them, the multi-level convolutional neural network module adopts a network structure based on Inception-Resnet, the network width is 38, the input and output of each layer of the network have the same resolution, and the class attention mechanism and spatial attention mechanism are introduced to improve the recognition accuracy.

[0092] Specifically, the beneficial effects of the present invention are mainly reflected in the following aspects:

[0093] The system of the present invention has many significant advantages and beneficial effects. From a macro perspective, the system realizes the organic combination of deep learning technology and medical expertise, providing a comprehensive and intelligent solution for the diagnosis of blood diseases. The system can not only automatically complete the entire process from image acquisition to diagnosis report generation, but also provide doctors with detailed analysis results and diagnostic suggestions, greatly improving the efficiency and accuracy of blood disease diagnosis.

[0094] Going deep into the core design of the system, the innovative architecture of the multi-level convolutional neural network module is a highlight of the present invention. By adopting at least two levels of convolutional neural network models, the system can gradually improve the accuracy of recognition and effectively deal with complex and diverse abnormal cell types. The information interaction mechanism between models at different levels enables the system to adaptively focus on difficult-to-identify cell areas, further improving the accuracy of recognition. At the same time, the application of multi-task learning strategies enables the system to simultaneously realize cell classification and positioning, providing more comprehensive and accurate information for clinical diagnosis.

[0095] At the algorithm level, the present invention introduces class attention and spatial attention mechanisms, which can help the model better focus on key cell features and image areas. Combined with the network structure based on Inception-Resnet, the system achieves good computational efficiency while maintaining high recognition accuracy. This design not only improves the performance of the system, but also enhances its practicality when processing a large number of blood smear images.

[0096] Another important innovation of the present invention is the integration of abnormal cell library and knowledge graph. By establishing a comprehensive abnormal cell library, the system can quickly identify abnormal cells of known types. The introduction of knowledge graph enables the system to provide diagnostic suggestions based on medical knowledge, which greatly enhances the practical value and credibility of the system. This close combination of image analysis and knowledge reasoning represents an important development direction of medical AI systems.

[0097] In terms of practical application, the system of the present invention also takes into account the issue of privacy protection. By introducing a dedicated privacy protection module, the system can effectively protect the patient's privacy information while providing high-quality diagnostic services. This design not only meets the increasingly stringent requirements for medical data protection, but also enhances the applicability of the system in actual clinical environments.

[0098] In general, the system of the present invention has achieved a significant improvement in the accuracy and efficiency of abnormal cell identification through the collaborative work of multiple modules, and also has important functions such as knowledge reasoning and privacy protection. The organic combination of these advantages enables the system to provide strong support for the early diagnosis and precise treatment of blood diseases, and is expected to play an important role in clinical practice and promote the advancement of blood disease diagnosis technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 It is the overall logic block diagram of the system of the present invention.

[0100] Figure 2 It is a logic block diagram of the multi-level convolutional neural network module of the present invention.

[0101] Figure 3This is a logic block diagram of the abnormal cell bank module of the present invention.

[0102] Figure 4 This is a logical block diagram of the knowledge graph module of the present invention.

[0103] Figure 5 It is a logic block diagram of the blood index calculation module of the present invention.

[0104] Figure 6 This is a logic block diagram of the privacy protection module of the present invention. DETAILED DESCRIPTION

[0105] See also Figure 1-5 The present invention provides a deep learning-based intelligent recognition system and method for abnormal cells in blood smears. The system uses advanced deep learning technology, combined with multi-level convolutional neural networks and knowledge graphs, to achieve efficient and accurate recognition of abnormal cells in blood smears.

[0106] Specifically, the system of the present invention includes the following modules:

[0107] The image acquisition module 1 is used to acquire blood smear microscopic images and pre-process the acquired images, including denoising and morphological processing. Preferably, the denoising process can use methods such as Gaussian filtering or median filtering to reduce noise interference in the image. Morphological processing can include opening and closing operations to smooth cell contours and fill small holes.

[0108] The multi-level convolutional neural network module 2 is data-connected to the image acquisition module 1 and is used to receive the preprocessed image. The module performs multi-task analysis on the received image through at least two levels of convolutional neural network models, and outputs the classification information of the cells in the image, including cell category and cell location. It is worth noting that there is information interaction between models at different levels. The high-level model will receive the output results of the low-level model, compare them with its own results, obtain the difference areas, merge and optimize these areas, and then proceed to the next level of recognition. This multi-level architecture design realizes the synergy and complementarity between models, which can gradually improve the accuracy of recognition.

[0109] The abnormal cell library module 3 is connected to the multi-level convolutional neural network module 2 for storing images and related information of known abnormal cells. This module receives the output of the multi-level convolutional neural network module 2 and matches it with the stored abnormal cell information. This process helps to quickly identify abnormal cells of known types and improve diagnostic efficiency.

[0110] The knowledge graph module 4 is connected to the data of the abnormal cell bank module 3, and provides relevant disease knowledge and diagnosis suggestions based on the matching results of the abnormal cell bank module 3. In one embodiment of the present invention, the knowledge graph may contain information such as symptoms, causes and treatment methods of various blood diseases, providing a comprehensive reference for doctors.

[0111] The blood index calculation module 5 is connected with the multi-level convolutional neural network module 2 and the abnormal cell library module 3. Based on the output results of these two modules, multiple blood indexes are calculated and graphical display results of blood indexes are generated. For example, red blood cell count, white blood cell classification count, platelet count and other indicators can be calculated and displayed intuitively in the form of a bar graph or a line graph.

[0112] The result output module 6 is connected with the knowledge graph module 4 and the blood index calculation module 5 to integrate the diagnosis suggestions and graphical display results to generate the final diagnosis report. This report will provide doctors with comprehensive and detailed blood analysis results to assist clinical diagnosis.

[0113] The multi-level convolutional neural network module 2 includes the following units:

[0114] The first-level convolutional neural network unit 21 is used to perform preliminary analysis on the input image and output preliminary cell classification and positioning information. In a preferred embodiment of the present invention, the unit can use a classic network structure such as VGG16 or ResNet50 as a basis to extract low-level features of the image.

[0115] The second level convolutional neural network unit 22 is data connected to the first level convolutional neural network unit 21, and is used to receive the output results of the first level unit, further analyze it, identify the difference area, and optimize the identified difference area. This level of network can focus on more complex features, such as cell morphology and texture.

[0116] The third level convolutional neural network unit 23 is connected to the second level convolutional neural network unit 22 for receiving the optimization results of the second level unit, performing a final analysis on it, and outputting the final cell classification and location information. This level of network can integrate the information of the first two levels to make a more accurate judgment.

[0117] It is worth noting that the first, second, and third level convolutional neural network units all use an Inception-Resnet-based network structure with a network width of 38, and the input and output of each layer of the network have the same resolution. This design not only retains the multi-scale feature extraction capability of the Inception structure, but also uses the residual connection of Resnet to alleviate the gradient vanishing problem, which is conducive to training deeper networks.

[0118] For example, for a typical blood smear image, its resolution may be 1000x1000 pixels. The first-level network may downsample it to 250x250 pixels for preliminary analysis, the second-level network maintains this resolution for deeper analysis, and the third-level network may slightly upsample to 300x300 pixels to obtain finer localization information.

[0119] The multi-level convolutional neural network module 2 also includes an attention mechanism unit 24, which is connected to the data of the first, second and third level convolutional neural network units. The unit introduces a class attention mechanism in each level of the convolutional neural network to integrate multi-category image classification tasks. A weight is set for each channel in the loss function of the cell category and the weight is trained. At the same time, a spatial attention mechanism is introduced to enhance the model's attention to key areas.

[0120] In one embodiment of the present invention, the class attention mechanism can be implemented by the following formula:

[0121] A c =σ(W c ·F+b c ),

[0122] Among them, A c is the class attention weight, σ is the sigmoid activation function, W c is the learnable weight matrix, F is the feature map, b c is the bias term. The spatial attention mechanism can be implemented by the following formula:

[0123] M s =σ(f 7×7 ([AvgPool(F);MaxPool(F)])),

[0124] Among them, M s is the spatial attention map, f 7×7 represents a 7x7 convolution operation, AvgPool and MaxPool represent average pooling and maximum pooling operations respectively.

[0125] The introduction of these attention mechanisms enables the model to pay more attention to important cell categories and image regions, thereby improving recognition accuracy. For example, when identifying leukemia cells, the model may pay more attention to the morphology and size of the cell nucleus, while when identifying anemia-related red blood cell abnormalities, it may pay more attention to the color and shape of the cell.

[0126] In general, the system of the present invention realizes in-depth analysis of blood smear images through the collaborative work of multi-level convolutional neural networks and the attention mechanism. At the same time, through the combination of abnormal cell library and knowledge graph, it can not only accurately identify abnormal cells, but also provide relevant medical knowledge support, and provide comprehensive and reliable auxiliary information for clinical diagnosis. This combination of deep learning and medical knowledge represents an important direction for the application of artificial intelligence in the medical field, and is expected to significantly improve the efficiency and accuracy of blood disease diagnosis. The system of the present invention further improves the structure and function of the abnormal cell library module.

[0127] Abnormal cell bank module 3 includes the following submodules:

[0128] The abnormal cell image submodule 31 is used to store standard images of various known abnormal cells. In one embodiment of the present invention, these standard images may include various types of leukemia cells, abnormal red blood cells related to anemia, platelet abnormalities, etc. These images are reviewed and annotated by professional doctors to ensure their representativeness and accuracy.

[0129] The abnormal cell feature submodule 32 is data-connected to the abnormal cell image submodule 31 and is used to store feature information corresponding to the abnormal cell image. These features may include cell size, shape, nuclear-cytoplasmic ratio, color features, etc. Preferably, these features can be represented in vector form to facilitate subsequent matching analysis. For example, the feature vector of an abnormal lymphoblast may be as follows:

[0130] f=[size,shape,nucleus_size,cytoplasm_color,granularity,...],

[0131] Each element is normalized to the interval [0,1] for easy comparison and analysis.

[0132] The matching analysis submodule 33 is data-connected to the abnormal cell image submodule 31 and the abnormal cell feature submodule 32, and is used to receive the output result of the multi-level convolutional neural network module 2, match and analyze the received result with the stored abnormal cell image and feature, and output the matching result. In a preferred embodiment of the present invention, the matching analysis can use cosine similarity to measure the similarity between the cell to be identified and the cell in the library. Specifically, the following formula can be used:

[0133]

[0134] Among them, f1 and f2 represent the feature vectors of the cells to be identified and the cells in the library, respectively. When the similarity exceeds a preset threshold (eg, 0.95), the system will consider that a matching abnormal cell has been found.

[0135] The knowledge graph module 4 of the present invention is also designed in detail. The knowledge graph module 4 includes the following submodules:

[0136] The disease knowledge base submodule 41 is used to store disease knowledge related to blood abnormalities. This knowledge base may contain information such as symptoms, causes, pathogenesis, treatment methods, etc. of various blood diseases. Preferably, this information is stored in a structured manner to facilitate rapid retrieval and reasoning.

[0137] The diagnosis rule submodule 42 is used to store diagnosis rules based on abnormal cell types. These rules may come from medical textbooks, clinical guidelines or expert experience. For example, a simple rule may be:

[0138] "If the proportion of blast cells in peripheral blood is >20%, acute leukemia is highly suspected"

[0139] The inference engine submodule 43 is data-connected with the disease knowledge base submodule 41 and the diagnosis rule submodule 42, and generates a diagnosis suggestion based on the matching result of the abnormal cell library module 3, combined with the stored disease knowledge and diagnosis rules. In one embodiment of the present invention, the inference engine may adopt a forward chaining reasoning strategy, starting from the known abnormal cell information, and deriving possible disease diagnoses by applying diagnosis rules.

[0140] The system of the present invention also makes a detailed design for the blood index calculation module 5. As described in claim 6, the blood index calculation module 5 includes the following submodules:

[0141] The cell counting submodule 51 is used to calculate the number of various types of cells based on the output results of the multi-level convolutional neural network module 2. Preferably, this calculation process will take into account the magnification and field of view of the image to obtain a standardized cell counting result.

[0142] The ratio calculation submodule 52 is connected to the cell counting submodule 51 to calculate the proportion of each type of cell. For example, the five-category ratio of white blood cells (neutrophils, lymphocytes, monocytes, eosinophils, and basophils) is calculated. These ratios are crucial for diagnosing a variety of blood diseases.

[0143] The morphological analysis submodule 53 is used to analyze the morphological characteristics of cells, such as the degree of deformity of red blood cells. In one embodiment of the present invention, the degree of morphological abnormality of red blood cells can be quantified by calculating the circularity of the red blood cells. The circularity can be calculated by the following formula:

[0144]

[0145] Among them, area is the cell area and perimeter is the cell circumference. The closer the circularity value is to 1, the closer the cell is to a perfect circle.

[0146] The index generation submodule 54 is connected with the cell counting submodule 51, the ratio calculation submodule 52 and the morphology analysis submodule 53 to generate a complete blood index report by integrating the calculation results of each submodule. This report may include multiple indicators such as red blood cell count (RBC), white blood cell count (WBC), platelet count (PLT), hemoglobin concentration (HGB), mean corpuscular volume (MCV), etc.

[0147] In order to further improve the accuracy and reliability of the system, the system of the present invention also introduces biochemical test results. The system also includes:

[0148] The biochemical test module 7 is used to perform biochemical tests on blood samples to obtain biochemical index data, which may include blood sugar, blood lipids, liver function, kidney function and other biochemical indexes.

[0149] The data fusion module 8 is connected to the multi-level convolutional neural network module 2 and the biochemical test module 7, and is used to receive the image analysis results of the multi-level convolutional neural network module 2 and the biochemical index data of the biochemical test module 7, and perform fusion analysis on the received data to improve the accuracy of abnormal cell detection. In a preferred embodiment of the present invention, data fusion can adopt weighted average or decision tree methods. For example, the following formula can be used to calculate the final abnormal probability:

[0150] P final =w1·P image +w2·P biochem ,

[0151] Among them, P image and P biochem They represent the abnormal probabilities based on image analysis and biochemical testing respectively, and w1 and w2 are the corresponding weights, which can be optimized by machine learning methods.

[0152] Through the above design, the system of the present invention realizes the organic combination of image analysis, knowledge reasoning and biochemical testing, and provides comprehensive and reliable technical support for the intelligent identification of abnormal blood cells. This fusion analysis of multi-source data not only improves the accuracy of diagnosis, but also provides doctors with a more comprehensive basis for disease assessment, which helps to formulate a more accurate treatment plan. The system of the present invention adopts a multi-task learning strategy in the design of the multi-level convolutional neural network module. The multi-level convolutional neural network module 2 simultaneously performs the two tasks of cell classification and cell localization, and realizes the coordinated optimization of the two tasks by sharing the feature extraction layer.

[0153] Specifically, the cell classification task is used to determine the category of cells. In one embodiment of the present invention, the cell categories may include normal red blood cells, normal white blood cells (further subdivided into neutrophils, lymphocytes, monocytes, etc.), normal platelets, and various abnormal cells such as leukemia cells, abnormal red blood cells, etc. The output of the cell classification task can be expressed as a probability distribution:

[0154] P(c|I)=softmax(f c (CNN(I))),

[0155] Where I is the input image, CNN(I) represents the feature extraction process of the convolutional neural network, and f c is the classification head network, and c represents the cell category. The cell localization task is used to determine the position and boundaries of the cell in the image. Preferably, this can be achieved by object detection or instance segmentation methods. For example, algorithms such as Faster R-CNN or Mask R-CNN are used. The output of cell localization can be represented as a series of bounding boxes and corresponding confidences:

[0156] B={(x i ,y i ,w i ,h i ,conf i )|i=1,2,...,N},

[0157] Among them, (x i ,y i ) represents the center coordinate of the bounding box, w i and h i Respectively represent width and height, conf i Indicates confidence.

[0158] By sharing the feature extraction layer, the system of the present invention can more efficiently utilize limited computing resources and improve the generalization ability of the model. This is because different tasks may require the extraction of features at different levels. Through multi-task learning, the model can learn richer and more robust feature representations.

[0159] In order to protect patient privacy and ensure the security of the system, the system of the present invention also includes a privacy protection module. The privacy protection module 9 is data-connected to the image acquisition module 1 and the result output module 6, and mainly performs the following functions:

[0160] First, the collected blood smear microscopic images are desensitized to remove any personal identifying information that may be contained. In a preferred embodiment of the present invention, this may include blurring or removing label information in the image, such as the patient's name, ID, etc. In addition, the system may also perform slight random transformations on the image, such as rotation, scaling, etc., to further reduce the risk of identification without affecting the image features relevant to diagnosis.

[0161] Secondly, the generated diagnostic report is encrypted and stored. In one embodiment of the present invention, the diagnostic report can be encrypted using the AES-256 encryption algorithm. The encryption process can be expressed as:

[0162] C=E k (M),

[0163] Among them, C is the ciphertext, M is the plaintext (i.e., the diagnosis report), E k represents an encryption function using key k.

[0164] Finally, the privacy protection module 9 is also responsible for controlling the access rights of the diagnostic report. Preferably, the system can adopt a role-based access control (RBAC) model. For example, the following roles and permissions can be defined:

[0165] 1. Doctor: can view the complete diagnosis report;

[0166] 2. Nurses: can view some diagnostic results, but cannot access detailed pathological analysis;

[0167] 3. Laboratory technicians: can only access information related to sample processing;

[0168] 4. Patients: can view simplified diagnosis results;

[0169] Through this fine-grained access control, the system of the present invention can protect the privacy of patients to the greatest extent while providing convenient diagnostic services.

[0170] Finally, the present invention provides a method for intelligently identifying abnormal cells in blood smears based on deep learning. This method includes the following steps:

[0171] S1: Collect blood smear microscopic images and preprocess the collected images. In one embodiment of the present invention, the preprocessing may include operations such as image denoising, contrast enhancement and color standardization. For example, Gaussian filtering can be used to remove noise, and then contrast limited adaptive histogram equalization (CLAHE) can be applied to enhance image contrast.

[0172] S2: Input the preprocessed image into a multi-level convolutional neural network module. The module includes at least two levels of convolutional neural network models, analyzes the image in a multi-task manner, and outputs classification information of cells in the image, including cell category and cell location.

[0173] S3: In the multi-level convolutional neural network analysis process, the models at different levels receive the output results of the previous level model and compare them with their own results to obtain the difference area, merge and optimize the difference area and then perform the next level of identification. Preferably, the definition of the difference area can be based on the difference in classification probability or location confidence. For example, the following formula can be used to define the difference area:

[0174] D={(x,y)||P1(c|I x,y )-P2(c|I x,y )|>τ}

[0175] Among them, P1 and P2 represent the classification probability outputs of the two-level models, respectively. x,y It represents the local area of ​​the image at the (x, y) position, and τ is the preset threshold, which can usually be set between 0.2 and 0.3.

[0176] S4: Match the output of the multi-level convolutional neural network module with the abnormal cell information stored in the abnormal cell library. This step can use the cosine similarity-based matching method described above.

[0177] S5: Based on the matching results of the abnormal cell library, the knowledge graph is used to provide relevant disease knowledge and diagnosis suggestions. In a preferred embodiment of the present invention, this step may involve a complex reasoning process, such as using a probabilistic graphical model or a Bayesian network to infer the possibility of a disease.

[0178] S6: Based on the output results of the multi-level convolutional neural network module and the matching results of the abnormal cell library, multiple blood indicators are calculated and graphical display results of the blood indicators are generated. These indicators may include the count, proportion, and morphological characteristics of various types of cells.

[0179] S7: Integrate the diagnostic suggestions provided by the knowledge graph and the graphical display results of the blood indicators to generate a final diagnostic report. Preferably, the generation of the diagnostic report can adopt a template method and combine natural language generation technology to provide a more humane description.

[0180] It is worth noting that the multi-level convolutional neural network module in the method of the present invention adopts a network structure based on Inceptio n-Resnet, with a network width of 38, and the input and output of each layer of the network have the same resolution. At the same time, the class attention mechanism and the spatial attention mechanism are introduced to improve the recognition accuracy. This design can not only extract multi-scale image features, but also effectively focus on important cell areas and categories, thereby achieving high-precision abnormal cell recognition.

[0181] Through the above steps, the method of the present invention realizes the automation of the whole process from image acquisition to diagnosis report generation, greatly improving the efficiency and accuracy of abnormal cell identification in blood smears. This intelligent diagnosis method based on deep learning is expected to play an important role in clinical practice and provide strong support for the early diagnosis and precise treatment of blood diseases.

[0182] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. The intelligent recognition system of abnormal cells in blood smears based on deep learning is characterized by: include: Image acquisition module for: Collect microscopic images of blood smears; Preprocess the collected images, including denoising and morphological processing; A multi-level convolutional neural network module is data-connected to the image acquisition module and is used to: Receiving the preprocessed image sent by the image acquisition module; Performing multi-task analysis on the received image through at least two-level convolutional neural network models, and outputting classification information of cells in the image, wherein the classification information includes cell category and cell location; Among them, the models at different levels receive the output results of the previous level model and compare them with their own results to obtain the difference areas, merge and optimize the difference areas before proceeding to the next level of recognition; The abnormal cell library module is connected to the multi-level convolutional neural network module data and is used to: Store images and related information of known abnormal cells; Receiving the output result of the multi-level convolutional neural network module and matching it with the stored abnormal cell information; The knowledge graph module is connected with the abnormal cell library module data and is used to: Based on the matching results of the abnormal cell bank module, provide relevant disease knowledge and diagnostic suggestions; The blood index calculation module is connected to the multi-level convolutional neural network module and the abnormal cell library module data, and is used to: Calculating a plurality of blood indicators based on the output results of the multi-level convolutional neural network module and the matching results of the abnormal cell library module; Generate graphical display results of blood indicators; The result output module is data-connected with the knowledge graph module and the blood index calculation module, and is used to: Integrate the diagnostic suggestions provided by the knowledge graph module and the graphical display results generated by the blood index calculation module; Generate a final diagnostic report.

2. The system according to claim 1, characterized in that The multi-level convolutional neural network module includes: The first level of convolutional neural network units is used to: Perform preliminary analysis on the input image and output preliminary cell classification and location information; The second-level convolutional neural network unit is data-connected to the first-level convolutional neural network unit, and is used to: Receiving the output result of the first-level convolutional neural network unit; Further analysis of the received results to identify areas of difference; Optimize the identified difference areas; The third-level convolutional neural network unit is data-connected to the second-level convolutional neural network unit, and is used to: Receiving the optimization result of the second-level convolutional neural network unit; Conduct a final analysis on the optimized results and output the final cell classification and localization information; Among them, the first-level, second-level and third-level convolutional neural network units all adopt a network structure based on Inception-Resnet, with a network width of 38, and the input and output of each layer of the network have the same resolution.

3. The system according to claim 2, characterized in that The multi-level convolutional neural network module also includes: The attention mechanism unit is connected to the first, second and third level convolutional neural network unit data to: Introducing a class-attention mechanism into each level of convolutional neural network to integrate multi-category image classification tasks; Set weights for each channel in the loss function of the cell class and train the weights; The spatial attention mechanism is introduced to enhance the model's attention to key areas.

4. The system according to claim 1, characterized in that The abnormal cell bank module comprises: Abnormal cell image submodule, used for: Store standard images of various known abnormal cells; The abnormal cell feature submodule is data-connected with the abnormal cell image submodule and is used to: storing feature information corresponding to the abnormal cell image; The matching analysis submodule is data-connected with the abnormal cell image submodule and the abnormal cell feature submodule, and is used to: Receiving the output result of the multi-level convolutional neural network module; matching and analyzing the received results with the stored abnormal cell images and features; Output the matching results.

5. The system according to claim 1, characterized in that The knowledge graph module includes: Disease knowledge base submodule, used for: Store disease knowledge related to blood abnormalities; Diagnostic rules submodule, used to: storing diagnostic rules based on abnormal cell types; The inference engine submodule is data-connected with the disease knowledge base submodule and the diagnosis rule submodule, and is used to: Based on the matching results of the abnormal cell bank module, combined with the stored disease knowledge and diagnosis rules; Generate diagnostic recommendations.

6. The system according to claim 1, characterized in that The blood index calculation module comprises: Cell counting submodule for: Calculating the number of each type of cells based on the output results of the multi-level convolutional neural network module; The ratio calculation submodule is data-connected with the cell counting submodule and is used for: Calculate the proportion of each type of cell; Morphological analysis submodule, used for: Analyze the morphological characteristics of cells, such as the degree of deformity of red blood cells; The index generation submodule is data-connected with the cell counting submodule, the ratio calculation submodule and the morphology analysis submodule, and is used to: The calculation results of each submodule are integrated to generate a complete blood index report.

7. The system according to claim 1, characterized in that Also includes: Biochemical test module for: Conduct biochemical tests on blood samples to obtain biochemical index data; A data fusion module is data-connected to the multi-level convolutional neural network module and the biochemical testing module, and is used to: Receiving the image analysis results of the multi-level convolutional neural network module and the biochemical index data of the biochemical testing module; The received data is fused and analyzed to improve the accuracy of abnormal cell detection.

8. The system according to claim 1, characterized in that The multi-level convolutional neural network module adopts a multi-task learning strategy to simultaneously perform cell classification and cell localization tasks, where: The cell classification task is used to determine the category of cells; The cell localization task is used to determine the position and boundaries of cells in the image; By sharing the feature extraction layer, the collaborative optimization of the two tasks is achieved.

9. The system according to claim 1, characterized in that Also includes: A privacy protection module is data-connected to the image acquisition module and the result output module, and is used to: Desensitize the collected blood smear microscopic images to remove any personal identifying information that may be contained; Encrypt and store the generated diagnostic report; Controls access to diagnostic reports.

10. A method for intelligently identifying abnormal cells in blood smears based on deep learning based on the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1: Collect blood smear microscopic images and preprocess the collected images; S2: inputting the preprocessed image into a multi-level convolutional neural network module, which includes at least two levels of convolutional neural network models, analyzes the image in a multi-task manner, and outputs classification information of cells in the image, wherein the classification information includes cell category and cell location; S3: In the multi-level convolutional neural network analysis process, the models at different levels receive the output results of the previous level model and compare them with their own results to obtain the difference areas, merge and optimize the difference areas before proceeding to the next level of recognition; S4: matching the output result of the multi-level convolutional neural network module with the abnormal cell information stored in the abnormal cell library; S5: Based on the matching results of the abnormal cell library, the knowledge graph is used to provide relevant disease knowledge and diagnostic suggestions; S6: Based on the output results of the multi-level convolutional neural network module and the matching results of the abnormal cell library, multiple blood indicators are calculated, and graphical display results of the blood indicators are generated; S7: Integrate the diagnostic suggestions provided by the knowledge graph and the graphical display results of blood indicators to generate the final diagnostic report; Among them, the multi-level convolutional neural network module adopts a network structure based on Inception-Resnet, the network width is 38, the input and output of each layer of the network have the same resolution, and the class attention mechanism and spatial attention mechanism are introduced to improve the recognition accuracy.

Images