Micro-nucleated blood cell recognition system based on generative model data augmentation

Abstract

The application provides a kind of micro nuclear blood cell identification system based on generative model data augmentation, adopts micro nuclear identification model, and the blood cell image in chromosome image database is input into generative model offline training module to complete distributed training respectively, and the image of specified cytoplasm and nucleus position and contour is generated;Wherein, the generative model offline training module trains the diffusion probability model based on control network, which is used to train the model that can generate a large number of negative and positive samples, and the negative micro nuclear and positive micro nuclear cell image generation module is used to generate a large number of negative and positive images by the model trained by the generative model offline training module, then the negative and positive images obtained are classified by the micro nuclear identification model training module, to obtain the binary classification result of negative and positive, and the probability value of micro nuclear cell is displayed through user interface.The application can quickly obtain more representative data set, and provide more accurate support for medical diagnosis and disease treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical artificial intelligence, and particularly relates to a micro-nucleus blood cell recognition system based on generative model data augmentation. BACKGROUND

[0002] Micro-nucleus recognition is an important research direction in the field of modern medicine, and has important background significance and challenges. Micro-nucleus is a kind of nuclear morphological abnormality, which is usually closely related to some diseases such as leukemia, occupational disease, anemia, etc. The goal of using computer algorithm to assist micro-nucleus recognition is to accurately detect and locate micro-nucleus from cell images, so as to assist doctors in disease diagnosis and treatment. This technology has important significance for improving the accuracy of early disease diagnosis, monitoring disease progression and guiding individualized treatment.

[0003] At present, the industry often uses labeled samples to train machine learning models. However, this also faces some key challenges, one of which is the lack of positive samples and the difficulty of data collection. Micro-nucleus is a relatively rare cell abnormality, and needs to be accurately detected through a complex and tedious experimental process. Therefore, it is extremely difficult to obtain a sufficient number of positive samples for training and verifying the algorithm. In addition, data collection involves cooperation with medical institutions, and needs to cover a wide range of cases and cell types, which increases the complexity and time cost of data collection.

[0004] This situation has certain limitations on the research and application of micro-nucleus recognition technology. The lack of sufficient positive samples may affect the training and performance of the algorithm, resulting in insufficient generalization ability of the model. At the same time, the difficulty of data collection may limit the promotion of the algorithm in different cases and clinical applications. Therefore, it is particularly important to use the current generative neural network technology to augment the positive samples. SUMMARY

[0005] The purpose of the present application is to provide a micro-nucleus blood cell recognition system based on generative model data augmentation. Through this system, researchers and doctors can use a large number of unlabeled cell images and a small amount of cell contour labeling to train a generative model. Subsequently, through the trained generative model, a large number of negative samples and positive samples can be easily generated according to the required contour information of the cell nucleus and cytoplasm. This function provides a more rich and diverse data set for training deep learning models, thereby improving the robustness and accuracy of the model. Through the introduction of the generative model, the system can generate micro-nucleus samples with different morphologies, sizes and rotation angles, thereby enhancing the adaptability of the model to complex situations in the real world. The data augmented by the above system can be automatically trained in the system to obtain a micro-nucleus recognition model with strong generalization, to determine the probability value of whether the cell image is a micro-nucleus cell.

[0006] The application provides a micronucleus blood cell recognition system based on generative model data augmentation, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the computer memory stores a trained micronucleus recognition model;

[0007] The micronucleus recognition model is based on a generative model data augmentation algorithm and specifically comprises a chromosome image database, a generative model offline training module, a negative micronucleus and positive micronucleus cell image generation module, and a micronucleus recognition model training module.

[0008] The computer processor implements the following steps when executing the computer program:

[0009] The blood cell images in the chromosome image database are input into the generative model offline training module to complete distributed training respectively, and images of specified cytoplasm and cell nucleus positions and contours are generated; wherein the generative model offline training module trains a diffusion probability model based on a control network, which is used to train a model capable of generating a large number of negative and positive samples, the negative micronucleus and positive micronucleus cell image generation module generates a large number of negative and positive images using the model trained by the generative model offline training module, and then classifies the obtained negative and positive images through the micronucleus recognition model training module to obtain a binary classification model of negative and positive, and displays the probability value of micronucleus cells through a user interaction interface.

[0010] Specifically, the chromosome image database stores millions of blood cell images for training the model proposed in the application.

[0011] The generative model offline training module is used to train the generative model proposed in the application, which can call data in multiple threads and complete distributed training on multiple GPUs.

[0012] The negative micronucleus and positive micronucleus cell image generation module is the core module of the application, which can be used for the generation of negative micronucleus images and positive micronucleus images after reading the model weight obtained from the generative model offline training module, to generate blood cell images using the input cytoplasm and cell nucleus contour as information.

[0013] The micronucleus recognition model training module uses a large number of blood cell images with clear micronucleus negative and positive labels generated to perform deep neural network binary classification modeling training.

[0014] The user interaction interface is used for the specific operation of the system proposed in the application by the user, and the probability value of micronucleus cells will be displayed by the system when importing custom blood cell image data.

[0015] The core of the generation model offline training module in the present application is to train a diffusion probability model based on a control network (ControlNet). The diffusion probability model is a widely used generation model based on Markov random process and deep neural network, which can generate target images through random sampling of Gaussian distribution. However, due to the randomness of the sampling of the model, the generated results cannot be controlled, and it cannot be determined whether the generated cell image is a positive micronucleus image, so it is meaningless for the subsequent micronucleus recognition model training. Therefore, the present application adds a control network to guide the cytoplasm and nucleus of the generated cell. The control network is a technology for guiding the direction of the data generated by the diffusion probability model, which is connected in parallel with the diffusion probability model. By inputting the guide information, the generation result of the diffusion probability model can meet the guide information. Since a large amount of data used in the present application is unlabeled blood cell images, the diffusion probability model can be directly trained to generate blood cell images with different shapes. Subsequently, using the control network module, only a small part of the cytoplasm and nucleus of the blood cells needs to be labeled, so that the model can generate images with specified cytoplasm and nucleus position and outline, so as to control the generation of non-micronucleus images and micronucleus images.

[0016] The negative micronucleus and positive micronucleus cell image generation module is based on the images trained by the generation model offline training module, combined with the specified cytoplasm outline, and according to the cell nucleus outline random generation algorithm based on medical knowledge to obtain the guide information for the generation model, and obtain the cell image meeting the guide information.

[0017] Further, the chromosome image database is derived from millions of blood cell images without any label, and the image encoding format is PNG.

[0018] The generation model offline training module adopts multi-thread data retrieval and completes distributed training on multiple GPUs. During the training process, the diffusion probability model without information guidance needs to be trained first. In one diffusion process, the input image is X t , the output image is X t-1 , and the intermediate is processed by a deep convolutional neural network. The loss function of the diffusion probability model is as follows:

[0019] ;

[0020] Where t represents the t-th step of the diffusion process, L t represents the loss function value at the t-th diffusion step, x0 is the target chromosome image during training, is a random sampling in the standard Gaussian distribution, is an artificially set hyperparameter;

[0021] Secondly, the control network needs to guide the diffusion probability model. First, label 100 cell images of cytoplasm and nucleus contours, and then load the frozen parameters of the diffusion probability model trained in the previous step into the control network. The mathematical expression of single diffusion is as follows:

[0022] ;

[0023] Wherein, is the generation model, and the input of the first diffusion is , is the output of the last diffusion, In the training process, that is, the cytoplasm and nucleus contour images of 100 cell images labeled in this step; the model parameters are divided into three parts, is the parameter of the diffusion probability model trained, which will be frozen in this step and will not participate in the training; is the encoder parameter of the control network, and its initial state is the trainable copy of is the parameter of the zero convolution layer in the control network, and the initialization is 0; is the diffusion probability model, is the zero convolution layer in the control network; it can be obtained from the formula that when the control network is added for the first round of training iteration,

[0024] Since are all initialized to 0,

[0025] Therefore ,

[0026] Further ;

[0027] The diffusion probability model containing the control network finally trained generates blood cell images that meet the guided information of cytoplasm and nucleus position and shape.

[0028] Further, the negative micronucleus and positive micronucleus cell image generation module generates a large number of control network guide information of different shapes under the given cytoplasm contour through a random generation algorithm.

[0029] Further, the control network guide information is obtained by doing OTSU segmentation on the images in the blood cell image database and taking the maximum connected domain, and a large number of cytoplasm contours are obtained. Using these cytoplasm contours as a reference, a large number of blood cell images are randomly generated, and the cytoplasm contour is randomly generated The process is as follows:

[0030] S1, first define the function , meaning to randomly select an integer between start and end according to a uniform distribution, defined as The image has rows and columns.

[0031] S2, randomly select the center point of the generated cell nucleus ( The specific formula is:

[0032] ;

[0033] This formula means in the cytoplasmic outline Within a certain range, a point is randomly selected as the center point of the cell nucleus;

[0034] S3, Calculation ( )and N points on the contour ( ), ... ,( distance set Where N is a configurable hyperparameter, referring to the... N points are uniformly sampled. The calculation method is as follows:

[0035] ;

[0036] This step is to calculate (the result is missing from the original text). Centered on ), the maximum distance that the cell nucleus boundary can reach in all directions;

[0037] S4. Calculate the baseline radius r of the roughly circular shape of the randomly generated cell nuclei, where... These are configurable hyperparameters, referring to the minimum baseline radius of the main core. If so, we need to go back to step 2 and regenerate, where min() is used to find the minimum value;

[0038] ;

[0039] S5. The baseline radius r generated in the previous step can already generate a circular cell nucleus. However, since the cell nucleus is not always circular, random perturbations need to be made at N points. Therefore, the random perturbation distances at N points are generated. ,in

[0040] ;

[0041] Where -r means that the boundary indentation cannot be less than the radius length. The boundary of the cell must not extend beyond the cytoplasm.

[0042] S6, the information obtained from the previous steps can be used to calculate the center of each point on the randomly generated cell nucleus contour with the reference radius r as the center (x, y) = (x0+ rcos(2pi i / N), y0+ rsin(2pi i / N)), i = 0, 1, 2,..., N-1, where N is the number of points on the contour of the randomly generated cell nucleus.

[0043]

[0044]

[0045] According to whether the generated sample is negative or positive, the number of cell nuclei that need to be generated is determined, and the generation process of each cell nucleus is repeated from step S2 to step S6, and finally a large number of .

[0046] Further, when a mononuclear cell needs to be generated, steps S2-S6 are executed only once; when a binuclear cell needs to be generated, steps S2-S6 are executed twice.

[0047] Further, the micronucleus recognition model training module uses a large number of blood cell images with clear micronucleus negative and micronucleus positive labels generated to perform deep neural network binary classification modeling training.

[0048] Further, the micronucleus recognition model training module performs binary classification training on the samples generated by the negative micronucleus and positive micronucleus cell image generation module using a constructed deep convolutional neural network ResNet-50, and the final obtained model can be used for the recognition task of whether the blood cell is a micronucleus cell.

[0049] Further, the user interaction interface is used for user-specific operation of the system, import of custom blood cell image data, and display of the recognition result of whether the blood cell is a micronucleus cell by the user interaction interface.

[0050] Further, the user interaction interface is used to assist doctors in screening micronucleus cells in patient blood samples, wherein the system main area lists all cells of the current slide for the user, and according to the recognition result of the system, the cells are divided into micronucleus cells and non-micronucleus cells, and relevant statistical information is presented on the right side of the interface to help doctors make diagnoses.

[0051] ​​​​The micro-nucleus blood cell recognition system based on the generative model data augmentation provided by the application is helpful to shorten the training time and improve the algorithm performance. The traditional method needs to spend a lot of time and manpower to mark and screen the existing data, and the samples generated by the generative model can reduce the workload and quickly obtain a more representative data set. The micro-nucleus recognition system based on the generative model for data augmentation has important invention significance in overcoming the limitations of traditional micro-nucleus recognition technology. Through the system, it can bring a more broad prospect to the micro-nucleus recognition field, provide more accurate support for medical diagnosis and disease treatment, and open up a new direction for the research and application of medical image analysis field. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to further understand the features and technical contents of the present application, please refer to the following detailed description of the present application and the accompanying drawings, however, the accompanying drawings are provided for reference and illustration only, and are not used to limit the present application.

[0053] Figure 1 The module structure diagram of the present application is shown in Figure 1.

[0054] Figure 2 The generative model structure diagram is shown in Figure 2.

[0055] Figure 3 The generated image example diagram is shown in Figure 3.

[0056] Figure 4 The interface diagram of the system is shown in Figure 4. DETAILED DESCRIPTION

[0057] In order to further illustrate the technical means adopted by the present application and its effects, the following will be described in detail in combination with the preferred embodiments of the present application and the accompanying drawings.

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0059] In the description of the present application, it needs to be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element indicated must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" are only for description purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can be explicitly or implicitly included one or more features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0060] In the present application, the word "exemplary" is used to mean "serving as an example, instance, or illustration." Any implementation described as "exemplary" in the present application is not necessarily to be construed as preferred or advantageous over other implementations. The following description is presented to enable any person skilled in the art to make and use the present application. In the following description, for the purposes of explanation, details are set forth in order to provide a thorough understanding of the present application. It is apparent to one skilled in the art that the present application can be practiced without the use of these specific details. In other instances, well-known structures and processes are not elaborated in order not to obscure the description of the present application with unnecessary details. Therefore, the present application is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein. Embodiment:

[0061] Please refer to Figures 1-4 The embodiments of the present application are divided into three parts, among which the blood cell image database and the offline training module of the generation model are one part; the negative micronucleus and positive micronucleus cell image generation module and the micronucleus recognition model training module are one part; and the user interaction interface is one part. The following will describe the three parts respectively:

[0062] The implementation method of the blood cell image database and the offline training module of the generation model is as follows:

[0063] The data source of the blood cell image database in the present application is millions of blood cell images without any annotation, and the image encoding format is PNG;

[0064] The offline training module of the generation model in the present application is used to train the model proposed in the present application, which can call data in multiple threads and complete distributed training on multiple GPUs. The following is the specific training process.

[0065] First, the diffusion probability model without guidance needs to be trained. The diffusion probability model is essentially a random Gaussian noise After T diffusion, the target image is obtained, where T is a hyperparameter that can be manually set, Figure 2 The diffusion process from two adjacent intermediate images to is described in dashed box A. In a diffusion process, the input image is , and the output image is , which is processed by a deep convolutional neural network, and the loss function of the diffusion probability model is as follows:

[0066]

[0067] where t represents the t-th step of the diffusion process, represents the loss function value at the t-th diffusion step, is the target chromosome image during training, is a random sampling in the standard Gaussian distribution, is a hyperparameter manually set.

[0068] Second, the control network guiding the diffusion probability model is trained. Since the diffusion probability model trained in the previous step cannot control the results when generating cell images, it cannot be used for subsequent classification tasks, so a control network is introduced. First, the cytoplasm and nucleus contours of 100 cell images need to be labeled. Second, the diffusion probability model trained in the previous step, i.e., dashed box A in Figure 2 , is loaded and its parameters are frozen so that it cannot be trained, at this time the diffusion probability model has the ability to generate blood cell images. Subsequently, the control network is incorporated into the diffusion probability model module, as shown in dashed box B in Figure 2 , the composition of the control network is similar to the diffusion probability model, which is also a combination of encoder and decoder, but its decoder is composed of zero convolution layers. Specifically, the loss function of the control network parallel to the diffusion probability model is consistent with step 3 described above, and its mathematical expression in a single diffusion is as follows:

[0069]

[0070] where refers to the entire model in Figure 2 , the input of a diffusion is , is the output of the last diffusion, The cytoplasm and nucleus contour images of 100 cell images labeled in the training process, i.e. this step. The model parameters are divided into three parts , The parameters of the diffusion probability model trained in step 3, which will be frozen in this step and will not participate in training; The encoder parameters of the control network, whose initial state is The trainable copy of The parameters of the zero convolution layer in the control network, all initialized to 0. Refers to the diffusion probability model, Refers to the zero convolution layer in the control network. According to the formula, when the control network is added for the first round of training iteration,

[0071] Since All initialized to 0,

[0072] Therefore ,

[0073] Further .

[0074] As can be seen, at the beginning of the training after adding the control network, the model is equivalent to the diffusion probability model trained in step 3, which means that the model is fine-tuned on the basis of the diffusion probability model trained in step 3 on millions of unlabeled blood cell images, so it can guarantee the basic ability of the generated model to generate cell images. Since are trainable, after training on 100 blood cell images containing cytoplasm and nucleus contour images in this step, the model will further obtain the ability to control the specific position and shape of the cytoplasm and nucleus in the generated blood cell image according to the specific cytoplasm and nucleus contour based on the ability to generate blood cell images obtained in step 3.

[0075] The final trained diffusion probability model containing the control network can generate blood cell images that meet the position and shape of the cytoplasm and nucleus in the guide information, i.e. it can generate negative micronucleus cell images and positive micronucleus cell images with different shapes in the subsequent generation. For example Figure 2 In The guide information, where the black part represents the background, the white part represents the nucleus (in this example, there is a large main nucleus and a small micronucleus), and the gray part represents the cytoplasm. Input the control network and then let the diffusion probability model generate, which can generate a blood cell (in this example, a micronucleus cell) that meets the guide information, as shown in Figure 2 In . Figure 3Four groups of input different generation results are shown in the figure, and it can be seen that the input of this model is flexible and diverse, and the morphology, size and position of the cytoplasm and nucleus can be customized at will, and the generated images are also very diverse in style, and the content is reasonable.

[0076] How to use the negative micronucleus and positive micronucleus cell image generation module to generate a large number of negative micronucleus and positive micronucleus images, and use the micronucleus recognition model training module to train, the specific steps are as follows:

[0077] As described above, it has been clarified how to generate cell images using a diffusion probability model containing a control network, so the key to generating a large number of images is how to obtain a variety of The present application proposes a random generation algorithm for the contour of the nucleus under the given contour of the cytoplasm, which can generate a large number of control network guide information of different morphologies.

[0078] First, by doing Otsu segmentation on the images in the local blood cell image database and taking the maximum connected domain, a large number of cytoplasm contours can be obtained, and a large number of blood cell images can be randomly generated using these cytoplasm contours as a reference, and a fixed cytoplasm contour will be given below to describe a random generation process.

[0079] S1, first define the function , which means to take an integer between start and end according to uniform distribution, define the number of rows of the image as rows and the number of columns as cols;

[0080] S2, randomly obtain the center point of the generated nucleus , the specific formula is:

[0081]

[0082] This formula means that a point is randomly obtained within the contour of the cytoplasm as the center point of the nucleus;

[0083] S3, calculate the distance set of N points on the contour,..., and , where N is a hyperparameter that can be set, indicating that N points are uniformly sampled on , The calculation method is as follows:

[0084] ​​

[0085] This step is to calculate the maximum distance that the nuclear boundary can take in each direction, with (r) as the center;

[0086] S4, calculate the reference radius r of the circle-like nucleus of this random generated nucleus, wherein, is a settable hyperparameter, indicating the minimum value of the reference radius of the main nucleus. If , it is necessary to return to step 2 to regenerate.

[0087]

[0088] S5, the reference radius r generated in the previous step can generate a circular nucleus. However, since the nucleus is not always circular, it is necessary to make random perturbations on N points, so as to generate random perturbation distances of N points , wherein

[0089]

[0090] wherein -r represents that the boundary cannot be reduced to less than the length of the radius, represents that the boundary cannot be extended beyond the cytoplasmic range;

[0091] S6, the information obtained from the previous steps can be used to calculate each point on the profile of the randomly generated nucleus (x1, y1),..., (xn, yn) with the reference radius r as the center, wherein:

[0092]

[0093]

[0094] S7, according to whether the generated sample is negative or positive, the number of nuclei to be generated is determined, and the generation process of each nucleus is repeated steps 2-6, and finally is obtained. For example, when a mononuclear cell needs to be generated, steps 2-6 need to be executed only once; and when a binuclear cell needs to be generated, steps 2-6 need to be executed twice.

[0095] Through the above method, the present application can randomly generate many , so as to generate a large number of negative and positive micronucleus samples. In order to identify the negative and positive, the present application constructs a deep convolutional neural network ResNet-50 to perform binary classification training on these generated samples, and the finally obtained model can be used for the identification task of whether the blood cell is a micronucleus cell.

[0096] ​​​The user interaction interface of the present application is used for assisting doctors to screen micronucleus cells in blood samples of patients, as shown in Figure 4 The system main area lists all cells of the current slide for the user, and according to the recognition result of the system, the cells are divided into micronucleus cells and non-micronucleus cells, and the relevant statistical information is presented on the right side of the interface to help the doctor to make a diagnosis.

[0097] The above-mentioned embodiments of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A microkernel blood cell recognition system based on generative model data augmentation, characterized in that, Computer storage, computer processor and computer program stored in the computer storage and executable on the computer processor, wherein the computer storage stores a trained micronucleus recognition model; The micronucleus recognition model is based on a generative model data augmentation algorithm and specifically includes a chromosome image database, a generative model offline training module, a negative micronucleus and positive micronucleus cell image generation module and a micronucleus recognition model training module. The computer processor implements the following steps when executing the computer program: The blood cell images in the chromosome image database are input into the generative model offline training module to complete distributed training respectively, and images of specified cytoplasm and nucleus positions and contours are generated; wherein the generative model offline training module trains a diffusion probability model based on a control network, which is used to train a model capable of generating a large number of negative and positive samples, the negative micronucleus and positive micronucleus cell image generation module generates a large number of negative and positive images using the model trained by the generative model offline training module, and then classifies the obtained negative and positive images through the micronucleus recognition model training module to obtain a binary classification model of negative and positive, and displays the probability value of micronucleus cells through a user interaction interface; The negative micronucleus and positive micronucleus cell image generation module generates a large number of different morphologies of control network guide information under a given cytoplasm contour through a random generation algorithm; The control network guidance information is obtained by doing Otsu segmentation on the images in the blood cell image database, and taking the maximum connected domain, obtaining a large number of cytoplasm contours, using these cytoplasm contours as a reference, randomly generating a large number of blood cell images, and the cytoplasm contour C mask_c is generated in the following process: msak ​ S1, first define the function Random(start, end), meaning to take an integer randomly between start and end according to uniform distribution, define C mask_c The image has rows rows and columns cols. S2, a center point (x k , y k ) of the generated nucleus is randomly obtained, and the specific formula is: x k = Random(0, cols), y k = Random(0, rows) where (x k , y k ) ∈C mask_c ； This formula means that a point is randomly taken as the center point of the nucleus within the range of the cytoplasm outline C mask_c ; S3, compute (x k , y k ) and the distance set GAP = (gap0,..., gap mask_c ) of N points (x0, y0),..., (x N , y N ) on the C N contour, where N is a hyperparameter that can be set, indicating that N points are uniformly sampled on the C mask_c , and gap i is calculated as follows: This step is to calculate the maximum distance that can be taken in each direction from the nucleus boundary, centered at (x k , y k ); S4, calculate the reference radius r of the circle-like shape of the randomly generated nucleus this time, wherein r minkm is a settable hyperparameter, indicating the minimum value of the reference radius of the main nucleus, if r min_km > min(GAP), then it is necessary to return to step 2 to regenerate, wherein min() is the minimum value; r = Random(r min_km min(GAP)); S5, the reference radius r generated by the previous step can generate a circular nucleus, but since the nucleus is not always circular, a random disturbance distance of N points needs to be generated DIS = (dis0,..., dis N ), where dis i = Random(-r, gap i -r); where -r means that the border cannot be retracted less than the radius length, gap i -r means that the border cannot be expanded beyond the cytoplasmic extent; S6、By the information obtained from the previous steps, the center of the reference radius r can be calculated at each point on the contour of the randomly generated cell nucleus this time, and each point (xk0, yk0),..., (xk N , yk N ) on the contour of the randomly generated cell nucleus is calculated, where: According to whether the generated sample is negative or positive, the number of nuclei to be generated is determined, and the generation process of each nucleus is repeated from step S2 to step S6, and finally a plurality of C mask .

2. The microkernel blood cell identification system based on generative model data augmentation of claim 1, wherein, The data of the chromosome image database is derived from millions of blood cell images without any annotation, and the image encoding format is PNG.

3. The microkernel blood cell identification system based on generative model data augmentation of claim 1, wherein, The generative model offline training module uses multi-threaded data retrieval and completes distributed training on multiple GPUs. During the training process, a diffusion probability model without information guidance needs to be trained. In a diffusion process, the input image is Xt, and the output image is Xt-1. After processing by a deep convolutional neural network, the loss function of the diffusion probability model is as follows: Wherein, t represents the tth step of the diffusion process, Lt represents the loss function value at the tth step of diffusion, x0 is the target chromosome image during training, ∈ is a random sampling in the standard Gaussian distribution, is a hyperparameter set artificially; Secondly, the diffusion probability model needs to be guided by a control network. The cytoplasm and nucleus contours of 100 cell images are labeled first, and then the frozen parameters of the diffusion probability model trained in the previous step are loaded into the control network. The mathematical expression of the single diffusion is as follows: f c (X t ,C mask ;θ,θ copy ,the z )=f(X t ;θ)+Z(f(X t +C mask ;θ copy );i z ); wherein f c (X t , C mask ; θ, θ copy , θ z ) denotes a generative model, the input of which is X t and C mask in one diffusion, X t is the output of the last diffusion, C mask is the image of the cytoplasm and the nucleus contour of the 100 cell images labeled in the training process, i.e. in the present step; the model parameters are divided into θ, θ copy , and θ z , θ is the parameter of the trained diffusion probability model, which will be frozen in the present step and will not participate in the training; θ copy is the parameter of the encoder of the control network, the initial state of which is a trainable copy of θ; θ z is the parameter of the zero convolution layer in the control network, which is initialized as 0; f denotes the diffusion probability model, and Z denotes the zero convolution layer in the control network; it can be obtained from the formula that when the control network is added for the first round of training iteration, since θ z is initialized as 0, Thus Z(f(X t +C mask ; θ copy ) ; θ z ) = 0, Further f c (X t ; θ, θ mask ; θ copy , θ z ) = f(X t ; θ) ; The final trained diffusion probability model containing the control network generates blood cell images with cytoplasm and nucleus positions and shapes consistent with the guide information.

4. The microkernel blood cell identification system based on generative model data augmentation of claim 1, wherein, When a mononuclear cell needs to be generated, steps S2-S6 need to be executed only once; when a binuclear cell needs to be generated, steps S2-S6 need to be executed twice.

5. The microkernel blood cell identification system based on generative model data augmentation of claim 1, wherein, The micronucleus recognition model training module uses a large number of blood cell images with clear micronucleus negative and positive labels to perform deep neural network binary classification modeling training.

6. The microkernel blood cell identification system based on generative model data augmentation of claim 5, wherein, The micro-nucleus recognition model training module trains a deep convolutional neural network ResNet-50 to classify the samples generated by the micro-nucleus cell image generation module, and finally obtains a model that can be used for the recognition task of whether blood cells are micro-nucleus cells.

7. The microkernel blood cell identification system based on generative model data augmentation of claim 1, wherein, The user interaction interface is used for the specific operation of the system by the user, and the user interaction interface displays the recognition result of whether the blood cell image data is a micro-nucleus cell.

8. The microkernel blood cell identification system based on generative model data augmentation of claim 1, wherein, The user interaction interface is used for assisting doctors in screening micro-nucleus cells in blood samples of patients, wherein the system main area lists all cells of the current slide for the user, and according to the recognition result of the system, the cells are divided into two kinds of micro-nucleus cells and non-micro-nucleus cells, and relevant statistical information is presented on the right side of the interface to help doctors make a diagnosis.

Images