Three-dimensional CT image lesion recognition method and system based on big data and deep learning

Abstract

The present invention provides a method and system for three-dimensional CT image lesion recognition based on big data and deep learning, the method comprising the following steps: obtaining a two-dimensional CT image data set and reconstructing it into a three-dimensional CT image data set; distributing and storing the three-dimensional CT image data set in a corresponding model container according to the organ image classification result; training a basic lesion recognition model using the three-dimensional CT image data to obtain a target lesion recognition model; obtaining a lesion recognition request sent by a client; extracting two-dimensional CT image data from the lesion recognition request and reconstructing three-dimensional CT image data to generate a target image classification result; calling the target lesion recognition model in the corresponding model container based on the target image classification result; generating a lesion recognition result through the target lesion recognition model, and feeding back the lesion recognition result to the client that sent the lesion recognition request. The present invention has the effect of quickly and accurately analyzing and processing three-dimensional CT image data.

Description

Technical Field

[0001] The present invention belongs to the technical field of neural network, and specifically relates to a three-dimensional CT image lesion recognition method and system based on big data and deep learning. Background Art

[0002] With the development of CT scanning technology, more and more medical institutions use three-dimensional CT images to provide doctors with auxiliary diagnosis and treatment methods for diseases. These images provide rich anatomical information, which can help doctors identify lesions and evaluate diseases more accurately. However, in the existing CT image analysis process, it is necessary to manually check each CT image one by one to find abnormalities, which is not only time-consuming but also easily affected by subjective factors. And with the rapid growth of medical imaging data, traditional methods are difficult to meet clinical needs, especially when a large number of cases need to be diagnosed quickly and processed. Traditional methods seem to be unable to cope with complex and large-scale data. Summary of the invention

[0003] The present invention provides a three-dimensional CT image lesion recognition method and system based on big data and deep learning to solve the problem of low efficiency and accuracy in three-dimensional CT image data analysis.

[0004] In a first aspect, the present invention provides a three-dimensional CT image lesion recognition method based on big data and deep learning, which is applied to a lesion recognition cloud server, wherein the lesion recognition cloud server is configured with a pre-trained organ image classification model and a CT image reconstruction model, and the lesion recognition cloud server uses containerization technology and is configured with multiple model containers according to organ types, and a basic lesion recognition model is uniformly deployed in each of the model containers, and the lesion recognition cloud server is connected to multiple clients in communication, and the method comprises the following steps:

[0005] Acquire a plurality of two-dimensional CT image data sets uploaded by the clients, and reconstruct the two-dimensional CT image data sets into three-dimensional CT image data sets using the CT image reconstruction model;

[0006] Inputting the three-dimensional CT image data set into the organ image classification model, and generating organ image classification results of all three-dimensional CT image data in the three-dimensional CT image data set through the organ image classification model;

[0007] According to the organ image classification result, all corresponding three-dimensional CT image data in the three-dimensional CT image data set are distributed and stored in the corresponding model container;

[0008] For each of the model containers, the basic lesion recognition model deployed in the model container is trained using the stored three-dimensional CT image data to obtain a target lesion recognition model;

[0009] Obtaining a lesion identification request sent by the client;

[0010] Extracting target two-dimensional CT image data from the lesion identification request, and reconstructing the target two-dimensional CT image data set into target three-dimensional CT image data using the CT image reconstruction model;

[0011] Inputting the target three-dimensional CT image data into the target lesion recognition model, and generating a target image classification result of the target three-dimensional CT image data through the organ image classification model;

[0012] Calling the target lesion recognition model in the corresponding model container based on the target image classification result;

[0013] The target two-dimensional CT image data and the target three-dimensional CT image data are input into the target lesion recognition model, a lesion recognition result is generated by the target lesion recognition model, and the lesion recognition result is fed back to the client that sent the lesion recognition request.

[0014] Optionally, the organ type includes multiple organ part categories and multiple organ blood vessel categories, and the step of distributing and storing all corresponding three-dimensional CT image data in the three-dimensional CT image data set into the corresponding model container according to the organ image classification result includes the following steps:

[0015] For any three-dimensional CT image data in the three-dimensional CT image data set, if the organ image classification result is any of the organ part categories, identifying whether there is a lesion annotation mask in the three-dimensional CT image data;

[0016] If the lesion annotation mask exists in the three-dimensional CT image data, preprocessing the three-dimensional CT image data based on the lesion annotation mask, and storing the preprocessed three-dimensional CT image data as annotated training set data in the model container corresponding to the same organ part category;

[0017] If the lesion annotation mask does not exist in the three-dimensional CT image data, the three-dimensional CT image data is preprocessed, and the preprocessed three-dimensional CT image data is stored as non-annotated training set data in the model container corresponding to the same organ part category.

[0018] Optionally, the method further comprises the following steps:

[0019] If the organ image classification result is any of the organ blood vessel categories, identifying whether there is a lesion annotation mask in the three-dimensional CT image data;

[0020] If the lesion annotation mask does not exist in the three-dimensional CT image data, the three-dimensional CT image data is not retained;

[0021] If the lesion marking mask exists in the three-dimensional CT image data, the three-dimensional CT image data is preprocessed based on the lesion marking mask, and the preprocessed three-dimensional CT image data is stored in the model container corresponding to the same organ and blood vessel category.

[0022] Optionally, the preprocessing of the three-dimensional CT image data based on the lesion annotation mask comprises the following steps:

[0023] Eliminate all null-value data voxels in the three-dimensional CT image data;

[0024] Identifying the annotated data voxels corresponding to the lesion annotation mask in the three-dimensional CT image data;

[0025] Calculating the distances between all the labeled data voxels, and calculating the average distance of all the labeled voxel distances;

[0026] An interpolation operation is performed on the lesion marking mask in the three-dimensional CT image data based on the spacing average value.

[0027] Optionally, for each of the model containers, using the stored three-dimensional CT image data to train the basic lesion recognition model deployed in the model container to obtain a target lesion recognition model comprises the following steps:

[0028] For each of the model containers corresponding to the organ part category, the stored annotated training set data is input into the basic lesion recognition model, and the basic lesion recognition model is trained using the annotated training set data to obtain an initial lesion recognition model;

[0029] Inputting the stored non-labeled training set data into the initial lesion recognition model, performing lesion recognition on the non-labeled training set data through the initial lesion recognition model, and generating a model labeling result;

[0030] Calculating the result confidence of the model annotation result based on the annotated training set data;

[0031] If the confidence of the result exceeds a preset confidence threshold, then generating new annotated training set data by combining the corresponding model annotation result and the unannotated training set data;

[0032] Using the expanded labeled training set data to train the initial lesion recognition model to obtain a target lesion recognition model;

[0033] For each of the model containers corresponding to the organ and blood vessel categories, all the stored three-dimensional CT image data are integrated into a training data set, and the number of data in the training data set is counted;

[0034] If the data quantity is greater than or equal to a preset quantity threshold, the basic lesion recognition model is trained using all the three-dimensional CT image data to obtain a target lesion recognition model;

[0035] If the amount of data is less than the amount threshold, the training data set is expanded using a data enhancement method;

[0036] The expanded training data set is used to train the basic lesion recognition model to obtain a target lesion recognition model.

[0037] Optionally, the step of expanding the training data set using a data enhancement method comprises the following steps:

[0038] Grid-encode all the training data in the original training data set into grid-encoded training data;

[0039] Divide all the grid coding training data equally into a first grid coding training data set and a second grid coding training data set according to the data quantity;

[0040] Copying the first grid coded training data set and the second grid coded training data set to obtain a first copied training data set and a second copied training data set;

[0041] For the first copy training data in the first copy training data set, selecting a grid code in the first copy training data as a first target grid code based on a preset first selection probability, and transforming all the first target grid codes using a preset coding vector change function;

[0042] For the second copy training data in the second copy training data set, selecting a grid code in the second copy training data as a second target grid code based on a preset second selection probability, and setting all the second target grid codes to zero;

[0043] Merging the first grid coded training data set, the second grid coded training data set, the first copy training data set and the second copy training data set to obtain the expanded training data set, and counting the number of expanded data of the expanded training data set;

[0044] If the amount of the expanded data is less than the data threshold, the above training data set expansion step is repeated until the amount of the expanded data in the training data set after expansion is greater than or equal to the data threshold.

[0045] Optionally, the method further comprises the following steps:

[0046] Creating a first shared volume among all first model containers corresponding to the organ part category;

[0047] Mounting the first shared volume to all the first model containers;

[0048] When the basic lesion recognition model in any one of the first model containers obtains the target lesion recognition model through model training, the model weight of the target lesion recognition model is saved to the first shared volume;

[0049] creating a second shared volume between all second model containers corresponding to the organ vessel class;

[0050] Mounting the second shared volume to all the second model containers;

[0051] When the basic lesion recognition model in any one of the second model containers obtains the target lesion recognition model through model training, the model weight of the target lesion recognition model is saved in the second shared volume.

[0052] Optionally, the step of inputting the three-dimensional CT image data set into the organ image classification model and generating organ image classification results of all three-dimensional CT image data in the three-dimensional CT image data set by the organ image classification model comprises the following steps:

[0053] For each three-dimensional CT image data in the three-dimensional CT image data set, counting the number of slices of the CT image slices in the three-dimensional CT image data;

[0054] If the number of slices is less than a preset slice number threshold, all the CT image slices are traversed, and the CT image slice with the highest average pixel value is input into the organ image classification model, and the organ image classification result of the three-dimensional CT image data is generated by the organ image classification model;

[0055] If the number of slices is greater than or equal to the slice number threshold, acquiring the data size of the three-dimensional CT image data;

[0056] Based on the data size, the three-dimensional CT image data is reduced in dimension into two-dimensional CT image data;

[0057] The two-dimensional CT image data is input into the organ image classification model, and the organ image classification result corresponding to the three-dimensional CT image data is generated by the organ image classification model.

[0058] Optionally, the dimensionality reduction processing of the three-dimensional CT image data into two-dimensional CT image data based on the data size comprises the following steps:

[0059] determining a specific axis of the three-dimensional CT image data;

[0060] If the data size is less than or equal to a preset size threshold, reducing the dimension of the three-dimensional CT image data into two-dimensional CT image data based on the specific axis and using a maximum intensity projection method;

[0061] If the data size is greater than the size threshold, the three-dimensional CT image data is reduced in dimension into two-dimensional CT image data based on the specific axis and using an axial slicing method.

[0062] In a second aspect, the present invention also provides a three-dimensional CT image lesion recognition system based on big data and deep learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the three-dimensional CT image lesion recognition method based on big data and deep learning as described in any one of the first aspects is implemented.

[0063] The beneficial effects of the present invention are:

[0064] 1. Automatically generate classification results through the organ image classification model, and distribute and store the 3D CT image data in the corresponding model container, which greatly reduces the time and error of manual operation and improves the efficiency of data processing. Use the automated process to train the basic lesion recognition model and generate the target lesion recognition model, realize the automatic update and optimization of the model, and further improve the recognition efficiency and accuracy.

[0065] Second, by preprocessing and storing the 3D CT image data with lesion annotation masks, the high quality of the training data is ensured, thereby improving the recognition accuracy of the model. The confidence of the model annotation results is calculated based on the annotated training set data, and the high-confidence results are selected to expand the training data set, further enhancing the reliability and stability of the model.

[0066] 3. Each model container is independently deployed and managed, allowing models of different organs or lesion types to be updated and optimized independently, enhancing the flexibility and scalability of the system. In the case of insufficient data, the training data set is expanded through data enhancement methods to ensure that the model can be effectively trained and applied under various data scales.

[0067] Fourth, the initial lesion recognition model is used to identify lesions in the unlabeled training set data and generate model annotation results, making full use of the unlabeled data and expanding the scale and diversity of the training data set. For the model container of the organ vascular category, all stored 3D CT image data are integrated, and corresponding training strategies are adopted according to the amount of data, which optimizes the utilization of data resources and the training effect of the model.

[0068] 5. Generate lesion recognition results through the target lesion recognition model and feed back the results to the client, providing fast and accurate diagnosis support, improving the efficiency and accuracy of clinical diagnosis. It can also process two-dimensional CT image data uploaded by multiple clients, support concurrent processing and multi-user access, and enhance the practicality and service capabilities of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a framework flow chart of a three-dimensional CT image lesion recognition method based on big data and deep learning in one of the embodiments of the present application.

[0070] Figure 2 This is a flowchart of the steps of a three-dimensional CT image lesion recognition method based on big data and deep learning in one embodiment of the present application. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0072] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0073] Reference Figure 1The three-dimensional CT image lesion recognition method based on big data and deep learning disclosed in the present invention is applied to a lesion recognition cloud server. The lesion recognition cloud server is configured with a pre-trained organ image classification model and a CT image reconstruction model. The lesion recognition cloud server uses containerization technology and is configured with multiple model containers according to the organ type. A basic lesion recognition model is uniformly deployed in each model container. The lesion recognition cloud server is connected to multiple clients for communication.

[0074] Reference Figure 2 , Figure 2 FIG. 1 is a flow chart of a method for identifying lesions in three-dimensional CT images based on big data and deep learning in one embodiment. It should be understood that although Figure 2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2 At least part of the steps in the above method may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps. Figure 2 As shown, the method for identifying lesions in three-dimensional CT images based on big data and deep learning disclosed in the present invention specifically includes the following steps:

[0075] S101. Acquire two-dimensional CT image datasets uploaded by multiple clients, and reconstruct the two-dimensional CT image datasets into three-dimensional CT image datasets using a CT image reconstruction model.

[0076] Among them, the client device is usually an imaging device or diagnostic workstation in a hospital, and multiple clients upload their collected two-dimensional CT image data to a centrally managed server. The two-dimensional CT image dataset uploaded by the client is usually stored in the DICOM (Digital Imaging and Communications in Medicine) format, which is a standard format for storing and transmitting medical images. The two-dimensional CT image dataset uploaded by each client usually contains multiple slices, which are scanned along a certain axis (usually the horizontal axis) of the patient's body. In order to ensure the integrity and consistency of the data, the uploaded data needs to be preprocessed, including checking the file format, verifying the data integrity, and preliminarily organizing and archiving the data. For example, you can write a Python script that uses the `pydicom` library to read and verify the metadata of each DICOM file to ensure that all slices are of the same size and arranged in the correct order.

[0077] Next, the CT image reconstruction model is used to reconstruct the 2D CT image dataset into a 3D CT image dataset. CT image reconstruction models are usually based on convolutional neural networks (CNN) or other deep learning techniques. These models have been trained with a large amount of training data and can effectively extract 3D structural information from 2D slices. The core of the reconstruction process is to combine multiple 2D slices according to their positions in space to generate a 3D volume dataset. Specifically, each 2D slice can be regarded as a matrix, and these matrices are stacked in the order of slices to form a 3D array. For example, assuming there are 100 256×256 pixel 2D CT slices, they can be stacked into a 256×256×100 3D array. At this point, the pre-trained CT image reconstruction model can be used to process this 3D array. The model will extract features through convolution operations and generate a high-resolution 3D CT image dataset through deconvolution operations. This process can be implemented using deep learning frameworks such as TensorFlow or PyTorch.

[0078] In the specific implementation process, the reconstructed three-dimensional CT image data set needs to be post-processed to ensure its quality and usability. The post-processing steps include image enhancement, noise removal, and contrast adjustment. Image enhancement technology can improve the clarity of three-dimensional CT images and make the details more prominent. Noise removal technology can reduce artifacts and noise generated during the reconstruction process, improve image quality and diagnostic accuracy. For example, median filtering or Gaussian filtering technology can be used to process three-dimensional CT image data. Contrast adjustment technology can make the boundaries of different tissues and organs clearer by adjusting the grayscale of the image.

[0079] S102. Input the three-dimensional CT image data set into the organ image classification model, and generate organ image classification results of all three-dimensional CT image data in the three-dimensional CT image data set through the organ image classification model.

[0080] Among them, the received three-dimensional CT image data set will be input into a pre-trained organ image classification model. This classification model is usually built based on deep learning technology, such as using convolutional neural network (CNN) to process three-dimensional image data. The classification model will process each three-dimensional CT image data, extract the features in the image, and classify the image into different organ categories based on these features. Specifically, the model will process the input image data layer by layer through multiple convolutional layers, pooling layers, and fully connected layers, and finally output a probability vector, indicating the probability that the image belongs to each organ category. Based on this probability vector, the category with the highest probability is selected as the classification result of the image. In this way, the organ category to which each three-dimensional CT image data belongs can be quickly and accurately determined, providing a basis for subsequent lesion identification.

[0081] S103. According to the organ image classification result, all corresponding three-dimensional CT image data in the three-dimensional CT image data set are distributed and stored in the corresponding model container.

[0082] Among them, according to the classification results generated by the organ image classification model, each data in the three-dimensional CT image data set is assigned to the corresponding model container. Each model container corresponds to a specific organ category, and a lesion recognition model for the organ is deployed inside the container. After the classification result is determined, the server will store each three-dimensional CT image data in the corresponding model container according to its classification result. In order to achieve efficient data distribution and storage, a distributed storage system can be used to distribute and store data on multiple physical nodes to ensure high availability and fast access to data. Each model container not only stores the three-dimensional CT image data of the corresponding category, but also stores related metadata, such as data source, timestamp, classification confidence, etc.

[0083] S104. For each model container, the basic lesion recognition model deployed in the model container is trained using the stored three-dimensional CT image data to obtain a target lesion recognition model.

[0084] Among them, a basic lesion recognition model is deployed in each model container. These models are usually built based on deep learning technology, such as using convolutional neural networks (CNN) or three-dimensional convolutional neural networks (3D-CNN). The basic lesion recognition model is trained using the three-dimensional CT image data stored in the model container. During the training process, the model will continuously adjust the parameters of the model according to the input data and the corresponding annotation information through the back propagation algorithm to minimize the error between the predicted results and the actual annotations. In order to improve the training efficiency and effect of the model, data enhancement techniques such as rotation, translation, and scaling can be used to expand the training data. In addition, cross-validation methods can also be used to ensure the generalization ability of the model on different data sets. After the training is completed, the basic lesion recognition model will be updated to the target lesion recognition model with higher recognition accuracy and stability.

[0085] S105. Obtain the lesion identification request sent by the client.

[0086] Among them, when the client needs to identify the lesion, it will send a lesion identification request to the server. This request usually contains the unique identifier of the target three-dimensional CT image data, as well as other relevant information, such as the request time, client ID, etc. After receiving the request, the server will perform preliminary verification and analysis to ensure the legitimacy and integrity of the request. If the request is valid, the server will extract the corresponding three-dimensional CT image data from the database and prepare to input it into the target lesion identification model for processing. In order to improve the efficiency of request processing, a queue mechanism can be used to queue all lesion identification requests in the order of arrival to ensure that each request can be responded to in a timely manner.

[0087] S106. Extract target two-dimensional CT image data from the lesion identification request, and reconstruct the target two-dimensional CT image data set into target three-dimensional CT image data using a CT image reconstruction model.

[0088] Among them, the first step of extracting the target two-dimensional CT image data is to parse the lesion identification request and obtain the patient's scanning parameters, including patient ID, scanning date, scanning part, etc. With this information, the corresponding two-dimensional CT image data set can be retrieved from the medical image storage system. The medical image storage system usually uses the DICOM standard to store image data, and each DICOM file contains a two-dimensional CT slice and related metadata. The retrieval process can obtain all relevant CT slices by querying the DICOM server, using the patient ID and scanning date as query conditions. The metadata of each slice contains information such as slice position and pixel spacing. Next, the target two-dimensional CT image data set is reconstructed into target three-dimensional CT image data using the CT image reconstruction model. The specific reconstruction process is the same as the detailed process of step S101.

[0089] S107. Input the target three-dimensional CT image data into the target lesion recognition model, and generate a target image classification result of the target three-dimensional CT image data through the organ image classification model.

[0090] The classification model processes the input 3D CT image data layer by layer through multiple convolutional layers, pooling layers, and fully connected layers, extracts features from the image, and classifies the image into different organ categories based on these features. The classification results are output in the form of a probability vector, indicating the probability that the image belongs to each organ category. Based on this probability vector, the category with the highest probability is selected as the classification result of the image, providing a basis for subsequent lesion identification.

[0091] S108. Call the target lesion recognition model in the corresponding model container based on the target image classification result.

[0092] The server will input the target 3D CT image data into the corresponding target lesion recognition model, and the model will process the input data to generate lesion recognition results. In order to ensure the accuracy of the recognition results, the output of the model can be post-processed, such as removing noise, smoothing boundaries, etc., to improve the quality of the final result.

[0093] S109. Input the target two-dimensional CT image data and the target three-dimensional CT image data into the target lesion recognition model, generate a lesion recognition result through the target lesion recognition model, and feed back the lesion recognition result to the client that sent the lesion recognition request.

[0094] Among them, the target lesion recognition model is usually a convolutional neural network (CNN) or a three-dimensional convolutional neural network (3D-CNN) based on deep learning. These models have been trained with a large amount of labeled data and have strong lesion recognition capabilities. Two-dimensional CT image data and three-dimensional CT image data are used as inputs of the model, and the model extracts features from these image data and analyzes them. The specific implementation method is to preprocess the two-dimensional CT image data first, including normalization and resizing, to ensure the consistency of the input data. Normalization is to scale the pixel values ​​to a fixed range (such as 0 to 1) to reduce the impact of data differences on model training. Resizing is to adjust all input images to the same size for batch processing. For three-dimensional CT image data, voxel (volume pixel) normalization is required, and cropping or padding is performed according to the input requirements of the model to ensure the consistency of the size of the input data.

[0095] After receiving the input data, the lesion recognition model will extract and analyze the data. Taking the convolutional neural network as an example, the model processes the input image through multiple convolutional layers and pooling layers. The convolution layer performs convolution operations on the image by sliding the convolution kernel to extract the local features of the image. The pooling layer reduces the size of the feature map by downsampling operations while retaining important features. For a three-dimensional convolutional neural network, the convolution and pooling operations are performed in a three-dimensional space, which can better capture three-dimensional structural information. After multiple layers of convolution and pooling operations, the model generates a high-dimensional feature map. Subsequently, the high-dimensional feature map is mapped to the classification space through a fully connected layer to generate a lesion recognition result. Specific implementation method The Softmax activation function can be used to convert the feature values ​​output by the model into a probability distribution, thereby obtaining the probability of each category. In this way, the category probability distribution of the lesion can be generated to determine the type and location of the lesion. The output results of the model usually include the category (such as tumor, nodule, etc.), location (such as coordinate information) and confidence (such as probability value) of the lesion.

[0096] After the lesion identification result is generated, the lesion identification result needs to be fed back to the client that sent the lesion identification request. The feedback process includes data format conversion, result encapsulation and data transmission. First, the lesion identification result is converted into a standardized data format, such as JSON or XML format, so that it can be parsed and displayed by the client. The specific implementation method is to organize the category, location and confidence information in the lesion identification result into the form of key-value pairs and encapsulate them in a standard format. For example, the lesion category can be used as the key and the category name as the value; the lesion location can be used as the key and the coordinate information as the value; the confidence is used as the key and the probability value as the value. In this way, the data format of the lesion identification result can be ensured to be consistent, which is convenient for the client to parse. Subsequently, the encapsulated lesion identification result is transmitted to the client via the network. The specific implementation method can use the HTTP protocol to send data to the URL address specified by the client through a POST request. In order to ensure the security of data transmission, SSL / TLS encryption technology can be used to encrypt the transmitted data to prevent data leakage.

[0097] After the client receives the lesion recognition result, it can parse and display the result. The parsing process includes extracting the lesion recognition results from a standardized data format and converting these results into a form that is easy to understand and display. The specific implementation method is to parse the data in JSON or XML format, extract the lesion category, location and confidence information, and display this information on the user interface. For example, the location of the lesion can be marked on a two-dimensional CT image or a three-dimensional CT image, and the category and confidence of the lesion can be displayed. In order to improve the user experience, a graphical user interface (GUI) technology can be used to display the lesion recognition results in a graphical way. For example, a bounding box of the lesion can be drawn on a CT image, and the category and confidence of the lesion can be displayed next to the bounding box. In this way, the doctor can intuitively understand the type and location of the lesion and improve the efficiency and accuracy of the diagnosis. The display of the lesion recognition results can also be combined with interactive functions, allowing the doctor to perform further operations on the lesion, such as zooming in, zooming out, rotating, etc. In this way, a more flexible and convenient diagnostic tool can be provided to doctors to improve the accuracy and efficiency of clinical diagnosis.

[0098] In one embodiment, the organ type includes multiple organ part categories and multiple organ blood vessel categories, and step S103 specifically includes the following steps:

[0099] For any three-dimensional CT image data in the three-dimensional CT image data set, if the organ image classification result is any organ part category, then identify whether there is a lesion annotation mask in the three-dimensional CT image data;

[0100] If there is a lesion annotation mask in the three-dimensional CT image data, the three-dimensional CT image data is preprocessed based on the lesion annotation mask, and the preprocessed three-dimensional CT image data is stored as the annotation training set data in the model container corresponding to the same organ part category;

[0101] If there is no lesion annotation mask in the three-dimensional CT image data, the three-dimensional CT image data is preprocessed, and the preprocessed three-dimensional CT image data is stored as non-annotated training set data in a model container corresponding to the same organ part category.

[0102] In this embodiment, the organ image classification model is usually constructed based on deep learning technology, such as using a convolutional neural network (CNN) to process three-dimensional image data. The model processes the input three-dimensional CT image data, extracts the features in the image, and classifies the image into different organ categories based on these features. Specifically, the model processes the input image data layer by layer through multiple convolutional layers, pooling layers, and fully connected layers, and finally outputs a probability vector, indicating the probability that the image belongs to each organ category. Based on this probability vector, the category with the highest probability is selected as the classification result of the image.

[0103] Once the organ image classification results are determined, it is necessary to further identify whether there is a lesion annotation mask in the 3D CT image data. The lesion annotation mask is a mark, usually manually annotated by professional doctors, used to indicate the specific location and morphology of the lesion in the image. The presence of a lesion annotation mask can be identified by checking the metadata of the image data. If there is a lesion annotation mask, the 3D CT image data needs to be preprocessed. The specific preprocessing method can be adjusted according to the characteristics of the image data and the requirements of the target model. The preprocessed 3D CT image data will be stored as annotated training set data in the model container corresponding to the same organ part category. The annotated training set data contains image data and the corresponding lesion annotation mask, which can be used to train the supervised learning model and improve the recognition accuracy of the model. If there is no lesion annotation mask in the 3D CT image data, the image data also needs to be preprocessed. The preprocessed 3D CT image data will be stored as unannotated training set data in the model container corresponding to the same organ part category. The unannotated training set data only contains image data and no lesion annotation mask. Through the above process, the existing image data can be fully utilized to improve the generalization ability and robustness of the model.

[0104] In this embodiment, the 3D CT image lesion recognition method based on big data and deep learning also includes the following steps:

[0105] If the organ image classification result is any organ vascular category, then identify whether there is a lesion annotation mask in the three-dimensional CT image data;

[0106] If there is no lesion annotation mask in the three-dimensional CT image data, the three-dimensional CT image data will not be retained;

[0107] If there is a lesion annotation mask in the three-dimensional CT image data, the three-dimensional CT image data is preprocessed based on the lesion annotation mask, and the preprocessed three-dimensional CT image data is stored in a model container corresponding to the same organ and blood vessel category.

[0108] In this embodiment, the step of preprocessing the three-dimensional CT image data based on the lesion annotation mask specifically includes the following steps:

[0109] Eliminate all the null value data voxels in the three-dimensional CT image data;

[0110] Identify the annotated data voxels corresponding to the lesion annotation mask in the three-dimensional CT image data;

[0111] Calculate the distance between all the labeled data voxels, and calculate the average distance of all the labeled voxel distances;

[0112] The lesion annotation mask in the 3D CT image data is interpolated based on the average value of the spacing.

[0113] In this embodiment, null data voxels refer to voxels without valid information in the three-dimensional CT image data, which may be caused by noise during the scanning process, data transmission errors or other factors. Eliminating these null voxels can improve the quality and processing efficiency of the data. In specific implementation, it can be determined whether it is a null voxel by checking the grayscale value of each voxel. Usually, the grayscale value of the null voxel will be a special value, such as zero or a negative number. By traversing all voxels in the three-dimensional CT image data and eliminating these null voxels, it can be ensured that the subsequent processing steps are only performed on valid data.

[0114] Then identify the annotated data voxels corresponding to the lesion annotation mask in the three-dimensional CT image data. The lesion annotation mask is used to indicate the specific location and morphology of the lesion in the image, and the annotated data voxels refer to the voxels marked as the lesion area in the lesion annotation mask. To identify these annotated data voxels, the data of the lesion annotation mask can be read, and all the voxels marked as lesions can be extracted. In specific implementation, the lesion annotation mask can be aligned with the original three-dimensional CT image data, and then all the voxels in the mask can be traversed to record the location and grayscale value of all the voxels marked as lesions.

[0115] Then, the labeled voxel spacing between all labeled data voxels is calculated, and the spacing average of all labeled voxel spacing is calculated. The labeled voxel spacing refers to the distance between two adjacent labeled data voxels in three-dimensional space. To calculate the labeled voxel spacing, all labeled data voxels can be traversed to calculate the Euclidean distance between each pair of adjacent voxels. In specific implementation, the following formula can be used to calculate the distance between two voxels:

[0116]

[0117] Among them, (x1, y1, z1) and (x2, y2, z2) are the coordinates of two voxels respectively. By calculating the distances between all the annotated data voxels and taking the average of these distances, the spacing average of the annotated voxel spacing can be obtained. This average value can reflect the density of the voxel distribution in the lesion area. Finally, the lesion annotation mask in the three-dimensional CT image data is interpolated based on the spacing average value. The purpose of the interpolation operation is to generate new voxels between the existing annotated data voxels to improve the spatial resolution of the lesion annotation mask. In specific implementation, a three-dimensional interpolation algorithm such as trilinear interpolation or cubic spline interpolation can be used. The interpolation algorithm generates new voxels based on the position and grayscale value of the existing annotated data voxels and inserts these voxels into the original data. The key to the interpolation operation is to determine the interpolation step size and range based on the spacing average value to ensure that the generated voxels can accurately reflect the morphology and characteristics of the lesion area. Through the interpolation operation, the spatial resolution of the lesion annotation mask can be significantly improved, making it more detailed and accurate.

[0118] In this embodiment, since there are significant differences in anatomical structure, lesion characteristics and diagnostic requirements between organ parts and blood vessels, data quality and analysis accuracy can be optimized more specifically by classifying and processing them separately; for organ part categories, the processing method may focus on the identification of the overall structure and the division of functional areas, while for organ blood vessel categories, the processing method pays more attention to the morphology, branches and fine identification of lesion areas of blood vessels. Through this classification and processing method, it can be ensured that each type of data can obtain the most suitable technical means and algorithm support during the preprocessing, labeling and analysis process, thereby improving the utilization efficiency of image data and the accuracy of the final diagnosis.

[0119] In one implementation manner, step S104 specifically includes the following steps:

[0120] For each model container corresponding to an organ part category, the stored labeled training set data is input into the basic lesion recognition model, and the basic lesion recognition model is trained using the labeled training set data to obtain an initial lesion recognition model;

[0121] Inputting the stored non-labeled training set data into the initial lesion recognition model, performing lesion recognition on the non-labeled training set data through the initial lesion recognition model, and generating a model annotation result;

[0122] Calculate the confidence of the model annotation results based on the annotated training set data;

[0123] If the confidence of the result exceeds the preset confidence threshold, the corresponding model annotation result and the unannotated training set data are combined to generate new annotated training set data;

[0124] Use the expanded labeled training set data to train the initial lesion recognition model to obtain the target lesion recognition model;

[0125] For each model container corresponding to each organ vascular category, all stored three-dimensional CT image data are integrated into a training data set, and the number of data in the training data set is counted;

[0126] If the data quantity is greater than or equal to a preset quantity threshold, all three-dimensional CT image data are used to train a basic lesion recognition model to obtain a target lesion recognition model;

[0127] If the amount of data is less than the quantity threshold, the training data set is expanded using data augmentation methods;

[0128] The expanded training data set is used to train the basic lesion recognition model to obtain the target lesion recognition model.

[0129] In this embodiment, for each model container corresponding to an organ part category, first, the stored labeled training set data is input into the basic lesion recognition model, and the basic lesion recognition model is trained using these labeled training set data to obtain an initial lesion recognition model. During specific implementation, the model parameters are gradually adjusted to improve the accuracy of lesion recognition by learning the lesion features in the labeled data. Through multiple iterations of training, the model can accurately identify the lesion area in the labeled training set, and finally obtain the initial lesion recognition model. Next, the stored non-labeled training set data is input into the initial lesion recognition model, and the non-labeled training set data is subjected to lesion recognition by the initial lesion recognition model, and the model annotation result is generated. The non-labeled training set data is a three-dimensional CT image data without lesion annotation information. After these data are input into the initial lesion recognition model, the model will perform lesion recognition on these data according to the features learned from the labeled training set data.

[0130] Then, the confidence of the model annotation results is calculated based on the annotated training set data. In specific implementation, the accuracy and confidence of the model annotation results can be calculated by comparing the model annotation results with the actual lesion area in the annotated training set data. A variety of methods can be used for confidence calculation, such as cross-validation or ROC curve analysis. Through these methods, the performance of the model on the non-annotated training set data can be evaluated, and the confidence value of each model annotation result can be calculated. The higher the confidence value, the higher the degree of trust the model has in its predicted results, and vice versa. The purpose of calculating the confidence of the results is to screen out high-confidence model annotation results to improve the quality of subsequent training data.

[0131] If the confidence of the result exceeds the preset confidence threshold, new annotated training set data is generated by combining the corresponding model annotation results and non-annotated training set data. The preset confidence threshold is a pre-set value used to screen out high-confidence model annotation results. In specific implementation, the model annotation results with confidence values ​​higher than the threshold can be combined with the corresponding non-annotated training set data to generate new annotated training set data. These new annotated training set data contain lesion information automatically annotated by the model, and this annotation information has a high confidence. In this way, the scale of the annotated training set data can be expanded, the diversity and representativeness of the data can be improved, and richer samples can be provided for subsequent training.

[0132] The initial lesion recognition model is trained using the expanded labeled training set data to obtain the target lesion recognition model. In specific implementation, the expanded labeled training set data can be input into the initial lesion recognition model again for multiple iterative training. After each iterative training, the parameters of the model are adjusted according to the new training data to improve the accuracy and robustness of lesion recognition. In this way, the model can gradually learn more lesion features and improve its generalization ability on different data sets. Finally, after multiple iterative training, a target lesion recognition model with better performance can be obtained, which has significant improvements in the accuracy and robustness of lesion recognition.

[0133] For each model container corresponding to an organ and blood vessel category, in specific implementation, all three-dimensional CT image data related to the organ and blood vessel category can be integrated to form a complete training data set. Then, the number of data in the training data set is counted to determine whether it meets the needs of subsequent training. The number of data can be counted through a simple counting operation, and the statistical results will be used to determine whether data enhancement operations are required.

[0134] If the amount of data is greater than or equal to the preset quantity threshold, all three-dimensional CT image data are used to train the basic lesion recognition model to obtain the target lesion recognition model. The preset quantity threshold is a pre-set value used to determine whether the size of the training data set is large enough. In specific implementation, if the amount of data is greater than or equal to the threshold, these data can be directly input into the basic lesion recognition model for training. The training process is similar to the training process of the aforementioned labeled training set data. Through multiple iterative training, the model can learn the lesion characteristics of the vascular category and finally obtain the target lesion recognition model.

[0135] If the number of data is less than the number threshold, the training data set is expanded by using a data enhancement method. The data enhancement method is a technique for generating more samples by transforming and processing existing data. In specific implementation, a variety of data enhancement methods can be used, such as rotation, translation, scaling, noise addition, etc. Through these methods, more samples can be generated to expand the size of the training data set. The purpose of data enhancement is to improve the diversity and representativeness of the training data to enhance the generalization ability of the model. The basic lesion recognition model is trained with the expanded training data set to obtain the target lesion recognition model. In specific implementation, the expanded training data set can be input into the basic lesion recognition model for multiple iterative training. The training process is similar to the training process of the aforementioned labeled training set data. Through multiple iterative training, the model can learn more lesion characteristics of vascular categories, and finally obtain a target lesion recognition model with better performance. In this way, it can be ensured that even in the case of insufficient data, the model can still obtain sufficient training samples, thereby improving its performance in practical applications.

[0136] In one embodiment, the step of expanding the training data set using a data augmentation method specifically includes the following steps:

[0137] Grid-encode all training data in the original training data set into grid-encoded training data;

[0138] According to the amount of data, all grid coding training data are evenly divided into a first grid coding training data set and a second grid coding training data set;

[0139] Copying the first grid coded training data set and the second grid coded training data set to obtain a first copied training data set and a second copied training data set;

[0140] For the first copy training data in the first copy training data set, a grid code in the first copy training data is selected as a first target grid code based on a preset first selection probability, and all first target grid codes are transformed using a preset coding vector change function;

[0141] For the second copy training data in the second copy training data set, selecting a grid code in the second copy training data as a second target grid code based on a preset second selection probability, and setting all second target grid codes to zero;

[0142] Merging the first grid coded training data set, the second grid coded training data set, the first copy training data set, and the second copy training data set to obtain an expanded training data set, and counting the number of expanded data in the expanded training data set;

[0143] If the amount of expanded data is less than the data threshold, the above training data set expansion steps are repeated until the amount of expanded data in the expanded training data set is greater than or equal to the data threshold.

[0144] In this embodiment, grid coding is a method of discretizing continuous spatial data by dividing the data into grid units of a fixed size and assigning a unique coding value to each grid unit. The original training data set contains three-dimensional CT image data, which can be divided into grid units of size 1X1X1, and a unique coding value is assigned to each grid unit. In this way, each data point in the original training data set will be converted into a corresponding grid code to form grid coded training data. Then, all grid coded training data are evenly divided into a first grid coded training data set and a second grid coded training data set according to the number of data. In specific implementation, the total number of grid coded training data can be counted first, and then these data can be evenly divided into two parts.

[0145] Next, all the data in the first grid code training data set and the second grid code training data set can be directly copied to generate two new training data sets, which are respectively called the first copy training data set and the second copy training data set. The copy operation can be implemented by a simple data copy method, such as using an array copy function in a programming language. For the first copy training data in the first copy training data set, a first selection probability value, such as 0.5, can be set first, and then a part of the grid codes in the first copy training data set can be randomly selected as the first target grid code according to this probability value. The selection operation can be implemented by a random number generation function, such as using a random number generation function in a programming language to generate a random number between 0 and 1. If the random number is less than or equal to 0.5, the corresponding grid code is selected as the target grid code. After the selection is completed, all selected target grid codes can be transformed using a preset code vector change function. The code vector change function can be a linear transformation function, such as multiplying each grid code value by a constant factor to change its value.

[0146] For the second copy training data in the second copy training data set, a second selection probability value, such as 0.3, can be set first, and then a part of the grid codes in the second copy training data set can be randomly selected as the second target grid codes according to the probability value. The selection operation can be implemented by a method similar to the above. After the selection is completed, the values ​​of all selected target grid codes can be set to zero. The zeroing operation can be implemented by a simple assignment operation, such as assigning the corresponding grid code value to 0 in a programming language.

[0147] The first grid coded training data set, the second grid coded training data set, the first copy training data set and the second copy training data set are merged to obtain an expanded training data set. The merging operation can be implemented by a simple data splicing method. After the merging is completed, the number of data in the expanded training data set can be counted to determine the scale of the expanded data. The statistical operation can be implemented by a simple counting function. If the number of data is less than the threshold, it is necessary to repeat the aforementioned training data set expansion steps, and continue to perform data copying, encoding transformation and zeroing operations until the number of data in the expanded training data set reaches or exceeds the threshold. In this way, it can be ensured that the final training data set contains enough data samples, thereby improving the effect and generalization ability of model training.

[0148] In one embodiment, the 3D CT image lesion recognition method based on big data and deep learning further includes the following steps:

[0149] creating a first shared volume between all first model containers corresponding to the organ part category;

[0150] Mount the first shared volume to all first model containers;

[0151] When the basic lesion recognition model in any first model container obtains the target lesion recognition model through model training, the model weight of the target lesion recognition model is saved to the first shared volume;

[0152] creating a second shared volume between all second model vessels corresponding to the organ vessel class;

[0153] Mount the second shared volume to all second model containers;

[0154] When the basic lesion recognition model in any second model container obtains the target lesion recognition model through model training, the model weight of the target lesion recognition model is saved in the second shared volume.

[0155] In this embodiment, if Figure 1As shown, first you need to determine which model containers belong to the organ part category. These containers are used to handle different organ parts, such as heart, liver, lungs, etc. Next, create a shared volume, which can be understood as a shared storage space that all related model containers can access. The creation of shared volumes can be achieved through container orchestration tools such as Docker Swarm or Kubernetes. This shared volume will be used to store and share the data and model weights generated during model training, thereby achieving data sharing and collaboration between model containers. The shared volume created earlier needs to be mounted to each model container so that they can access the data in the shared volume. In Docker, you can mount a shared volume to a specified directory of the container using the `-v` option when starting the container. For example, assuming there are three model containers, namely `model_container1`, `model_container2`, and `model_container3`, you can use the following command to start and mount the shared volume:

[0156] docker run-d --name model_container1-v first_shared_volume: / mnt / shared_volume model_image

[0157] docker run-d --name model_container2-v first_shared_volume: / mnt / shared_volume model_image

[0158] docker run-d --name model_container3-v first_shared_volume: / mnt / shared_volume model_image

[0159] In this way, all model containers can read and write data in the shared volume by accessing the / mnt / shared_volume directory. This approach ensures data consistency and sharing between various model containers, thereby improving the efficiency and collaboration of model training.

[0160] When the basic lesion recognition model in a model container completes training and generates a target lesion recognition model, the model's weight file needs to be saved to the shared volume. A model weight file is usually a file containing model parameters, such as a file in `.h5` format in TensorFlow. In the model training code, this can be achieved by specifying the file save path as a directory in the shared volume. After training is completed, the model weight file will be saved in the shared volume, and other model containers can access and use this file. This approach ensures the sharing and reuse of model weights, enables model transfer learning between different model containers, and improves the efficiency of model training and reasoning.

[0161] Similar to creating the first shared volume, it is necessary to determine which model containers belong to the organ vascular category. These containers are used to process different vascular sites, such as liver vessels, heart vessels, and brain vessels. Next, create a new shared volume, which is used to store and share data and model weights between the vascular category model containers. This shared volume will be used for data sharing and collaboration between the organ vascular category model containers to ensure that each container can access and share training data and model weights. Mount the second shared volume created earlier to each vascular category model container so that they can access the data in the shared volume. Similar to the first shared volume, in Docker, you can mount a shared volume to the specified directory of the container through the `-v` option when starting the container. When the basic lesion recognition model in a vascular category model container completes training and generates a target lesion recognition model, the model weight file needs to be saved to the second shared volume. After training is completed, the model weight file will be saved in the second shared volume, and other vascular category model containers can access and use this file.

[0162] In one implementation manner, step S102 specifically includes the following steps:

[0163] For each three-dimensional CT image data in the three-dimensional CT image data set, counting the number of slices of the CT image slices in the three-dimensional CT image data;

[0164] If the number of slices is less than a preset slice number threshold, all CT image slices are traversed, and the CT image slice with the highest average pixel value is input into the organ image classification model, and the organ image classification result of the three-dimensional CT image data is generated by the organ image classification model;

[0165] If the number of slices is greater than or equal to the slice number threshold, the data size of the three-dimensional CT image data is obtained;

[0166] Based on the data size, the three-dimensional CT image data is reduced in dimension into two-dimensional CT image data;

[0167] The two-dimensional CT image data is input into the organ image classification model, and the organ image classification results corresponding to the three-dimensional CT image data are generated by the organ image classification model.

[0168] In this embodiment, the three-dimensional CT image data is composed of multiple continuous two-dimensional CT image slices, and the process of counting the number of slices is to calculate the total number of these two-dimensional slices. The specific implementation method is to read the metadata or header file of the three-dimensional CT image data, which usually contains the basic information of the image data, including the number of slices, the size of each slice, the pixel spacing, etc. After reading this information, the total number of slices can be directly obtained. If the number of slices is less than the preset slice number threshold, all CT image slices are traversed, and the CT image slice with the highest average pixel value is input into the organ image classification model. The preset slice number threshold is a value set based on experience or actual needs, which is used to determine whether the three-dimensional CT image data requires special processing.

[0169] The process of traversing all CT image slices is to read each slice one by one and calculate its average pixel value. The calculation formula for the average pixel value is: add all the pixel values ​​in the slice and then divide it by the total number of pixels. By comparing the average pixel values ​​of all slices, the slice with the highest average pixel value can be found. This slice usually contains the clearest and most representative image information. The slice is input into a pre-trained organ image classification model, and the model will analyze and identify the input slice to generate the organ image classification result of the three-dimensional CT image data. The specific implementation method is to normalize the slice data, scale the pixel value to the range required by the model input, and then input it into the model. The model extracts image features and classifies them through deep learning technologies such as convolutional neural networks, and outputs the probability distribution of organ categories.

[0170] If the number of slices is greater than or equal to the slice number threshold, the data size of the three-dimensional CT image data is obtained. The data size includes the width, height and total number of slices of each slice. The specific implementation method is to obtain these size information by reading the metadata or header file of the three-dimensional CT image data. Assuming that a certain three-dimensional CT image data contains 200 slices, and the size of each slice is 512x512 pixels, the data size is (200, 512, 512). Then, based on the data size, the three-dimensional CT image data is reduced to two-dimensional CT image data. The specific implementation method is to project the three-dimensional data along a certain axis through the maximum projection method or the average projection method. The maximum projection method is to select the maximum value at each pixel position along the axial direction to form a two-dimensional image; the average projection method is to calculate the average value at each pixel position along the axial direction to form a two-dimensional image. For example, for three-dimensional CT image data containing 200 slices, maximum projection can be performed along the slice axis direction, and the maximum value of the 200 pixel values ​​at each pixel position is used as the projected pixel value to generate a 512x512 two-dimensional image.

[0171] The two-dimensional CT image data after dimensionality reduction is input into the organ image classification model, and the organ image classification results corresponding to the three-dimensional CT image data are generated by the organ image classification model. After normalization, the two-dimensional CT image data is input into the pre-trained organ image classification model. The model extracts and analyzes the input two-dimensional image through deep learning technologies such as convolutional neural networks to generate the probability distribution of organ categories. The specific implementation method is to input the two-dimensional image into the convolution layer of the model, and the convolution layer extracts the local features of the image through a sliding convolution kernel, and then downsamples through the pooling layer to reduce the size of the feature map and retain important features. After multiple layers of convolution and pooling operations, the model will generate a high-dimensional feature map. The high-dimensional feature map is mapped to the classification space through the fully connected layer to generate the probability distribution of the organ category. Through the Softmax activation function, the output feature value is converted into a probability value to obtain the probability of each category. Finally, the model outputs the organ image classification result, including the category of the organ and the corresponding probability value.

[0172] In one implementation manner, the step of reducing the dimension of the three-dimensional CT image data into two-dimensional CT image data based on the data size specifically includes the following steps:

[0173] Determine a specific axis of three-dimensional CT image data;

[0174] If the data size is less than or equal to a preset size threshold, the 3D CT image data is reduced in dimension into 2D CT image data based on a specific axis and using a maximum intensity projection method;

[0175] If the data size is greater than the size threshold, the three-dimensional CT image data is reduced to two-dimensional CT image data based on a specific axis and using an axial slicing method.

[0176] In this embodiment, the selection of a specific axis generally depends on the needs of the medical image and the purpose of analysis, for example, the Z axis is selected for projection or slicing. If the data size is less than or equal to a preset size threshold, the three-dimensional CT image data is reduced in dimension to two-dimensional CT image data based on the specific axis and using the maximum intensity projection method. Maximum intensity projection (MIP) is a commonly used dimensionality reduction method that projects all voxel values ​​along a specific axis and selects the maximum value to generate a two-dimensional image. If the data size is greater than the size threshold, the three-dimensional CT image data is reduced in dimension to two-dimensional CT image data based on the specific axis and using the axial slicing method. The axial slicing method is to cut the three-dimensional data into multiple two-dimensional slices along a specific axis and process them separately. The two-dimensional CT image data after dimensionality reduction is input into a pre-trained organ image classification model to generate the corresponding organ image classification results.

[0177] Different data sizes correspond to different processing complexities. For smaller data sets, the maximum intensity projection (MIP) method can quickly generate a two-dimensional image, because MIP only needs to traverse the data set once and select the maximum value, and the computational complexity is low. For larger data sets, the axial slicing method can avoid processing a large amount of data at one time, split it into multiple smaller slices and process them one by one, reduce the load of a single calculation, and improve the overall processing efficiency. On the other hand, the maximum intensity projection method is suitable for situations where the data size is small, and can retain key high-intensity information. The generated two-dimensional image has a high contrast and is suitable for quick preview and preliminary analysis. For larger data sets, the axial slicing method can retain more detailed information, avoid information loss, and the generated two-dimensional slices are more suitable for detailed analysis and diagnosis.

[0178] The present invention also discloses a three-dimensional CT image lesion recognition system based on big data and deep learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the three-dimensional CT image lesion recognition method based on big data and deep learning described in any one of the above-mentioned embodiments is implemented.

[0179] Among them, the processor can adopt a central processing unit (CPU). Of course, according to actual usage, other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be adopted. The general-purpose processor can adopt a microprocessor or any conventional processor, etc., and this application does not impose any restrictions on this.

[0180] Among them, the memory can be an internal storage unit of a computer device, such as a hard disk or memory of a computer device, or an external storage device of a computer device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD) or a flash memory card (FC) equipped on the computer device, etc., and the memory can also be a combination of an internal storage unit and an external storage device of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or is to be output, and this application does not impose any restrictions on this.

[0181] A person skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as above, which are not provided in detail for the sake of simplicity.

[0182] One or more embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A 3D CT image lesion recognition method based on big data and deep learning, characterized in that: Applied to a lesion recognition cloud server, the lesion recognition cloud server is configured with a pre-trained organ image classification model and a CT image reconstruction model, the lesion recognition cloud server uses containerization technology and is configured with multiple model containers according to organ types, each of the model containers is uniformly deployed with a basic lesion recognition model, the lesion recognition cloud server is connected to multiple clients, and the method comprises the following steps: Acquire a plurality of two-dimensional CT image data sets uploaded by the clients, and reconstruct the two-dimensional CT image data sets into three-dimensional CT image data sets using the CT image reconstruction model; Inputting the three-dimensional CT image data set into the organ image classification model, and generating organ image classification results of all three-dimensional CT image data in the three-dimensional CT image data set through the organ image classification model; According to the organ image classification result, all corresponding three-dimensional CT image data in the three-dimensional CT image data set are distributed and stored in the corresponding model container; For each of the model containers, the basic lesion recognition model deployed in the model container is trained using the stored three-dimensional CT image data to obtain a target lesion recognition model; Obtaining a lesion identification request sent by the client; Extracting target two-dimensional CT image data from the lesion identification request, and reconstructing the target two-dimensional CT image data set into target three-dimensional CT image data using the CT image reconstruction model; Inputting the target three-dimensional CT image data into the target lesion recognition model, and generating a target image classification result of the target three-dimensional CT image data through the organ image classification model; Calling the target lesion recognition model in the corresponding model container based on the target image classification result; The target two-dimensional CT image data and the target three-dimensional CT image data are input into the target lesion recognition model, a lesion recognition result is generated by the target lesion recognition model, and the lesion recognition result is fed back to the client that sent the lesion recognition request.

2. The method for lesion recognition based on three-dimensional CT images based on big data and deep learning according to claim 1, characterized in that: The organ type includes multiple organ part categories and multiple organ blood vessel categories, and the step of distributing and storing all corresponding three-dimensional CT image data in the three-dimensional CT image data set into the corresponding model container according to the organ image classification result includes the following steps: For any three-dimensional CT image data in the three-dimensional CT image data set, if the organ image classification result is any of the organ part categories, identifying whether there is a lesion annotation mask in the three-dimensional CT image data; If the lesion annotation mask exists in the three-dimensional CT image data, preprocessing the three-dimensional CT image data based on the lesion annotation mask, and storing the preprocessed three-dimensional CT image data as annotated training set data in the model container corresponding to the same organ part category; If the lesion annotation mask does not exist in the three-dimensional CT image data, the three-dimensional CT image data is preprocessed, and the preprocessed three-dimensional CT image data is stored as non-annotated training set data in the model container corresponding to the same organ part category.

3. The method for lesion recognition based on three-dimensional CT images based on big data and deep learning according to claim 2, characterized in that: The method further comprises the steps of: If the organ image classification result is any of the organ blood vessel categories, identifying whether there is a lesion annotation mask in the three-dimensional CT image data; If the lesion annotation mask does not exist in the three-dimensional CT image data, the three-dimensional CT image data is not retained; If the lesion marking mask exists in the three-dimensional CT image data, the three-dimensional CT image data is preprocessed based on the lesion marking mask, and the preprocessed three-dimensional CT image data is stored in the model container corresponding to the same organ and blood vessel category.

4. The method for lesion recognition based on three-dimensional CT images based on big data and deep learning according to claim 3, characterized in that: The preprocessing of the three-dimensional CT image data based on the lesion annotation mask comprises the following steps: Eliminate all null-value data voxels in the three-dimensional CT image data; Identifying the annotated data voxels corresponding to the lesion annotation mask in the three-dimensional CT image data; Calculating the distances between all the labeled data voxels, and calculating the average distance of all the labeled voxel distances; An interpolation operation is performed on the lesion marking mask in the three-dimensional CT image data based on the spacing average value.

5. The method for lesion recognition based on three-dimensional CT images based on big data and deep learning according to claim 3, characterized in that: For each of the model containers, using the stored three-dimensional CT image data to train the basic lesion recognition model deployed in the model container to obtain a target lesion recognition model comprises the following steps: For each of the model containers corresponding to the organ part category, the stored annotated training set data is input into the basic lesion recognition model, and the basic lesion recognition model is trained using the annotated training set data to obtain an initial lesion recognition model; Inputting the stored non-labeled training set data into the initial lesion recognition model, performing lesion recognition on the non-labeled training set data through the initial lesion recognition model, and generating a model labeling result; Calculating the result confidence of the model annotation result based on the annotated training set data; If the confidence of the result exceeds a preset confidence threshold, then generating new annotated training set data by combining the corresponding model annotation result and the unannotated training set data; Using the expanded labeled training set data to train the initial lesion recognition model to obtain a target lesion recognition model; For each of the model containers corresponding to the organ and blood vessel categories, all the stored three-dimensional CT image data are integrated into a training data set, and the number of data in the training data set is counted; If the data quantity is greater than or equal to a preset quantity threshold, the basic lesion recognition model is trained using all the three-dimensional CT image data to obtain a target lesion recognition model; If the amount of data is less than the amount threshold, the training data set is expanded using a data enhancement method; The expanded training data set is used to train the basic lesion recognition model to obtain a target lesion recognition model.

6. The method for lesion recognition based on three-dimensional CT images based on big data and deep learning according to claim 5, characterized in that: The method of using data enhancement to expand the training data set includes the following steps: Grid-encode all the training data in the original training data set into grid-encoded training data; Divide all the grid coding training data equally into a first grid coding training data set and a second grid coding training data set according to the data quantity; Copying the first grid coded training data set and the second grid coded training data set to obtain a first copied training data set and a second copied training data set; For the first copy training data in the first copy training data set, selecting a grid code in the first copy training data as a first target grid code based on a preset first selection probability, and transforming all the first target grid codes using a preset coding vector change function; For the second copy training data in the second copy training data set, selecting a grid code in the second copy training data as a second target grid code based on a preset second selection probability, and setting all the second target grid codes to zero; Merging the first grid coded training data set, the second grid coded training data set, the first copy training data set and the second copy training data set to obtain the expanded training data set, and counting the number of expanded data of the expanded training data set; If the amount of the expanded data is less than the quantity threshold, the above training data set expansion step is repeated until the amount of the expanded data in the training data set after expansion is greater than or equal to the quantity threshold.

7. The method for lesion recognition based on three-dimensional CT images based on big data and deep learning according to claim 5, characterized in that: The method further comprises the steps of: Creating a first shared volume among all first model containers corresponding to the organ part category; Mounting the first shared volume to all the first model containers; When the basic lesion recognition model in any one of the first model containers obtains the target lesion recognition model through model training, the model weight of the target lesion recognition model is saved to the first shared volume; creating a second shared volume between all second model containers corresponding to the organ vessel class; Mounting the second shared volume to all the second model containers; When the basic lesion recognition model in any one of the second model containers obtains the target lesion recognition model through model training, the model weight of the target lesion recognition model is saved in the second shared volume.

8. The method for lesion recognition based on three-dimensional CT images based on big data and deep learning according to claim 1, characterized in that: The step of inputting the three-dimensional CT image data set into the organ image classification model and generating organ image classification results of all three-dimensional CT image data in the three-dimensional CT image data set by the organ image classification model comprises the following steps: For each three-dimensional CT image data in the three-dimensional CT image data set, counting the number of slices of the CT image slices in the three-dimensional CT image data; If the number of slices is less than a preset slice number threshold, all the CT image slices are traversed, and the CT image slice with the highest average pixel value is input into the organ image classification model, and the organ image classification result of the three-dimensional CT image data is generated by the organ image classification model; If the number of slices is greater than or equal to the slice number threshold, acquiring the data size of the three-dimensional CT image data; Based on the data size, the three-dimensional CT image data is reduced in dimension into two-dimensional CT image data; The two-dimensional CT image data is input into the organ image classification model, and the organ image classification result corresponding to the three-dimensional CT image data is generated by the organ image classification model.

9. The method for lesion recognition based on three-dimensional CT images based on big data and deep learning according to claim 8, characterized in that: The dimensionality reduction processing of the three-dimensional CT image data into two-dimensional CT image data based on the data size comprises the following steps: determining a specific axis of the three-dimensional CT image data; If the data size is less than or equal to a preset size threshold, reducing the dimension of the three-dimensional CT image data into two-dimensional CT image data based on the specific axis and using a maximum intensity projection method; If the data size is greater than the size threshold, the three-dimensional CT image data is reduced in dimension into two-dimensional CT image data based on the specific axis and using an axial slicing method.

10. A three-dimensional CT image lesion recognition system based on big data and deep learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the three-dimensional CT image lesion recognition method based on big data and deep learning as described in any one of claims 1 to 9.

Images