Morphology and timing prediction based vascular feature extraction and analysis apparatus and method

Abstract

The application discloses a blood vessel feature extraction and analysis device and method based on morphology and time sequence prediction, and the device comprises a single-frame image processing module, a time sequence prediction module and an intelligent diagnosis module. The single-frame image processing module can divide blood vessel images into single frames for processing and classify blood vessel feature points. The time sequence prediction module performs time sequence analysis and prediction analysis on the classified blood vessel feature points. The intelligent diagnosis module compares the prediction analysis result with a stored knowledge graph and gives a result suggestion. In summary, the application provides a blood vessel feature extraction device and method based on blood vessel morphology and time sequence prediction, which have higher accuracy and stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, specifically to a device and method for extracting vascular features based on vascular morphology and temporal prediction. Background Technology

[0002] Cardiovascular diseases are a group of conditions affecting the heart and blood vessels, including coronary artery disease, hypertension, valvular heart disease, and arrhythmias. Cardiovascular diseases often lead to changes in the morphology of blood vessels, such as arteriosclerosis, stenosis, and dilation. Therefore, vascular morphology is closely related to the diagnosis of cardiovascular diseases.

[0003] Currently, in clinical diagnosis, medical imaging techniques (such as CT and MRI) are commonly used to observe the morphology of the heart and blood vessels, thereby diagnosing diseases. Simultaneously, experienced physicians compare relevant feature points in the images at different time points to achieve disease analysis and diagnosis.

[0004] For example, in the diagnosis of coronary heart disease, doctors can use imaging techniques such as coronary angiography to observe the narrowing of blood vessels, and information such as the degree and location of the narrowing can be obtained through vascular morphology analysis. Furthermore, by registering and comparing images from different time points, doctors can observe the progression of the disease and make further diagnostic and treatment decisions.

[0005] Therefore, vascular morphology and temporal prediction technologies play an important role in the diagnosis and early warning of cardiovascular diseases. Through research on vascular morphology and analysis of temporal predictions, the early diagnosis rate and treatment outcomes of cardiovascular diseases can be improved.

[0006] Currently, most patents in the field of cardiovascular disease prediction rely on patient parameters and blood flow for diagnosis. Chinese invention patent application CN201810655810.9 proposes a cardiovascular disease detection device that detects parameters such as blood pressure, blood oxygen, pulse, heart sounds, and body temperature. After the sensors collect this data, the system classifies and diagnoses the data.

[0007] Chinese invention patent application number CN202210591301.0 proposes a method for assessing vulnerable vascular plaques based on multimodal imaging. Using multimodal image data, it constructs a plaque-related mechanical calculation method, including a two-dimensional fast mechanical calculation method and a mechanical calculation method based on three-dimensional vascular and plaque models, thereby realizing model-based plaque prediction.

[0008] Therefore, it is conceivable to utilize technologies such as computer vision and machine learning to automatically extract vascular feature points from angiography images. These feature points can be used for vascular morphological analysis and temporal prediction, thereby predicting the progression of vascular lesions and treatment outcomes. This would help doctors make diagnoses and treatments more quickly and accurately, improving treatment effectiveness and patient survival rates. Summary of the Invention

[0009] To address the aforementioned problems, this invention provides a vascular feature extraction device and method based on vascular morphology and temporal prediction with improved accuracy and stability.

[0010] To achieve this objective, the present invention provides the following technical solution:

[0011] In a first aspect, the present invention provides a device for extracting and analyzing vascular features based on morphology and temporal prediction, including a single-frame image processing module, a temporal prediction module, and an intelligent diagnosis module.

[0012] The single-frame image processing module can segment blood vessel images into single frames for processing and classify blood vessel feature points.

[0013] The temporal prediction module performs temporal analysis and prediction analysis on the classified vascular feature points;

[0014] The intelligent diagnostic module compares the results of the predictive analysis with the stored knowledge graph and provides suggestions.

[0015] Preferably, the single-frame image processing module includes a preprocessing module, a segmentation module, a SIFT feature point extraction algorithm module, and a CNN network.

[0016] The preprocessing module preprocesses a single frame of image;

[0017] The segmentation module separates the vascular region and the non-vascular region in the image processed by the preprocessing module.

[0018] The SIFT feature point extraction algorithm module extracts blood vessel feature points from the image;

[0019] The CNN network classifies vascular feature points, which can divide the feature points on the vascular contour into different categories, including but not limited to vascular edge points and vascular intersection points.

[0020] Preferably, the time series prediction module uses an LSTM model.

[0021] Preferably, vascular feature points include: vascular edge points and vascular intersection points.

[0022] Preferably, the preprocessing includes: performing image preprocessing by sequentially binarizing, mean filtering, and Gaussian filtering, and then using a ResUNet neural network for semantic segmentation of the blood vessel image.

[0023] Preferably, the vascular images include DSA images.

[0024] A second aspect of the present invention provides a method for extracting and analyzing vascular features based on morphology and temporal prediction, comprising the following steps:

[0025] S1. Imported vascular images from the system;

[0026] S2. Use the SIFT algorithm to extract feature points from vascular images;

[0027] S3. Use the feature points extracted in step S2 as input to the CNN model. By training the CNN model, the feature points on the blood vessel contour are divided into different categories.

[0028] S4. Use the LSTM model to perform temporal prediction of vascular feature points;

[0029] S5. Based on the results of the predictive analysis and the stored knowledge graph, provide suggestions.

[0030] Preferably, in step S3, the feature points extracted by SIFT are classified using CNN, including the following steps:

[0031] S31. Data preparation: Use the SIFT algorithm to extract feature points from the blood vessel contour image and save these points to a file. Each feature point should include its position, scale, and orientation, as well as the corresponding feature descriptor vector.

[0032] S32. Data preprocessing: Load the file containing feature points into memory, preprocess the data as needed, and adjust the image size to an appropriate value.

[0033] S33. Data partitioning: Divide the data into training and testing sets;

[0034] S34. Construct a CNN model;

[0035] S35. Model Training: Train the CNN model using the training data. After each training batch, evaluate the model's performance on the test data and record the training error and test error.

[0036] S36. Model Evaluation: The performance of the trained CNN model is evaluated using test data, and the accuracy is used for evaluation. Accuracy is defined as follows:

[0037]

[0038] The convention is as follows: TP (True Positives): True Positives, which are predicted to be positive and are actually positive.

[0039] FP (False Positives): False positives are examples that are predicted to be positive but are actually negative.

[0040] FN(false Negatives): False negatives, which are predicted to be negative but are actually positive.

[0041] TN (True Negatives): True negatives are those that are predicted to be negative and are indeed negative.

[0042] S37. Model Application: Use the trained CNN model to classify new blood vessel contour image data, extract SIFT feature points from the new image, and use the trained model to predict the category of each feature point.

[0043] S38. Generate labels or classification information for blood vessel contour images based on the classification results of feature points.

[0044] Preferably, step S4 includes:

[0045] S41. Data Processing: The input data of the model needs to be processed: the information of feature point position, scale and orientation is encoded into vector form and standardized preprocessing is performed.

[0046] S42. Constructing the LSTM Model: Based on the characteristics of the input data and task requirements, design the structure of the LSTM model, including the number and size of the LSTM layers, activation functions, and optimizer parameters. The model consists of two LSTM layers and one fully connected layer. The input data is a binary image of a set size. The activation function for both the LSTM and fully connected layers is Sigmoid. The optimizer is Adam, and the loss function is binary cross-entropy. Batch gradient descent is used for optimization during model training.

[0047] S43. Model Training: Use the prepared SIFT feature point location dataset as the training set to train the LSTM model. During the training process, cross-validation can be used to tune and optimize the model parameters.

[0048] S44. Model Evaluation: Evaluate the LSTM model using the test dataset, calculate the error between the predicted and actual locations, and use mean squared error (MSE) or other metrics to evaluate the model's performance.

[0049] S45. Model Application: For new image data, input it into the SIFT feature extraction algorithm to obtain the location, scale, and orientation information of SIFT feature points, and input this information into the trained LSTM model for location prediction.

[0050] Preferably, the knowledge graph in step S5 is obtained using the following method:

[0051] R1. Determine the scope and theme of the knowledge graph: First, clarify that the theme of the knowledge graph is vascular-related diseases, including knowledge of cardiovascular diseases, atherosclerosis, hypertension, and cerebrovascular diseases; at the same time, define the scope of the knowledge graph, including information on diseases, symptoms, examinations, and treatments.

[0052] R2. Collect knowledge and build ontology: Collect relevant medical knowledge and data, such as medical literature, medical records, doctors' experience and expertise. The collected knowledge needs to be classified and summarized to build ontology, that is, to define the formal specifications of concepts, attributes, relationships and constraints. Ontology is the core of knowledge graph and can help achieve semantic consistency and accuracy.

[0053] R3. Building Knowledge Graphs: Based on ontology, knowledge graphs can be built, which are knowledge networks formed by organizing concepts, entities, attributes, and relational elements. Knowledge graphs need to achieve a certain level of structuring and semantics to support automated machine processing and reasoning; entities and relations can be described in the form of RDF triples.

[0054] R4. Design query interface and algorithm: In order to realize intelligent diagnosis and treatment strategies, it is necessary to design query interface and algorithm. The query interface can extract relevant information based on the medical record data input by the user.

[0055] R5. Validation and Optimization of the System: Finally, the system is validated and optimized. Validation can be performed by testing actual cases examined by doctors, comparing the consistency and accuracy of the system's diagnostic and treatment outcomes with the doctors' predictions.

[0056] This invention provides a method for vascular feature extraction and analysis based on morphology and temporal prediction, not for disease diagnosis. The purpose of this method is to extract features from sample blood vessels and compare and evaluate them. The results obtained can be used for disease prediction and to provide suggestions for disease prevention.

[0057] Compared with the prior art, the beneficial effects and significant progress of applying the technical solution of the present invention are as follows:

[0058] 1. The vascular feature extraction device and method based on vascular morphology and temporal prediction of the present invention can not only analyze single-frame images, but also analyze and predict multiple frames of images in time sequence, so as to more comprehensively reflect the development of vascular diseases.

[0059] 2. The vascular feature extraction device and method based on vascular morphology and temporal prediction of the present invention can perform high-precision analysis and diagnosis of vascular morphological features. Compared with traditional medical image processing and diagnostic techniques, it has higher diagnostic accuracy and stability.

[0060] 3. The vascular feature extraction device and method based on vascular morphology and temporal prediction of the present invention combines the diagnostic results of vascular lesions with a medical knowledge graph, which can improve the interpretability and reliability of the diagnosis. The medical knowledge graph can provide doctors with more detailed diagnostic information, helping them to make more accurate diagnoses. Attached Figure Description

[0061] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly described below.

[0062] Figure 1 This is a flowchart of the present invention;

[0063] Figure 2 This is a schematic diagram of the ResUNet network structure described in this invention;

[0064] Figure 3 This is a schematic diagram of the registration described in this invention;

[0065] Figure 4 This is a flowchart of the LSTM model described in this invention;

[0066] Figure 5 This is a flowchart illustrating the construction of the knowledge graph for this invention. Detailed Implementation

[0067] Example 1

[0068] A device for extracting and analyzing vascular features based on morphology and temporal prediction, comprising a single-frame image processing module, a temporal prediction module, and an intelligent diagnosis module;

[0069] The single-frame image processing module can segment blood vessel images into single frames for processing and classify blood vessel feature points.

[0070] The temporal prediction module performs temporal analysis and prediction analysis on the classified vascular feature points;

[0071] The intelligent diagnostic module compares the results of the predictive analysis with the stored knowledge graph and provides suggestions.

[0072] In this embodiment, the single-frame image processing module includes a preprocessing module, a segmentation module, a SIFT feature point extraction algorithm module, and a CNN network.

[0073] The preprocessing module preprocesses a single frame of image;

[0074] The segmentation module separates the vascular region and the non-vascular region in the image processed by the preprocessing module.

[0075] The SIFT feature point extraction algorithm module extracts blood vessel feature points from the image;

[0076] The CNN network classifies vascular feature points, which can divide the feature points on the vascular contour into different categories, including but not limited to vascular edge points and vascular intersection points.

[0077] In this embodiment, the time series prediction module uses an LSTM model.

[0078] In this embodiment, the vascular feature points include: vascular edge points and vascular intersection points.

[0079] In this embodiment, the preprocessing includes: performing image preprocessing such as binarization, mean filtering, and Gaussian filtering in sequence, and then using the ResUNet neural network for semantic segmentation of the blood vessel image.

[0080] In this embodiment, the vascular images include DSA images.

[0081] Example 2

[0082] A method for extracting and analyzing vascular features based on morphology and temporal prediction includes the following steps:

[0083] S1. Imported vascular images from the system;

[0084] S2. Use the SIFT algorithm to extract feature points from vascular images;

[0085] S3. Use the feature points extracted in step S2 as input to the CNN model. By training the CNN model, the feature points on the blood vessel contour are divided into different categories.

[0086] S4. Use the LSTM model to perform temporal prediction of vascular feature points;

[0087] S5. Based on the results of the predictive analysis and the stored knowledge graph, provide suggestions.

[0088] In this embodiment, step S3 uses CNN to classify the feature points extracted by SIFT, including the following steps:

[0089] S31. Data preparation: Use the SIFT algorithm to extract feature points from the blood vessel contour image and save these points to a file. Each feature point should include its position, scale, and orientation, as well as the corresponding feature descriptor vector.

[0090] S32. Data preprocessing: Load the file containing feature points into memory, preprocess the data as needed, and adjust the image size to an appropriate value.

[0091] S33. Data partitioning: Divide the data into training and testing sets;

[0092] S34. Construct a CNN model;

[0093] S35. Model Training: Train the CNN model using the training data. After each training batch, evaluate the model's performance on the test data and record the training error and test error.

[0094] S36. Model Evaluation: The performance of the trained CNN model is evaluated using test data, and the accuracy is used for evaluation. Accuracy is defined as follows:

[0095]

[0096] The convention is as follows: TP (True Positives): True Positives, which are predicted to be positive and are actually positive.

[0097] FP (False Positives): False positives are examples that are predicted to be positive but are actually negative.

[0098] FN(false Negatives): False negatives, which are predicted to be negative but are actually positive.

[0099] TN (True Negatives): True negatives are those that are predicted to be negative and are indeed negative.

[0100] S37. Model Application: Use the trained CNN model to classify new blood vessel contour image data, extract SIFT feature points from the new image, and use the trained model to predict the category of each feature point.

[0101] S38. Generate labels or classification information for blood vessel contour images based on the classification results of feature points.

[0102] In this embodiment, step S4 includes:

[0103] S41. Data Processing: The input data of the model needs to be processed: the information of feature point position, scale and orientation is encoded into vector form and standardized preprocessing is performed.

[0104] S42. Constructing the LSTM Model: Based on the characteristics of the input data and task requirements, design the structure of the LSTM model, including the number and size of the LSTM layers, activation functions, and optimizer parameters. The model consists of two LSTM layers and one fully connected layer. The input data is a binary image of a set size. The activation function for both the LSTM and fully connected layers is Sigmoid. The optimizer is Adam, and the loss function is binary cross-entropy. Batch gradient descent is used for optimization during model training.

[0105] S43. Model Training: Use the prepared SIFT feature point location dataset as the training set to train the LSTM model. During the training process, cross-validation can be used to tune and optimize the model parameters.

[0106] S44. Model Evaluation: Evaluate the LSTM model using the test dataset, calculate the error between the predicted and actual locations, and use mean squared error (MSE) or other metrics to evaluate the model's performance.

[0107] S45. Model Application: For new image data, input it into the SIFT feature extraction algorithm to obtain the location, scale, and orientation information of SIFT feature points, and input this information into the trained LSTM model for location prediction.

[0108] In this embodiment, the knowledge graph in step S5 is obtained using the following method:

[0109] R1. Determine the scope and theme of the knowledge graph: First, clarify that the theme of the knowledge graph is vascular-related diseases, including knowledge of cardiovascular diseases, atherosclerosis, hypertension, and cerebrovascular diseases; at the same time, define the scope of the knowledge graph, including information on diseases, symptoms, examinations, and treatments.

[0110] R2. Collect knowledge and build ontology: Collect relevant medical knowledge and data, such as medical literature, medical records, doctors' experience and expertise. The collected knowledge needs to be classified and summarized to build ontology, that is, to define the formal specifications of concepts, attributes, relationships and constraints. Ontology is the core of knowledge graph and can help achieve semantic consistency and accuracy.

[0111] R3. Building Knowledge Graphs: Based on ontology, knowledge graphs can be built, which are knowledge networks formed by organizing concepts, entities, attributes, and relational elements. Knowledge graphs need to achieve a certain level of structuring and semantics to support automated machine processing and reasoning; entities and relations can be described in the form of RDF triples.

[0112] R4. Design query interface and algorithm: In order to realize intelligent diagnosis and treatment strategies, it is necessary to design query interface and algorithm. The query interface can extract relevant information based on the medical record data input by the user.

[0113] R5. Validation and Optimization of the System: Finally, the system is validated and optimized. Validation can be performed by testing actual cases examined by doctors, comparing the consistency and accuracy of the system's diagnostic and treatment outcomes with the doctors' predictions.

[0114] Example 3

[0115] like Figure 1 As shown, this embodiment uses a specific example to further illustrate the invention.

[0116] Step 1: In this system, DSA images will be imported in real time during the surgery.

[0117] Step 2: After entering the single-frame image processing module, the DSA image will be segmented into single frames for processing. First, preprocessing of the DSA single-frame image is performed, including binarization, mean filtering, and Gaussian filtering. After preprocessing, as shown... Figure 2 As shown, the ResUNet neural network is used for semantic segmentation of blood vessel images. The segmented images will be processed using the SIFT algorithm to extract relevant feature points, and then CNN will be used to classify these feature points.

[0118] Step 3: Enter the temporal prediction module. In the temporal prediction module, the LSTM model is used to perform vascular morphology prediction analysis on multiple frames of images over time, and the development of corresponding vascular lesions is obtained based on the changes in vascular morphology.

[0119] Based on this information, the intelligent diagnostic module of this invention will perform real-time intelligent diagnosis and provide corresponding effective treatment strategies according to the knowledge graph of vascular diseases and related morphologies within the system.

[0120] like Figure 3 As shown, CNN is used to classify the feature points extracted by SIFT, which can divide the feature points on the blood vessel contour into different categories, such as blood vessel edge points and blood vessel intersections. This classification can help the system better understand the morphology and structure of blood vessels and provide better feature representations for subsequent tasks.

[0121] The following are the steps to accomplish this task:

[0122] Data preparation: Feature points are extracted from the blood vessel contour image using the SIFT algorithm, and these points are saved to a file. Each feature point should include its location, scale, and orientation, as well as its corresponding feature descriptor vector.

[0123] Data preprocessing: The file storing feature points is loaded into memory, and the data is preprocessed as needed. In this embodiment, the input image size is adjusted to 256×256.

[0124] Data partitioning: The data is divided into training and testing sets. In this embodiment, 80% of the data is used for training and 20% for testing.

[0125] Building a CNN model: In this embodiment, the LeNet network model is used.

[0126] Model Training: Train the CNN model using the training data. After each training batch, evaluate the model's performance on the test data and record the training error and test error.

[0127] Model Evaluation: The performance of the trained CNN model is evaluated using test data. This example uses Accuracy for evaluation, which is defined as follows:

[0128]

[0129] The convention is as follows: TP (True Positives): True Positives, which are predicted to be positive and are actually positive.

[0130] FP (False Positives): False positives are examples that are predicted to be positive but are actually negative.

[0131] FN(false Negatives): False negatives, which are predicted to be negative but are actually positive.

[0132] TN (True Negatives): True negatives are those that are predicted to be negative and are indeed negative.

[0133] Model Application: A trained CNN model is used to classify new blood vessel contour image data. SIFT feature points are extracted from the new image, and the trained model is used to predict the category of each feature point. Finally, labels or classification information for the blood vessel contour images can be generated based on the feature point classification results.

[0134] When using the LSTM model to predict the temporal location of SIFT feature points, this invention can be performed according to the following steps, as follows: Figure 4 As shown:

[0135] Data processing: The input data of the model needs to be processed, such as encoding information such as the location, scale and orientation of feature points into vector form and performing preprocessing operations such as standardization.

[0136] Building an LSTM model: Based on the characteristics of the input data and the task requirements, design the structure of the LSTM model, including the number and size of LSTM layers, activation functions, optimizers, and other parameters.

[0137] In this embodiment, the model consists of two LSTM layers and one fully connected layer. The input data is a 256x256 binary image. The activation function for both the LSTM and fully connected layers is sigmoid. The optimizer is Adam, and the loss function is binary cross-entropy. Batch gradient descent is used for optimization during model training.

[0138] Model Training: The LSTM model is trained using the prepared SIFT feature point location dataset. During training, methods such as cross-validation can be used for parameter tuning and optimization.

[0139] Model Evaluation: The LSTM model is evaluated using a test dataset, and the error between the predicted and actual locations is calculated. Mean squared error (MSE) or other metrics can be used to evaluate the model's performance.

[0140] Model application: For new image data, it is input into the SIFT feature extraction algorithm to obtain information such as the position, scale and orientation of SIFT feature points, and this information is then input into the trained LSTM model for position prediction.

[0141] The steps for constructing the knowledge graph of this invention are as follows, specifically as follows: Figure 5 As shown:

[0142] Define the scope and theme of the knowledge graph: First, clarify that the theme of the knowledge graph is vascular-related diseases, including knowledge about cardiovascular diseases, atherosclerosis, hypertension, and cerebrovascular diseases. Simultaneously, define the scope of the knowledge graph, including information on diseases, symptoms, examinations, and treatments.

[0143] Knowledge Collection and Ontology Building: This involves collecting relevant medical knowledge and data, such as medical literature, medical records, and physician experience and expertise. The collected knowledge needs to be categorized and summarized to build an ontology, which is a formal specification defining concepts, attributes, relationships, and constraints. Ontology is the core of a knowledge graph, helping to achieve semantic consistency and accuracy.

[0144] Building a knowledge graph: Based on ontology, a knowledge graph can be built, which is a knowledge network formed by organizing elements such as concepts, entities, attributes, and relationships. A knowledge graph needs to achieve a certain level of structuring and semantic representation to support automated machine processing and reasoning. For example, entities and relationships can be described using RDF triples (Subject-Predicate-Object).

[0145] Designing a query interface and algorithm: To achieve intelligent diagnosis and treatment strategies, a query interface and algorithm need to be designed. The query interface can extract relevant information from the user-input medical record data; in this embodiment, this involves the changes in feature points on blood vessels, thereby inferring changes in vascular morphology. The algorithm can automatically infer possible diagnostic results and treatment suggestions based on relationships and rules in a knowledge graph, and provide corresponding confidence levels and supporting evidence.

[0146] System Validation and Optimization: Finally, the system is validated and optimized. Validation can be performed by testing actual cases examined by doctors, comparing the consistency and accuracy of the system's diagnostic and treatment outcomes with those of the doctors.

[0147] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] The applicant declares that, in the process of describing the above-mentioned specification:

[0149] The terms "this embodiment," "an embodiment of the present invention," "as shown," "further," and "further improved technical solutions," etc., indicate that the specific features, structures, materials, or characteristics described in the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms are not necessarily directed at the same embodiment or example, and the specific features, structures, materials, or characteristics described can be combined or combined in any suitable manner in one or more embodiments or examples. Furthermore, without causing contradiction, those skilled in the art can combine or combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.

[0150] Finally, it should be noted that:

[0151] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them;

[0152] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention. Non-essential improvements, adjustments or substitutions made by those skilled in the art based on the content of this specification are all within the scope of protection claimed by the present invention.

Claims

1. A morphological and timing prediction based vascular feature extraction and analysis device, characterized in that, The single-frame image processing module, the time sequence prediction module and the intelligent diagnosis module are included. The single-frame image processing module can separate the blood vessel image into single frames for processing and classify the blood vessel feature points. The time sequence prediction module performs time sequence analysis and prediction analysis on the classified blood vessel feature points. The intelligent diagnosis module compares the prediction analysis result with the stored knowledge graph to give a result suggestion. The steps of time sequence analysis and prediction analysis include: Data processing: For the input data of the model, the feature point position, scale and direction information needs to be encoded into vector form and standardized preprocessing operation is performed; Constructing the LSTM model: According to the characteristics of the input data and the task requirements, the structure of the LSTM model is designed, including the number and size of the LSTM layers, the activation function, the parameters of the optimizer; the model includes two layers of LSTM and one fully connected layer, wherein the input data is a binary image with a set size, the activation function of the LSTM layer is Sigmoid, the activation function of the fully connected layer is Sigmoid, the optimizer of the model is Adam, and the loss function is binary cross entropy; in the model training, batch gradient descent is used for optimization; Model training: using the prepared SIFT feature point position data set as the training set, the LSTM model is trained, and in the training process, the cross-validation method can be used for model parameter optimization; Model evaluation: using the test data set to evaluate the LSTM model, calculating the error between the predicted position and the actual position, and using mean square error MSE or other indicators to evaluate the performance of the model; Model application: for new image data, input it into the SIFT feature extraction algorithm to obtain the position, scale and direction information of the SIFT feature points, and input these information into the trained LSTM model for position prediction.

2. A morphological and temporal prediction based vessel feature extraction and analysis device as claimed in claim 1, characterized in that, The single-frame image processing module includes a preprocessing module, a segmentation module, a SIFT feature point extraction algorithm module and a CNN network. The preprocessing module pre-processes the single-frame image. The segmentation module separates the blood vessel region and the non-blood vessel region in the image processed by the preprocessing module. The SIFT feature point extraction algorithm module extracts the blood vessel feature points in the image. The CNN network classifies the blood vessel feature points, which can divide the feature points on the blood vessel contour into different categories, including but not limited to blood vessel edge points and blood vessel intersection points.

3. The morphological and temporal prediction based vessel feature extraction and analysis device as claimed in claim 1, wherein, The time sequence prediction module adopts the LSTM model.

4. The morphological and temporal prediction based vessel feature extraction and analysis device as claimed in claim 1, wherein, The blood vessel feature points include blood vessel edge points and blood vessel intersection points.

5. A morphological and temporal prediction based vessel feature extraction and analysis device as claimed in claim 2, characterized in that, The preprocessing includes sequentially completing the image preprocessing of binarization, mean filtering and Gaussian filtering, and then using the ResUNet neural network for semantic segmentation of the blood vessel image.

6. The morphological and temporal prediction based vessel feature extraction and analysis device as claimed in claim 1, wherein, The blood vessel image includes DSA image.

7. A method for morphological and timing prediction based vessel feature extraction and analysis, the method comprising: receiving a medical image; identifying a region of interest in the medical image; extracting a plurality of features from the region of interest; and determining a plurality of vessel features from the plurality of features. The steps include: S1, the system imports the acquired blood vessel image; S2, extracting feature points from the blood vessel image using SIFT algorithm; S3, using the feature points extracted in step S2 as the input of the CNN model, dividing the feature points on the blood vessel contour into different categories by training the CNN model; S4, using LSTM model to predict the time sequence of the feature points of blood vessels; S5, using the results of prediction analysis and the stored knowledge graph to give result suggestions; Step S4 includes: S41, data processing: for the input data of the model, processing is needed: encode the information of feature point position, scale and direction into vector form, and perform standardization preprocessing operation; S42, construct LSTM model: design the structure of LSTM model according to the characteristics of input data and task requirements, including the number and size of LSTM layers, activation function, parameters of optimizer; The model includes two layers of LSTM and one fully connected layer, wherein the input data is a binary image with a set size, the activation function of LSTM layer is Sigmoid, the activation function of fully connected layer is Sigmoid, the optimizer of the model is Adam, and the loss function is binary cross entropy; In model training, batch gradient descent is used for optimization; S43, model training: using the prepared SIFT feature point position data set as the training set, and then training the LSTM model, in the training process, cross validation method can be used for model parameter optimization; S44, model evaluation: use the test data set to evaluate the LSTM model, calculate the error between the predicted position and the actual position, and use mean square error MSE or other indicators to evaluate the performance of the model; S45, model application: for new image data, input it into the SIFT feature extraction algorithm to obtain the position, scale and direction information of the SIFT feature points, and input these information into the trained LSTM model for position prediction.

8. A method for vessel feature extraction and analysis based on morphological and temporal prediction as claimed in claim 7, wherein, In step S3, CNN is used to classify the feature points extracted by SIFT, including the following steps: S31, data preparation, using SIFT algorithm to extract feature points from blood vessel contour image, and saving these points to file, each feature point should contain its position, scale and direction, and corresponding feature descriptor vector; S32, data preprocessing, load the file saving feature points into memory, and preprocess the data as needed, adjust the size of the picture to a suitable size; S33, data division, divide the data into training set and test set; S34, construct CNN model; S35, model training, train the CNN model using training data, evaluate the performance of the model on test data after each training batch, and record the training error and test error; S36, model evaluation, evaluate the performance of the trained CNN model using test data, and use Accuracy for evaluation, Accuracy is defined as follows: ， Wherein, TP (True Positives): true positive, predicted as positive and actually positive; FP (False Positives): false positive, predicted as positive but actually negative; FN (false Negatives): false negative, predicted as negative but actually positive; TN (True Negatives): True negatives, predicted as negative and actually negative; S37, model application, using the trained CNN model to classify new blood vessel contour image data, extracting SIFT feature points from new images, and using the trained model to predict the category of each feature point; S38, generating the label or classification information of the blood vessel contour image according to the classification result of the feature points.

9. The method of claim 7, wherein the method further comprises: determining a plurality of morphological features of the blood vessel based on the plurality of blood vessel images; and determining a plurality of temporal features of the blood vessel based on the plurality of blood vessel images. The knowledge graph in step S5 is obtained by the following method: R1, determine the scope and theme of the knowledge graph: first, define the theme of the knowledge graph as blood vessel related diseases, including cardiovascular diseases, atherosclerosis, hypertension, and cerebrovascular diseases; at the same time, define the scope of the knowledge graph, including the information of diseases, symptoms, examinations and treatments; R2, collect knowledge and establish ontology: collect relevant medical knowledge and data, such as medical literature, medical record data, doctors' experience and professional knowledge, and classify and summarize the collected knowledge to establish ontology, that is, define the formal specification of concepts, attributes, relationships and constraints. Ontology is the core of knowledge graph, which can help achieve semantic consistency and accuracy; R3, establish knowledge graph: based on ontology, the knowledge graph can be established, that is, the knowledge network formed by organizing concepts, entities, attributes and relationship elements. Knowledge graph needs to be structured and semantized to support machine automation processing and reasoning; RDF triple can be used to describe entities and relationships; R4, design query interface and algorithm: in order to realize intelligent diagnosis and treatment strategy, query interface and algorithm need to be designed. Query interface can extract relevant information according to user input medical record data; R5, verify and optimize the system: finally, verify and optimize the system. Verification can be tested by actual doctor's case, and the consistency and accuracy of the diagnosis and treatment results of the system with the prediction results of the doctor are compared.

Images