Cardiac magnetic resonance movie image classification method and system based on time frame expansion and feature selection intersection

Abstract

The invention discloses a cardiac magnetic resonance movie image classification method and system based on time frame expansion and feature selection intersection, and the method comprises the steps: firstly obtaining a cardiac magnetic resonance movie image, carrying out the preprocessing of the cardiac magnetic resonance movie image, and dividing the region of interest; secondly, image omics feature extraction is carried out on the image after the region of interest is divided, and image omics features are integrated based on time sequence information; and finally, constructing a feature selection cross model and performing training to obtain a trained feature selection cross model, evaluating the classification performance of the trained feature selection cross model, and performing cardiac magnetic resonance movie image classification based on the evaluated model. According to the method, the image omics characteristics of the myocardial section in the Cine-MRI are extracted, the time sequence information of the sample is integrated, and the time-space characteristic potential of the sample is mined, so that comprehensive description of the focus part is realized, and the classification accuracy and robustness of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a cardiac magnetic resonance movie image classification method and system based on the crossover of time frame expansion and feature selection. Background Art

[0002] Myocardial perfusion imaging is considered the gold standard method for diagnosing coronary microcirculatory disorders (CMD). However, this imaging method relies on stress and the use of contrast agents, which limits its application in clinical diagnosis. Cardiac magnetic resonance imaging provides dynamic images of the heart during motion, which is of high value for evaluating cardiac function and identifying cardiac abnormalities. Therefore, it has shown the potential to diagnose CMD without contrast and stress.

[0003] In recent years, with the development of medical image processing technology and machine learning algorithms, the emerging radiomics technology can quantify the image information of the lesion site by extracting a large number of image features, thereby revealing potential biomarkers of the disease. This method has been studied and applied in the cardiovascular field to a certain extent, providing a new perspective for disease assessment and prognosis analysis. However, radiomics still faces considerable challenges when processing Cine-MRI data, mainly in the following aspects:

[0004] Complexity of time series. Cine-MRI sequence images reflect the dynamic changes of the heart throughout its entire motion cycle, that is, there are multiple time frames that need to be analyzed in each cardiac cycle. Compared with traditional static medical image analysis, time series images not only need to capture spatial features, but also need to consider information in the time dimension, which significantly increases the complexity of image analysis. Therefore, how to effectively extract high-dimensional image features that reflect subtle changes in myocardial motion in time series and accurately distinguish pathological changes from normal physiological fluctuations is a major challenge in processing Cine-MRI data.

[0005] The robustness and generalization ability of the model. Existing methods based on radiomics modeling are mainly based on traditional machine learning models, ignoring the full use of radiomics features. How to deeply mine the spatiotemporal information in Cine-MRI sequence images to effectively reflect the state of myocardial ischemia is another major challenge in modeling.

[0006] In general, although radiomics technology has shown great potential in Cine-MRI data processing, in-depth research is still needed in time series feature extraction, model construction, etc. to adapt to the solution of clinical problems. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a cardiac magnetic resonance movie image classification method and system based on the intersection of time frame expansion and feature selection to solve the problems existing in the above-mentioned prior art.

[0008] To achieve the above object, the present invention provides a cardiac magnetic resonance movie image classification method based on time frame expansion and feature selection crossover, comprising the following steps:

[0009] Acquiring cardiac magnetic resonance movie images, preprocessing the cardiac magnetic resonance movie images and dividing regions of interest;

[0010] Extract radiomic features from images after dividing the region of interest, and integrate the radiomic features based on time series information;

[0011] A feature selection cross model is constructed and trained to obtain a trained feature selection cross model, the classification performance of the trained feature selection cross model is evaluated, and cardiac magnetic resonance movie image classification is performed based on the evaluated model.

[0012] Optionally, the process of preprocessing the cardiac magnetic resonance movie image and dividing the region of interest includes:

[0013] Standardize cardiac magnetic resonance cine image data to eliminate resolution differences caused by different imaging devices;

[0014] The Min-Max method was used to standardize the brightness and contrast of cardiac magnetic resonance movie images to reduce the differences between images;

[0015] The region of interest of cardiac magnetic resonance cine images was determined by manual delineation.

[0016] Optionally, the process of extracting radiomics features from the image after dividing the region of interest includes:

[0017] Features are extracted from the region of interest, wherein the imaging features include shape features, first-order statistical features, and texture features.

[0018] Optionally, the process of integrating radiomics features based on time series information includes:

[0019] Using the time frame expansion (EFU) method, the radiomics features of the same sample at different time frames were treated as independent samples to obtain a quantitatively expanded dataset.

[0020] Optionally, the feature selection cross model includes a feature selection module based on an attention mechanism, an explicit feature cross module and an implicit feature cross module;

[0021] The process of training the feature selection crossover model includes:

[0022] The expanded data set is input into the feature selection cross model for training. In the feature selection module, the contribution of each feature is gradually adjusted according to the attention weight, and the weight is updated by back propagation. Then, the multi-order cross features generated by the explicit feature cross module are adjusted, and the nonlinear relationship between feature combinations is optimized through the multi-layer perceptron network of the implicit feature cross module.

[0023] Optionally, the process of evaluating the classification performance of the trained feature selection crossover model includes:

[0024] The model was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC) to quantify the classification accuracy of the model;

[0025] The performance of the model in the cardiac magnetic resonance movie image classification task was verified by comparing the AUC value with the existing method.

[0026] The present invention also provides a cardiac magnetic resonance movie image classification system based on time frame expansion and feature selection crossover, which is used to implement a cardiac magnetic resonance movie image classification method based on time frame expansion and feature selection crossover, and the system comprises:

[0027] An image acquisition module, used for acquiring cardiac magnetic resonance movie images;

[0028] A region of interest division module, used for preprocessing the cardiac magnetic resonance movie image and dividing the region of interest;

[0029] The radiomics feature processing module is used to extract radiomics features from the images after dividing the regions of interest, and integrate the radiomics features based on time series information;

[0030] A model training module is used to construct and train a feature selection cross model to obtain a trained feature selection cross model;

[0031] A model evaluation module, used to evaluate the classification performance of the trained feature selection cross model and obtain a model that passes the evaluation;

[0032] A classification module for classifying cardiac magnetic resonance cine images based on the evaluated model.

[0033] Optionally, the model training module includes a model building unit and a model training unit, wherein:

[0034] The model building unit is used to build a feature selection crossover model;

[0035] The model training unit is used to train the feature selection cross model to obtain a trained feature selection cross model;

[0036] The feature selection cross model includes a feature selection module based on an attention mechanism, an explicit feature cross module and an implicit feature cross module;

[0037] The feature selection module is used to allocate attention weights through a two-layer fully connected network to screen radiomics features and reduce noise feature interference;

[0038] The dominant feature crossover module is used to generate new crossover features by crossover processing the radiomics features;

[0039] The implicit feature cross module is used to capture the nonlinear relationship between feature combinations based on a multi-layer perceptron.

[0040] The present invention also provides a computer terminal device, comprising:

[0041] one or more processors;

[0042] A memory, coupled to the processor, for storing one or more programs;

[0043] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of a cardiac magnetic resonance movie image classification method based on the intersection of time frame expansion and feature selection.

[0044] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of a cardiac magnetic resonance movie image classification method based on the intersection of time frame expansion and feature selection are implemented.

[0045] Compared with the prior art, the present invention has the following advantages and technical effects:

[0046] The present invention provides a cardiac magnetic resonance movie image classification method and system based on time frame expansion and feature selection crossover, firstly obtaining cardiac magnetic resonance movie images, preprocessing cardiac magnetic resonance movie images and dividing regions of interest; secondly performing radiomics feature extraction on the images after dividing the regions of interest, and integrating radiomics features based on time series information; finally constructing a feature selection crossover model and training it, obtaining the trained feature selection crossover model, evaluating the classification performance of the trained feature selection crossover model, and classifying cardiac magnetic resonance movie images based on the evaluated model. The present invention extracts radiomics features of myocardial segments in Cine-MRI, integrates the time series information of samples and mines the temporal and spatial feature potential of samples, thereby achieving a comprehensive description of the lesion site and improving the classification accuracy and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0048] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0049] Figure 2 : is a specific algorithm structure diagram of an embodiment of the present invention, wherein (A) is a structural diagram of the model, (B) is a structural diagram of a feature selection module based on an attention mechanism, (C) is a structural diagram of an explicit feature cross module, and (D) is a structural diagram of an implicit feature cross module;

[0050] Figure 3 A comparison chart of classification performance of the method according to an embodiment of the present invention;

[0051] Figure 4 The system structure and flow chart of the embodiment of the present invention are shown in FIG. DETAILED DESCRIPTION

[0052] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0053] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0054] The cardiac magnetic resonance movie image analysis method based on time frame expansion and feature selection crossover is a method that uses the time series information of Cine-MRI and applies feature selection and feature crossover techniques to radiomic features to improve the accuracy of sample classification. The method first extracts the radiomic features of the ROI area from the preprocessed and standardized Cine-MRI images, then independently processes the feature data of different time frames of the same sample (EFU) to achieve the effect of enhancing the training data set, and finally trains the model based on the enhanced training set to obtain a stable and accurate classifier.

[0055] Embodiment 1

[0056] like Figure 1-2 As shown, this embodiment provides a cardiac magnetic resonance movie image classification method based on time frame expansion and feature selection crossover, including:

[0057] Acquire cardiac magnetic resonance movie images, pre-process the cardiac magnetic resonance movie images and divide the regions of interest;

[0058] Extract radiomic features from images after dividing the region of interest, and integrate the radiomic features based on time series information;

[0059] A feature selection crossover model is constructed and trained to obtain the trained feature selection crossover model, the classification performance of the trained feature selection crossover model is evaluated, and cardiac magnetic resonance movie image classification is performed based on the evaluated model.

[0060] Optionally, the process of preprocessing the cardiac magnetic resonance cine images and dividing the region of interest includes:

[0061] Standardize cardiac magnetic resonance cine image data to eliminate resolution differences caused by different imaging devices;

[0062] The Min-Max method was used to standardize the brightness and contrast of cardiac magnetic resonance movie images to reduce the differences between images;

[0063] The region of interest of cardiac magnetic resonance cine images was determined by manual delineation.

[0064] As an implementation of this embodiment, the specific logical sequence of the processing flow for radiomics feature extraction is as follows:

[0065] 1. Image preprocessing and ROI area division:

[0066] First, the Cine-MRI images are standardized to ensure that the image data is consistent in brightness and resolution. Then, on the pre-processed images, the region of interest (ROI) is delineated by manual delineation or automatic segmentation methods, such as the myocardial segment in the left ventricle area of ​​the heart, to provide a regional basis for subsequent feature extraction.

[0067] 2. Extract low- and high-level radiomics features from original and derived images:

[0068] For the delineated ROI area, multi-dimensional imaging features are extracted from the Cine-MRI original image and different derived images. Derived images refer to the image type obtained after the original image is derived and processed by applying wavelet transform, filtering and other techniques to enrich and refine the image information and increase the capture of image details. Subsequently, these imaging features are extracted and calculated, and are divided into low-order features and high-order features, mainly including the following categories:

[0069] Shape characteristics: describe the geometric characteristics of the myocardial region, such as the area, perimeter, aspect ratio, etc.

[0070] First-order statistical features: These features include mean, variance, skewness, and kurtosis, which reflect the basic statistical information of pixel intensity distribution.

[0071] Texture features: mainly through grayscale co-occurrence matrix, grayscale gradient method and other methods, the microstructural changes of the tissue are analyzed from the distribution pattern and arrangement of pixels to reveal the heterogeneity information within the tissue.

[0072] 3. Feature output and integration:

[0073] After the above feature extraction is completed, all feature data are integrated into a set of radiomics feature sets to serve as the basis for the input data of the subsequent classification model.

[0074] Among them, "derived image" refers to a new image generated by applying various image processing algorithms or transformations to the original image. Its purpose is to further explore the potential information in the image through filtering, wavelet transform and other methods, and increase the fine-grained capture of structural features. For example:

[0075] Wavelet transform image: Decomposes the image into parts of different resolutions so that features of different frequencies can be more clearly displayed, which helps to highlight slight changes in edges or textures in the image.

[0076] Filtered image: The filtering operation can highlight certain specific frequency information in the image, thereby helping to identify the target area in a complex background.

[0077] Among them, the calculation methods of low-order and high-order imaging omics features are explained:

[0078] The calculation of low-level radiomics features and high-level radiomics features has the following differences:

[0079] 1. Low-level radiomics features:

[0080] Low-order features are usually directly calculated statistical or shape features. These features are based on simple statistical descriptions of pixel intensity and geometric properties, usually obtained through basic statistical methods or simple shape analysis, mainly including:

[0081] Shape features: obtained by calculating the geometric parameters of ROI (such as area, perimeter, aspect ratio, etc.).

[0082] First-order statistical characteristics: obtained directly by calculating the mean, variance, skewness, kurtosis and other indicators of the pixel intensity in the ROI area, reflecting the overall distribution of pixel intensity.

[0083] 2. Advanced imaging features:

[0084] Higher-order features are calculated using more complex statistical or mathematical methods to capture spatial structures or texture patterns in images to characterize the microstructure of tissues. Such features can reveal the complexity and heterogeneity of pixel arrangements. The main calculation methods include:

[0085] Texture features: obtained through methods such as gray-level co-occurrence matrix (GLCM) and gray-level run length matrix (GLRLM), which are used to describe the spatial relationship and arrangement between pixels. For example, GLCM can calculate indicators such as contrast, homogeneity, and energy of grayscale values ​​to quantify the smoothness, roughness, or stripes of image texture.

[0086] Wavelet transform features: Wavelet transform decomposes the image into different frequency components and extracts detailed information of different frequency bands. These features are suitable for revealing local changes in tissue structure.

[0087] Therefore, the entire imaging feature extraction process is centered on the ROI region, calculating low-order and high-order features from the original image and multiple derived images, providing rich basic data for the subsequent deep learning classification model. These feature combinations can significantly improve the model's accuracy in identifying myocardial lesions in Cine-MRI.

[0088] As an implementation of this embodiment, step Sa: preprocessing of Cine-MRI data and delineation of ROI area.

[0089] Specifically, step Sa: First, resample the cardiac magnetic resonance movie data of different patients to eliminate the resolution differences caused by different devices and maintain the stability and consistency of feature extraction. Secondly, the Min-Max method is used to standardize the movie images to reduce the brightness and contrast changes caused by differences in imaging parameters between images. Finally, the ROI area corresponding to each myocardial segment of the left ventricle in the patient's Cine-MRI is determined by manual delineation to prepare for the subsequent feature extraction steps.

[0090] Optionally, the process of extracting radiomics features from the image after dividing the region of interest includes:

[0091] Features are extracted from the region of interest, and the radiomics features include shape features, first-order statistical features, and texture features.

[0092] As an implementation of this embodiment, step Sb: extraction of imaging genomics features.

[0093] Specifically, step Sb: for the delineated ROI area, low-order and high-order imaging omics features are calculated from the original Cine-MRI image and its various derivative images. Feature types include shape features, such as the geometric morphology and contour characteristics of the myocardial segment; first-order statistical features, which reflect the basic statistical information of pixel intensity, including mean, variance, skewness and kurtosis; and texture features, which reveal the microstructural changes of myocardial tissue by analyzing the complexity of pixel arrangement patterns. In addition, this module also considers features in derived images, such as features of images after wavelet transform or filtering, to capture more information about myocardial tissue. These comprehensive and large number of imaging omics features provide the basis for the construction of subsequent feature cross-models, ensuring that the classification model can fully understand and utilize the complex image information contained in the myocardial region in Cine-MRI.

[0094] Optionally, the process of integrating radiomics features based on time series information includes:

[0095] Using the time frame expansion (EFU) method, the radiomics features of the same sample at different time frames were treated as independent samples to obtain a quantitatively expanded dataset.

[0096] As an implementation of this embodiment, step Sc: integration of sample time series information.

[0097] Specifically, step Sc: further process the features extracted in the previous step using the time frame expansion (EFU) method. The imaging features of the same sample in different time frames in the Cine-MRI sequence are treated as independent samples and expanded to the sample dimension, thereby expanding the scale of the training data set, enhancing the model's ability to understand the dynamic changes of myocardial segments, enabling the model to capture subtle changes in myocardial motion, providing richer information for identifying myocardial lesions, and ultimately improving the stability and accuracy of model classification.

[0098] Optionally, the feature selection cross model includes a feature selection module based on an attention mechanism, an explicit feature cross module, and an implicit feature cross module;

[0099] The process of training a feature selection crossover model includes:

[0100] The expanded data set is input into the feature selection cross model for training. In the feature selection module, the contribution of each feature is gradually adjusted according to the attention weight, and the weight is updated through back propagation. Then, the multi-order cross features generated by the explicit feature cross module are adjusted, and the nonlinear relationship between feature combinations is optimized through the multi-layer perceptron network of the implicit feature cross module.

[0101] As an implementation method of this embodiment, the process of constructing a feature selection cross model includes:

[0102] 1. Expand the training set: Use the time frame expansion (EFU) method to independently process different time frames of the time series image and expand them into multiple samples to increase the size of the training set. This data expansion strategy can capture the temporal dynamic changes in cardiac magnetic resonance movie images and provide richer training data for the model.

[0103] 2. Model structure design: The feature selection cross model is designed as a hierarchical structure, which mainly includes a feature selection module based on the attention mechanism and an explicit / implicit feature cross module:

[0104] Feature selection module based on attention mechanism:

[0105] The first layer of fully connected network: All imaging features are input, and the first layer of fully connected network reduces the dimension of features through dimensionality reduction to extract key information.

[0106] The second layer of fully connected network: assigns an attention weight of 0 to 1 to each feature, which represents the importance of the model to each feature. This can give low weights to unimportant features, thereby reducing the interference of noise features and improving the signal-to-noise ratio of the features.

[0107] Explicit feature crossover module:

[0108] Feature crossover: The explicit feature crossover module combines different radiomics features to generate crossover features of multiple orders. By increasing the number of explicit feature crossover modules, the model can obtain feature combinations of different orders to form more representative composite features, thereby better revealing lesion information.

[0109] Hidden feature cross module:

[0110] Implicit nonlinear combination: This module inputs features into a multi-layer perceptron (MLP) network and captures the implicit complex combination relationship between image features through the nonlinear fitting ability of the neural network, thereby improving the model's nonlinear representation ability for features.

[0111] 3. Model construction output: Through the combination of the above modules, the feature selection cross model is constructed. The purpose of the model structure is to reduce the interference of noise features while maximizing the effective combination of image features to enhance the model's ability to classify lesions.

[0112] As an implementation of this embodiment, the process of training the feature selection cross model includes:

[0113] 1. Training set input: Use the training set expanded by the time frame expansion (EFU) method to input the image features into the constructed feature selection cross model. The expanded sample set provides the model with data samples of multiple time frames, enabling the model to adapt to the characteristics of dynamic changes in the heart.

[0114] 2. Model training:

[0115] Training process: In the feature selection module, the contribution of each feature is gradually adjusted according to the attention weight, and the weight is updated through backpropagation to reduce the loss.

[0116] Tuning of explicit and implicit feature cross modules: During the training process, the multi-order cross features generated by the explicit feature cross module are adjusted multiple times, and the nonlinear relationship between feature combinations is continuously optimized through the multi-layer perceptron network of the implicit feature cross module.

[0117] 3. Performance evaluation:

[0118] Model performance evaluation indicators: After training, the model is validated on an independent test set, and the accuracy of model classification is evaluated by generating ROC curves and calculating AUC values ​​to ensure its applicability in cardiac magnetic resonance movie image classification tasks.

[0119] Comparison of results: The AUC value of the model was compared with other methods to verify the advantages of the model in radiomics feature classification and ensure that the model has good robustness and high diagnostic accuracy.

[0120] 4. Training output:

[0121] The optimized and trained feature selection crossover model was obtained, which has efficient classification ability and can accurately distinguish the lesion characteristics reflected in cardiac magnetic resonance movie images.

[0122] As an implementation method of this embodiment, step Sd: construction and training of a deep learning model.

[0123] Specifically, step Sd: Based on the expanded training set, a deep learning model including feature selection and feature crossover functions is constructed. The structure of the model is as follows: Figure 2 As shown in A in the figure. The model is mainly divided into a feature selection module based on the attention mechanism and an explicit / implicit feature cross module.

[0124] The structure of the feature selection module based on the attention mechanism is as follows Figure 2As shown in Figure 1, the module screens radiomic features through a two-layer fully connected network. The first layer of the fully connected network is responsible for reducing the feature dimension to extract key information, and the second layer of the fully connected network assigns an attention weight between 0 and 1 to each feature, indicating the importance of the model to each feature, thereby reducing the interference of noise features and improving the signal-to-noise of radiomic features.

[0125] The structure of the dominant feature cross module is as follows Figure 2 As shown in C in , this module crosses different features of the sample and generates corresponding cross coefficients, so that the newly generated cross features have higher lesion representation capabilities. The feature cross formula of the (l+1)th order is as follows, where x l represents the l-order cross feature of the input, x l+1 represents the (l+1)th order feature after feature crossover, matrix W l Cross the input features and store the cross coefficients, b l represents the bias term, ⊙ Represents the Hadamard product. By setting the number of explicit feature cross modules in the model, the model can obtain new features of different orders after cross-combining all imaging omics. These newly generated cross features have better representation ability and enhance the classification performance of the model.

[0126] x l+1 =x 0⊙ (W l x l +b l )+x l

[0127] The structure of the implicit feature cross module is as follows Figure 2 As shown in D in Figure 1, this module inputs the features of the samples into a multi-layer perceptron (MLP) and uses the fitting ability of the neural network to capture the implicit nonlinear relationship between feature combinations.

[0128] Optionally, the process of evaluating the classification performance of the trained feature selection crossover model includes:

[0129] The model was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC) to quantify the classification accuracy of the model;

[0130] The performance of the model in the cardiac magnetic resonance movie image classification task was verified by comparing the AUC value with the existing method.

[0131] As an implementation of this embodiment, step Se: evaluation of model classification performance.

[0132] Specifically, step Se: after the model is built, the model is trained on the training set processed by the EFU method, and its classification performance on the independent test set is comprehensively evaluated.

[0133] The effect of the cardiac magnetic resonance movie image analysis method based on time frame expansion and feature selection crossover of the present invention can be illustrated by real data:

[0134] (1) Real data experimental process

[0135] To demonstrate the effect of the present invention, a real data set of 1878 cardiac magnetic resonance movie images of myocardial segments from 93 patients was used for testing, of which 1370 were used for training the model and performing data enhancement using the EFU method, and 342 were used for testing the classification effect of the present invention.

[0136] The deep selection crossover model (method A) of the time series integration method EFU described in the present invention is compared with the deep selection crossover model (method B) without using the time series integration method EFU, the existing machine learning algorithm model (method C) and the simple neural network (method D), and the receiver operating characteristic (ROC) curve and the area under the curve (AUC) of the classification method are obtained, and the ROC curve and AUC are used as measures of classifier performance.

[0137] (2) Experimental results

[0138] The classification ROC curves of the two methods on the real experimental data set are Figure 3 The true positive rate refers to the percentage of positive cases that are actually positive and are correctly judged as positive using this method, and the false positive rate refers to the percentage of negative cases that are actually negative and are incorrectly judged as positive using this method. Figure 3 As shown in the figure, the ROC curve of this method is higher than that of the other two methods in most threshold ranges; the AUC value comparison: the AUC value of this method is 0.767, the AUC value of method B is 0.699, the AUC value of method C is 0.596, and the AUC value of method D is 0.614. The area under the curve (AUC) can measure the overall classification performance. The larger the AUC value, the better the overall performance of the classification method. Therefore, this method is better than the other two methods.

[0139] The experimental results show that the cardiac magnetic resonance movie image analysis method based on the crossover of time frame expansion and feature selection described in the present invention effectively improves the classification performance.

[0140] Based on this, the embodiment of the present invention introduces an innovative method for classifying sample imaging features extracted from cardiac magnetic resonance film (Cine-MRI) images through time series information integration and feature selection cross-processing technology. The core steps include: first, preprocessing the patient's Cine-MRI images to obtain standardized image data. Secondly, manually outline and segment the left ventricular myocardial segment to clarify the specific object of feature extraction. Subsequently, the feature extraction module mines imaging features of different dimensions from the processed image. On this basis, the present invention proposes a unique data processing strategy to process the imaging features of different time frames of the sample: the time frame expansion (Expanded Frame Unfolding, EFU) method, by independently processing the feature data belonging to different time frames of the same sample, so as to expand the scale of the training data set. In addition, the present invention designs a feature selection cross-model, which selects and reconstructs sample image features to construct more expressive features to achieve accurate classification. This technical solution of the present invention not only shows innovation in the unique processing of different time frames of Cine-MRI images, but also significantly improves the accuracy of sample classification through the efficient use of sample imaging features.

[0141] Embodiment 2

[0142] Based on the same general inventive concept, Figure 4 As shown, the present invention also provides a cardiac magnetic resonance movie image classification system based on time frame expansion and feature selection crossover. The cardiac magnetic resonance movie image classification system based on time frame expansion and feature selection crossover provided by the present invention is described below. The cardiac magnetic resonance movie image classification system based on time frame expansion and feature selection crossover described below and the cardiac magnetic resonance movie image classification method based on time frame expansion and feature selection crossover described above can be referred to each other. The system includes:

[0143] An image acquisition module, used for acquiring cardiac magnetic resonance movie images;

[0144] A region of interest division module is used to pre-process cardiac magnetic resonance movie images and divide regions of interest;

[0145] The radiomics feature processing module is used to extract radiomics features from the images after dividing the regions of interest, and integrate the radiomics features based on time series information;

[0146] A model training module is used to construct and train a feature selection cross model to obtain a trained feature selection cross model;

[0147] Model evaluation module, used to evaluate the classification performance of the trained feature selection cross model and obtain a model that passes the evaluation;

[0148] A classification module for classifying cardiac magnetic resonance cine images based on the evaluated model.

[0149] Optionally, the model training module includes a model building unit and a model training unit, wherein:

[0150] The model building unit is used to build a feature selection crossover model;

[0151] The model training unit is used to train the feature selection cross model to obtain the trained feature selection cross model;

[0152] The feature selection cross model includes a feature selection module based on the attention mechanism, an explicit feature cross module, and an implicit feature cross module;

[0153] The feature selection module is used to allocate attention weights through a two-layer fully connected network to screen radiomics features and reduce noise feature interference;

[0154] The dominant feature crossover module is used to generate new crossover features by cross-processing the radiomics features;

[0155] The implicit feature cross module is used to capture the nonlinear relationship between feature combinations based on the multi-layer perceptron.

[0156] It should be understood that the cardiac magnetic resonance movie image classification system based on time frame expansion and feature selection cross-talk provided by the embodiment of the present invention has all the advantages of the cardiac magnetic resonance movie image classification method based on time frame expansion and feature selection cross-talk provided by the above-mentioned embodiment.

[0157] The present invention also provides a computer terminal device, comprising:

[0158] one or more processors;

[0159] A memory, coupled to the processor, for storing one or more programs;

[0160] When one or more programs are executed by one or more processors, the one or more processors implement the steps of a cardiac magnetic resonance movie image classification method based on the intersection of time frame expansion and feature selection.

[0161] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of a cardiac magnetic resonance movie image classification method based on the intersection of time frame expansion and feature selection are implemented.

[0162] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A cardiac magnetic resonance movie image classification method based on the intersection of time frame expansion and feature selection, characterized in that: The following steps are involved: Acquiring cardiac magnetic resonance movie images, preprocessing the cardiac magnetic resonance movie images and dividing regions of interest; Extract radiomic features from images after dividing the region of interest, and integrate the radiomic features based on time series information; A feature selection cross model is constructed and trained to obtain a trained feature selection cross model, the classification performance of the trained feature selection cross model is evaluated, and cardiac magnetic resonance movie image classification is performed based on the evaluated model.

2. The method according to claim 1, characterized in that The process of preprocessing the cardiac magnetic resonance movie image and dividing the region of interest includes: Standardize cardiac magnetic resonance cine image data to eliminate resolution differences caused by different imaging devices; The Min-Max method was used to standardize the brightness and contrast of cardiac magnetic resonance movie images to reduce the differences between images; The region of interest of cardiac magnetic resonance cine images was determined by manual delineation.

3. The method according to claim 1, characterized in that: The process of extracting radiomics features from the image after dividing the region of interest includes: Features are extracted from the region of interest, wherein the imaging features include shape features, first-order statistical features, and texture features.

4. The method according to claim 1, characterized in that: The process of integrating radiomics features based on time series information includes: Using the time frame expansion (EFU) method, the radiomics features of the same sample at different time frames were treated as independent samples to obtain a quantitatively expanded dataset.

5. The method according to claim 4, characterized in that The feature selection cross model includes a feature selection module based on an attention mechanism, an explicit feature cross module and an implicit feature cross module; The process of training the feature selection crossover model includes: The expanded data set is input into the feature selection cross model for training. In the feature selection module, the contribution of each feature is gradually adjusted according to the attention weight, and the weight is updated by back propagation. Then, the multi-order cross features generated by the explicit feature cross module are adjusted, and the nonlinear relationship between feature combinations is optimized through the multi-layer perceptron network of the implicit feature cross module.

6. The method according to claim 1, characterized in that The process of evaluating the classification performance of the trained feature selection crossover model includes: The model was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC) to quantify the classification accuracy of the model; The performance of the model in the cardiac magnetic resonance movie image classification task was verified by comparing the AUC value with the existing method.

7. A cardiac magnetic resonance movie image classification system based on the intersection of time frame expansion and feature selection, characterized in that: The system for implementing the cardiac magnetic resonance movie image classification method based on time frame expansion and feature selection crossover according to any one of claims 1 to 6 comprises: An image acquisition module, used for acquiring cardiac magnetic resonance movie images; A region of interest division module, used for preprocessing the cardiac magnetic resonance movie image and dividing the region of interest; The radiomics feature processing module is used to extract radiomics features from the images after dividing the regions of interest, and integrate the radiomics features based on time series information; A model training module is used to construct and train a feature selection cross model to obtain a trained feature selection cross model; A model evaluation module, used to evaluate the classification performance of the trained feature selection cross model and obtain a model that passes the evaluation; A classification module for classifying cardiac magnetic resonance cine images based on the evaluated model.

8. The system according to claim 7, characterized in that The model training module includes a model building unit and a model training unit, wherein: The model building unit is used to build a feature selection crossover model; The model training unit is used to train the feature selection cross model to obtain a trained feature selection cross model; The feature selection cross model includes a feature selection module based on an attention mechanism, an explicit feature cross module and an implicit feature cross module; The feature selection module is used to allocate attention weights through a two-layer fully connected network to screen radiomics features and reduce noise feature interference; The dominant feature crossover module is used to generate new crossover features by crossover processing the radiomics features; The implicit feature cross module is used to capture the nonlinear relationship between feature combinations based on a multi-layer perceptron.

9. A computer terminal device, characterized in that: include: one or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the various steps of the cardiac magnetic resonance movie image classification method based on the intersection of time frame expansion and feature selection as described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the method for classifying cardiac magnetic resonance movie images based on the intersection of time frame expansion and feature selection as described in any one of claims 1 to 6 is implemented.

Images