Heart failure combined left bundle branch conduction block classification method based on feature fusion

Abstract

The invention discloses a heart failure combined left beam branch conduction block classification method based on feature fusion, and the method comprises the steps: extracting radiomics features and deep learning features through a heart magnetic resonance image, carrying out the feature fusion and selection, constructing a classification model through a machine learning algorithm, and carrying out the classification of the heart failure combined left beam branch conduction block. According to the method, the advantages of deep learning in the aspect of image processing and the high efficiency of machine learning in a classification task are combined, and complex information of a heart image can be reflected more comprehensively and accurately by combining radiomics features and deep learning features; an automatic and intelligent feature extraction and classification method is utilized, a large amount of medical image data can be rapidly processed and analyzed, accurate classification of the heart failure combined left bundle branch conduction block is achieved, the method has higher classification accuracy and higher robustness, and in the field of medical image classification, the method has the advantages that the classification accuracy is improved, and the classification efficiency is improved. And the efficiency and the quality of medical services can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image classification, and in particular to a classification method for heart failure combined with left bundle branch block based on feature fusion. Background Art

[0002] Heart failure (HF) is the terminal stage of various cardiovascular diseases, characterized by high morbidity, high hospitalization rate, poor quality of life and high medical costs. Left bundle branch block (LBBB) is common in HF patients. Whether LBBB and non-LBBB can be correctly classified plays a vital role in decision-making and prognosis. At present, LBBB is mainly classified clinically by electrocardiogram and imaging examinations, but this method has the problems of strong subjectivity and insufficient accuracy, which may lead to inconsistent classification results. In order to improve the classification accuracy of LBBB, cardiovascular imaging technology has received extensive attention, especially cardiac magnetic resonance (CMR), which can evaluate myocardial deformation in detail and provide new possibilities for the accurate classification of LBBB.

[0003] Machine learning and deep learning are developing rapidly in the medical field. Medical image classification is one of the important applications. Radiomics combines medical imaging and machine learning. It uses computer algorithms and image processing technology to extract a large number of quantitative features from medical images, including morphological features, texture features, density features, edge features, etc. These features can be used to classify different diseases, predict disease progression, and evaluate treatment response. Deep learning, as a branch of machine learning, can automatically learn and extract high-level features from data by building deep neural network models. It is particularly suitable for processing complex medical image data. Therefore, how to combine CMR with machine learning and deep learning to accurately classify patients with HF and LBBB is still a problem that needs to be solved urgently. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a classification method for heart failure combined with left bundle branch block based on feature fusion. The method combines the radiomics features and deep learning features of CMR images, and can accurately classify HF combined with LBBB through a machine learning algorithm, providing a reliable basis for clinical diagnosis and treatment decisions. The method specifically includes the following steps:

[0005] Step a: constructing a data set, the data set including cardiac magnetic resonance images of heart failure combined with left bundle branch block, and preprocessing the medical image data;

[0006] Step b: extracting a large number of radiomic features from the preprocessed images, including morphological features, texture features, density features, and edge features;

[0007] Step c: Using the pre-trained deep learning model and fine-tuning it, feature extraction is performed on the pre-processed image to obtain deep learning features;

[0008] Step d: Combining radiomics features and deep learning features to perform feature fusion, obtain the fused feature vector, and perform feature selection;

[0009] Step e: Model the filtered features, conduct model evaluation training to verify the performance of the fusion feature model, and output the model evaluation results;

[0010] Step f: Use the model evaluation result to classify the input heart failure combined with left bundle branch block image, and obtain and output the classification result.

[0011] Preferably, in step a, image preprocessing includes denoising, contrast enhancement, myocardial contour extraction, and generation of a mask image of a region of interest, and the image size is unified to 256*256.

[0012] Preferably, in step b, radiomics feature extraction is implemented through the PyRadiomics open source library, and these features include gradient features, shape features, texture features, intensity features and wavelet features, as well as image derivative processing such as gradient, exponential, logarithm, square, square root and wavelet filtering.

[0013] Preferably, in step c, extracting deep learning features specifically includes: using a ResNet50 model pre-trained on the ImageNet dataset, fine-tuning the ResNet50 model on the ACDC2017 heart dataset to optimize the model's classification ability for cardiac image features, using the fine-tuned ResNet50 model to extract features from the input cardiac image, selecting the output of the penultimate average pooling layer as the image feature, the average pooling layer is responsible for adjusting the output of the previous layer to an average pooling feature map of a fixed size, and its output features can capture key morphology, texture and spatial relationship information in the cardiac image, and using the extracted features in subsequent analysis and classification tasks to achieve accurate diagnosis of heart disease.

[0014] Preferably, in step d, feature selection is achieved through a series of statistical analysis and machine learning techniques, including: using t-test to screen the initial fusion features, selecting features with p-values ​​less than 0.05 as important features, then applying Pearson correlation coefficient analysis to further evaluate the correlation between these features, removing redundant features with correlation coefficients greater than 0.9 and retaining the most representative feature subset, and then applying the LASSO model to compress and select the remaining features. The model can automatically select the features that contribute the most to the classification task while ensuring model performance by applying the L1 regularization term, thereby obtaining the final non-zero coefficient feature set.

[0015] Preferably, in step e, the established model includes LR, Bayes, SVM and KNN,stratified sampling was used to divide the data set into training set and test set in a ratio of 7:3. The training set was used to build and train the prediction model, while the test set was used to evaluate the performance and generalization ability of the model.

[0016] Preferably, in the model evaluation, after training is completed, the model is evaluated using a test set, and the evaluation indicators include ROC curve and AUC value, accuracy, sensitivity, specificity and F1 score, and decision curve analysis (DCA) is also used to evaluate the net benefit of the model.

[0017] Preferably, in the verification that the fusion feature model is superior to the radiomics feature model and the deep learning model, the radiomics features and the deep learning model are separately modeled and evaluated, and compared with the fusion feature model to verify whether they can enrich the feature representation and enable the model to capture more diverse information.

[0018] Preferably, in the step f, the method of using the model evaluation result to classify the input heart failure combined with left bundle branch block image specifically includes: inputting the heart failure combined with left bundle branch block cardiac magnetic resonance image into the trained fusion feature model, using the fusion feature model to classify and predict the heart failure combined with left bundle branch block cardiac magnetic resonance image, and outputting the classification result of the classification prediction.

[0019] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0020] 1. By combining radiomics features and deep learning features, it can more comprehensively and accurately reflect the complex information of cardiac images and improve the classification accuracy of heart failure combined with left bundle branch block;

[0021] 2. Using automated and intelligent feature extraction and classification methods, it can quickly process and analyze large amounts of medical imaging data and improve diagnostic efficiency;

[0022] 3. Provide objective and accurate diagnostic basis for clinicians, assist them in developing personalized treatment plans, and improve treatment effects and patients’ quality of life. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0024] In the attached picture:

[0025] Figure 1 A schematic diagram of the overall process provided for an embodiment of the present invention;

[0026] Figure 2 A schematic diagram of data set division provided by an embodiment of the present invention;

[0027] Figure 3 A schematic diagram of feature extraction provided by an embodiment of the present invention;

[0028] Figure 4 A schematic diagram of feature selection results provided by an embodiment of the present invention;

[0029] Figure 5 A schematic diagram showing a comparison of ROC curves of various models provided in an embodiment of the present invention;

[0030] Figure 6 A schematic diagram for comparing DCA curves of various models provided in an embodiment of the present invention;

[0031] Figure 7 A schematic diagram of the confusion matrix classification results of the fusion feature model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] The present invention combines Figures 1 to 7 , providing the following technical solution: a classification method for heart failure combined with left bundle branch block based on feature fusion, the specific operation process of the method includes:

[0034] Step a: construct a data set including cardiac magnetic resonance images of heart failure combined with left bundle branch block, and preprocess the medical image data in the data set;

[0035] In this embodiment, image preprocessing includes denoising, contrast enhancement, myocardial contour extraction, and generation of a mask image of the region of interest. The image size is unified to 256*256. The region of interest is segmented first to segment the region of interest from the original image to represent the myocardial part. In this embodiment, the region of interest is segmented to represent the myocardial part. Figure 1 , Figure 1 The processing flow of the present invention is shown in Figure 1 In the figure, two different images represent the original image and the mask image respectively. The accuracy of ROI segmentation is crucial for subsequent feature extraction and model building.

[0036] Step b: extracting radiomic features from the preprocessed images, including morphological features, texture features, density features, and edge features;

[0037] In this embodiment, after the segmentation of the region of interest is completed, the next step is to extract radiomic features and deep learning features. The radiomic features include shape, texture and first-order statistics features, that is, the radiomic feature extraction is implemented through the PyRadiomics open source library. These features include gradient features, shape features, texture features, intensity features and wavelet features, as well as image derivative processing such as gradient, exponential, logarithm, square, square root and wavelet filtering.

[0038] Step c: extract features from the preprocessed image using a pretrained deep learning model after fine-tuning to obtain deep learning features;

[0039] In this embodiment, for deep learning features, a ResNet50 model pre-trained on the ImageNet dataset is selected, and after fine-tuning on the ACDC2017 heart dataset, the output of the penultimate average pooling layer is selected as the image feature. These features can reflect the shape, size, texture, spatial relationship and other attributes of the region of interest;

[0040] Exemplarily, extracting deep learning features specifically includes: using a ResNet50 model pre-trained on the ImageNet dataset, fine-tuning the ResNet50 model on the ACDC2017 heart dataset to optimize the model's classification ability for cardiac image features, using the fine-tuned ResNet50 model to extract features from the input cardiac image, and selecting the output of the penultimate average pooling layer as the image feature, wherein the average pooling layer is responsible for adjusting the output of the previous layer to an average pooling feature map of a fixed size, and its output features can capture key morphological, texture, and spatial relationship information in the cardiac image, and the extracted features are used in subsequent analysis and classification tasks to achieve accurate diagnosis of heart disease.

[0041] Step d: Combining radiomics features and deep learning features to perform feature fusion, obtain the fused feature vector, and perform feature selection;

[0042] In this embodiment, after the two types of features are fused, the next step is to perform feature selection. The purpose of feature selection is to select the most representative features that best reflect disease information from a large number of features to reduce the complexity of the model and improve the generalization ability of the model. Here, t-test, Pearson and LASSO algorithms are used for feature selection. T-test will preliminarily screen features based on the p-value of the feature. The Pearson correlation coefficient can be used to evaluate the linear correlation between the feature and the target variable, and LASSO is a linear regression method that can achieve feature selection and weight sparsification by introducing the L1 regularization term.

[0043] Exemplarily, feature selection is achieved through a series of statistical analysis and machine learning techniques, including: using t-test to screen the initial fusion features, selecting features with p-values ​​less than 0.05 as important features, and then applying Pearson correlation coefficient analysis to further evaluate the correlation between these features, removing redundant features with correlation coefficients greater than 0.9 and retaining the most representative feature subset, and then applying the LASSO model to compress and select the remaining features. By applying the L1 regularization term, the model can automatically select the features that contribute the most to the classification task while ensuring the performance of the model, and obtain the final non-zero coefficient feature set.

[0044] Step e: Model the filtered features, conduct model evaluation training to verify the performance of the fusion feature model, and output the model evaluation results.

[0045] In this embodiment, the last step is to build a prediction model and use the test set data to evaluate the model, respectively building LR, Bayes, SVM and KNN models use stratified sampling to divide the data set into training set and test set in a ratio of 7:3. The training set is used to build and train the prediction model, while the test set is used to evaluate the performance and generalization ability of the model and evaluate the model. In the model evaluation, after the training is completed, the test set is used to evaluate the model. The evaluation indicators include ROC curve and AUC value, accuracy, sensitivity, specificity and F1 score. Decision curve analysis (DCA) is also used to evaluate the net benefit of the model and select the best performing model. Finally, the fusion feature model is verified to verify that the fusion feature model is superior to the radiomics feature model and the deep learning model. The radiomics features and deep learning models are modeled and evaluated separately and compared with the fusion feature model to verify whether they can enrich the feature representation and enable the model to capture more diverse information.

[0046] In this embodiment, combined with Figure 2 , Figure 2 Schematic diagram of data set division. The data set used in this embodiment is HF merged LBBB data instances from 3 centers, with a total of 200 HF instances, of which 73 instances showed LBBB mode and 127 instances showed non-LBBB mode.

[0047] In this embodiment, combined with Figure 3 , Figure 3 The figure is a schematic diagram of feature extraction. By constructing radiomics features, deep learning features and fusion features respectively and then comparing them, it is verified that the fusion features have the best model prediction ability;

[0048] Exemplarily, manually extracted radiomic features can be divided into three categories: geometry, intensity, and texture. Geometric features are signal-independent features that describe the two-dimensional and three-dimensional shape of the myocardium. Intensity features depict the first-order statistical distribution of voxel intensity within the myocardium. Texture features describe the pattern, i.e., the second- and higher-order spatial distribution of intensity. Texture features are derived using several different methods, including gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-level difference matrix (NGTDM).

[0049] Preferably, the CMR images are firstly subjected to derivation, including gradient, exponential, logarithmic, square, square root and wavelet filtering (HHH, HHL, HLH, HLL, LHH, LHL, LLH, LLL), and then 1288 manual features are extracted, including 306 first-order features, 14 shape features and 968 texture features, and all the extracted features are normalized using the Z-Score method;

[0050] Furthermore, deep learning features are extracted through the ResNet50 network pre-trained on the ImageNet dataset. In order to improve the accuracy and pertinence of feature extraction, the ACDC2017 heart dataset is used to fine-tune ResNet50. The dataset contains CMR images of 150 patients, and each patient has several CMR images. After screening out all images containing myocardium, image preprocessing operations such as resizing and normalization are performed. A total of 2817 images with a size of 256*256 are obtained. The images are input into ResNet50 for training for 200 cycles with a BatchSize of 32. After the training is completed, deep learning feature extraction is performed and the corresponding deep learning results are obtained. Finally, the two categories are fused and feature selection is performed to screen out the most representative features for subsequent classification tasks.

[0051] Preferably, since many of the initially extracted features may be redundant or meaningless, the first step is to reduce the feature dimension using t-test, and a significance level of p = 0.05 is selected. If the p-value of a feature is less than 0.05, it is considered to be significantly different, so it is retained as an important feature for further analysis; on the contrary, if the p-value is greater than 0.05, the feature is deleted; the next feature selection step is Pearson, and the correlation coefficient threshold is 0.9. If the correlation coefficient of two features is greater than the threshold, one of them is retained and the other is discarded. Finally, based on the regularization weight λ, the LASSO regression method compresses all regression coefficients to zero and assigns zero coefficients to many irrelevant features. The optimal λ is found through 10-fold cross validation, and the features with non-zero coefficients obtained from this process are selected for classification model fitting.

[0052] In this embodiment, combined with Figure 4 , Figure 4 This is a schematic diagram of feature selection results. After the fusion features were selected, 11 features were obtained, including 2 radiomics features and 9 deep learning features. Each feature corresponds to a coefficient value, and its size and positive or negative reflects the importance of the feature in the model.

[0053] Preferably, four classification models, LR, Naive Bayes, KNN and SVM, were constructed and trained respectively. Subsequently, these models were evaluated using the test data set. The evaluation indicators included AUC, accuracy, sensitivity, specificity, F1 score and decision curve analysis (DCA) to verify the prediction ability and reliability of the model. The comparison results of some indicators are shown in Table 1. It can be seen that the SVM model has the highest accuracy, specificity and F1 score on the test set, which are 0.883, 0.900 and 0.829 respectively, and the AUC value is also the highest, which is 0.949;

[0054] Table 1: The performance of the four models on the training and test sets is shown in the following table;

[0055]

[0056] In this embodiment, combined with Figure 5 , Figure 5 The figure is a schematic diagram for comparing the ROC curves of various models, where A is the training set result and B is the test set result. The curve area enclosed by the SVM model in the training set is the largest and it is also the best in the test set.

[0057] In this embodiment, combined with Figure 6 , Figure 6 The DCA curve comparison diagram of each model, where A, B, C, and D are LR, Bayes, SVM and KNN models, the decision curve takes into account the patient's risks and benefits to evaluate whether the clinical decision is feasible. The horizontal axis is the risk threshold and the vertical axis is the net benefit. The larger the area enclosed by the model and the treat all and treat none curves, the higher the clinical benefit of the model. As can be seen from the figure, the SVM model has the largest area and the highest net benefit.

[0058] For example, after evaluation, SVM outperformed other models among four different machine learning models. These results indicate that SVM has better classification ability;

[0059] Specifically, the comparison results of the three feature models are shown in Table 2. The radiomics feature is modeled using the SVM model. It can be seen that the performance of the radiomics feature model on the test set has dropped significantly, and the overall performance index is relatively low, so it is not the best choice. The deep learning model has shown high performance on both the training set and the test set, and the performance drop on the test set is relatively small, indicating that the model has good generalization ability. However, the fusion feature model has the highest AUC value (0.996) on the training set because it combines the advantages of multiple features, which is 0.046 higher than the deep learning model, and it also maintains a high performance on the test set (AUC = 0.949), showing a strong comprehensive prediction ability. The fusion feature model integrates the information of radiomics features and deep learning features, and is better at capturing the complexity and nuances of the data, thus showing better performance;

[0060] Table 2: The comparison results of the three feature models are shown in the following table;

[0061]

[0062] Step f: Use the model evaluation result to classify the input heart failure combined with left bundle branch block image, and obtain and output the classification result.

[0063] In this embodiment, combined with Figure 7 , Figure 7 The figure is a schematic diagram of the confusion matrix classification results of the fusion feature model. The cardiac magnetic resonance images of heart failure combined with left bundle branch block are input into the trained fusion feature model. The fusion feature model is used to classify and predict the cardiac magnetic resonance images of heart failure combined with left bundle branch block, and the classification results of the classification prediction are output. Compared with other methods, 3 LBBB examples were wrong with a sensitivity of 0.959, and 5 non-LBBB examples were wrong with a specificity of 0.961. The overall accuracy is 0.960, which shows that the fusion feature model can effectively distinguish the two categories, can quickly process and analyze a large amount of medical imaging data, and improve diagnostic efficiency.

[0064] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0065] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A classification method for heart failure combined with left bundle branch block based on feature fusion, characterized by: The following steps are involved: Step a: construct a data set including cardiac magnetic resonance images of heart failure combined with left bundle branch block, and preprocess the medical image data in the data set; Step b: extracting radiomic features from the preprocessed image, wherein the radiomic features include morphological features, texture features, density features, and edge features; Step c: Using the pre-trained deep learning model and fine-tuning it, feature extraction is performed on the pre-processed image to obtain deep learning features; Step d: combining the radiomics feature and the deep learning feature to perform feature fusion, obtain a fused feature vector, and perform feature selection; Step e: Model the filtered features, conduct model evaluation training to verify the performance of the fusion feature model, and output the model evaluation results; Step f: Use the model evaluation result to classify the input heart failure combined with left bundle branch block image, and obtain and output the classification result.

2. According to claim 1, a classification method for heart failure combined with left bundle branch block based on feature fusion, characterized in that: The preprocessing process of the medical image data in the data set includes: denoising, contrast enhancement, myocardial contour extraction, and generation of a mask image of the region of interest. The image size is unified to 256*256. By performing region of interest segmentation, the region of interest is segmented from the original image to represent the myocardial part.

3. The method for classifying heart failure combined with left bundle branch block based on feature fusion according to claim 2, characterized in that: The extracting of radiomics features further comprises: After the segmentation of the region of interest is completed, radiomics features and deep learning features are extracted. Radiomics features include shape, texture and first-order statistics features. Radiomics feature extraction is achieved through the PyRadiomics open source library. The radiomics features include gradient features, shape features, texture features, intensity features and wavelet features, as well as gradient, exponential, logarithmic, square, square root and wavelet filter image derivation processing.

4. The method for classifying heart failure combined with left bundle branch block based on feature fusion according to claim 3, characterized in that: The method further comprises: using a pre-trained deep learning model and fine-tuning it to extract features from the pre-processed image, and fine-tuning the ResNet50 model on the ACDC2017 heart dataset to optimize the model's ability to classify cardiac image features; Using the fine-tuned ResNet50 model, feature extraction is performed on the input heart image, and the output of the penultimate average pooling layer is selected as the image feature; The average pooling layer is responsible for adjusting the output of the previous layer to an average pooling feature map of a fixed size, and its output features can capture key morphological, texture and spatial relationship information in the heart image; The extracted features are used in subsequent analysis and classification tasks to diagnose heart disease.

5. The method for classifying heart failure combined with left bundle branch block based on feature fusion according to claim 4, characterized in that: The image features are used to reflect the attributes of the shape, size, texture, and spatial relationship of the region of interest.

6. The method for classifying heart failure combined with left bundle branch block based on feature fusion according to claim 5, characterized in that: The feature selection includes: performing feature selection by using t-test, Pearson correlation coefficient and LASSO algorithm, wherein the t-test preliminarily screens features according to the p-value of the features, the Pearson correlation coefficient is used to evaluate the linear correlation between the features and the target variable, and the LASSO algorithm realizes feature selection and weight sparsification by introducing an L1 regularization term.

7. The method for classifying heart failure combined with left bundle branch block based on feature fusion according to claim 6, characterized in that: The feature selection further includes: using t-test to screen the initial fusion features, selecting features with p-values ​​less than 0.05 as important features, applying Pearson correlation coefficient analysis to evaluate the correlation between features, removing redundant features with correlation coefficients greater than 0.9, and retaining the most representative feature subset, and then applying the LASSO model to compress and select the remaining features. By applying the L1 regularization term, the model can automatically select the features that contribute the most to the classification task while ensuring the performance of the model, and obtain the final non-zero coefficient feature set.

8. The method for classifying heart failure combined with left bundle branch block based on feature fusion according to claim 7, characterized in that: The step e specifically includes: constructing a prediction model, evaluating the model using the test set data, and constructing LR, Bayes, SVM and KNN models use stratified sampling to divide the data set into a training set and a test set in a ratio of 7:

3. The training set is used to build and train the prediction model, and the test set is used to evaluate the performance and generalization ability of the model. The evaluation of the model includes: after the training is completed, the model is evaluated using the test set. The evaluation indicators include ROC curve and AUC value, accuracy, sensitivity, specificity and F1 score. The net benefit of the model is evaluated by using decision curve analysis (DCA), and the best performing model is selected. Finally, the fusion feature model is verified to verify that the fusion feature model is superior to the radiomics feature model and the deep learning model. The radiomics features and deep learning models are modeled and evaluated separately and compared with the fusion feature model to verify whether they can enrich the feature representation and enable the model to capture more diverse information.

9. The method for classifying heart failure combined with left bundle branch block based on feature fusion according to claim 8, characterized in that: The method for classifying the input heart failure combined with left bundle branch block images using the model evaluation results specifically includes: inputting the heart failure combined with left bundle branch block cardiac magnetic resonance images into the trained fusion feature model, using the fusion feature model to classify and predict the heart failure combined with left bundle branch block cardiac magnetic resonance images, and outputting the classification results of the classification prediction.

Images