A convolutional neural network-based ACS prediction method

Abstract

The present invention provides an ACS prediction method based on a convolutional neural network, and relates to the field of artificial intelligence computer technology. The method first collects segment data of three main branches of the coronary artery on CCTA data, including segment data at the opening and segment data at the lesion, and divides the data into two categories according to whether an ACS event occurs, and forms a training data set and a verification data set; then a CNN network model for ACS classification is constructed, and the model is trained using the training data set to obtain model parameters; finally, the trained model parameters are loaded into the constructed CNN network model for ACS classification, and then the verification data set is used to verify the model, so as to select the optimal CNN network model for ACS classification, thereby accurately predicting the occurrence of future ACS events, and being able to assist clinical workers in more accurately and objectively predicting future ACS events.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence computer technology, and in particular to an ACS prediction method based on a convolutional neural network. Background Art

[0002] Cardiovascular disease is one of the major causes of death in the world, and acute coronary syndrome (ACS) is its common first manifestation. With the rapid development of medical imaging technology, coronary CT angiography (CCTA) as a non-invasive imaging mode has been widely used in the evaluation and diagnosis of suspected cardiovascular patients.

[0003] Vascular wall inflammation is the main factor leading to the instability of atherosclerotic plaques, which can promote the progression and rupture of coronary atherosclerosis, thereby inducing the occurrence of ACS. In the process of atherosclerosis formation, the inflammatory vascular wall secretes a variety of proinflammatory factors to the pericoronary adipose tissue (PCAT) in a paracrine form, resulting in an imbalance in the water / fat ratio of PCAT. Based on this, the perivascular fat density index (Fat attenuation index, FAI) as a sensitive biomarker that can dynamically reflect coronary inflammation can be used together with vascular plaques to predict ACS. However, changes in the composition of PACT are not only related to vascular inflammation, but also closely related to the dysfunction of adipose tissue caused by fibrosis and microvascular remodeling. Therefore, it is necessary to combine multi-level and multi-dimensional PCAT characteristics to better predict ACS.

[0004] Convolutional neural networks (CNN) are a type of multi-layer neural network. They are mainly composed of convolutional layers, pooling layers, and fully connected layers. Trainable filters and local neighborhood pooling operations are alternately applied to the original input image to form increasingly complex feature hierarchical results. Currently, CNN has been widely used in disease risk prediction, such as benign and malignant assessment of pulmonary nodules, and has achieved good results.

[0005] Although CNN has achieved good results in disease risk prediction, it has not been applied in ACS prediction. Many classic CNN networks based on classification, such as Inception network, VGG network and ResNet network, have achieved good performance in ImageNet challenge, but their performance in ACS prediction is unsatisfactory. Some scholars have developed and designed networks based on disease risk prediction, such as CNN networks for predicting benign and malignant lung nodules. Although they have achieved good results, they cannot effectively predict ACS. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention proposes an ACS prediction method based on convolutional neural network.

[0007] An ACS prediction method based on a convolutional neural network specifically comprises the following steps:

[0008] Step 1: Collecting the vascular segment data required for ACS classification; the data used for ACS classification comes from the vascular segments at the lesions and openings of the three main branch vessels of the coronary artery in CCTA;

[0009] Step 2: Preprocess the acquired vascular segment data; since the segment data are from the three main branches of the coronary artery and are from different parts of the branches, it is necessary to first splice these segment data so that the lesion segments and opening segments of the three main branches of the coronary artery can become separate complete data. Considering the different CCTA scanning methods, the segment data are normalized to the same physical size to ensure the uniformity of the data scale;

[0010] Step 3: First, build a CNN model, and then use labeled vascular segment data to train and verify the CNN model to obtain a CNN model with good performance;

[0011] Step 3.1: Construct a CNN model consisting of two backbone networks, ResNet and SPP-net, with the same structure but different parameters, as well as a channel feature fusion network and a fully connected prediction layer, namely, the FC prediction layer;

[0012] The ResNet network in the backbone network is an improvement on the existing ResNet-50 network; it is composed of a convolutional module, two maximum pooling layers Max Pooling and two Residual modules, namely ResBlock, which are similar to the first three modules of the ResNet-50 network, namely: the convolutional module is composed of a three-dimensional Convolutional layer, BatchNormalization, namely BN and ReLU functions; the first ResBlock1 and the second ResBlock2 are both composed of three sub-convolutional modules and a Shortcut connection, and the sub-convolutional module is composed of a three-dimensional Convolutional layer, BN and ReLU functions; there is a Max Pooling layer between the first convolutional module and the first ResBlock1, and there is a Max Pooling layer between the first ResBlock1 and the second ResBlock2;

[0013] The channel feature fusion network is used to convert the feature map with 2 channels into 1 channel, increase the nonlinearity between channels, and reduce the redundancy of parameters; the channel feature fusion network is composed of two consecutive sub-network modules connected by a Dropout layer, where the sub-network consists of a one-dimensional Convolutional layer, a one-dimensional BN and a ReLU function;

[0014] The FC prediction layer is a fully connected layer used to predict the ACS. It consists of two consecutive sub-nonlinear FC layers and a Linear layer. The sub-nonlinear FC layer consists of a Linear layer, a ReLU function, and a Dropout layer.

[0015] Step 3.2: Train the CNN model based on the training set data to obtain model parameters, thereby building a CNN model with excellent performance;

[0016] The training set data is composed of the input vascular segment data processed by step 2 and label data corresponding to the data, the label means whether the real result corresponding to the data is ACS, and the label is a value of 0 or 1; 0 represents not ACS, and 1 represents ACS; the training set data is input into the constructed CNN model, and the CNN model automatically optimizes its own model parameters by comparing the difference between the output result and the label value, and performing cyclic calculations until the difference between the training output result and the label result is minimized, and the model reaches the optimal state; when the model reaches the optimal state, the optimal model parameters are obtained, and a CNN model with excellent performance is obtained; in the process of model training, the cross entropy function is used as the loss function to perform back propagation of the model;

[0017] Step 3.3: Use validation set data to verify the trained CNN model to ensure that the CNN model has excellent performance;

[0018] The validation set data is composed of the input vascular segment data processed in step 2 and the label data corresponding to the data, and is a group of data that has not participated in the CNN model training; the validation set is input into the CNN model with excellent performance in step 3.2 for verification, and the prediction result of the validation set data is obtained. If the difference between the prediction result and the label of the validation set data is small, the model parameters obtained in step 3.2 are valid model parameters, which are saved and used for the prediction of whether ACS occurs in the vascular segment data; if the difference between the prediction result and the label of the validation set data is large, it is necessary to return to step 3.2 and retrain the CNN model until the CNN model using the model parameters saved in step 3.2 can well predict the validation set data; the model parameters at this time are saved and used for the prediction of whether ACS occurs in the vascular segment data;

[0019] Step 4: Input the vascular segment data required for ACS classification after step 2 into the constructed CNN model, and output the prediction result of whether ACS occurs. The vascular segment data here refers to the data that does not participate in the training and verification of the CNN model, that is, the data without corresponding labels.

[0020] Beneficial technical effects of the present invention:

[0021] The present invention proposes an ACS prediction method based on a convolutional neural network. The method uses a multi-path input method to predict ACS. In the construction of the CNN model, an improved ResNet network and an SPP-net joint network are used as the backbone network to solve the problem of different sizes of multi-path input data; a channel feature fusion network is used to enhance the nonlinear characteristics between channels when fusing multi-channel features; the FC prediction layer designed by the present invention can further reduce redundant parameters, prevent overfitting, and improve the generalization ability of network prediction. The present invention proposes an ACS prediction method based on a convolutional neural network, which analyzes the fat and plaques at the lesions of the three main branch vessels of the coronary artery in the CCTA image, as well as the fat around the blood vessels at the openings of the three main branch vessels, so as to accurately predict the occurrence of ACS.

[0022] The CNN prediction method based on convolutional neural network described in the present invention was verified and tested on CCTA data. The results showed that the CNN network model designed by the present invention can explore the deep features of the segment data of coronary artery lesions and the segment data of coronary artery openings on CCTA data, and can assist clinical workers in more accurately and objectively predicting future ACS events. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Flow chart of the ACS prediction method based on convolutional neural network in an embodiment of the present invention;

[0024] Figure 2 Schematic diagram of a blood vessel segment according to an embodiment of the present invention;

[0025] Figure 3 Flowchart of the CNN model for training and verifying ACS prediction in an embodiment of the present invention;

[0026] Figure 4 CNN model framework diagram for ACS prediction in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The present invention will be further described in detail below through specific embodiments with reference to the accompanying drawings:

[0028] The present invention proposes a convolutional neural network-based ACS prediction method, and verifies the accuracy, specificity and sensitivity of the CNN model designed by the present invention through training and testing, thereby efficiently assisting clinical workers and achieving accurate risk prediction of ACS.

[0029] A convolutional neural network-based ACS prediction method is shown in the attached Figure 1 As shown, the specific steps include:

[0030] Step 1: Collecting the vascular segment data required for ACS classification; the data used for ACS classification comes from the vascular segments at the lesions and openings of the three main branch vessels of the coronary artery in CCTA;

[0031] In CCTA data, the changes of plaques at coronary artery lesions and the fat around the blood vessels and the fat around the blood vessels at the opening of the coronary artery can effectively predict the occurrence of future ACS events. Therefore, the present invention uses the blood vessel segments at the lesions and the blood vessel segments at the opening of the three main branches of the coronary artery (Right Coronary Artery, RCA, Left Circumflex, LCX and Left Anterior Descending, LAD) in CCTA data as input data, such as Figure 2 shown.

[0032] Step 2: Preprocess the acquired vascular segment data; since the segment data are from the three main branches of the coronary artery and are from different parts of the branches, it is necessary to first splice these segment data so that the lesion segments and opening segments of the three main branches of the coronary artery can become separate complete data. Considering the different CCTA scanning methods, the segment data are normalized to the same physical size to ensure the uniformity of the data scale;

[0033] The vascular segment data input in step one are spliced ​​and normalized. There are 6 sets of vascular segment data input in step one, namely, RCA lesion segment data, RCA opening segment data, LCX lesion segment data, LCX opening segment data, LAD lesion segment data, and LAD opening segment data. These 6 segment data need to jointly predict whether the corresponding cases will have ACS events in the future. Considering the input method of the CNN model, the present invention splices the RCA lesion segment data, the LCX lesion segment data and the LAD lesion segment data into a new lesion segment data according to the X-axis direction of the data. Similarly, the RCA opening segment data, the LCX opening segment data and the LAD opening segment data are spliced ​​into a new opening segment data according to the X-axis direction of the data. Finally, the new lesion segment data and the new opening segment data are normalized in physical pixel size to ensure that the input data of all different cases are consistent in physical size.

[0034] Step 3: First, build a CNN model, and then use labeled vascular segment data to train and verify the CNN model to obtain a CNN model with good performance;

[0035] Input the preprocessed data into the constructed CNN model for calculation. Before the CNN model is calculated, it is necessary to first build the CNN model, and then use the training set data to train the model to obtain the model parameters. Use the validation data set data to verify the model to ensure the validity of the model parameters, such as Figure 3 shown.

[0036] Step 3.1: Construct a CNN model consisting of two backbone networks, ResNet and SPP-net, with the same structure but different parameters, as well as a channel feature fusion network and a fully connected prediction layer (Fully Connected Layer), namely, the FC prediction layer; Figure 4 shown.

[0037] The ResNet network in the backbone network of the present invention is an improvement on the existing ResNet-50 network. It is composed of a convolutional module, two maximum pooling layers (Max Pooling) and two Residual modules (ResBlock), which are similar to the first three modules of the ResNet-50 network, namely: the convolutional module is composed of a three-dimensional Convolutional layer, BatchNormalization (BN) and ReLU function; the first ResBlock1 and the second ResBlock2 are both composed of three sub-convolutional modules and a Shortcut connection, and the sub-convolutional module is composed of a three-dimensional Convolutional layer, BN and ReLU function. There is a Max Pooling layer between the first convolutional module and the first ResBlock1, and there is a Max Pooling layer between the first ResBlock1 and the second ResBlock2. Table 1 shows the main structural parameters of the ResNet network in the present invention.

[0038] Table 1 Main structural parameters of ResNet network under different input data

[0039]

[0040]

[0041] The SPP-net in the backbone network of the present invention is a multi-scale network, which can solve the problem of different segment data sizes for each case and extract data features from a multi-scale perspective at the same time. The feature map output size of the SPP-net of the present invention is (1×1×1+2×2×2+4×4×4)*128=9344.

[0042] When the new lesion segment data and the new opening segment data are input into the two backbone networks respectively, feature maps of size [1×9344] will be output respectively. By concatenating the two output feature maps in a concatenated manner, a feature map of size [2×9344] can be obtained.

[0043] The channel feature fusion network is a network designed by the present invention, which is used to convert a feature map with 2 channels into 1 channel, increase the nonlinearity between channels, and reduce the redundancy of parameters.

[0044] The channel feature fusion network designed by the present invention is composed of two consecutive sub-network modules connected by a Dropout layer, wherein the sub-network is composed of a one-dimensional Convolutional layer, BN and ReLU functions. Table 2 shows the parameters of the channel feature fusion network;

[0045] Table 2 Channel feature fusion network parameters

[0046]

[0047] The FC prediction layer is a fully connected layer used to predict the final ACS. It consists of two consecutive sub-nonlinear FC layers plus a Linear layer, where the sub-nonlinear FC layer consists of a Linear layer, a ReLU function, and a Dropout layer. Table 3 shows the network parameters of the FC prediction layer;

[0048] Table 3 Network parameters of FC prediction layer

[0049]

[0050] Step 3.2: Train the CNN model based on the training set data to obtain model parameters, thereby building a CNN model with excellent performance;

[0051] After building the CNN model, it is necessary to use the training set data to train the CNN model so that the model reaches the optimal state and obtains the optimal model parameters. The training set data consists of the input vascular segment data processed in step 2 and the label data corresponding to the data. The label means whether the real result corresponding to the data is ACS. The label is a value of 0 and 1. Generally, 0 represents not ACS and 1 represents ACS. The data obtained by the method in step 1 and its corresponding label data are input into the constructed CNN model. The CNN model automatically optimizes its own model parameters by comparing the difference between the output result and the label value, and cyclically calculates until the difference between the output result of the training and the label result is minimized, then the model reaches the optimal state. When the model reaches the optimal state, the optimal model parameters can be obtained. In the process of model training, the cross entropy function is used as the loss function to perform back propagation of the network.

[0052] Step 3.3: Use validation set data to verify the trained CNN model to ensure that the CNN model has excellent performance;

[0053] The optimal model parameters obtained in step 3.2 are optimal for the training set data, but they may not have generalization capabilities, so it is necessary to verify the model parameters. The verification method is to use the validation set data to compare the output results. The composition and acquisition method of the validation set data are the same as those of the training set data, except that the validation set data is a new set of data that does not participate in the CNN model training.

[0054] The method of verifying the CNN model with validation set data is as follows: input the optimal model parameters obtained in step 3.2 into the constructed CNN model to obtain the prediction results of the validation set data. If the prediction results have little difference with the labels of the validation set data, the model parameters obtained in step 3.2 are valid model parameters and can be saved for prediction of new data. If the prediction results have a large difference with the labels of the validation set data, it is necessary to return to step 3.2 and retrain the CNN model until the CNN model using the model parameters saved in step 3.2 can predict the validation set data well. The model parameters at this time can be saved for prediction of data.

[0055] Step 4: Input the vascular segment data required for ACS classification after step 2 into the constructed CNN model, and output the prediction result of whether ACS occurs. The vascular segment data here refers to the data that does not participate in the training and verification of the CNN model, that is, the data without corresponding labels.

[0056] The present invention adopts the Softmax function as the classification function to convert the input result of the CNN model into a number between 0 and 1, thereby predicting the risk probability of ACS events occurring in the future for the case corresponding to the input data.

Claims

1. An ACS prediction method based on convolutional neural network, characterized in that: The specific steps include: Step 1: Collecting the vascular segment data required for ACS classification; the data used for ACS classification comes from the vascular segments at the lesions and openings of the three main branch vessels of the coronary artery in CCTA; Step 2: Preprocess the acquired vascular segment data; since the segment data are from the vascular segments at the lesions and openings of the three main branches of the coronary artery, it is necessary to first splice these segment data so that the lesion segments and opening segments of the three main branches of the coronary artery can become separate complete data. Considering the different CCTA scanning methods, the segment data are physically normalized to ensure the uniformity of the data scale; Step 3: First, a CNN model is constructed, and then the labeled vascular segment data processed in step 2 is used to train and verify the CNN model to obtain a CNN model with good performance; Step 4: Input the vascular segment data required for ACS classification after step 2, which refers to the data that has not participated in the training and verification of the CNN model, that is, the data without corresponding labels, into the constructed CNN model, and output the prediction result of whether ACS occurs; The step 3 is specifically as follows: Step 3.1: Construct a CNN model consisting of two backbone networks, ResNet and SPP-net, with the same structure but different parameters, as well as a channel feature fusion network and a fully connected prediction layer, namely, the FC prediction layer; Step 3.2: Train the CNN model based on the training set data to obtain model parameters; Step 3.3: Use the validation set data to verify the trained CNN model; The channel feature fusion network is used to convert a feature map with 2 channels into 1 channel, increase the nonlinearity between channels, and reduce the redundancy of parameters; The channel feature fusion network consists of two consecutive sub-network modules connected by a Dropout layer, where the sub-network consists of a one-dimensional Convolutional layer, a one-dimensional BN and a ReLU function.

2. The ACS prediction method based on convolutional neural network according to claim 1, characterized in that: The ResNet network in the backbone network is an improvement on the existing ResNet-50 network; it consists of a convolutional module, two maximum pooling layers Max Pooling and two Residual modules, namely ResBlock, which are similar to the first three modules of the ResNet-50 network, namely: the convolutional module consists of a three-dimensional Convolutional layer, Batch Normalization, namely BN and ReLU function; the first ResBlock1 and the second ResBlock2 are both composed of three sub-convolutional modules and a Shortcut connection, and the sub-convolutional module consists of a three-dimensional Convolutional layer, BN and ReLU function; there is a Max Pooling layer between the first convolutional module and the first ResBlock1, and there is a Max Pooling layer between the first ResBlock1 and the second ResBlock2.

3. The ACS prediction method based on convolutional neural network according to claim 1, characterized in that: The FC prediction layer is a fully connected layer used to finally predict the ACS, which is composed of two consecutive sub-nonlinear FC layers plus a Linear layer, wherein the sub-nonlinear FC layer is composed of a Linear layer, a ReLU function and a Dropout layer.

4. The ACS prediction method based on convolutional neural network according to claim 1, characterized in that: The step 3.2 is specifically as follows: The training set data is composed of the input vascular segment data processed in step 2 and the label data corresponding to the data. The label means whether the actual result corresponding to the data is ACS. The label is a value of 0 or 1. 0 represents not ACS, and 1 represents ACS; the training set data is input into the constructed CNN model, and the CNN model automatically optimizes its own model parameters by comparing the difference between the output result and the label value, and performing cyclic calculations until the difference between the training output result and the label result is minimized, and the model reaches the optimal state; when the model reaches the optimal state, the optimal model parameters are obtained, and a CNN model with excellent performance is obtained; in the process of model training, the cross entropy function is used as the loss function to perform back propagation of the model.

5. The ACS prediction method based on convolutional neural network according to claim 1, characterized in that: The step 3.3 is specifically as follows: The validation set data is composed of the input vascular segment data processed in step 2 and the label data corresponding to the data. It is a set of data that does not participate in the CNN model training. The validation set is input into the CNN model with excellent performance in step 3.2 for verification to obtain the prediction results of the validation set data. If the prediction results are slightly different from the labels of the validation set data, the model parameters obtained in step 3.2 are valid model parameters and are saved for prediction of whether ACS occurs in the vascular segment data. If the prediction results are significantly different from the labels of the validation set data, it is necessary to return to step 3.2 to retrain the CNN model until the CNN model using the model parameters saved in step 3.2 can well predict the validation set data. The model parameters at this time are saved for prediction of whether ACS occurs in the vascular segment data.

Images