Calcification section coronary artery risk prediction method and device based on chest CT (Computed Tomography) plain scanning

Abstract

The invention provides a calcification section coronary artery risk prediction method and device based on chest CT (Computed Tomography) plain scanning. The method comprises the following steps: firstly, acquiring and screening chest CT (Computed Tomography) plain-scan image data; secondly, performing calcification section coronary artery labeling on the image data; thirdly, performing CCTA image interpretation and grouping on the marked areas; and finally, collecting a CT image and a classification label model corresponding to the CT image, using a ResTransNet network to extract features from the calcification section coronary artery label, training the ResTransNet model, and predicting a risk state. According to the method, the deep learning method is applied to the coronary artery of the calcification section of chest CT flat scanning, the plaque type and the blood vessel stenosis degree are accurately predicted, and effective information is provided for subclinical coronary atherosclerosis risk screening and prevention; constructing plaque types and coronary artery stenosis degrees gt; and the prediction device has great clinical significance when being used for screening the risk of subclinical coronary atherosclerosis.

Description

Technical Field

[0001] The present invention belongs to the field of medical auxiliary diagnosis, and in particular relates to a method and device for predicting the risk of calcified coronary arteries based on chest CT plain scan. Background Art

[0002] Subclinical coronary atherosclerosis is very common in the general population, and atherosclerotic cardiovascular disease (ASCVD) events often occur in these patients who do not receive preventive treatment because they have no obvious symptoms and are easily misunderstood as low risk by traditional risk factors. In a large-scale study of the incidence of subclinical coronary atherosclerosis in the general population, 42.1% (10603 / 25182) of the participants had asymptomatic coronary atherosclerosis, of which about 5.2% (1317 / 25182) had a stenosis of more than 50%.

[0003] In the healthcare field, the focus is often placed on patients who show obvious symptoms, thereby ignoring asymptomatic patients who may benefit from preventive treatment. Although the process of subclinical coronary atherosclerosis developing into ASCVD events is inevitable, it can be regulated through intervention measures. Intensive lipid-lowering therapy can effectively control plaque progression and thus reduce the risk of ASCVD events. Therefore, screening high-risk populations for subclinical coronary atherosclerosis is crucial for preventing ASCVD events.

[0004] Nowadays, deep learning technology has made significant progress in the field of computer-aided detection and diagnosis, and is widely used in many tasks such as disease classification, image segmentation and lesion detection. In the field of cardiovascular imaging, deep learning technology can calculate the partial flow reserve fraction and coronary artery calcification (CAC) score through CT images; the deep learning algorithm based on late gadolinium-enhanced magnetic resonance imaging (MRI) can effectively evaluate ventricular tachycardia; the deep learning classification algorithm based on CCTA can predict the reperfusion effect after intravascular intervention. These achievements can be attributed to the use of convolutional neural networks (CNNs), which extract key features with limited parameters through convolution operations. Compared with traditional fully connected networks, CNNs are more relaxed in data requirements, thus promoting innovation in the field of image and signal processing.

[0005] However, deep learning faces a major challenge when analyzing coronary arteries on chest CT plain scan images - artifacts caused by heart pulsation. In conventional chest CT scans, heart rate control and electrocardiogram synchronization technology are usually not used, so artifacts caused by heart pulsation are more common in chest CT plain scan images. However, this phenomenon makes a large amount of chest CT plain scan data unable to be applied to deep learning analysis of coronary arteries. However, deep learning analysis focusing on CAC segment vessels can effectively reduce the impact of this problem.

[0006] Vulnerable plaques are not fixed, and up to three-quarters of vulnerable plaques may regain stability over time with appropriate medical therapy. Calcified and non-calcified components play a key role in plaque stability. Calcium load is higher in stable plaques than in vulnerable plaques and is inversely proportional to the area of ​​necrotic core. Plaque volume, especially non-calcified plaque volume, is a more powerful predictor of cardiovascular events than luminal stenosis and clinical risk factors. Subclinical atherosclerotic plaques found on CCTA, both obstructive and non-obstructive, are potential targets for treatment. Although obstructive lesions are generally considered to be more likely to cause clinical events, in the SCOT-HEART and PROMISE trials, subclinical non-obstructive plaques accounted for as much as 42-66% of future myocardial infarction events. When CCTA finds coronary artery stenosis >50%, it is necessary to consider the use of CT blood flow reserve fraction, CT perfusion, or stress testing to document or exclude the presence of ischemia, and some patients may even require medical treatment or coronary angiography. Summary of the invention

[0007] In view of the shortcomings of the prior art, on the one hand, the present invention provides a method for predicting the risk of coronary artery calcification segments based on chest CT plain scan, comprising the following steps:

[0008] Step 1: Collect and screen chest CT plain scan image data;

[0009] Step 2: marking the calcified coronary artery on the image data;

[0010] Step 3: Interpret and group the CCTA images of the marked areas;

[0011] Step 4: Collect CT images and their corresponding classification label models, use the ResTransNet network to extract features from the calcified coronary artery annotations, train the ResTransNet model, and predict the risk status.

[0012] Furthermore, in step 1, the image data inclusion criteria are:

[0013] (1) All CAC segments that are not affected by pulsation artifacts on chest CT scans;

[0014] (2) The interval between chest CT scan and CCTA examination should not exceed 1 month, and the CCTA images corresponding to the CAC segment vessels can be accurately interpreted.

[0015] Furthermore, in step 2, the calcified coronary artery region not affected by pulsation is selected as the annotation object, and the annotation range extends 3 mm along the long axis of the coronary artery at the proximal and distal edges of the calcification, and the coronary artery and calcification region within the range are annotated.

[0016] Furthermore, the grouping in step 3 is specifically as follows:

[0017] (1) Based on the type of plaque, the patients were divided into two groups: the pure calcification group, which contained only calcified plaques; and the non-pure calcification group, which contained mixed plaques or both calcified and non-calcified plaques;

[0018] (2) According to the degree of stenosis, the patients were divided into two groups: non-significant stenosis group and significant stenosis group.

[0019] Further, the step 4 specifically includes:

[0020] Step 4-1: In the data processing stage, collect CT images and their corresponding classification labels;

[0021] Step 4-2: Use the ResTransNet network to automatically extract image features;

[0022] Step 4-3: Train the ResTransNet model using the cross entropy loss function;

[0023] Step 4-4: Predict the risk status using the trained ResTransNet model.

[0024] Furthermore, in step 4-1, the data is divided into a training set, a validation set and an external test set, wherein the ratio of the training set, the validation set and the external test set is 3:1:1.

[0025] Furthermore, the ResTransNet model includes a CNN module and a Transformer module. The CNN module uses convolution to identify high-level internal features of an image, divides the feature map into n patch sequences, and then feeds them into the Transformer module for global feature learning.

[0026] In particular, the Transformer module contains 12 repeated Transformer units, each of which is composed of a multi-head self-attention module and a multi-layer perceptron module.

[0027] Furthermore, the multi-head self-attention module uses a self-attention mechanism to capture the interactions between all n patches by encoding the global context information of each patch; projects the input sequence onto a weight matrix, fuses the position and classification information into the self-attention feature map, and then inputs the self-attention feature map into the multi-layer perceptron module after it is processed by a linear layer.

[0028] On the other hand, the present invention provides a device for predicting the risk of coronary artery calcification segment based on chest CT plain scan, comprising:

[0029] An acquisition unit, used for acquiring and screening chest CT plain scan image data;

[0030] A labeling unit, used for labeling the calcified coronary artery of the image data;

[0031] An interpretation unit is used to interpret and group the CCTA images of the marked areas;

[0032] The prediction unit is used to collect CT images and their corresponding classification label models, extract features from the calcified coronary artery annotations using the ResTransNet network, and predict the risk status after training the ResTransNet model.

[0033] Beneficial effects of the invention: The invention provides a method and device for predicting the risk of calcified coronary arteries based on chest CT plain scan, applies a deep learning method to the calcified coronary arteries of chest CT plain scan, and is used to accurately predict the type of plaque and the degree of vascular stenosis, providing effective information for subclinical coronary atherosclerosis risk screening and prevention; constructs a prediction device for plaque type and coronary artery stenosis degree >50%, which is of great clinical significance for subclinical coronary atherosclerosis risk screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a conceptual diagram of an embodiment of the present application showing a deep learning model based on chest CT scan for screening high-risk groups for subclinical coronary atherosclerosis.

[0035] Figure 2 This is a schematic diagram of the CAC segment blood vessel marking in an embodiment of the present application.

[0036] Figure 3 This is the ResTransNet model of the embodiment of this application. DETAILED DESCRIPTION

[0037] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than 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.

[0038] Figure 1 FIG. 4 is a flow chart of a method for predicting coronary artery risk in calcified segments based on chest CT plain scan according to an embodiment of the present invention. Figure 1 As shown, the method for predicting the risk of coronary artery in calcified segments based on chest CT plain scan includes the following steps:

[0039] S1. Collect and screen chest CT plain scan image data.

[0040] Patients who underwent chest CT scan and CCTA examinations, and the time interval between CT scan and CCTA was no more than 1 month.

[0041] Exclusion criteria were set as follows: (1) all CAC segment vessels in chest CT plain scan were affected by pulsation artifacts; (2) poor CCTA image quality resulted in the inability to accurately interpret the images of the CAC segment vessels.

[0042] The chest CT plain scan examinations of the training set and validation set of the embodiment of the present application were completed by 4 multi-row CT scanners, specifically 2 Brilliance iCT, Philips; 1 Definition AS, Siemens; 1 Aquilion ONE, Toshiba. The data of the external test set were obtained from another 3 multi-row CT scanners, specifically 1 Optima CT680, GE; 2 Definition AS, Siemens. The main scanning parameters of the Brilliance iCT scanner are as follows: tube voltage 120kVp; tube current, automatic tube current modulation; matrix size, 512×512; reconstruction filter, filtered back projection; layer thickness, 1mm. Except that the reconstruction filter of the Definition AS CT scanner is "B70f", the imaging parameters of other machines such as tube voltage, tube current, matrix and layer thickness are consistent with Brilliance iCT.

[0043] S2. Marking the calcified coronary arteries on the chest CT plain scan image data.

[0044] Since the non-calcified components of mixed plaques may be mainly located around calcifications, in order to ensure that deep learning analysis can extract necessary information in the labeled area. Figure 2As shown, the embodiment of the present application extends 3 mm along the long axis of the blood vessel on the proximal and distal edges of the calcification based on the CAC range. The annotated object includes the calcification and the corresponding blood vessel, but does not include the radiation hardening artifacts around the calcification.

[0045] S3: Interpret the CCTA images of the marked areas and divide them into groups.

[0046] CCTA images were interpreted for the CAC segment vessel marking area on chest CT scan to determine the following two imaging indicators:

[0047] (1) Based on the type of plaque, the patients were divided into two groups: the pure calcification group, which only contained calcified plaques; and the non-pure calcification group, which contained mixed plaques or both calcified and non-calcified plaques.

[0048] (2) The degree of stenosis of the CAC segment was evaluated, and the samples were divided into a non-significant stenosis group and a significant stenosis group according to whether the stenosis degree exceeded 50%.

[0049] S4: Collect CT images and their corresponding classification label models, collect CT images and their corresponding classification label models, use the ResTransNet network to extract features from the calcified coronary artery annotations, train the ResTransNet model, and predict the risk status.

[0050] The embodiment of the present application designs a deep learning prediction model based on chest CT plain scan, and the deep learning neural network used can optionally include DenseNet, ResNet and Transformer, etc.

[0051] Taking the ResTransNet network architecture as an example, this architecture combines the advantages of CNNs and self-attention mechanisms, can efficiently extract key features in images, and then use nonlinear classifiers for accurate predictions.

[0052] Based on ResTransNet, the following two prediction models are proposed to be constructed: (1) a prediction model for the type of calcified plaques; and (2) a prediction model for the degree of coronary artery stenosis >50%.

[0053] S4.1: In the data processing stage, CT images and their corresponding classification labels are collected.

[0054] S4.2: Automatically extract image features using the ResTransNet model.

[0055] S4.3: Train the ResTransNet model using the cross entropy loss between the predicted and true values.

[0056] S4.4: Predict risk status using a nonlinear classifier.

[0057] S4.1 divides the data into a training set, a validation set, and an external test set, where the ratio of the training set, validation set, and external test set is 3:1:1. The reason for selecting this ratio is to ensure the adequacy of the training set while retaining enough data for validation and external testing to adjust model parameters.

[0058] In order to more effectively extract global features from CT plain scan data sets, the embodiment of the present application incorporates a Transformer module based on the traditional convolutional network to enhance the performance of the model. Figure 3 The overall architecture is described in detail, which includes CNN modules and Transformer modules. In the process, CNN first automatically identifies the high-level internal features of the image, and then these features are fed into the Transformer module for global feature learning.

[0059] In the ResTransNet model, the ResNet architecture is used to refer to the fact that as the network depth increases, the resolution of the feature map decreases while the number of channels increases. Specifically, in this network, the output channels of each layer are 16, 32, and 64, respectively.

[0060] The feature map X with c channels and size h×w is used as the input of Transformer. PatchEmbedding is the process of converting features into patches. The feature map X∈R c×h×w is divided into n patch sequences X = {x1, x2, ..., x n}(X∈R c×h×w ).

[0061] The Transformer module contains 12 repeated Transformer units. Figure 3 As shown in the figure, each Transformer unit consists of a Multi-Head Self-Attention (MSA) module and a Multi-Layer Perceptron (MLP) module.

[0062] The multi-head self-attention module uses a self-attention mechanism to capture the interactions between all n patches by encoding the global context information of each patch. Q , W K and W V are three learnable weights used to automatically learn the importance of each patch. The input sequence X is first projected onto these weight matrices to obtain Q = XW Q , K = XW K and V=XW V .

[0063] The output Z of the MSA layer is:

[0064] Z = MSA(LN(Attention(X+X class +X position )))+X

[0065] Among them, the attention function is calculated as In the attention function calculation formula, d k is the dimension of the K vector, providing the appropriate normalization required to make the gradient more stable; LN represents linear normalization; X class and X position are random parameters representing the classification and location information, which are fused into the self-attention feature map to capture the structure of the object image.

[0066] After the self-attention feature map Z is processed by the linear layer, it is input into the MLP and we get:

[0067] Y=MLP(LN(Z))+Z

[0068] Among them, the MLP module consists of two linear layers and a GELU activation function.

[0069] Using the ViT method, the classification information Y(0) (i.e. X class ) rather than the extracted image features as the input of the classifier. The nonlinear classifier is composed of a linear layer and a ReLu activation function.

[0070] The network is trained using the cross-entropy (CE) loss between the predicted and true values, which can be expressed as:

[0071]

[0072] Among them, y is the true label, is the predicted probability. This loss function reflects the difference between the true label and the predicted probability.

[0073] In the final stage of the ResTransNet model, class tokens are adopted instead of traditional features to achieve accurate prediction of classification labels, thereby effectively predicting risk status.

[0074] S5: Model evaluation: The embodiment of the present application uses the area under the curve (AUC) and receiver operator characteristic (ROC), accuracy, sensitivity and specificity to measure the performance of the ResTransNet model.

[0075] This embodiment further provides a device for predicting the risk of coronary arteries in calcified segments based on chest CT plain scan, and applies the method for predicting the risk of coronary arteries in calcified segments based on chest CT plain scan described in the above embodiment, including:

[0076] The acquisition unit is used to acquire and screen chest CT plain scan image data.

[0077] The labeling unit is used to label the calcified coronary artery of the image data.

[0078] The interpretation unit is used to interpret and group the CCTA images of the marked areas.

[0079] The prediction unit is used to collect CT images and their corresponding classification label models, extract features from the calcified coronary artery annotations using the ResTransNet network, and predict the risk status after training the ResTransNet model.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting the risk of coronary artery calcification based on chest CT scan, characterized in that: The following steps are involved: Step 1: Collect and screen chest CT plain scan image data; Step 2: marking the calcified coronary artery on the image data; Step 3: Interpret and group the CCTA images of the marked areas; Step 4: Collect CT images and their corresponding classification label models, use the ResTransNet network to extract features from the calcified coronary artery annotations, train the ResTransNet model, and predict the risk status.

2. The method for predicting coronary artery risk in calcified segments based on chest CT scan according to claim 1, characterized in that: In step 1, the image data inclusion criteria are: (1) All CAC segments that are not affected by pulsation artifacts on chest CT scans; (2) The interval between chest CT scan and CCTA examination should not exceed 1 month, and the CCTA images corresponding to the CAC segment vessels can be accurately interpreted.

3. The method for predicting coronary artery risk in calcified segments based on chest CT scan according to claim 1, characterized in that: In step 2: The calcified coronary artery area not affected by pulsation was selected as the annotation object. The annotation range was 3 mm along the long axis of the coronary artery at the proximal and distal edges of the calcification, and the coronary artery and calcification area within the range were annotated.

4. The method for predicting coronary artery risk in calcified segments based on chest CT scan according to claim 1, characterized in that: The grouping in step 3 is specifically as follows: (1) Based on the type of plaque, the patients were divided into two groups: pure calcification group, which only contained calcified plaques; and non-pure calcification group, which contained mixed plaques or both calcified and non-calcified plaques. (2) According to the degree of stenosis, the patients were divided into two groups: non-significant stenosis group and significant stenosis group.

5. The method for predicting coronary artery risk of calcified segments based on chest CT scan according to claim 1, characterized in that: The step 4 specifically includes: Step 4-1: In the data processing stage, collect CT images and their corresponding classification labels; Step 4-2: Use the ResTransNet network to automatically extract image features; Step 4-3: Train the ResTransNet model using the cross entropy loss function; Step 4-4: Predict the risk status using the trained ResTransNet model.

6. The method for predicting coronary artery risk in calcified segments based on chest CT scan according to claim 5, characterized in that: In step 4-1, the data is divided into a training set, a validation set and an external test set, wherein the ratio of the training set, the validation set and the external test set is 3:1:

1.

7. The method for predicting coronary artery risk of calcified segments based on chest CT plain scan according to claim 5, characterized in that: The ResTransNet model includes a CNN module and a Transformer module; The CNN module uses convolution to identify high-level internal features of the image, divides the feature map into n patch sequences and feeds them into the Transformer module for global feature learning.

8. The method for predicting coronary artery risk in calcified segments based on chest CT scan according to claim 7, characterized in that: The Transformer module includes 12 repeated Transformer units, each of which is composed of a multi-head self-attention module and a multi-layer perceptron module.

9. The method for predicting coronary artery risk in calcified segments based on chest CT scan according to claim 8, characterized in that: The multi-head self-attention module uses a self-attention mechanism to capture the interactions between all n patches by encoding the global context information of each patch; The input sequence is projected onto the weight matrix, and the position and classification information are fused into the self-attention feature map. After the self-attention feature map is processed by the linear layer, it is input into the multi-layer perceptron module.

10. A device for predicting the risk of coronary artery calcification based on chest CT scan, characterized in that: include: An acquisition unit, used for acquiring and screening chest CT plain scan image data; A labeling unit, used for labeling the calcified coronary artery of the image data; An interpretation unit is used to interpret and group the CCTA images of the marked areas; The prediction unit is used to collect CT images and their corresponding classification label models, extract features from the calcified coronary artery annotations using the ResTransNet network, and predict the risk status after training the ResTransNet model.

Images