A method for identifying coronary artery stenosis based on deep autoencoder imaging

Abstract

The present invention provides a method for identifying coronary artery stenosis based on deep autoencoder composition, which belongs to the field of deep learning technology. In order to solve the problem that the existing graph neural network, when applied to the pathological identification of coronary artery stenosis, often has poor recognition effect due to redundant interference in the image, and requires a large number of samples to build a network model, the current pathological identification of coronary artery stenosis still needs to rely too much on expert experience. The method of the present invention performs superpixel segmentation on the overall CTA image of the coronary artery, and roughly identifies the obtained superpixel segmentation map to determine the suspected position of the coronary artery stenosis; then, the cross-sectional CTA image corresponding to the suspected position is precisely identified by the coronary artery stenosis recognition model based on the deep autoencoder network to obtain the stenosis level result of the coronary artery, and the coronary artery stenosis recognition model includes a superpixel segmentation module, a feature extraction module and a classification module. The method of the present invention can achieve a high accuracy in the recognition of coronary artery stenosis.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method for identifying coronary artery stenosis based on deep autoencoder composition. Background Art

[0002] In recent years, cardiovascular disease has become a major factor threatening human health. Coronary arteriosclerotic heart disease (CAD), referred to as coronary heart disease, is a disease caused by coronary atherosclerosis causing lumen stenosis and occlusion, resulting in myocardial hypoxia, ischemia, and necrosis. With the continuous improvement of coronary CT angiography (CTA) technology, CTA has been widely used in early screening and early diagnosis of suspected CAD patients. In recent years, medical artificial intelligence has become a research hotspot, providing help and convenience for imaging physicians to diagnose diseases. At present, there are few applications of artificial intelligence in coronary CTA in China, mainly focusing on removing background interference, improving image quality, and image segmentation of coronary artery vessels. However, in the diagnosis of coronary artery pathology, it still relies too much on the experience of doctors.

[0003] Recently, deep learning methods represented by graph neural networks have been widely used in many fields due to their outstanding performance in sample similarity relations. However, if graph neural networks are applied to the field of coronary artery pathology recognition, there are still some problems: on the one hand, graph neural networks are sensitive to noise, which will affect the recognition effect; on the other hand, due to the high cost of CTA image acquisition, how to use a small number of samples to accurately identify pathology is also an urgent problem to be solved. Summary of the invention

[0004] The technical problems to be solved by the present invention are:

[0005] When existing graph neural networks are applied to the identification of coronary artery stenosis pathology, the recognition effect is often poor due to redundant interference in the images, and a large number of samples are required to build the network model. Currently, the identification of coronary artery stenosis pathology still relies too much on expert experience.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The present invention provides a method for identifying coronary artery stenosis based on deep autoencoder composition, which essentially provides an image processing and recognition technology to identify coronary artery stenosis in CTA images, including the following steps:

[0008] S1. Collect coronary artery CTA images, including whole coronary artery CTA images and corresponding coronary artery cross-sectional CTA images, and randomly divide the coronary artery CTA images into a training set and a test set, and annotate the coronary artery stenosis grade of the cross-sectional CTA images of the training set;

[0009] S2, performing superpixel segmentation on the whole CTA image of the coronary artery in the test set, and performing rough recognition on the obtained superpixel segmentation map to determine the suspected location of coronary artery stenosis;

[0010] S3. constructing a coronary artery stenosis recognition model based on a deep autoencoder network, wherein the coronary artery stenosis recognition model includes a superpixel segmentation module, a feature extraction module and a classification module;

[0011] The cross-sectional CTA images corresponding to the suspicious positions obtained by S2 are combined with the cross-sectional CTA images of the training set and input into the coronary artery stenosis recognition model for learning, so as to obtain the stenosis level of the coronary artery of the test set samples, and the accuracy of the output results is analyzed by a specialist to obtain the accuracy of the model recognition;

[0012] S4. Determine the suspected stenosis location of the whole coronary artery CTA image to be identified according to the method described in S2, merge the coronary artery cross-sectional CTA image corresponding to the suspected location with the cross-sectional CTA image of the training set and input them into the learned coronary artery stenosis identification model for identification, and obtain the coronary artery stenosis grade result.

[0013] Furthermore, the superpixel segmentation in S2 adopts a watershed algorithm based on gradient descent.

[0014] Furthermore, the coarse recognition of the obtained superpixel segmentation map described in S2 includes the following process:

[0015] First, the relative width of the blood vessel is calculated along the extension direction of the blood vessel based on the segmentation map of the whole coronary artery CTA. Then, the mode of the relative width of the blood vessel is taken as the normal width of the blood vessel. The minimum value of the relative width of the blood vessel is compared with the normal width of the blood vessel, and the degree of stenosis at the minimum value of the blood vessel width is calculated to determine whether the location is a suspected location of coronary artery stenosis.

[0016] Furthermore, the superpixel segmentation module denoises the arterial cross-sectional image and makes the image boundary clear based on a progressive superpixel segmentation algorithm based on a graph structure;

[0017] Each superpixel in the segmentation map corresponds to a unique node in the structure map. The segmentation map of this level is generated by merging some adjacent superpixels in the previous level segmentation map. Correspondingly, the structure map of this level is obtained by merging the related nodes in the previous level structure map. A total of n-1 superpixel segmentations are performed, and the obtained hierarchical matrix and adjacency matrix are respectively represented by H (l) and A (l) Represents, where l=2, 3, ..., n.

[0018] Furthermore, the feature extraction module is based on a deep autoencoder network, and uses a convolution layer to perform a graph convolution operation on each structure graph. In order to improve the representation ability of the features for the original image, a cross entropy loss function is used to optimize the model;

[0019] Encoding process: In the first-level encoder, a shallow feature extraction mechanism is designed to replace the graph convolution operation. The graph convolution is used to learn the features of the superpixel images in the second to nth level autoencoders. The learned features are expressed as E (l) express:

[0020] E (l) =f G (BN(Pool(H (l-1) ),A (l) )),2≤l≤n

[0021] Among them, the level matrix H (l) Used to connect the features of each level in the SAE network; the adjacency matrix A (l) The graph structure represented is used to construct the autoencoder, BN(·) represents the batch normalization function, Pool(·) represents the pooling function, and f G (·) represents the graph convolution process, which is defined as:

[0022]

[0023] in, Where σ(·) represents the Leaky ReLU activation function, and M is The degree matrix, W, W θ , B, λ are learnable parameters, Θ represents the Hadamard product; I is the identity matrix;

[0024] Decoding process: Use graph convolution to reconstruct features and fuse features from coarse to fine. Feature F (l) The reconstruction process is:

[0025] F (l) =f G (BN(Unpool(F (l+1) ),A (l) )),2≤l <n

[0026] Among them, Unpool(·) represents the unpooling function;

[0027] Corresponding to the first-level encoder, the structure of the last-level decoder is designed to reconstruct the features F obtained by graph convolution (n-1) Input the last level decoder to obtain the feature F of the coronary artery cross section O =F (n) .

[0028] Further, the classification module constructs a graph structure based on a sparse representation algorithm and classifies the sample stenosis level based on a sparse graph regularization algorithm;

[0029] The process of constructing the graph structure of the sparse expression algorithm is as follows: the coronary artery cross-sectional features F extracted by the deep autoencoder network are transformed into o Divide into t labeled samples feature F t and the features F of u unlabeled samples u , the pathological labels of the labeled samples constitute the label matrix Y; the features of all labeled samples F t Construct a dictionary D, for each unlabeled sample feature x u ∈F u can all be expressed as non-negative linear combinations of dictionaries D:

[0030]

[0031] Among them, β is the regularization parameter and Z is the score matrix.

[0032] Furthermore, in the process of classifying samples by the sparse graph regularization algorithm, the score matrix Z can be obtained by solving the variable splitting and augmented Lagrangian algorithm;

[0033] The similarity matrix W is defined as:

[0034] W=Z tT Δ -1 Z t

[0035] Among them, Δ is the element The diagonal matrix, Z t represents the training data, and Z t ∈ Z;

[0036] By subtracting the first derivative of the formula,

[0037]

[0038] Solve to get the class probability:

[0039]

[0040] in, is the simplified graph Laplacian matrix, γ is the regularization parameter;

[0041] The pathological label of the unlabeled sample is obtained by the following formula:

[0042]

[0043] in, Represents the score matrix A * The u-th column of .

[0044] A system for identifying coronary artery stenosis based on deep autoencoder composition, the system having a program module corresponding to the steps of any one of the above technical solutions, and executing the steps of the above coronary artery stenosis identification method based on deep autoencoder composition during operation.

[0045] A computer-readable storage medium stores a computer program, wherein the computer program is configured to implement the steps of the method for identifying coronary artery stenosis based on deep autoencoder composition described in any one of the above technical solutions when called by a processor.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention discloses a method for identifying coronary artery stenosis based on deep autoencoder composition. First, the overall CTA image of the coronary artery is roughly identified, and then the cross-sectional CTA image is precisely identified through a coronary artery stenosis identification model based on a deep autoencoder network, so as to more comprehensively identify the coronary artery stenosis from these two different perspectives. The coronary artery stenosis identification model of the present invention combines deep learning theory and uses a deep autoencoder network to extract coronary artery features, which can extract deeper pathological features that are difficult to detect by human observation; the proposed shallow feature extraction mechanism as the first-level encoder can not only remove redundant interference in the original image, but also improve the ability of pathological features to characterize such pathologies, making it possible to identify coronary artery stenosis pathology using a small number of pathological samples; the pathological classifier based on the sparse graph regularization algorithm is used, which fully considers the similarity relationship between case samples and can achieve a higher accuracy rate under small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A structural diagram of a method for identifying coronary artery stenosis based on deep autoencoder mapping in an embodiment of the present invention Figure 1 ;

[0049] Figure 2 is an example diagram of coronary artery CTA image data used in an embodiment of the present invention;

[0050] Figure 3 is a binary segmentation map of the whole CTA image of the coronary artery in an embodiment of the present invention;

[0051] Figure 4 A schematic diagram of constructing a multi-level graph using a progressive superpixel segmentation technique in an embodiment of the present invention;

[0052] Figure 5 Schematic diagram of a graph convolution operation in an embodiment of the present invention;

[0053] Figure 6 Schematic diagram of the structure of the shallow feature extraction mechanism of the first-stage encoder in an embodiment of the present invention;

[0054] Figure 7 Schematic diagram of the convolutional network structure of the last stage decoder in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In the description of the present invention, it should be noted that the terms "first", "second", and "third" mentioned in the embodiments of the present invention are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second", and "third" may explicitly or implicitly include one or more of the features.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0057] like Figure 1 As shown, a method for identifying coronary artery stenosis based on deep autoencoder composition is used. The method adopts a "two-step method" to identify coronary artery stenosis, that is, firstly, the stenosis position in the coronary artery is roughly identified, and then the roughly identified stenosis position is precisely identified to determine the stenosis result, including the following steps:

[0058] S1. Collect 100 sets of coronary artery whole body CTA images and corresponding coronary artery cross-sectional CTA images from the Department of Cardiology of the First Affiliated Hospital of Harbin Medical University, and divide the coronary artery CTA images into training set X t (10%) and the test set X u(90%), where t is the number of samples in the training set, u is the number of samples in the test set, and t+u=100. Referring to the Coronary Disease Reporting and Data System (CAD-RADS), the coronary artery stenosis grades of the coronary artery cross-sectional CTA images of the training set were standardized into 6 grades: no stenosis (0% stenosis), mild stenosis (1%-24% stenosis), mild stenosis (25%-49% stenosis), moderate stenosis (50%-74% stenosis), severe stenosis (≥75% stenosis) and occlusion (100% stenosis).

[0059] S2, such as Figure 2 and Figure 3 As shown, superpixel segmentation is performed on the whole coronary artery CTA image of the test set, and the obtained superpixel segmentation map is roughly identified to determine the suspected location of coronary artery stenosis, including the following steps:

[0060] S21, performing superpixel segmentation on the whole coronary artery CTA image of the test set, including the following steps:

[0061] S211, using the watershed algorithm based on gradient descent, the gray value of each pixel is simulated as the altitude, and the whole picture is described as a topographic map;

[0062] S212, finding the local minimum gray value pixel in the topographic map, the pixel where the local minimum gray value is located and its surrounding pixels form a water-logged depression, each water-logged depression is a cluster, and the boundary of the depression forms a watershed, and superpixel segmentation is performed with the watershed as the boundary;

[0063] S213, perform multiple superpixel segmentation until the boundary between the blood vessel and the background is clear, and obtain the following Figure 3 The binary segmentation map of the whole coronary artery CTA is shown.

[0064] Since this step is to roughly identify the suspected stenosis location, the superpixel segmentation of the whole coronary artery CTA image requires the algorithm to have high recall rate, simplicity and low complexity. Therefore, the watershed algorithm based on gradient descent in the classic superpixel segmentation method is selected.

[0065] S22, calculating the relative width of the blood vessel along the extension direction of the blood vessel according to the segmentation image of the whole coronary artery CTA, comprising the following steps:

[0066] S221, constructing a coordinate system based on the whole coronary artery CTA segmentation image;

[0067] S222, randomly select a group of pixel points in the coronary artery (the white area in the binary image) along the horizontal axis of the coordinate system; for each vascular pixel, find a non-vascular pixel point a closest to it according to the vascular pixel coordinates, and find a non-vascular pixel point b closest to it on the other side;

[0068] S223, calculating the Euclidean distance between the non-vascular pixel points a and b, that is, the relative width of the blood vessel at the location of the vascular pixel point;

[0069] S23. Determining the suspected location of coronary artery stenosis based on the blood vessel width includes the following steps:

[0070] S231, taking the mode of the relative widths of the blood vessel as the normal width of the blood vessel;

[0071] S232. Compare the minimum value of the relative width of the blood vessel with the normal width of the blood vessel. If the degree of stenosis at the minimum value of the blood vessel width is greater than or equal to 1% (in order to calculate the evaluation index of the method of the present invention, the position where the degree of blood vessel stenosis is (1% to 49%) is determined as non-significant stenosis, and the position where the degree of blood vessel stenosis is (≥50%) is determined as significant stenosis), then the position is determined to be a suspected position of coronary artery stenosis.

[0072] S3. constructing a coronary artery stenosis recognition model based on a deep autoencoder network, wherein the coronary artery stenosis recognition model includes a superpixel segmentation module, a feature extraction module and a classification module;

[0073] Compare the cross-sectional CTA images corresponding to the suspicious positions obtained by S2 with the training set X t The cross-sectional CTA images are combined and input into the coronary artery stenosis recognition model for learning to obtain the stenosis level of the coronary artery of the test set samples. The accuracy of the output results is analyzed by specialists to obtain the accuracy of the model recognition.

[0074] The superpixel segmentation module removes noise from the arterial cross-sectional image and makes the image boundary clear based on the progressive superpixel segmentation algorithm based on the graph structure;

[0075] Each superpixel in the segmentation map corresponds to a unique node in the structure graph, such as Figure 4 As shown, S (l) and G (l) Respectively represent the lth superpixel and its structure diagram, S (l) Each superpixel in G corresponds to (l) A unique node in (l-1) Some neighboring superpixels in S (l) , correspondingly, by merging G (l-1) The relevant nodes in G (l) A total of 4 superpixel segmentations were performed, and the resulting hierarchical matrix and adjacency matrix were represented by H (l) and A (l) Indicates, where l = 2, 3, 4, 5;

[0076] The superpixel segmentation algorithm used in this step is slightly different from the superpixel segmentation method based on gradient descent in S2. In order to obtain the nodes of the encoder and decoder and facilitate the graph convolution operation, a progressive superpixel segmentation technology based on graph structure is adopted to perform superpixel segmentation on each cross-sectional image separately, and construct a graph structure with each pixel as a node.

[0077] like Figure 5 As shown, the feature extraction module is based on a deep autoencoder network. The deep autoencoder network has 9 layers, and a convolutional layer is used to extract each structure graph G. (l) Perform graph convolution operations. In order to improve the representation ability of features for the original image, the cross entropy loss function is used to optimize the model.

[0078] Coding process: Figure 6 As shown in Figure 2, in order to reduce the amount of computation of graph convolution, in the first-level encoder, a shallow feature extraction mechanism is designed to replace the graph convolution operation. Graph convolution is used to learn the features of the superpixel images in the second to fifth-level autoencoders. The learned features are represented by E (l) express:

[0079] E (l) =f G (BN(Pool(H (l-1) ),A (l) )),2≤l≤5

[0080] Among them, the level matrix H (l) Used to connect the features of each level in the SAE network, the adjacency matrix A (l) The graph structure represented is used to construct the autoencoder, BN(·) represents the batch normalization function, Pool(·) represents the pooling function, and f G (·) represents the graph convolution process, which is defined as:

[0081]

[0082] in, Where σ(·) represents the Leaky ReLU activation function, and M is The degree matrix, W, W θ , B, λ are learnable parameters, Θ represents the Hadamard product, and I is the identity matrix;

[0083] Decoding process: Use graph convolution to reconstruct features and fuse features from coarse to fine. Feature F (l) The reconstruction process is as follows:

[0084] F (l) =f G (BN(Unpool(F (l+1) ),A (l))),2≤l<5

[0085] Among them, Unpool(·) represents the unpooling function.

[0086] Corresponding to the first-level encoder, the structure of the last-level decoder is as follows Figure 7 As shown, the reconstructed feature F obtained by graph convolution (4) Input the last level decoder to obtain the feature F of the coronary artery cross section O =F (5) .

[0087] The classification module constructs a graph structure based on a sparse expression algorithm and classifies the sample stenosis level based on a sparse graph regularization algorithm;

[0088] The process of constructing the graph structure of the sparse expression algorithm is as follows: the coronary artery cross-sectional features F extracted by the deep autoencoder network are transformed into o Divide into t labeled samples feature F t and the features F of u unlabeled samples u , the pathological labels of the labeled samples constitute the label matrix Y. The features F of all labeled samples t Construct a dictionary D, for each unlabeled sample feature x u ∈F u can all be expressed as non-negative linear combinations of dictionaries D:

[0089]

[0090] Among them, β is the regularization parameter and Z is the score matrix.

[0091] In the process of classifying samples by the sparse graph regularization algorithm, the score matrix Z can be obtained by solving the variable splitting and augmented Lagrangian algorithm;

[0092] The similarity matrix W is defined as:

[0093] W=Z tT Δ -1 Z t

[0094] Among them, Δ is the element The diagonal matrix, Z t represents the training data, and Z t ∈ Z;

[0095] By subtracting the first derivative of the formula,

[0096]

[0097] Solve to get the class probability:

[0098]

[0099] in, is the simplified graph Laplacian matrix, and γ is the regularization parameter.

[0100] The pathological label of the unlabeled sample is obtained by the following formula:

[0101]

[0102] in, Represents the score matrix A * The u-th column of .

[0103] S4, determine the suspected stenosis position of the coronary artery in the whole CTA image to be identified according to the method described in S2, and compare the coronary artery cross-sectional CTA image corresponding to the suspected position with the training set X t The cross-sectional CTA images are merged and input into the learned coronary artery stenosis recognition model for recognition, and the coronary artery stenosis grade results are obtained.

[0104] The accuracy of the above recognition results is judged by professional doctors, the accuracy of the method of the present invention is calculated, and the evaluation index of the method of the present invention is obtained.

[0105] In order to eliminate the interference of experimental randomness, the present invention conducts 10 rounds of comparative experiments to obtain the average value, and uses the diagnostic accuracy, missed diagnosis rate and misdiagnosis rate to evaluate the diagnostic effect. Table 1 shows the accuracy of the method of the present invention for the rough identification of the suspected stenosis position in the overall CTA image of the coronary artery. It can be seen from Table 1 that the method of the present invention can accurately identify the existence of stenosis in the coronary artery. The recognition accuracy for the existence of stenosis (≥1%) can reach 92.20%, and the diagnostic accuracy for the significant stenosis (≥50%) that seriously affects the health of the patient can reach 95.24%. Table 2 shows the accuracy of the identification of the stenosis of the coronary artery cross-sectional CTA image corresponding to the suspicious position. It can be seen from Table 2 that for the precise identification of the cross-sectional CTA image, the accuracy of the method of the present invention can reach more than 92%, and the missed diagnosis rate and misdiagnosis rate can be controlled within 5%, achieving a high accuracy rate.

[0106] Table 1

[0107]

[0108] Table 2

[0109]

[0110] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A method for identifying coronary artery stenosis based on deep autoencoder composition, characterized in that The following steps are involved: S1. Collect coronary artery CTA images, including whole coronary artery CTA images and corresponding coronary artery cross-sectional CTA images, and randomly divide the coronary artery CTA images into a training set and a test set, and annotate the coronary artery stenosis grade of the cross-sectional CTA images of the training set; S2, performing superpixel segmentation on the whole CTA image of the coronary artery in the test set, and performing rough recognition on the obtained superpixel segmentation map to determine the suspected location of coronary artery stenosis; S3. constructing a coronary artery stenosis recognition model based on a deep autoencoder network, wherein the coronary artery stenosis recognition model includes a superpixel segmentation module, a feature extraction module and a classification module; The cross-sectional CTA images corresponding to the suspicious positions obtained by S2 are combined with the cross-sectional CTA images of the training set and input into the coronary artery stenosis recognition model for learning, so as to obtain the stenosis level of the coronary artery of the test set samples, and the accuracy of the output results is analyzed by a specialist to obtain the accuracy of the model recognition; S4, using the whole CTA image of the coronary artery to be identified to determine the suspected stenosis position according to the method described in S2, combining the coronary artery cross-sectional CTA image corresponding to the suspected position with the cross-sectional CTA image of the training set and inputting them into the learned coronary artery stenosis identification model for identification, and obtaining the coronary artery stenosis grade result; The superpixel segmentation module removes noise from the arterial cross-sectional image and makes the image boundary clear based on the progressive superpixel segmentation algorithm based on the graph structure; Each superpixel in the segmentation map corresponds to a unique node in the structure map. The segmentation map of this level is generated by merging some adjacent superpixels in the previous level segmentation map. Correspondingly, the structure map of this level is obtained by merging the related nodes in the previous level structure map. A total of n-1 superpixel segmentations are performed, and the obtained hierarchical matrix and adjacency matrix are respectively represented by H (l) and A (l) It means, where l = 2, 3, ..., n; The feature extraction module is based on a deep autoencoder network, and uses a convolution layer to perform graph convolution operations on each structure graph. In order to improve the ability of features to represent the original image, a cross entropy loss function is used to optimize the model; Encoding process: In the first-level encoder, a shallow feature extraction mechanism is designed to replace the graph convolution operation. The graph convolution is used to learn the features of the superpixel images in the second to nth level autoencoders. The learned features are expressed as E (l) express: HAVE BEEN (l) =f G (BN(Pool(H (l-1) ),A (l) ),2≤l≤n Among them, the level matrix H (l) Used to connect the features of each level in the SAE network; the adjacency matrix A (l) The graph structure represented is used to construct the autoencoder, BN(·) represents the batch normalization function, Pool(·) represents the pooling function, and f G (·) represents the graph convolution process, which is defined as: in, Where σ(·) represents the Leaky ReLU activation function, and M is The degree matrix, W, W θ , B, λ are learnable parameters, Θ represents the Hadamard product; I is the identity matrix; Decoding process: Use graph convolution to reconstruct features and fuse features from coarse to fine. Feature F (l) The reconstruction process is: F (l) =f G (BN(Unpool(F (l+1) ),A (l) )),2≤l<n Among them, Unpool(·) represents the unpooling function; Corresponding to the first-level encoder, the structure of the last-level decoder is designed to reconstruct the features F obtained by graph convolution (n-1) Input the last level decoder to obtain the feature F of the coronary artery cross section O =F (n) ; The classification module constructs a graph structure based on a sparse expression algorithm and classifies the sample stenosis level based on a sparse graph regularization algorithm; The process of constructing the graph structure of the sparse expression algorithm is as follows: the coronary artery cross-sectional features F extracted by the deep autoencoder network are transformed into o Divide into t labeled samples feature F t and the features F of u unlabeled samples u , the pathological labels of the labeled samples constitute the label matrix Y; the features of all labeled samples F t Construct a dictionary D, for each unlabeled sample feature x u ∈F u are all expressed as non-negative linear combinations of dictionaries D: Among them, β is the regularization parameter and Z is the score matrix.

2. A method for identifying coronary artery stenosis based on deep autoencoder composition according to claim 1, characterized in that The superpixel segmentation described in S2 adopts the watershed algorithm based on gradient descent.

3. The method for identifying coronary artery stenosis based on deep autoencoder mapping according to claim 1, characterized in that the step of performing coarse recognition on the obtained superpixel segmentation map in S2 comprises the following process: First, the relative width of the blood vessel is calculated along the extension direction of the blood vessel based on the segmentation map of the whole coronary artery CTA. Then, the mode of the relative width of the blood vessel is taken as the normal width of the blood vessel. The minimum value of the relative width of the blood vessel is compared with the normal width of the blood vessel, and the degree of stenosis at the minimum value of the blood vessel width is calculated to determine whether the location is a suspected location of coronary artery stenosis.

4. The method for identifying coronary artery stenosis based on deep autoencoder composition according to claim 1, characterized in that In the process of classifying samples by the sparse graph regularization algorithm, the score matrix Z is obtained by solving the variable splitting and augmented Lagrangian algorithm; The similarity matrix W is defined as: In=From tT Δ -1 WITH t Among them, Δ is the element The diagonal matrix, Z t represents the training data, and Z t ∈ Z; By subtracting the first derivative of the formula, Solve to get the class probability: in, is the simplified graph Laplacian matrix, γ is the regularization parameter; The pathological label of the unlabeled sample is obtained by the following formula: in, Represents the score matrix A * The u-th column of .

5. A coronary artery stenosis recognition system based on deep autoencoder composition, characterized in that The system has a program module corresponding to the steps of any one of claims 1 to 4 above, and executes the steps in the above-mentioned coronary artery stenosis identification method based on deep autoencoder composition when running.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the method for identifying coronary artery stenosis based on deep autoencoder mapping according to any one of claims 1 to 4 when called by a processor.

Images