A method and apparatus for diagnosing aortic dissection

Abstract

This invention discloses a method and apparatus for diagnosing aortic dissection. The method involves registering CTA images and plain CT images to obtain a patient's aortic CT image dataset and a corresponding aortic CTA image dataset. The aortic centerline is extracted from both the CTA and CT image datasets. A model for extracting the aortic centerline based on the CT image centerline extraction results is then trained. This model is used to extract the aortic centerline from the patient's CT images. Simultaneously, based on the extracted results and corresponding patient labels, the method provides auxiliary diagnosis of aortic dissection. This invention addresses the problem of diagnosing aortic dissection based on extracting the vascular centerline from plain CT images.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging diagnostic technology, and in particular relates to a method and device for diagnosing aortic dissection based on the vascular centerline of plain CT images. Background Technology

[0002] The aorta is the main pathway for the heart to pump blood throughout the body, playing a vital role in human blood circulation. The aorta consists of three layers: the intima, media, and adventitia. When aortic dissection occurs, the intima tears, allowing blood to enter the media, forming a separate channel—a true and false lumen—from the original aortic conduit. After dissection, the enormous blood flow and pressure can quickly cause the aortic intima to rupture, leading to death. Therefore, aortic dissection is an acute, high-risk cardiovascular disease. Due to limited medical resources in primary hospitals and the diverse symptoms presenting at onset, misdiagnosis and subsequent deaths are common in my country. Therefore, accurate diagnosis and treatment of aortic dissection is a crucial issue in the medical field.

[0003] Currently, the most important auxiliary diagnostic method for aortic dissection is CT angiography (CTA). However, this imaging method is expensive, time-consuming, generates large amounts of data, and has low diagnostic efficiency. Furthermore, some patients cannot undergo CTA due to physical conditions (such as liver or kidney failure, or allergies to contrast agents), and many primary and secondary hospitals lack the medical equipment and facilities to perform CTA, relying solely on analysis of plain CT image data. Therefore, the accuracy of doctors diagnosing aortic dissection based on plain CT images remains limited.

[0004] The aortic centerline is a crucial feature of the aorta and a vital clinical reference for physicians in determining whether, how severe, and where the aorta is diseased. However, in practice, subjective factors hinder physicians from efficiently and accurately extracting the aortic centerline for quantitative analysis. Therefore, accurately extracting the aortic centerline and using it to aid diagnosis is a pressing issue. Currently, centerline extraction methods are mainly categorized into topology-based, tracking-based, and minimum path-based methods. Topology-based extraction algorithms refine the topology of the target. While meeting topological conditions, redundant points around the target are gradually removed to obtain its skeleton, the centerline. Topology-based algorithms offer high accuracy, but are time-consuming when the target is large, making them unsuitable for extracting the centerline of large targets like the aorta. Tracking-based extraction algorithms use local image information to determine the centerline point, then adjust its position based on its relationship with nearby points, iteratively tracking all points on the centerline. This algorithm requires a pre-defined center point and cannot achieve fully automated centerline extraction. Furthermore, this algorithm only uses local operators. Although it is fast, it cannot consider the overall topological structure of the image, making it unsuitable for areas with large vessel curvature and numerous bifurcations. The minimum cost path extraction algorithm primarily calculates a path that minimizes cost between a given start and end point. This algorithm has low computational complexity and fast processing speed, but sometimes it yields results too close to the vessel edge, often requiring post-processing correction. In addition, this algorithm also depends on the image quality of the target image. Therefore, while the minimum cost path algorithm has achieved good results in aortic vessel centerline extraction based on CTA images, its application in aortic vessel centerline extraction based on plain CT images remains limited. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention proposes a method and apparatus for diagnosing aortic dissection based on the vascular centerline of plain CT images, aiming to solve the problem that existing technologies cannot diagnose aortic dissection by extracting the vascular centerline from plain CT images.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A diagnostic method for aortic dissection includes the following steps:

[0008] Step S1: Register the CTA images and CT plain scan images to obtain the patient's aortic CT plain scan image dataset and the corresponding aortic CTA image dataset;

[0009] Step S2: Extract the aortic vessel centerline from the aortic CTA image dataset;

[0010] Step S3: Based on the CTA image centerline extraction results, extract the aortic vessel centerline from the CT plain scan image dataset;

[0011] Step S4: Train the model based on the centerline extraction results of CT plain scan images to obtain the aortic vessel centerline extraction model based on CT plain scan images;

[0012] Step S5: Extract the aortic vessel centerline from the patient's CT plain scan image based on the aortic vessel centerline extraction model; simultaneously, perform auxiliary diagnosis of aortic dissection based on the extraction results and corresponding patient labels.

[0013] Preferably, image registration is performed using a 3D U-Net network in step S1.

[0014] Preferably, in step S2, the aortic vessel centerline is extracted from the aortic CTA image dataset using a minimum cost path algorithm.

[0015] Preferably, in step S3, a cost function based on Hausdorff distance is used to extract the aortic vessel centerline from the CT plain scan image dataset using the minimum cost path algorithm.

[0016] Preferably, in step S4, the centerline extraction result of CT plain scan image is used as the gold standard to perform supervised training based on convolutional neural network to obtain the aortic vessel centerline extraction model based on CT plain scan image.

[0017] Preferably, step S5 uses a transformer binary classification model based on a self-attention mechanism to determine whether the patient's aortic vessel centerline is abnormal in order to diagnose whether the patient has aortic dissection.

[0018] The present invention also provides a diagnostic device for aortic dissection, comprising:

[0019] The registration module is used to register CTA images and CT plain scan images to obtain the patient's aortic CT plain scan image dataset and the corresponding aortic CTA image dataset;

[0020] The first extraction module is used to extract the aortic vessel centerline from the aortic CTA image dataset.

[0021] The second extraction module is used to extract the aortic vessel centerline from the CT plain scan image dataset based on the CTA image centerline extraction results.

[0022] The training module is used to train the aortic vessel centerline extraction model based on the centerline extraction results of CT plain scan images.

[0023] The diagnostic module is used to extract the aortic vessel centerline from the patient's CT plain scan image based on the aortic vessel centerline extraction model; at the same time, it performs auxiliary diagnosis of aortic dissection based on the extraction results and the corresponding patient labels.

[0024] Preferably, the second extraction module uses a cost function based on Hausdorff distance to extract the aortic vessel centerline from the CT plain scan image dataset using a minimum cost path algorithm.

[0025] As a preferred option, the training module uses the centerline extraction results of CT plain scan images as the gold standard to perform supervised training on the 3D U-Net neural network, thereby obtaining an aortic vessel centerline extraction model based on CT plain scan images.

[0026] As a preferred option, the diagnostic module uses the relative positional relationship between the aortic vessel boundary and the centerline to determine whether the centerline is abnormal, thereby diagnosing whether the patient has aortic dissection.

[0027] Since the minimum cost path algorithm performs poorly in centerline extraction from plain CT images, this invention first extracts the centerline from CTA images that are related to the plain CT images. Then, based on this extraction result, the coarse centerline extraction result of the plain CT images is corrected, resulting in a gold standard for centerline extraction from plain CT images that is entirely machine-based and does not rely on manual annotation. On this basis, a convolutional neural network is used for supervised training of the plain CT images and their coarse centerline extraction results to obtain a corrected centerline extraction model based on plain CT images. After accurately extracting the centerline, considering that the presence of a false lumen in aortic dissection can affect the extraction result of the aortic vessel centerline, i.e., produce abnormal deviations, this invention uses a transformer based on a self-attention mechanism to obtain a binary classification model. The abnormality of the extraction result is used as the criterion to diagnose whether the patient has aortic dissection, thereby assisting doctors in diagnosis. Attached Figure Description

[0028] Figure 1 It is a 3D U-Net network structure;

[0029] Figure 2 The registration process for CT plain scan images and CTA images;

[0030] Figure 3 This is a flowchart of the aortic dissection diagnosis method of the present invention;

[0031] Figure 4This is a structural diagram of the aortic dissection diagnostic device of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0033] Example 1:

[0034] like Figure 3 As shown, this invention provides a method for diagnosing aortic dissection based on the vascular centerline of plain CT images, comprising the following steps:

[0035] Step 1: Aortic multimodal image registration based on 3D U-Net network

[0036] This step involves registering two different modalities of CT images: plain scan and contrast-enhanced CT. First, the raw image data is acquired using medical imaging equipment and converted into the medically standard DICOM format, resulting in each patient's aortic plain scan image and aortic CT angiography image. Hereinafter, the aortic plain scan CT image will be referred to as a CT plain scan image, and the aortic CT angiography image as a CTA image.

[0037] After obtaining the original CT and CTA images, preprocessing is required to obtain preprocessed image data. The same preprocessing is performed on both the CTA and CT images. Because medical imaging is characterized by relatively small datasets and high similarity in some data, random horizontal and vertical rotations are first applied to the images to prevent overfitting during subsequent model training. Simultaneously, CT slices should be normalized to ensure a uniform image size. This size depends on the specific CT imaging equipment. The following standardization formula is applied during image preprocessing:

[0038]

[0039] Where μ is the mean, σ is the variance, b is the offset, and g is the scaling factor. In addition, considering the inevitable noise interference during image generation and transmission, noise suppression measures should be taken. This invention uses the Wiener filtering algorithm to achieve adaptive filtering and eliminate noise. On the other hand, considering the low contrast of the aorta and surrounding heart organs in CT plain scan images, the HU values ​​of voxels with excessively high (e.g., bones) or excessively low (e.g., the heart) values ​​are forcibly converted to zero during preprocessing. The voxel threshold should be set according to the CT equipment; generally, a positive threshold can be set to 150 HU, and a negative threshold can be set to -150 HU. The same operation is performed in CTA images. This makes the edges of blood vessels more prominent in the image and weakens the edges of bones and heart organs, avoiding interference with centerline extraction.

[0040] Because CTA images require the injection of contrast agents during imaging, differences in brightness, contrast, and resolution can occur between CTA images and CT scan images. Furthermore, imaging conditions such as CT capture time, CT imaging parameters, and camera position and angle can also lead to differences between CT and CTA images. On the other hand, changes in the patient's physical condition between imaging sessions or image distortion caused by pathology can result in a non-one-to-one correspondence between CTA and CT images. This invention employs a 3D U-Net-based CT image registration method to register CTA and CT images. During registration, the CT scan image is used as the reference image, and the CTA image is a floating image. The convolutional neural network used in the registration process is a 3D U-Net. Figure 1 The 3D U-Net network structure is shown, which is the same as the standard U-Net, with an encoder on the left and a decoder on the right. The entire registration process is as follows: Figure 2 As shown.

[0041] The objective function of 3D U-Net includes two terms: a loss function based on the difference between the reference image and the floating image, and a term primarily used to prevent parameter overfitting and limit the registration transformation displacement. The loss function is typically based on the negative value of a similarity metric between two images used in the registration problem, while the other term needs to be selected based on the specific CT equipment used. The objective function is as follows:

[0042] G(φ,F,W)=λL(φ)+D(F,W)

[0043] Where φ represents the registration displacement field, λ represents the regularization parameter, F represents the reference image, and W represents the registration image. D(F, W) is the loss function based on the difference between the reference image and the floating image, L(φ) represents the regularization term used to prevent parameter overfitting and limit the amount of registration transformation displacement, and G(φ, F, W) is the objective function.

[0044] After setting the objective function, the neural network is trained to obtain the transformation matrix from the floating image to the reference image, thus completing the registration.

[0045] This process establishes the basic dataset required by the method of the present invention, namely, the patient's aortic CT plain scan image dataset and the one-to-one aortic CTA image dataset.

[0046] Step 2: CTA aortic centerline extraction based on minimum cost path algorithm

[0047] The aortic centerline is extracted from the CTA image dataset described in Step 1. This step employs a minimum-cost path algorithm-based extraction method for the aortic centerline. First, aortic seed points are marked on the CTA dataset established in Step 1, with one point each at the ascending aorta, aortic arch, and descending aorta. Then, the Dijkstras algorithm is applied to these three seed points to obtain the minimum-cost path. This algorithm searches for the path between the starting and ending seed points that minimizes the cumulative cost, where the cumulative cost is defined as:

[0048] E(C)=∫P(C p )C′ p |d p θ′|

[0049] Where P(x) represents the cost function value at position x, and the cost function is defined as follows:

[0050]

[0051] Where ∈ is a very small positive number to prevent singular values ​​of P(x). V(x) is the vascular characteristic function, defined as:

[0052] V(x) = max V σ (α), σ min ≤σ≤σ max

[0053]

[0054] Among them, R B =α1 / α2, β and c are two parameters. α1 and α2 are the eigenvalues ​​of the Hessian matrix at the point (x, y) in the image, satisfying |α1|≤|α2| and α1≈0. Given a starting seed point, the algorithm can search for all paths with minimum cost, and finally backtrack to find the path with the minimum cumulative cost, which is the vessel centerline. The above vessel centerline extraction is performed on the aortic CTA vessels. After the vessel centerline is extracted, it is marked on the CT plain scan image as the training basis for subsequent vessel centerline extraction on CT plain scan images. At the same time, patients with aortic dissection and normal individuals in the entire dataset are marked as the basis for the formation of subsequent aortic dissection diagnostic models.

[0055] Step 3: Extraction of the aortic vessel centerline from CT plain scan images based on CTA extraction results

[0056] Based on the CTA image centerline extraction results described in step two, the aortic vessel centerline is extracted from the CT plain scan image dataset described in step one. First, the minimum cost path algorithm involved in step two is used to coarsely extract the vessel centerline from the CT plain scan image. However, the starting and ending seed points during extraction should be set based on the starting and ending points of the extraction results in step two. Considering that the minimum cost path algorithm may extract curves that deviate from the actual center and are biased towards the vessel boundary due to inaccurate initial point settings, the Euclidean distance between the starting and ending points of the coarsely extracted vessel centerline in step two and the vessel boundary image is calculated. If the distance is too close to the vessel boundary on one side, a starting point is randomly selected on the other side for extraction using the minimum cost path algorithm. Simultaneously, a term is added to the cost function during extraction, namely...

[0057]

[0058] Where θ represents the aortic centerline extracted in step three, θ′ represents the newly generated path, and ||θ-θ′|| haus Let be the Hausdorff distance between the two paths. The Hausdorff distance, or Maximum Symmetric Surface Distance (MSD), is used to calculate the similarity between any two sets of points in an image. Its mathematical definition is as follows: Let S(A) be the set of surface elements of set A in space, then the shortest distance from any point v in space to set S(A) is:

[0059]

[0060] Where ||·|| represents the Euclidean distance, and based on the above definition, MSD is defined as:

[0061]

[0062] In practical applications, Hausdorff distance is often calculated in the following simple form, where p and q represent the gold standard and the algorithm extraction result, respectively. and This represents the distance transformation between the gold standard and the algorithm extraction results.

[0063]

[0064] Because CTA images have clearer vessel boundaries and higher image contrast, the extraction of the aortic vessel centerline from CTA images is more accurate and instructive. Adding this step ensures that the aortic vessel centerline generated from the CT scan image does not deviate excessively from the vessel centerline in the CTA image. After the above steps, the aortic vessel centerline from the CT scan image is finally generated.

[0065] Step 4: Train a model for extracting the aortic vessel centerline based on CT scan images.

[0066] Using the centerline extraction result of the CT scan image described in step three as the gold standard, supervised aortic vessel centerline extraction based on a convolutional neural network is performed. The input of this network is the coarse extraction result of the minimum cost path algorithm on the CT scan image. The neural network architecture is 3D U-Net, and the loss function is designed based on the Dice coefficient and Hausdorff distance. The loss function based on Hausdorff distance has already been introduced in step three; here, only the loss function based on the Dice coefficient is introduced, and its definition is:

[0067]

[0068] Here, ε is a very small smoothing coefficient used to prevent the denominator from being zero, and also to prevent excessively steep changes in the loss and gradient. I and U are obtained by multiplying and adding the output probability map with the corresponding matrix of the gold standard; this is the loss value based on the Dice coefficient (Dice Loss). The loss function used in this step is the average of the above two:

[0069]

[0070] The training of the 3D U-Net network has been described in step one, and the training process and hyperparameter tuning will not be repeated here. After training, this invention obtains a model for extracting the aortic vessel centerline based on CT plain scan images. It should be noted that 3D U-Net is a commonly used network in the field of medical image processing; other neural networks capable of achieving the same purpose can also be used here.

[0071] Step 5: Aortic dissection assisted diagnosis based on the centerline of CT plain scan images

[0072] Pathologically, aortic dissection forms a false lumen adjacent to the true lumen of the aorta. The boundary of the false lumen is similar to that of the true lumen, thus interfering with the extraction of the centerline. This step, based on this principle, diagnoses aortic dissection by determining whether the aortic centerline in the patient's CT scan is abnormal. First, based on the aforementioned aortic centerline extraction results from the CT scan and the corresponding patient labels, a binary classification model using a transformer based on a self-attention mechanism is obtained. That is, the CT scan of aortic dissection patients indicates an abnormal aortic centerline, while the CT scan of normal individuals indicates a normal aortic centerline. For patients with aortic dissection, the false vessel boundary created by the false lumen will shift the aortic centerline extraction results. Therefore, a series of normal vectors connecting the aortic vessel boundary and the aortic centerline are used as input, and the classification probability is output through a sigmoid function. If the aortic centerline is determined to be abnormal, the patient is very likely to have aortic dissection, thus assisting doctors in diagnosing and treating aortic dissection.

[0073] The present invention has the following technical effects:

[0074] 1. This invention addresses the issue of poor performance of the minimum cost path algorithm in extracting the aortic vessel centerline from CT scan images by taking the following measures: First, in step two, the minimum cost path algorithm is used to extract the aortic vessel centerline from CTA images, obtaining a relatively accurate extraction result based on the computer itself. Then, in step three, this extraction result is used to guide the centerline extraction from CT scan images. This is mainly reflected in two aspects: first, it provides better starting and ending seed points for the minimum cost path algorithm to generate new aortic vessel centerlines; second, it sets up a correction mechanism based on Hausdorff distance to prevent points that are too far from the aortic vessel boundary from becoming seed points. This approach effectively utilizes the higher quality of CTA images, which is more conducive to centerline extraction, resulting in better centerline extraction results from CT scan images. This is beneficial for subsequent training to obtain an aortic vessel centerline extraction model based on CT scan images.

[0075] 2. This invention addresses the problem of poor performance of a single CT plain image centerline extraction model trained on a convolutional neural network. It builds upon the centerline extraction results from the plain image, which are entirely based on the computer's own capabilities, as described in step three. This annotation result completely avoids the subjectivity of doctors during annotation, facilitates the convergence of the convolutional neural network, and reduces the cost of implementing this invention under different medical conditions.

[0076] 3. According to existing technology, it is difficult for doctors to directly determine whether the aorta has a true or false lumen from CT scan images, thus determining whether a patient has aortic dissection. Therefore, in step five, this invention diagnoses aortic dissection by judging whether the aortic vessel centerline is abnormal, that is, using a transformer based on a self-attention mechanism to obtain a binary classification model to assist in diagnosis. Considering that the aortic vessel centerline extracted from CT scan images of different patients is specific, the line connecting the aortic vessel boundary and the centerline is used as input when designing the transformer input, that is, the relative positional relationship between the aortic vessel boundary and the centerline is used to determine whether the centerline is abnormal. Due to the interference of the false lumen in aortic dissection, this relative positional relationship will change due to the false lumen boundary, so this diagnostic method can accurately determine whether a patient has aortic dissection.

[0077] Example 2:

[0078] like Figure 4 As shown, the present invention also provides a diagnostic device for aortic dissection, comprising:

[0079] The registration module is used to register CTA images and CT plain scan images to obtain the patient's aortic CT plain scan image dataset and the corresponding aortic CTA image dataset;

[0080] The first extraction module is used to extract the aortic vessel centerline from the aortic CTA image dataset.

[0081] The second extraction module is used to extract the aortic vessel centerline from the CT plain scan image dataset based on the CTA image centerline extraction results.

[0082] The training module is used to train the aortic vessel centerline extraction model based on the centerline extraction results of CT plain scan images.

[0083] The diagnostic module is used to extract the aortic vessel centerline from the patient's CT plain scan image based on the aortic vessel centerline extraction model; at the same time, it performs auxiliary diagnosis of aortic dissection based on the extraction results and the corresponding patient labels.

[0084] As one implementation method of this embodiment, the second extraction module uses a cost function based on Hausdorff distance to extract the aortic vessel centerline from the CT plain scan image dataset using a minimum cost path algorithm.

[0085] As one implementation method of this embodiment, the training module uses the centerline extraction result of CT plain scan image as the gold standard to perform supervised training on the 3D U-Net neural network to obtain the aortic vessel centerline extraction model based on CT plain scan image.

[0086] As one implementation method of this embodiment, the diagnostic module uses the relative positional relationship between the aortic vessel boundary and the centerline to determine whether the centerline is abnormal, thereby diagnosing whether the patient has aortic dissection.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A diagnostic device for aortic dissection, characterized in that, include: The registration module is used to register CTA images and CT plain scan images based on a 3D U-Net network to obtain the patient's aortic CT plain scan image dataset and the corresponding aortic CTA image dataset. The first extraction module is used to extract the aortic vessel centerline from the aortic CTA image dataset. The second extraction module is used to extract the aortic vessel centerline from the CT plain scan image dataset based on the centerline extraction results of the CTA image, using a cost function based on Hausdorff distance and a minimum cost path algorithm. The training module is used to perform supervised training of the 3DU-Net neural network with the centerline extraction results of CT plain scan images as the gold standard, so as to obtain the aortic vessel centerline extraction model based on CT plain scan images. The diagnostic module is used to extract the aortic vessel centerline from the patient's CT plain scan image based on the aortic vessel centerline extraction model; at the same time, based on the patient's extraction results and corresponding patient labels, it performs auxiliary diagnosis of aortic dissection; in particular, it uses the relative positional relationship between the aortic vessel boundary and the centerline to determine whether the centerline is abnormal, thereby diagnosing whether the patient has aortic dissection. The second extraction module performs the following process: Label aortic seed points on the CTA dataset, with one point each at the ascending aorta, aortic arch, and descending aorta. Apply Dijkstra's algorithm to these seed points to find the minimum-cost path, searching between the starting and ending seed points for the path that minimizes the cumulative cost. The cumulative cost is defined as: E(C)=∫P(C p )C′ p |d p θ′| Where P(x) represents the cost function value at position x, and the cost function is defined as follows: Where ∈ is a very small positive number. To prevent singular values ​​of P(x), V(x) is the vascular characteristic function, defined as: V(x)=maxV σ (a),s min ≤σ≤σ max Among them, R B =α1 / α2, β and c are two parameters, α1 and α2 are the eigenvalues ​​of the Hessian matrix at point (x,y) in the image, and satisfy |α1|≤|α2|, α1≈0; given the starting seed point, search for all paths with minimum cost, and finally backtrack to find the path with minimum cumulative cost, which is the blood vessel centerline; Calculate the Euclidean distance between the starting and ending points of the coarsely extracted vessel centerline and the vessel boundary image. If the distance is too close to the boundary on one side, randomly select a starting point on the other side and perform extraction using the minimum cost path algorithm. Simultaneously, add a term to the cost function during extraction, i.e. Where θ represents the aortic centerline, θ′ represents the newly generated path, and ||θ-θ′|| haus The distance between the two paths is the Hausdorff distance.

Images