Aortic dissection disease classification method and device, storage medium

By adopting multi-task cascade training and attention mechanisms in aortic dissection disease, combining the center point sequence and geometric characteristics, extracting the aortic centerline and segmentation operation, the low accuracy in segmentation and typing of aortic dissection diseases is solved. The problem of this problem realizes automated typing and accurate lesion type judgment.

CN114332072BActive Publication Date: 2024-11-22BEIJING INST OF TECH
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202210009427.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-05
Publication Date
2024-11-22
Estimated Expiration
2042-01-05

AI Technical Summary

Technical Problem

The prior art has low accuracy in the segmentation and typing of aortic dissection diseases, especially due to the poor segmentation effect caused by excessive difference in true and false cavity volumes, and deep learning methods have limited data volume and complex structure. The effect is poor in the case.

Method used

Using a multi-task cascade training strategy based on semantic segmentation and geometric characteristics, combined with attention mechanism and center point sequence method, the A and B aortic dissections are performed by extracting the aortic centerline and segmentation operations, and combined with Stanford typing to perform A and B aortic dissections. classification.

Benefits of technology

Automatic classification of aortic dissection diseases is realized, the accuracy of the true and false cavity segmentation algorithm is improved, and the segmentation effect is enhanced through attention mechanism and geometric feature extraction methods is solved, solving the problem of low classification accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114332072B_ABST
    Figure CN114332072B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and device for aortic dissection disease classification, and a storage medium, comprising: obtaining CT image data of aortic dissection disease patients; performing true and false lumen segmentation of aortic dissection according to the CT image data to obtain the aorta; extracting the aortic centerline from the aorta; performing segmentation operations on the aorta according to the aortic centerline to obtain the ascending aorta part, the aortic arch part, and the descending aorta part; and classifying the ascending aorta part, the aortic arch part, and the descending aorta part into type A and type B based on the Stanford classification. The technical solution of the present invention is adopted to solve the problem of low segmentation accuracy caused by the large difference in true and false lumen volume when segmenting aortic dissection, and on this basis, the aortic dissection disease classification is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a method and device for classifying aortic dissection disease based on semantic segmentation and geometric characteristics, and a storage medium. Background Art

[0002] With the development of artificial intelligence in recent years, the use of computers to assist doctors in clinical diagnosis and decision-making has become a hot topic in the field of medical imaging, and medical imaging, which is often used in clinical diagnosis, has also become a research hotspot, such as computer CT scans and MRI. Aortic dissection disease refers to the rupture of the intima of the aortic tissue when high-pressure and high-speed blood flows into it, forming a true lumen and a false lumen. The incidence of the disease is very low, but the mortality rate is high, so early medical imaging examinations and diagnosis are very necessary. The following introduces the research status of several common research directions in the field of aortic dissection disease.

[0003] 1. Aortic dissection segmentation research

[0004] In the existing technology, there are many methods for aortic dissection segmentation, from mature traditional image processing technology to emerging methods such as machine learning and deep convolutional neural networks, all of which are used to segment the true and false lumens of aortic dissection. So far, there are three main types of aortic dissection segmentation algorithms, namely model-based algorithms, wavelet analysis, and deep learning research.

[0005] Model-based algorithms exploit the prior knowledge of the circular shape of the aorta and Hough transform to further refine the segmented aorta using an elastic deformable model (DM). To a large extent, model-based methods rely on prior assumptions about the morphology and properties of aortic dissections in CTA images, which makes them prone to failure on irregularly shaped arteries (such as arteries with aneurysms and severe stenosis).

[0006] Wavelet analysis algorithm is a method that combines multi-scale wavelet analysis with generative model matching to distinguish and segment the true lumen and false lumen. Based on the semi-automatic method to detect the complete boundary of the aorta, wavelet analysis is used to detect the edge of the dissected aorta, and then discriminative learning is used to refine the edge. Finally, generative learning methods are used to distinguish the true lumen and false lumen. However, this method is usually used to segment the normal aorta, and the segmentation effect for uncommon aortic dissection is poor.

[0007] In recent years, with the development of deep learning, some algorithms have been gradually used in the field of aortic dissection segmentation. In recent years, there are two main related algorithms. One is to segment AD by extracting the entire visual area commonly seen in the field of image segmentation, including two 3D segmentation networks: the first network is used to segment the entire aorta, and the second network is used to simultaneously segment the true lumen and the false lumen; the other is a new image segmentation method recently proposed. This method believes that the segmentation of aortic dissection only requires the extraction of the lumen contour, and only some simple processing is required to obtain the entire lumen area. It uses two independent two-dimensional cascade CNNs to extract the contours of the entire abdominal aorta lumen and the true lumen respectively. It uses a fixed number of points to annotate the contours (edges) of the entire aorta and the true lumen, and then simply converts the loops of the entire aortic lumen and the true lumen to obtain the loops of the false lumen, which means that the segmentation of aortic dissection is achieved.

[0008] 2. Aortic centerline extraction research

[0009] Centerline extraction is a very important issue in the field of aortic dissection research. Since the aorta can be regarded as a curved tubular structure, aortic centerline extraction can be considered as tubular object centerline extraction. There are four main methods for extracting the centerline of tubular objects, namely, topology refinement-based, distance transformation-based, path planning-based, and tracking-based methods.

[0010] The method based on topological thinning is to use morphological corrosion operations to continuously thin the object and finally leave the skeleton to obtain the center line. First, a thinning operation is performed in the initial blood vessel area to preliminarily obtain the blood vessel center line. At the same time, in order to prevent holes and additional lines from being generated after thinning, morphological closure is applied to the initial blood vessel area before thinning. Iterative calculations are usually performed in serial or parallel mode. Although iteration makes the calculation process complicated, this method can well maintain the geometric connectivity structure of the object and has high accuracy.

[0011] The distance transform-based method is a commonly used method for extracting the center line. The center of the sphere, that is, the center point, is obtained by determining the largest inscribed sphere of the object. Continuously performing this operation on the entire three-dimensional object can obtain a series of sphere centers, and finally a complete center line can be obtained.

[0012] Path planning based methods are usually implemented based on vision. The method based on positioning the virtual camera is more computationally efficient based on the medial axis transformation. By iteratively correcting the path towards the medial axis, the necessity of evaluating whether simple points are standard during the morphological thinning process is eliminated. The result of this algorithm is a smooth, focused path that facilitates navigation in virtual endoscopic exploration of 3D medical images.

[0013] The tracking-based method is to determine the local direction of the tubular object to obtain the directional change trend of the center line, and calculate the center point of the cross section. Repeat this process for the entire object, and finally get a center line. In this method, the judgment of the local direction is very important. The method based on the Hessian matrix determines the direction of the center line according to the direction of the eigenvector of the minimum eigenvalue; the ridge-based method determines the direction of the tubular by calculating the gradient; there is also a contour-based model, which gives the initial center line shape and then calculates the energy function to make the center line move continuously. When the gradient change is the smallest, the center line is obtained. As for the classification of aortic dissection, the traditional classification of dissection in the past mainly relied on manual diagnosis by doctors. This method requires doctors to observe a large number of CT images every day, which is a huge workload and low efficiency, and may lead to misdiagnosis due to doctor fatigue and other conditions.

[0014] In recent years, classification methods based on deep learning have begun to be applied in aortic dissection classification. Some deep learning classification algorithms have been used for aortic dissection classification. However, due to the limited amount of aortic dissection data, the poor network learning effect, and the large number of structures in the human body environment, the different shapes of the aorta, and the ever-changing distribution of true and false lumens, deep learning is even more unable to accurately learn accurate features, resulting in low classification accuracy. Therefore, it is difficult to achieve good results by relying solely on deep learning methods. Therefore, a reasonable and innovative aortic dissection classification method is very important for the classification of aortic dissection. Summary of the invention

[0015] The technical problem to be solved by the present invention is to provide a method and device, and a storage medium for aortic dissection disease classification based on semantic segmentation and geometric characteristics. In order to solve the problem of low segmentation accuracy caused by the large difference in the volume of the true and false cavities during aortic dissection segmentation, a new channel attention mechanism is integrated into the multi-task cascade training strategy, each channel is given its own weight, and feature fusion is performed, thereby achieving the purpose of using the results of complete aortic segmentation to guide the true and false cavity segmentation; on this basis, a classification study is carried out, firstly, the center line is extracted by a method based on the center point sequence, and then the key points are extracted and segmented based on the geometric characteristics of the aorta, and finally the aortic dissection disease classification is achieved.

[0016] To achieve the above object, the present invention adopts the following technical solution:

[0017] A method for classifying aortic dissection disease, comprising the following steps:

[0018] Step S1, obtaining CT image data of a patient with aortic dissection disease;

[0019] Step S2, segmenting the true and false lumens of aortic dissection according to the CT image data to obtain the aorta;

[0020] Step S3, extracting the aortic centerline of the aorta;

[0021] Step S4, segmenting the aorta according to the aortic centerline to obtain an ascending aorta portion, an aortic arch portion, and a descending aorta portion;

[0022] Step S5: classify the ascending aorta, aortic arch and descending aorta into type A and type B based on the Stanford classification.

[0023] Preferably, in step S2, a multi-task cascade strategy is used to segment the true and false lumens of aortic dissection, and the connection between tasks is achieved by sharing network parameters based on an attention mechanism.

[0024] Preferably, in step S3, the boundary points of the aorta are first found, and then the center point is found by using the intersection of the discrete cross section and the aorta, and finally the center line of the aorta is extracted.

[0025] Preferably, in step S4, the aorta is segmented based on the discrete point curvature of the aorta.

[0026] Preferably, in step S5, the model obtained by true and false lumen segmentation training is used to establish a classifier to test the ascending and descending aorta parts, obtain the dice index of the true and false lumens and the predicted volumes of the true and false lumens, and then use probability to determine whether the patient's aortic dissection is type A or type B; let the probability that the aorta segment A contains the true lumen be P diceA1 , the probability that segment A contains a false lumen is P dicea2 , the volume of the true lumen of segment A is Num A1 , the volume of the aortic false lumen in segment A is Num A2 , the probability that segment C contains a true cavity is P diceC1 , the probability that segment C contains a false lumen is P diceC2 , the volume of the true lumen of the aorta in segment C is Num C1 , the volume of the C segment aortic false lumen is Num C2 .

[0027] The probability of aortic dissection being type A or type B was calculated using the following formula:

[0028]

[0029]

[0030] The classifier is used to evaluate the calculated probability that the aortic dissection is type A and the probability that the aortic dissection is type B, and then outputs the judgment of the aorta type.

[0031] The present invention also provides an aortic dissection disease classification device, comprising:

[0032] An acquisition module, used for acquiring CT image data of a patient with aortic dissection disease;

[0033] A segmentation module, used for segmenting the true and false lumens of aortic dissection according to the CT image data to obtain the aorta;

[0034] An extraction module, used for extracting the aorta centerline from the aorta;

[0035] A segmentation module, used for segmenting the aorta according to the aortic centerline to obtain an ascending aorta portion, an aortic arch portion and a descending aorta portion;

[0036] The classification module is used to classify the ascending aorta, aortic arch and descending aorta into type A and type B based on the Stanford classification.

[0037] Preferably, the segmentation module adopts a multi-task cascade strategy to perform true and false lumen segmentation of aortic dissection, and realizes the connection between tasks by sharing network parameters based on the attention mechanism.

[0038] Preferably, the extraction module extracts the aorta centerline based on a center point sequence.

[0039] Preferably, the segmentation module segments the aorta based on the discrete point curvature of the aorta.

[0040] The present invention also provides a storage medium storing a computer program executable by an electronic device, and when the program is run on the electronic device, the electronic device executes an aortic dissection disease classification method.

[0041] The present invention has the following technical effects:

[0042] 1. The present invention sequentially performs segmentation of the true and false lumens of aortic dissection, extraction of the centerline, aortic segmentation and classification of aortic dissection, thereby realizing automatic classification of aortic dissection disease.

[0043] 2. The present invention proposes a new attention mechanism, through which the results of the first task can be used to guide the training of the second task, thereby improving the segmentation effect of the true and false lumen segmentation algorithm of aortic dissection.

[0044] 3. The present invention proposes an aorta centerline extraction method based on a center point sequence, which first finds the boundary points of the aorta, then uses the intersection of the discrete cross section and the aorta to find the center point, and finally draws the centerline of the aorta.

[0045] 4. The present invention proposes an aortic segmentation method based on geometric characteristics. Based on a series of discrete center point sequences obtained previously, the curvature can be used to find the position with the maximum curvature. This position is the dividing point of the segmentation, and the three segments of the ascending aorta, descending aorta, and aortic arch can be obtained.

[0046] 5. The present invention proposes an aortic dissection classification method based on the Stanford criteria. First, a classifier is established using a model trained by true and false lumen segmentation. Then, aortic dissection is classified based on whether each segment in the Stanford classification criteria contains a true and false lumen, thereby ultimately solving the problem of poor classification effect of current aortic dissection. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flow chart of the aortic dissection disease classification method of the present invention;

[0048] Figure 2 Schematic diagram of multi-task cascade strategy;

[0049] Figure 3 This is the schematic diagram of the attention module;

[0050] Figure 4 is a schematic diagram of the shape of the aorta;

[0051] Figure 5 is a schematic diagram of the circumscribed circle of an obtuse triangle;

[0052] Figure 6 It is a structural diagram of the aortic dissection disease classification method of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0054] Embodiment 1:

[0055] like Figure 1 As shown, the present invention provides a method for classifying aortic dissection disease, comprising:

[0056] Step S1: Obtain CT image data of a patient with aortic dissection

[0057] Medical images have various formats. Commonly used 3D image formats include nii, nrrd, Nifti, etc., and 2D image formats include DICOM, jpg, png, gif, etc. The original CT image format provided by the hospital is DICOM format. Each patient has hundreds of DICOM slices. However, considering that CT images are 3D images, the 3D spatial information in the image is very rich, and the 2D segmentation network ignores the positional relationship between different layers and cannot use this information. In addition, the images at different positions are often quite different, which is not conducive to the network learning common features. Therefore, it is necessary to convert the format first, that is, convert the data from the 2D DICOM format to the 3D nii format.

[0058] In addition, after using 3D Slicer software to mark, the format of the generated marked image is nrrd, and the different suffix names will cause trouble for subsequent program writing. Therefore, in order to facilitate subsequent program writing, it is necessary to keep the suffix name unified, so the suffix name of the marked file is changed from nrrd to nii.

[0059] In order to facilitate subsequent training, the present invention divides the CT image data of 100 patients with aortic dissection into training sets and test sets in a ratio of 3:1, and in order to ensure the randomness of the data, the data distribution order is not interfered by any human factors. In addition, data augmentation is performed using operations such as translation and rotation, and grayscale transformation is performed to increase the overall brightness of the CT image, which helps to improve the effect of aortic true and false lumen segmentation.

[0060] Step 2: Segmentation of the true and false lumens of aortic dissection based on attention mechanism

[0061] The segmentation algorithm of aortic dissection adopts a multi-task cascade strategy. First, the complete aorta is extracted from the CT image, and then the result is input into the network again to guide the segmentation of the true and false cavities of aortic dissection. The algorithm adopted by the present invention is an improved algorithm of the 3DUnet algorithm. However, 3D Unet has some shortcomings. For example, some features will be lost in the process of downsampling and then upsampling, resulting in inaccurate feature learning. Therefore, the present invention is improved on the basis of the original network. First, the data is preprocessed by grayscale transformation, and then the attention module is added on the basis of the original algorithm, which becomes a 3D ModifiedUnet algorithm. Finally, in order to improve the segmentation effect, the present invention changes the loss function.

[0062] (1) Multi-task cascade strategy

[0063] Usually in the field of deep learning, when multiple tasks need to be performed in sequence, a cascade strategy is generally adopted. Since the present invention needs to perform complete aortic segmentation first and then true and false lumen segmentation, a multi-task cascade strategy is adopted, and its principle diagram is shown in FIG. Figure 2As shown in the figure. Multi-task cascading refers to combining multiple related prediction tasks and using one of the tasks to improve the accuracy of other tasks. Since the networks of different tasks may be different, the present invention usually achieves the connection between tasks by sharing network parameters. For the aortic dissection segmentation task, the present invention adopts the attention mechanism to share parameters. The principle diagram is shown in the attached figure. Figure 2 shown.

[0064] (2) Attention Mechanism

[0065] The attention mechanism was first used in machine translation tasks and has been widely used in computer vision classification and semantic segmentation tasks in recent years. It is mainly composed of convolution operations. Combining this module helps to refine the segmentation of image details and parts that are difficult to segment.

[0066] The principle diagram of the attention module used in the present invention is shown in the attached figure. Figure 3 As shown in the figure, it can use the result of the first task to guide the training of the second task, that is, after the first task segments the complete aorta, the second task directly performs the segmentation of the true and false cavities on this basis, so as to achieve better results. Its structure consists of two parts, namely the Category Channel Weight (CCW) module and the Complementary Segmentation (CS) module, which are introduced separately below.

[0067] For the CCW module, the probability that each voxel in the entire CT image belongs to the background and the complete aorta has been obtained in the first task. Then, the present invention uses the following formulas (1) and (2) to calculate P a and P n Parameter,P a and P n Represents the probability distribution of all voxels belonging to the aorta and non-aorta, respectively.

[0068]

[0069]

[0070] in, C aorta and C non-aorta represent the pixel sets belonging to the aorta and non-aorta respectively, and C aorta It contains the true lumen and false lumen of aortic dissection. i, j, k represent the x, y, and z axis coordinates of the pixel respectively, and c represents a single voxel.

[0071] Next, the present invention converts P a and Pn The dimension of is reduced to a one-dimensional vector for subsequent calculations, that is, the transformed Where N = W × H × L, represents the total number of voxels. Similarly, F is also converted into 32 one-dimensional vectors, that is, where i∈[1,32].

[0072] On this basis, the channel weight W can be a and W n Calculate, W a and W n They represent the weight of each channel for segmenting the aorta and the background, respectively, and the calculation formula is shown below:

[0073]

[0074]

[0075] Among them, W a and W n All meet And i∈[1,32].

[0076] For the CS module, the block contains two paths that can make full use of the channel importance information of specific categories (W a and W n ) to improve segmentation performance. Figure 3 As shown in the right part, these two pathways focus on the segmentation of tumor and non-tumor voxels respectively. The CompSeg block can be described in more detail as follows. First, W a and W n are used to recalibrate each channel in F separately.

[0077]

[0078]

[0079] Use a 1×1×1 convolution kernel to a and U n Perform convolution operation and get S a and S n , and then perform weighted averaging to obtain Then pass it through a 1×1×1 convolution kernel output to get the result.

[0080]

[0081] (3) Loss Function

[0082] In the field of medical image segmentation, the present invention usually uses the DICE index to evaluate the accuracy of segmentation, and its calculation expression is as follows:

[0083]

[0084] Among them, |X∩Y| represents the intersection between X and Y, |X| and |Y| represent the number of elements of X and Y respectively, and the coefficient is multiplied in the numerator because there is repeated calculation in the denominator. This processing can ensure that the dice is between [0, 1]. The dice coefficient is usually used in medical image segmentation to measure the degree of overlap between the predicted labeled part and the true labeled part.

[0085] The corresponding loss function expression is: Loss = 1-dice. Since there is a class imbalance problem in the complete aorta segmentation of the first task and the true and false cavity segmentation of the second task, that is, the volume of the foreground is too small compared to the background volume, this will lead to reduced segmentation accuracy.

[0086] Now we make improvements based on formula (8). Since the essence of the class imbalance problem is that the volume difference between the foreground and the background is too large, we consider assigning different weights to each class according to the volume size, which can alleviate the problem of poor segmentation effect caused by the large volume difference between classes. That is, we improve the expression to the following expression:

[0087]

[0088] Where N is the number of categories in the segmentation task.

[0089] For the research problem of the present invention, both the complete aorta segmentation and the true and false lumen segmentation are two categories, so the expression can be specifically as follows:

[0090]

[0091] Since the number of true and false cavity categories is 2, N is 2, and α t Used to control the weight of the positive and negative samples on the total loss. t Take a smaller value to reduce the weight of the large part, considering α t Associated with dice, that is That is, weights are assigned by considering the volumes of the two categories. Generally speaking, the weight of the category with a larger volume is smaller, which can make up for the problem of category imbalance.

[0092] Step 3: Aortic centerline extraction based on center point sequence

[0093] The segmentation part in step 2 has segmented the complete aorta, and on this basis, the aortic centerline can be extracted. The present invention divides a CT image into all pixel points with the aortic boundary as the dividing line. Divided into two parts, the points inside the aorta are called inner points, and the points outside the aorta are called outer points. They are divided into two sets and named and At this point, the boundary points can be defined as the following set:

[0094]

[0095] Here, R(voxel) represents a sphere with a radius of 2 voxels and the voxel as the center. This expression can be used to find the boundary contour with a thickness of 2. Then we can start looking for the center point one by one. For any point in the aorta, the minimum distance from the point to the boundary can be calculated. The Euclidean distance is used as follows:

[0096] d=||x1-x2|| (11)

[0097] Among them, d represents the shortest distance from the point to the boundary, and x1 is the set A point in the set, that is, a point on the boundary, x2 is a The previous point is a point inside the aorta.

[0098] Then find the center point and consider using a cross section to truncate the CT image to obtain one or two circular cross sections. When there is only one circular cross section, it is obvious that the radius of the circle can be calculated using this cross section; however, when there are two cross sections, it is impossible to calculate directly, so consider using the maximum and minimum intersection points of the cross section and the aorta, and then connect the two points. After calculating the other two intersection points of this line with the aorta, the radii of the two circles can be calculated respectively. The intersection point of the cross section and the aorta is set as r in this invention. ij , where i represents the i-th cross section and j is equal to 1 or 2, indicating that there is one circular cross section or two circular cross sections, respectively.

[0099] Then, the set of center points can be obtained using the following formula:

[0100] χ={x i |int(d)=r ij} (12)

[0101] After the center point set is calculated, the centerline extraction begins. In the present invention, the process of obtaining the aortic centerline is analogized to obtaining the shortest distance from the vertex of the descending aorta to the vertex of the ascending aorta at the other end. By consulting the literature, it can be known that the Dijkstra algorithm is usually used to find the minimum path from the starting point to the final point, also known as the minimum path algorithm. Therefore, the Dijkstra algorithm can be used to extract the centerline, and the extracted minimum path can be smoothed by interpolation, that is, the centerline extraction is completed.

[0102] Step 4: Aortic segmentation based on aortic geometric characteristics

[0103] After the centerline is extracted, the aorta can be segmented. By consulting the literature and observing CT images, it can be seen that the actual aorta is approximately an arched cylindrical structure, and the curvature is the largest at the top of the aortic arch. The shape of the aorta is shown in the appendix. Figure 4 As shown, this is used as a priori condition to facilitate subsequent segmentation and classification. It can also be seen from the figure that the aorta is divided into three segments, namely the ascending aorta (segment A), the aortic arch (segment B) and the descending aorta (segment C).

[0104] We have already obtained the set of central points χ, and we can further calculate the curvature of each point. Usually, curvature refers to the rotation rate of the tangent direction angle of a point on a continuous curve to the arc length.

[0105] However, the present invention studies discrete center points, so the curvature cannot be calculated using a continuous calculation method. Considering geometric knowledge can also be used to find the curvature of discrete points. According to geometric knowledge, the coordinates of three known points can determine a triangle, and any triangle can determine a complete circumscribed circle. The radius of curvature can be obtained from the radius of the circle, and the curvature can be obtained by taking the inverse of it. Therefore, three adjacent center points can form a triangle, and it must be an obtuse triangle, as shown in the figure below. Figure 5 shown.

[0106] The radius R of the circumscribed circle is deduced to obtain the equation shown in formula (13):

[0107]

[0108] The relationship between the radius of curvature and the curvature is as follows:

[0109]

[0110] That is, the magnitude of the curvature K is:

[0111]

[0112] Formula (15) can be used to calculate the curvature of all points except the start and end positions. The start and end positions are definitely not the positions of the aortic arch and can therefore be ignored.

[0113] Then all the curvatures are screened. Considering the prior knowledge that the curvature of the top of the aortic arch is the largest in the entire aorta, the coordinates of the point with the largest curvature (smallest curvature radius) can be obtained by comparing all the curvatures (radii of curvature), namely, Figure 4 Point P. Then, from point P, find the lower boundary line of the aorta downward, get the coordinates that intersect with the boundary line, and cut along the z coordinate of this point to get the upper and lower parts. The upper part is the aortic arch part (segment B), and the lower part is the ascending aorta and descending aorta. Then further segment the lower part to get the ascending aorta (segment A) and descending aorta (segment C). Perform the same processing on all CT images, record the boundary data of the aorta in an Excel table, and write a program to automatically segment the original aortic image and the annotated image.

[0114] Step 5: Classification of aortic dissection based on Stanford

[0115] According to the location of the intimal tear and the distribution of the false lumen, aortic dissection can be divided into different types, corresponding to different treatment plans. There are two commonly used classification bases for aortic dissection disease, namely Stanford and Debakey. Among them, Stanford classification is more commonly used. It classifies aortic dissection according to the location of the intimal tear, that is, all ascending aortas containing false lumens are called type A (TAAD), and all ascending aortas that do not contain false lumens are called type B (TBAD). The aortic typing of the present invention mainly adopts the Stanford typing basis.

[0116] First, in order to determine the stability of the model, the present invention uses a model for segmenting true and false lumens to segment the healthy aorta. In the final result, the dice index of the false lumen is 0. Therefore, it can be considered that the trained model will not produce incorrect segmentation results due to special inputs. For example, when the input ascending aorta image does not have a false lumen (in this case, it is type B aortic dissection), it will not be mistakenly considered to have a true lumen. That is, the model can be used for subsequent classification operations.

[0117] The main idea of ​​this part is to build a classifier to classify the ascending and descending aorta into type A and type B. The classifier is improved on the basis of segmentation. The segmentation trained model is used to test the two parts of the aorta to obtain the dice index of the true and false lumens and the predicted volume of the true and false lumens. Then, the probability is used to determine whether the patient's aortic dissection is type A or type B. The following is the probability calculation:

[0118] We have previously obtained the probability that segment A of the aorta contains the true lumen. diceA1 , the probability that segment A contains a false lumen is P diceA2 , the volume of the true lumen of the aorta in segment A is Num A1 , the volume of the aortic false lumen in segment A is Num A2 , the probability that segment C contains a true cavity is P diceC1 , the probability that segment C contains a false lumen is P diceC2 , the volume of the true lumen of the aorta in segment C is Num C1 , the volume of the C segment aortic false lumen is Num C2 .

[0119] Using the above data, the probability of aortic dissection being type A or type B can be calculated using the following expressions:

[0120]

[0121]

[0122] The classifier evaluates the calculated probability that the aortic dissection is type A and the probability that it is type B, respectively, and then outputs a judgment on the aortic type.

[0123] Embodiment 2:

[0124] like Figure 6 As shown, the present invention also provides an aortic dissection disease classification device, comprising:

[0125] An acquisition module, used for acquiring CT image data of a patient with aortic dissection disease;

[0126] A segmentation module, used for segmenting the true and false lumens of aortic dissection according to the CT image data to obtain the aorta;

[0127] An extraction module, used for extracting the aorta centerline from the aorta;

[0128] A segmentation module, used for segmenting the aorta according to the aortic centerline to obtain an ascending aorta portion, an aortic arch portion and a descending aorta portion;

[0129] The classification module is used to classify the ascending aorta, aortic arch and descending aorta into type A and type B based on the Stanford classification.

[0130] As an implementation method of this embodiment, the segmentation module adopts a multi-task cascade strategy to perform true and false lumen segmentation of aortic dissection, and realizes the connection between tasks by sharing network parameters based on the attention mechanism.

[0131] As an implementation of this embodiment, the extraction module extracts the aorta centerline based on a center point sequence.

[0132] As an implementation of this embodiment, the segmentation module performs a segmentation operation on the aorta based on the discrete point curvature of the aorta.

[0133] Embodiment 3:

[0134] The present invention also provides a storage medium storing a computer program executable by an electronic device, and when the program is run on the electronic device, the electronic device executes an aortic dissection disease classification method.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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 method for classifying aortic dissection disease, characterized in that: The following steps are involved: Step S1, obtaining CT image data of a patient with aortic dissection disease; Step S2, segmenting the true and false lumens of aortic dissection according to the CT image data to obtain the aorta; Step S3, extracting the aortic centerline of the aorta; Step S4, segmenting the aorta according to the aortic centerline to obtain an ascending aorta portion, an aortic arch portion, and a descending aorta portion; Step S5, classifying the ascending aorta, the aortic arch and the descending aorta into type A and type B based on the Stanford classification; In step S2, a multi-task cascade strategy is used to segment the true and false lumens of aortic dissection, and the connection between tasks is achieved by sharing network parameters based on the attention mechanism; In step S3, the boundary points of the aorta are first found, and then the center point is found by using the intersection of the discrete cross section and the aorta, and finally the center line of the aorta is extracted; In step S4, the aorta is segmented based on the discrete point curvature of the aorta; In step S5, the model obtained by true and false lumen segmentation training is used to establish a classifier, and the ascending and descending aorta parts are tested to obtain the dice index of the true and false lumens and the predicted volumes of the true and false lumens. Then, the probability is used to determine whether the patient's aortic dissection is type A or type B. Let the probability that the aorta segment A contains the true lumen be P diceA1 , the probability that segment A contains a false lumen is P diceA2 , the volume of the true lumen of segment A is Num A1 , the volume of the aortic false lumen in segment A is Num A2 , the probability that segment C contains a true cavity is P dicec1 , the probability that segment C contains a false lumen is P dicec2 , the volume of the true lumen of the aorta in segment C is Num C1 , the volume of the C segment aortic false lumen is Num C2 ; The probability of aortic dissection being type A or type B was calculated using the following formula: The classifier is used to evaluate the calculated probability that the aortic dissection is type A and the probability that the aortic dissection is type B, and then outputs the judgment of the aorta type.

2. An aortic dissection disease typing device for implementing the aortic dissection disease typing method according to claim 1, characterized in that: include: An acquisition module, used for acquiring CT image data of a patient with aortic dissection disease; A segmentation module, used for segmenting the true and false lumens of aortic dissection according to the CT image data to obtain the aorta; An extraction module, used for extracting the aorta centerline from the aorta; A segmentation module, used for segmenting the aorta according to the aortic centerline to obtain an ascending aorta portion, an aortic arch portion and a descending aorta portion; The classification module is used to classify the ascending aorta, aortic arch and descending aorta into type A and type B based on the Stanford classification.

3. A storage medium, characterized in that: It stores a computer program executable by an electronic device, and when the program runs on the electronic device, the electronic device executes the steps of the method according to claim 1.

Citation Information

Patent Citations

  • A segmentation method of dissected aorta in CT images based on a convolution neural network

    CN109035255A

  • Dissection artery segmentation method and device

    CN110796670A

  • Full-type aortic dissection true and false cavity image segmentation method and system

    CN113674291A

  • Method and System for Automatic Aorta Segmentation

    US20100239148A1