Aortic dissection boundary determination method, system, electronic device and medium

Abstract

The present invention discloses a method, system, electronic device, and medium for determining the boundary of an aortic dissection, relating to the field of image segmentation. The method comprises: obtaining a CTA image for determining the boundary of an aortic dissection; preprocessing the CTA image for determining the boundary of an aortic dissection to obtain a preprocessed aortic CTA image; and determining the boundary of an aortic dissection in the preprocessed aortic CTA image using an aortic dissection boundary segmentation model. The present invention improves the accuracy of aortic dissection boundary segmentation.

Description

Technical Field

[0001] The present invention relates to the field of image segmentation, and in particular to a method, system, electronic equipment and medium for determining the boundary of an aortic dissection. Background Art

[0002] Methods for segmenting aortic dissection images include intensity-based semi-automatic algorithms such as level sets, active shape models, and graph cut models combined with shape priors for semi-automatic segmentation. Purely intensity-based techniques cannot accurately detect the boundaries of aortic dissections. For example, if organs with similar intensity values surround the lesion area, similar areas will often be classified as lesions. However, most proposed algorithms require user interaction or pre-computed lumen segmentation and centerline extraction. Furthermore, algorithms requiring user intervention compromise their robustness and applicability in clinical practice.

[0003] Traditional medical image segmentation algorithms include a graph-cut-based method. By designing a discriminant integrated feature and a new adaptive smoothing constraint, it ensures that the segmented and extracted aortic outer wall achieves a relatively smooth effect. Finally, the graph-cut optimization cost function is used to obtain the final aortic region segmentation result. This algorithm improves the segmentation accuracy of aortic dissection to a certain extent.

[0004] Since the advent of deep learning, it has been widely used in the field of organ segmentation due to its advantages such as high automation and high accuracy. DCAN is a multi-task framework that leverages contextual information at different levels to efficiently separate lesion contours, improving segmentation accuracy compared to U-net. A cascaded fully convolutional neural network (CFCN) and a dense 3D conditional random field (CRF) perform staged segmentation of CT abdominal images with high accuracy and speed. This demonstrates that a multi-stage, multi-level, cascaded architecture can help improve segmentation accuracy.

[0005] In the field of image segmentation, it has been theoretically proven that deeper networks have a larger receptive field and higher learning ability. However, deeper networks can also lead to vanishing gradients and activation function failure due to their complex structure. Increasing network layers also poses the problem of network parameter degradation. To achieve deeper networks, residual neural networks (ResNets) can be used. Unlike continuous convolutional layers, the ResNet model uses multiple convolutional layers to form residual convolutional blocks. This residual design enables ResNet models to maintain a deeper network layer while effectively improving segmentation accuracy. This residual structure has been applied to medical and natural image segmentation tasks with promising results. However, because aortic dissection is a rare disease, imaging datasets are typically small, which can lead to insufficient generalization of residual neural networks. Furthermore, residual neural networks require significant computing resources for training and inference, resulting in relatively long processing times and low accuracy. Summary of the Invention

[0006] The purpose of the present invention is to provide a method, system, electronic equipment and medium for determining the boundary of aortic dissection, so as to improve the segmentation accuracy of the boundary of aortic dissection.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A method for determining the boundary of an aortic dissection, comprising:

[0009] Acquire an aortic CTA image; the aortic CTA image is a CTA image for determining the boundary of the aortic dissection;

[0010] Preprocessing the aorta CTA image to obtain a preprocessed aorta CTA image; the preprocessed aorta CTA image is an aorta CTA enhanced image having a half envelope; the half envelope is a half envelope of the pulmonary artery to the aorta in the aorta CTA image;

[0011] The aortic dissection boundary of the preprocessed aortic CTA image is determined by using an aortic dissection boundary segmentation model; wherein, the aortic dissection boundary segmentation model is based on the L-BFGS algorithm and is obtained by training a cavity dense Y-type network using a first training data set; the first training data set includes historical aortic CTA images with marked true boundaries of aortic dissection; the cavity dense Y-type network includes a first encoder, a second encoder, a feature fusion module and a decoder; the first encoder and the second encoder are both connected to the feature fusion module; the first encoder, the second encoder and the feature fusion module are all connected to the decoder; the first encoder is a feature fusion module that is used to identify the aortic dissection boundary. The DenseNet network is trained using a training data set; the second training data set is historical aortic CTA images obtained in the ImageNet data set; the second encoder is a cascade structure consisting of four improved DenseBlock layers and a dilated spatial pyramid pooling module, wherein the improved DenseBlock layer is a standard convolution layer in the DenseBlock layer followed by two dilated convolution layers connected in sequence; the decoder includes multiple improved upsampling modules; the improved upsampling module is obtained by replacing the second standard convolution layer of the upsampling module of the U-Net network with the improved DenseBlock layer.

[0012] Optionally, preprocessing the aorta CTA image to obtain a preprocessed aorta CTA image specifically includes:

[0013] Performing a convolution operation on the aorta CTA image to obtain a first feature map; the grayscale value of all pixels in the first feature map is 0;

[0014] Determine a Hessian matrix for each pixel in the aorta CTA image; the Hessian matrix includes a first eigenvalue, a second eigenvalue, a first eigenvector, and a second eigenvector;

[0015] Determine a second feature map based on the first eigenvalue and the second eigenvalue of each pixel; the second feature map is a feature map having the half envelope; the grayscale value of the pixel at the half envelope is 1;

[0016] Setting the grayscale values of the pixels in the direction of the first eigenvector of the pixels with grayscale values of 1 in the second feature map to 1, to obtain a third feature map;

[0017] The third feature map is used as a mask and mapped onto the aorta CTA image to obtain a preprocessed aorta CTA image.

[0018] Optionally, determining a second feature map according to the first eigenvalue and the second eigenvalue of each pixel point specifically includes:

[0019] Determine whether the absolute value of the first eigenvalue of any pixel point is less than the absolute value of the second eigenvalue;

[0020] If yes, set the grayscale value of the pixel to 1;

[0021] If not, keep the gray value of the pixel at 0;

[0022] Traverse all pixels and get the second feature map.

[0023] Optionally, based on the L-BFGS algorithm, the first training data set is used to train the dilated dense Y-shaped network, specifically including:

[0024] Preprocessing the historical aorta CTA image to obtain a preprocessed historical aorta CTA image;

[0025] Inputting the preprocessed historical aortic CTA image into the cavity dense Y-shaped network to obtain the aortic dissection prediction boundary of the historical aortic CTA image;

[0026] Determining a loss function value between the predicted boundary of the aortic dissection and the actual boundary of the aortic dissection;

[0027] Determine whether the loss function value meets a preset condition; the preset condition is that the maximum number of iterations is reached or the loss function value is less than a preset value;

[0028] If yes, stop training and use the current dense hole Y-shaped network as the aortic dissection boundary segmentation model;

[0029] If not, determining the gradient value of each parameter in the void dense Y-type network with respect to the loss function value;

[0030] According to the gradient value, the L-BFGS algorithm is used to update the parameters of the void dense Y-type network, and the void dense Y-type network after the updated parameters is used as the void dense Y-type network, and the process returns to the step of "inputting the historical aortic CTA image with the true boundary of aortic dissection marked into the void dense Y-type network to obtain the predicted boundary of aortic dissection in the historical aortic CTA image".

[0031] Optionally, the loss function value is the average of the focal loss function value and the improved Dice loss function value; the improved Dice loss function value is determined based on the improved Dice coefficient; the improved Dice coefficient is determined based on the binary classification prediction probability of each pixel point, the gold standard and the hyperparameters.

[0032] Optionally, determining the improved Dice loss function value specifically includes:

[0033] Using the formula L Dice =1-Doce modified (p, g) determines the improved Dice loss function value; wherein, Dice modified (p,g) is the improved Dice coefficient.

[0034] Optionally, the determination of the improved Dice coefficient specifically includes:

[0035] Using the formula Determine the improved Dice coefficient; wherein, p i is the binary classification prediction probability value of the i-th pixel; g i is the gold standard; k is a hyperparameter; N is the number of pixels in the preprocessed historical aortic CTA image.

[0036] A system for determining the boundary of an aortic dissection, applied to the above-mentioned method for determining the boundary of an aortic dissection, comprises:

[0037] An image acquisition module is used to acquire an aortic CTA image; the aortic CTA image is a CTA image of the aortic dissection boundary to be determined;

[0038] a preprocessing module, configured to preprocess the aortic CTA image to obtain a preprocessed aortic CTA image; the preprocessed aortic CTA image is an enhanced aortic CTA image having a half envelope; the half envelope is a half envelope of the pulmonary artery relative to the aorta in the aortic CTA image;

[0039] A boundary determination module is used to determine the aortic dissection boundary of the preprocessed aortic dissection image using an aortic dissection boundary segmentation model; wherein the aortic dissection boundary segmentation model is based on the L-BFGS algorithm and is obtained by training a cavity dense Y-type network using a first training data set; the first training data set includes historical aortic CTA images with marked true boundaries of aortic dissections; the cavity dense Y-type network includes a first encoder, a second encoder, a feature fusion module and a decoder; the first encoder and the second encoder are both connected to the feature fusion module; the first encoder, the second encoder and the feature fusion module are all connected to the decoder; the first encoder is The DenseNet network is trained using a second training data set; the second training data set is historical aortic CTA images obtained in the ImageNet data set; the second encoder is a cascade structure consisting of four improved DenseBlock layers and a dilated spatial pyramid pooling module, wherein the improved DenseBlock layer is a standard convolution layer in the DenseBlock layer followed by two dilated convolution layers connected in sequence; the decoder includes multiple improved upsampling modules; the improved upsampling module is obtained by replacing the second standard convolution layer of the upsampling module of the U-Net network with the improved DenseBlock layer.

[0040] An electronic device includes: a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-mentioned aortic dissection boundary determination method.

[0041] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for determining the boundary of an aortic dissection is implemented.

[0042] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0043] The aortic dissection boundary determination method, system, electronic device, and medium of the present invention preprocess the CTA image whose aortic dissection boundary is to be determined to obtain an enhanced CTA image of the aorta with a semi-envelope, thereby achieving aortic separation, effectively maintaining the independence of the aortic target, reducing the possibility of obtaining false-positive target areas, and improving segmentation accuracy. Furthermore, the aortic dissection boundary segmentation model is used to determine the aortic dissection boundary in the enhanced CTA image of the aorta with a semi-envelope, capturing richer local and global information in both shallow and deep layers, and integrating semantic features from low to high levels, thereby reducing feature information loss, expanding the receptive field, and achieving the extraction and learning of multi-scale feature information, thereby improving segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 A flow chart of the method for determining the boundary of aortic dissection provided by the present invention;

[0046] Figure 2 The aorta CTA image in Example 1 of the present invention;

[0047] Figure 3 Schematic diagram of characteristic vectors in embodiment 1 of the present invention;

[0048] Figure 4 This is a structural diagram of the ASPP module in Example 1 of the present invention;

[0049] Figure 5 This is a diagram of a densely populated Y-shaped network structure provided in Example 1 of the present invention;

[0050] Figure 6 This is a structural diagram of the decoder provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] The purpose of the present invention is to provide a method, system, electronic equipment and medium for determining the boundary of aortic dissection, so as to improve the segmentation accuracy of the boundary of aortic dissection.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Example 1

[0055] like Figure 1 As shown, the method for determining the boundary of aortic dissection provided by the present invention includes:

[0056] Step 101: Acquire an aortic CTA image; the aortic CTA image is a CTA image for determining the boundary of an aortic dissection.

[0057] Step 102: Preprocess the aortic CTA image to obtain a preprocessed aortic CTA image; the preprocessed aortic CTA image is an enhanced aortic CTA image having a half envelope; the half envelope is the half envelope of the pulmonary artery to the aorta in the aortic CTA image.

[0058] In practical applications, different vascular structures are displayed in CTA image sequences at different levels. Even the same vascular structure can show differences in morphology, range, and position in images at different levels. In certain levels of the CTA image sequence, the distance between the aorta and the superior vena cava and the pulmonary artery in the mediastinum is very small. For image segmentation, the pixels and connected domains of the target area are two very important pieces of information for distinguishing foreground from background. Therefore, if target segmentation is performed directly on the original image, there is a high probability that the superior vena cava or pulmonary artery will be segmented as targets, resulting in false positive areas and greatly reducing segmentation accuracy. Therefore, it is very necessary to preprocess the CTA image before segmenting the aorta target.

[0059] In the CTA image sequence, the positional relationship and distribution characteristics between the aorta and other blood vessels in the mediastinum are observed, such as Figure 2 As shown in the figure, SVC stands for superior vena cava, AO stands for ascending aorta, and MPA stands for pulmonary artery. It can be seen that the ascending aorta is adjacent to the pulmonary artery and superior vena cava, the superior vena cava is located to the upper left of the ascending aorta, and the pulmonary artery is located to the lower right of the ascending aorta. The pulmonary artery also has a semi-envelope tendency towards the ascending aorta and superior vena cava. Therefore, there is a narrow gap area between the three. If this area can be extended along the tangent direction of the semi-envelope, the gap between the ascending aorta, superior vena cava, and pulmonary artery can be increased, thereby achieving separation of the three. Therefore, determining this semi-envelope is a very important step in CTA image preprocessing. To achieve this goal, an aortic segmentation method based on the Hessian matrix is adopted.

[0060] The present invention uses a filter based on the Hessian matrix to detect the half-envelope in the image; the implementation idea is to construct a feature map of the half-envelope structure in the image through the meaning of the eigenvalues and eigenvectors of the Hessian matrix, thereby determining the above-mentioned narrow and long gap area.

[0061] The Hessian matrix H of the two-dimensional image f(x,y) (aortic CTA image) is determined by its second-order partial derivative:

[0062]

[0063] Where x and y are the horizontal and vertical coordinates of the pixel point in the two-dimensional image respectively.

[0064] In a two-dimensional image, H is a two-dimensional positive definite matrix with two distinct eigenvalues, λ1 (the first eigenvalue) and λ2 (the second eigenvalue). These two eigenvalues correspond to two eigenvectors, e1 (the first eigenvector) and e2 (the second eigenvector). For different structures in the two-dimensional image, the magnitude relationship between λ1 and λ2 exhibits different characteristics. In particular, for curved structures, the eigenvalues satisfy |λ1| < |λ2|.

[0065] The first-order derivative of each pixel in a two-dimensional image represents the grayscale gradient of that point, while the second-order derivative represents the rate of change of the gradient at that point. For a curve, the second-order derivative represents its curvature, and the greater the curvature, the more curved the curve. For a curve structure, the two eigenvalues of its Hessian matrix show anisotropy in the direction of their respective corresponding eigenvectors, such as Figure 3 As shown in the figure, eigenvector e1 points in the direction of the curve's tangent, and its corresponding eigenvalue λ1 is small. Eigenvector e2 is perpendicular to the curve's tangent, and its corresponding eigenvalue λ2 is large. In other words, λ1 and e1 represent the strength and direction of the curve's small curvature, while λ2 and e2 represent the strength and direction of the curve's large curvature.

[0066] Based on the above analysis, the present invention proposes the following steps for aortic CTA image preprocessing:

[0067] 1) performing a convolution operation on the aorta CTA image to obtain a first feature map; the grayscale value of all pixels in the first feature map is 0.

[0068] In practical applications, the original CTA image is convolved with a 3×3 filter to construct a feature map f0 (first feature map) that is as large as the two-dimensional image f(x, y), and the grayscale values of all pixels in the feature map f0 are initialized to 0.

[0069] 2) Determine the Hessian matrix of each pixel in the aorta CTA image; the Hessian matrix includes a first eigenvalue, a second eigenvalue, a first eigenvector, and a second eigenvector.

[0070] In practical applications, the Hessian matrix corresponding to each pixel in the two-dimensional image f(x,y) is obtained, and the eigenvalues (first eigenvalue and second eigenvalue) and eigenvectors (first eigenvector and second eigenvector) of each Hessian matrix are obtained.

[0071] 3) Determine a second feature map based on the first eigenvalue and the second eigenvalue of each pixel; the second feature map is a feature map having the half envelope; the grayscale value of the pixel at the half envelope is 1.

[0072] In practical applications, the eigenvalues corresponding to each pixel are traversed, and the grayscale values of the corresponding positions of the pixels satisfying the condition |λ1|<|λ2 in the feature map f0 are set to 1, thereby obtaining the feature map f1 (second feature map) that can indicate the half-envelope structure in the two-dimensional image.

[0073] 4) The grayscale values of the pixels in the direction of the first eigenvector of the pixels with grayscale values of 1 in the second feature map are set to 1 to obtain a third feature map.

[0074] In practical applications, for a pixel point with a grayscale value of 1 in the feature map f1, starting from this point, the pixels in the direction of the feature vector e1 are set to 1, thereby obtaining the enhanced feature map f2 (the third feature map) after region expansion.

[0075] 5) Mapping the third feature map as a mask to the aorta CTA image to obtain a preprocessed aorta CTA image.

[0076] In practical applications, the enhanced feature map f2 is used as a mask to map it to the two-dimensional image f(x, y) to obtain the preprocessed aortic CTA image.

[0077] Step 103: Determine the aortic dissection boundary of the preprocessed aortic CTA image using the aortic dissection boundary segmentation model; wherein the aortic dissection boundary segmentation model is based on the L-BFGS algorithm and is obtained by training the cavity dense Y-type network using the first training data set; the first training data set includes historical aortic CTA images with the real boundary of the aortic dissection marked; the cavity dense Y-type network includes a first encoder, a second encoder, a feature fusion module and a decoder; the first encoder and the second encoder are both connected to the feature fusion module; the first encoder, the second encoder and the feature fusion module are all connected to the decoder; the first encoder is used to The DenseNet network is trained using a second training data set; the second training data set is historical aortic CTA images obtained in the ImageNet data set; the second encoder is a cascade structure consisting of four improved DenseBlock layers and a dilated spatial pyramid pooling module, wherein the improved DenseBlock layer is a standard convolution layer in the DenseBlock layer followed by two dilated convolution layers connected in sequence; the decoder includes multiple improved upsampling modules; the improved upsampling module is obtained by replacing the second standard convolution layer of the upsampling module of the U-Net network with the improved DenseBlock layer.

[0078] In practice, patients with aortic dissection have both a true lumen and a false lumen. Clinically, the true lumen is often small, while the false lumen is large and irregular in shape, making the segmentation task more challenging. As the aorta progresses from proximal to distal, the vessel diameter gradually decreases, and different aortic branches emerge. This creates a sample imbalance between the true and false lumens in two-dimensional images, posing a challenge to the loss function. Furthermore, the gradual shift in size between different slices and the resulting size instability of the true and false lumens in the segmentation task also need to be addressed, necessitating the extraction and learning of multi-scale feature information.

[0079] To solve these problems, the neural network must first face the problem of aortic span and parameter quantity. The DenseNet network can effectively reduce the parameter quantity and have better convergence at the same depth. The DenseBlock can effectively realize the reuse of features through dense connection design. The introduction of void convolution is conducive to expanding the receptive field of each layer and capturing more features without increasing the amount of additional calculations. In order to make full use of high-resolution and low-resolution feature information, jump connections can also be introduced to splice shallow semantic features and deep semantic feature information to achieve feature fusion. Based on the above ideas, the present invention proposes a void dense Y-type network (AtrousDenseY-Net, referred to as AD Y-Net).

[0080] The construction process of AD Y-Net is as follows:

[0081] 1. Introduce the DenseNet architecture.

[0082] First, we introduce the residual neural network ResNet. ResNet solves the degradation problem by introducing deep residual learning. Let x l is the output of the lth layer, through the output x of the previous layer l-1 Apply a nonlinear transformation function to calculate x l :

[0083] x l =F l (x l-1 )+x l-1 .

[0084] F l It represents the nonlinear transformation function to be learned, which consists of three layers: the first is the batch normalization layer (BN) and the nonlinear activation layer (ReLU), and the last is a 3×3 convolution layer.

[0085] F l The shapes of (x) and x must be consistent, otherwise a shortcut connection W is required. l to match the size.

[0086] x l =F l (x l-1 )+W l (x l-1 ).

[0087] DenseNet proposes a more radical dense connection mechanism: by improving information exchange between layers, all layers are interconnected, that is, each layer receives all previous layers as additional input. This structure allows the network to reuse features and improve efficiency.

[0088] x l =F l ([x0,x1,…,x l-1 ]).

[0089] Where, […] represents a concatenation operation. l It is a nonlinear conversion function that uses the structure of BN layer + ReLU layer + 3×3 convolution layer.

[0090] The DenseNet network structure is mainly composed of DenseBlock and Transition layers. The feature map sizes of different DenseBlock layers are inconsistent. In order to achieve a densely connected network structure, each DenseBlock is connected through the Transition layer to match the feature map sizes. The nonlinear transformation function in DenseBlock is F l ,Different from residual learning, the Transition layer of DenseNet consists of a BN layer, a nonlinear activation layer (ReLU), a 1×1 convolution layer, and a 2×2 average pooling layer.

[0091] DenseNet can be considered an extension of ResNet, establishing dense connections between all previous layers and subsequent layers and enabling feature reuse by connecting features at the channel level. The advantages of these modifications are: 1) fewer parameters and more efficient computation; 2) dense connections improve gradient backpropagation, making the network easier to train; and 3) alleviating the vanishing gradient problem.

[0092] 2. Introduce the ASPP module.

[0093] First, we introduce dilated convolution. Dilated convolution expands the kernel elements by adding points with zero weight between consecutive kernel elements. Compared with standard convolution, dilated convolution introduces an additional hyperparameter r, called the dilation rate, which determines the step size for sampling the input pixels. Assuming that the original convolution kernel size is K, the receptive field size d of the convolution kernel after the introduction of dilated convolution is:

[0094] d = K + (K-1) × (r-1).

[0095] The spatial resolution of the feature map after the introduction of the dilated convolution is o w 、o h for:

[0096]

[0097]

[0098] Among them, m×n is the convolution kernel size, w×h is the input image size, r is the void ratio, p is the padding, and s is the step size.

[0099] From the above formula, we can conclude that the higher the dilation rate, the wider the receptive field and the more comprehensive the global information extracted; on the contrary, the smaller the dilation rate, the more local the features it contains. The choice of dilation rate is also very important. The dilation rate is not constant for each layer. If the feature maps extracted by simply superimposing the same dilation rate multiple times will cause local information loss. Too high a dilation rate will also make the two features obtained have no correlation. To avoid this problem, the dilation rate during multiple dilation convolutions can be set to a sawtooth shape and ensure that its common divisor is not greater than 1. This can improve the segmentation performance of the network. For example:

[0100] r i =[1,2,3,4,1,2,3,4...](i=1,2...,q).

[0101] where r i It represents the dilation rate of the i-th dilated convolution in the q-th dilated convolution.

[0102] Large-sized convolution kernels result in a large amount of computation and a significant loss of resolution during pooling. Compared to standard convolution, dilated convolution can expand the filter's receptive field, helping to capture more feature information without reducing spatial resolution or increasing the amount of computation, thus resolving the conflict between a large receptive field and resolution loss. Therefore, based on the Pyramid Pooling Module (PPM), the present invention replaces the convolutional layers with dilated convolutions and introduces the Dilated Spatial Pyramid Pooling Module (ASPP) to achieve encoder feature extraction and fusion of multi-scale targets.

[0103] The structure of the ASPP module is as follows Figure 4 As shown in the figure, the first branch is a 1×1 standard convolution, aiming to maintain the original receptive field; the second to fourth branches are dilated convolutions with different dilation rates, aiming to extract features and obtain different receptive fields; the fifth branch performs global average pooling on the input to obtain global features. Finally, the feature maps of the five branches are stacked in the channel dimension and fused through a 1×1 standard convolution to fuse information at different scales. Unlike the PPM module, the ASPP module uses dilated convolutions with different dilation rates, which facilitates obtaining contextual feature information of the image at different scales. It then concatenates feature maps of different sizes in the channel dimension, ultimately achieving multi-scale feature fusion.

[0104] 3. Build the ADY-Net network.

[0105] like Figure 5 As shown in Figure 2, the AD Y-Net network uses two encoders and one decoder network.

[0106] 3.1 Encoder structure

[0107] Encoder 1 (the first decoder) is a pretrained DenseNet-169 (a densely connected convolutional network with 169 layers). The training process is as follows: after obtaining the ImageNet dataset, all network weights are initialized. The input data is forward-propagated through the convolutional layer, downsampling layer, and fully connected layer to obtain the network output. The error between the network output and the target value is calculated and passed back to the network. The error of each layer in the network is calculated in turn, and the weights of each layer are updated based on the calculated error. Training is repeated until the error between the network output and the target value is less than or equal to the expected value. Max pooling layers are used during the downsampling process to preserve more texture information. This encoder can avoid the model from falling into local optimal points and requires a smaller dataset, which greatly alleviates the problem of small biomedical image datasets.

[0108] Encoder 2 (the second encoder) utilizes a cascaded structure consisting of four densely connected blocks (DenseBlocks) and an atrous spatial pyramid pooling (ASPP) module. Two atrous convolutional layers are added sequentially after the standard convolutional layer in DenseBlock for feature extraction. This means that the nonlinear transformation function structure of DenseBlock becomes a batch normalization layer + ReLU layer + 3×3 standard convolutional layer + 3×3 atrous convolutional layer + 3×3 atrous convolutional layer. This improved densely connected block is called an atrous densely connected block (ADB). Compared to the original densely connected block, the ADB block preserves resolution and expands the receptive fields of both low- and high-level layers. The feature extraction process is as follows: First, in the ADB block, one standard convolution is followed by two atrous convolutions. The outputs of all atrous convolutions are concatenated as the input to the next operation. This cascade process gradually increases the receptive field and aggregates multiple receptive fields within the same resolution. This allows Encoder 2 to perceive the different receptive fields corresponding to different resolutions. Secondly, an average pooling operation is performed between each of the first three ADB blocks to compress image features, reduce computational overhead, and preserve more background information. To maximize the preservation of high-dimensional feature image information and facilitate subsequent multi-scale feature extraction, no pooling operation is performed before the fourth ADB block. Finally, using the fourth ADB block and the ASPP module, dilated convolutions with appropriate dilation rates and global average pooling are used to stack the feature maps of the five branches in the channel dimension. Information at different scales is then fused through a 1×1 standard convolution.

[0109] ADB contains K dilated convolutional layers with dense connectivity, where the forward model of the kth layer can be expressed as:

[0110] x t =H(y1,…,y t-1 ),y t=D(x t )

[0111] where x t and y t are the input and output of the tth layer respectively, the function H() is the nonlinear combination function of the output of the previous t-1 layer, and the function D() is the void convolution operation.

[0112] 3.2 Decoder structure.

[0113] like Figure 6 As shown in FIG, the present invention redesigns the upsampling path based on the U-Net upsampling module, which can be well adapted to the bidirectional encoder structure. Figure 6 It represents the process of an upsampling operation during the decoding process. Figure 6 The “+” and “×” in the figure represent feature concatenation and element-wise multiplication in the channel dimension. The bottom layer, from left to right, represents the feature map output by encoder 2, the feature fusion map of encoder 1 and encoder 2 (obtained through the feature fusion module), and the feature map output by encoder 1.

[0114] The upsampling layer consists of a transposed convolution with a stride of 2 and a kernel of 3×3, a BN layer, a ReLU layer, and a standard convolution to reduce the channel dimension.

[0115] To restore the spatial resolution of the input image, the feature maps output by encoder 1 and the fused feature maps of encoders 1 and 2 are first upsampled using transposed convolution to obtain feature maps with twice the resolution. The feature map of encoder 2, which has the same resolution as the upsampled fused feature map, is concatenated with the upsampled fused feature map. A standard convolutional layer is then used to reduce the number of concatenated feature maps to 1 / 2, so that the number of feature maps matches that of the feature map output by encoder 1.

[0116] The ADB block replaces the standard convolutional layer used in the U-Net upsampling process. Compared to standard convolution, this expands the receptive field of the convolution kernel while preserving resolution. Finally, the feature map generated by the ADB block through dilated convolution is multiplied element-wise along the channel dimension with the feature map obtained after upsampling by encoder 1 to achieve feature self-correction. After five similar operations, the resolution of the feature map is restored to the original image size for output.

[0117] The last three layers of the network use three standard convolutional layers to reduce the channel dimension of the output image, and finally use a sigmoid function nonlinear activation to calculate the classification of each pixel. Except for the last convolutional layer, all other convolutional layers use batch normalization layers and ReLU layers simultaneously.

[0118] The entire network always adopts a void dense connection method to learn semantic information, which allows the overall structure of the network to be deeper, realizes the expansion of the receptive field, reduces the loss of resolution, obtains high-level, multi-scale feature extraction, effectively introduces global information, and does not increase the number of parameters and computational costs, resulting in better segmentation performance.

[0119] 4. Determination of hybrid loss function.

[0120] As the aorta image changes with the body part, the size of the aorta blood vessels usually decreases as it moves from the chest to the abdomen. Since the image size remains unchanged, the number of pixels of the true and false cavities gradually decreases in the entire image, resulting in sample imbalance. At the same time, in the binary classification task, for multi-scale segmentation targets, the traditional loss function may cause the network algorithm to fall into a local optimal value during the learning process, resulting in a segmentation result that is biased towards the background. Based on the above analysis, the present invention adopts a hybrid loss function as the loss function for the segmentation task. The loss function value is determined by the focal loss function and the improved Dice loss function.

[0121] 4.1 Focal Loss Function

[0122] Sometimes the target area accounts for a much smaller proportion than the background area. In this case, the number of negative samples is much larger than the number of positive samples. Too many negative samples will cause the loss function to be too large, which is not conducive to the convergence of the positive sample objective function. Therefore, based on the cross entropy loss function, weights and adjustment factors are introduced to form the focus loss function, which is in the form of:

[0123]

[0124] Among them, G i,j is the groundtruth, M is the size of the image, set to W×H, S i,j is the predicted probability of the pixel point, w is a balanced weight given between the salient feature area and the non-salient feature area; λ is the adjustment factor, which is used to adjust the formula to the value of 1 if the network misclassifies the pixel point, otherwise λ tends to 0, which greatly reduces the loss weight of the background information.

[0125] 4.2 Improved Dice loss function.

[0126] The Dice coefficient is an important indicator for evaluating image segmentation. It evaluates the similarity between two groups of pixels and is usually used to calculate the similarity between the expected prediction image and the gold standard. The range is 0 to 1. The larger the value, the more similar the two are. The Dice coefficient expression is:

[0127]

[0128] In image segmentation, A is the network prediction result and B is the gold standard. If the network prediction result is completely consistent with the gold standard, its Dice coefficient is 1, but it is generally impossible to achieve such accuracy in practice; if the two have no overlap at all, its Dice coefficient is 0.

[0129] Considering the potential imbalance in the dataset being tested, the present invention improves the Dice loss function by increasing the weight ratio of misclassified pixels to increase the penalty for misclassification. Compared to the basic Dice loss function, when a pixel is misclassified, the gap between the predicted value of the improved Dice loss function and the true value of the label will be further widened, allowing the network to continue to train and learn the pixel in depth, and narrowing the gap in the loss function. The improved Dice coefficient and the improved Dice loss function are:

[0130]

[0131]

[0132] The sum of the loss function calculations includes N pixels, which is calculated based on the predicted i-th pixel; p i With g i Denote the binary prediction probability value and gold standard of the i-th pixel respectively; the hyperparameter k ranges from (0, 1] and is used to change the weight of the number of correctly classified pixels and the number of incorrectly classified pixels in the foreground of the segmented target. The improved Dice loss function does not need to assign weights to the foreground and background of samples of different categories as the traditional loss function based on sample reweighting does.

[0133] Based on the above analysis, the loss function proposed in this paper is the average of the focal loss function and the improved Dice loss function, which is expressed as:

[0134]

[0135] As an optional implementation, based on the L-BFGS algorithm, the first training data set is used to train the dilated dense Y-shaped network, specifically including:

[0136] The historical aortic CTA image is preprocessed to obtain a preprocessed historical aortic CTA image. In practical applications, in order to train the cavity dense Y-type network, a sufficient number of labeled data sets (first training data sets) need to be prepared. The present invention selects one of the largest aortic dissection databases: International Registry of Acute Aortic Dissection (IRAD). On the basis of constructing ADY-Net, the weights and biases of the AD Y-Net model are initialized using the normal distribution random initialization method. The hybrid loss function proposed in step 4 is used. Considering that AD Y-Net is essentially a nonlinear model and the data set of medical images is relatively small, the present invention uses the L-BFGS optimization algorithm to train AD Y-Net. L-BFGS (Limited-memory Broyden-Fletcher-Goldfarb-Shanno) is a gradient-based optimization algorithm used to solve unconstrained nonlinear optimization problems. In particular, for small data sets, its effect is often better than traditional optimization algorithms such as stochastic gradient descent.

[0137] The preprocessed historical aortic CTA image is input into the cavity dense Y-shaped network to obtain the aortic dissection prediction boundary of the historical aortic CTA image.

[0138] Determine a loss function value between the predicted boundary of the aortic dissection and the actual boundary of the aortic dissection.

[0139] Determine whether the loss function value meets a preset condition; the preset condition is that the maximum number of iterations is reached or the loss function value is less than a preset value.

[0140] If so, the training is stopped and the current hollow dense Y-shaped network is used as the aortic dissection boundary segmentation model.

[0141] If not, determine the gradient value of each parameter in the void dense Y-type network with respect to the loss function value.

[0142] According to the gradient value, the L-BFGS algorithm is used to update the parameters of the void dense Y-type network, and the void dense Y-type network after the updated parameters is used as the void dense Y-type network, and the process returns to the step of "inputting the historical aortic CTA image with the true boundary of aortic dissection marked into the void dense Y-type network to obtain the predicted boundary of aortic dissection in the historical aortic CTA image".

[0143] In practical applications, the model is trained using the prepared first training data set. In each training batch, the gradient of each parameter in the model with respect to the loss function value is calculated using the standard back-propagation algorithm. Next, the L-BFGS optimization algorithm is used to update the parameters in the model based on the gradient information. After multiple trainings, the loss function reaches the optimal value.

[0144] Use the gold standard as the test data set to evaluate the trained model and calculate the model's accuracy, recall rate, F1 value and other indicators.

[0145] Precision represents the probability of being correctly predicted in the test sample, that is, the ratio of the sample predicted by the classification model to be a positive sample and actually being a positive sample. The calculation formula is:

[0146]

[0147] The meaning of recall is the proportion of samples that are actually positive and are judged as positive by the classification model, that is, the proportion of the number of correctly predicted positive samples to the total number of positive samples. The calculation formula is:

[0148]

[0149] The F1 value is an indicator that combines precision and recall. Its essence is the harmonic mean. In this indicator, the importance of precision is considered to be equivalent to the importance of recall. Its calculation formula is:

[0150]

[0151] Based on the model evaluation results, methods such as Gaussian process are used to establish a mapping model between hyperparameters and segmentation effects. Based on this model, combined with the experience of multiple training and multiple parameter adjustments, the next hyperparameter combination is selected for training and verification to improve the network segmentation effect.

[0152] As an optional embodiment, based on the segmented image obtained in step 103, the diagnosis of aortic dissection is performed by determining whether a true lumen and a false lumen are formed in the segmented image. If a CTA image of a patient with aortic dissection contains both a true lumen and a false lumen, the patient is classified as abnormal; if a normal patient's image does not contain either a true lumen and a false lumen, the patient is classified as normal. A binary classification model is constructed based on the CTA image segmentation results and the gold standard.

[0153] The feature fusion map of the dual encoder in step 103 is used as input data for learning, the least squares support vector machine algorithm (LSSVM) is used as the mapping from the input end to the output end, and the radial basis kernel function (RBF) is used at the output end to obtain the binary classification probability.

[0154] After the binary classification model is built, it needs to be evaluated. The evaluation indicators include:

[0155] Classification accuracy (Accuracy) represents the ratio of correctly classified samples to all samples, and its calculation formula is:

[0156]

[0157] Among them, TP is the number of samples that are judged to be positive and are actually positive samples; TN is the number of samples that are judged to be negative and are actually negative samples; FP is the number of samples that are judged to be positive and are actually negative samples; FN is the number of samples that are judged to be negative and are actually positive samples.

[0158] The ROC curve is a widely used metric for evaluating the performance of classification models. The further the ROC curve deviates toward the Y-axis, the better the classifier. The horizontal axis of the ROC curve represents the false positive rate, and the vertical axis represents the true positive rate. Because the performance of different experimental results must be compared, a simple ROC curve cannot easily distinguish between good and bad classifiers. Therefore, the AUC value can be calculated as a classification model evaluation metric. The AUC value represents the area under the ROC curve, which represents the expected classification performance of a classifier. It ranges from [0, 1], with larger AUC values indicating better classifier performance.

[0159] Precision represents the probability of being correctly predicted in the test sample, that is, the ratio of the sample predicted by the classification model to be a positive sample and actually being a positive sample. The calculation formula is:

[0160]

[0161] The meaning of recall is the proportion of samples that are actually positive and are judged as positive by the classification model, that is, the proportion of the number of correctly predicted positive samples to the total number of positive samples. The calculation formula is:

[0162]

[0163] The F1 value is an indicator that combines precision and recall. Its essence is the harmonic mean. In this indicator, the importance of precision is considered to be equivalent to the importance of recall. Its calculation formula is:

[0164]

[0165] Comprehensively calculating the above evaluation metrics indicates that the model performs well in classifying CTA images. During application, if the model determines that true and false lumen segmentation exists, the patient is likely to have aortic dissection. The model then provides feedback to physicians, assisting them in diagnosing and treating patients with aortic dissection.

[0166] The aortic dissection boundary determination method, system, electronic device, and medium of the present invention have the following advantages:

[0167] 1. The present invention adopts a CTA image preprocessing method based on the Hessian matrix to discover the positional relationship between the aorta, pulmonary artery and superior vena cava. The properties of the Hessian matrix eigenvalues and eigenvectors are used to construct a feature map of the curved structure in the image and an enhanced feature map after regional expansion. The enhanced feature map is then mapped into the original image as a mask, thereby achieving the separation of the aorta, well maintaining the independence of the aorta target, reducing the possibility of obtaining false positive target areas, and improving the segmentation accuracy.

[0168] 2. An aortic true and false lumen segmentation algorithm based on a dense void Y-shaped network is proposed, which captures richer local and global information in the shallow and deep layers, and integrates semantic features from low-level to high-level, reducing the loss of feature information, expanding the receptive field, and realizing the extraction and learning of multi-scale feature information, which can achieve higher segmentation accuracy.

[0169] 3. An auxiliary diagnosis method is proposed. By evaluating multiple indicators of the segmented image and training a classification model, the judgment results are obtained and fed back to the doctor for auxiliary diagnosis, so that the cause of the disease can be found quickly and accurately, and time for treatment can be gained for the patient.

[0170] Example 2

[0171] In order to execute the method corresponding to the above embodiment 1 and achieve the corresponding functions and technical effects, a system for determining the boundary of an aortic dissection is provided below, including:

[0172] The image acquisition module is used to acquire an aortic CTA image; the aortic CTA image is a CTA image for determining the boundary of the aortic dissection.

[0173] A preprocessing module is used to preprocess the aortic CTA image to obtain a preprocessed aortic CTA image; the preprocessed aortic CTA image is an aortic CTA enhanced image with a half envelope; the half envelope is the half envelope of the pulmonary artery to the aorta in the aortic CTA image.

[0174] A boundary determination module is used to determine the aortic dissection boundary of the preprocessed aortic dissection image using an aortic dissection boundary segmentation model; wherein the aortic dissection boundary segmentation model is based on the L-BFGS algorithm and is obtained by training a cavity dense Y-type network using a first training data set; the first training data set includes historical aortic CTA images with marked true boundaries of aortic dissections; the cavity dense Y-type network includes a first encoder, a second encoder, a feature fusion module and a decoder; the first encoder and the second encoder are both connected to the feature fusion module; the first encoder, the second encoder and the feature fusion module are all connected to the decoder; the first encoder is The DenseNet network is trained using a second training data set; the second training data set is historical aortic CTA images obtained in the ImageNet data set; the second encoder is a cascade structure consisting of four improved DenseBlock layers and a dilated spatial pyramid pooling module, wherein the improved DenseBlock layer is a standard convolution layer in the DenseBlock layer followed by two dilated convolution layers connected in sequence; the decoder includes multiple improved upsampling modules; the improved upsampling module is obtained by replacing the second standard convolution layer of the upsampling module of the U-Net network with the improved DenseBlock layer.

[0175] Example 3

[0176] The present invention provides an electronic device, comprising: a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the aortic dissection boundary determination method of embodiment 1.

[0177] Example 4

[0178] The present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the method for determining the boundary of an aortic dissection according to the first embodiment is implemented.

[0179] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0180] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for determining the boundary of aortic dissection, characterized in that: include: Acquire an aortic CTA image; the aortic CTA image is a CTA image for determining the boundary of the aortic dissection; Preprocessing the aorta CTA image to obtain a preprocessed aorta CTA image; the preprocessed aorta CTA image is an aorta CTA enhanced image with a half envelope; The semi-envelope is the semi-envelope of the pulmonary artery to the aorta in the aortic CTA image; The aortic dissection boundary of the preprocessed aortic CTA image is determined by using an aortic dissection boundary segmentation model; wherein, the aortic dissection boundary segmentation model is based on the L-BFGS algorithm and is obtained by training a cavity dense Y-type network using a first training data set; the first training data set includes historical aortic CTA images with marked true boundaries of aortic dissection; the cavity dense Y-type network includes a first encoder, a second encoder, a feature fusion module and a decoder; the first encoder and the second encoder are both connected to the feature fusion module; the first encoder, the second encoder and the feature fusion module are all connected to the decoder; the first encoder is a feature fusion module that is used to identify the aortic dissection boundary. The DenseNet network is trained using a training data set; the second training data set is historical aortic CTA images obtained in the ImageNet data set; the second encoder is a cascade structure consisting of four improved DenseBlock layers and a dilated spatial pyramid pooling module, wherein the improved DenseBlock layer is a standard convolution layer in the DenseBlock layer followed by two dilated convolution layers connected in sequence; the decoder includes multiple improved upsampling modules; the improved upsampling module is obtained by replacing the second standard convolution layer of the upsampling module of the U-Net network with the improved DenseBlock layer.

2. The method for determining the boundary of aortic dissection according to claim 1, characterized in that: Preprocessing the aorta CTA image to obtain a preprocessed aorta CTA image specifically includes: Performing a convolution operation on the aorta CTA image to obtain a first feature map; the grayscale value of all pixels in the first feature map is 0; Determine a Hessian matrix for each pixel in the aorta CTA image; the Hessian matrix includes a first eigenvalue, a second eigenvalue, a first eigenvector, and a second eigenvector; Determine a second feature map based on the first eigenvalue and the second eigenvalue of each pixel; the second feature map is a feature map having the half envelope; the grayscale value of the pixel at the half envelope is 1; Setting the grayscale values of the pixels in the direction of the first eigenvector of the pixels with grayscale values of 1 in the second feature map to 1, to obtain a third feature map; The third feature map is used as a mask and mapped onto the aorta CTA image to obtain a preprocessed aorta CTA image.

3. The method for determining the boundary of aortic dissection according to claim 2, characterized in that: Determining a second feature map according to the first eigenvalue and the second eigenvalue of each pixel point specifically includes: Determine whether the absolute value of the first eigenvalue of any pixel point is less than the absolute value of the second eigenvalue; If yes, set the grayscale value of the pixel to 1; If not, keep the gray value of the pixel at 0; Traverse all pixels and get the second feature map.

4. The method for determining the boundary of aortic dissection according to claim 1, wherein: Based on the L-BFGS algorithm, the first training dataset is used to train the dilated dense Y-shaped network, specifically including: Preprocessing the historical aorta CTA image to obtain a preprocessed historical aorta CTA image; Inputting the preprocessed historical aortic CTA image into the cavity dense Y-shaped network to obtain the aortic dissection prediction boundary of the historical aortic CTA image; Determining a loss function value between the predicted boundary of the aortic dissection and the actual boundary of the aortic dissection; Determine whether the loss function value meets a preset condition; the preset condition is that the maximum number of iterations is reached or the loss function value is less than a preset value; If yes, stop training and use the current dense hole Y-shaped network as the aortic dissection boundary segmentation model; If not, determining the gradient value of each parameter in the void dense Y-type network with respect to the loss function value; According to the gradient value, the parameters of the void dense Y-type network are updated using the L-BFGS algorithm, and the void dense Y-type network after the updated parameters is used as the void dense Y-type network. The process returns to the step of "inputting the historical aortic CTA image with the true boundary of aortic dissection marked into the void dense Y-type network to obtain the predicted boundary of aortic dissection in the historical aortic CTA image." 5. The method for determining the boundary of aortic dissection according to claim 4, characterized in that: The loss function value is the average of the focus loss function value and the improved Dice loss function value; the improved Dice loss function value is determined based on the improved Dice coefficient; the improved Dice coefficient is determined based on the binary classification prediction probability of each pixel point, the gold standard and the hyperparameters.

6. The method for determining the boundary of aortic dissection according to claim 5, characterized in that: The determination of the improved Dice loss function value specifically includes: Using the formula L Dice =1-Dice modified (p, g) Determine the improved Dice loss function value; wherein, Dice modified (p, g) is the improved Dice coefficient.

7. The method for determining the boundary of aortic dissection according to any one of claim 5 or claim 6, characterized in that: The determination of the improved Dice coefficient specifically includes: Using the formula Determine the improved Dice coefficient; wherein, p i is the binary classification prediction probability value of the i-th pixel; g i is the gold standard; k is a hyperparameter; N is the number of pixels in the preprocessed historical aortic CTA image.

8. A system for determining the boundary of aortic dissection, characterized in that: include: An image acquisition module is used to acquire an aortic CTA image; the aortic CTA image is a CTA image of the aortic dissection boundary to be determined; a preprocessing module, configured to preprocess the aortic CTA image to obtain a preprocessed aortic CTA image; wherein the preprocessed aortic CTA image is an enhanced aortic CTA image having a half envelope; The semi-envelope is the semi-envelope of the pulmonary artery to the aorta in the aortic CTA image; A boundary determination module is used to determine the aortic dissection boundary of the preprocessed aortic dissection image using an aortic dissection boundary segmentation model; wherein the aortic dissection boundary segmentation model is based on the L-BFGS algorithm and is obtained by training a cavity dense Y-type network using a first training data set; the first training data set includes historical aortic CTA images with marked true boundaries of aortic dissections; the cavity dense Y-type network includes a first encoder, a second encoder, a feature fusion module and a decoder; the first encoder and the second encoder are both connected to the feature fusion module; the first encoder, the second encoder and the feature fusion module are all connected to the decoder; the first encoder is The DenseNet network is trained using a second training data set; the second training data set is historical aortic CTA images obtained in the ImageNet data set; the second encoder is a cascade structure consisting of four improved DenseBlock layers and a dilated spatial pyramid pooling module, wherein the improved DenseBlock layer is a standard convolution layer in the DenseBlock layer followed by two dilated convolution layers connected in sequence; the decoder includes multiple improved upsampling modules; the improved upsampling module is obtained by replacing the second standard convolution layer of the upsampling module of the U-Net network with the improved DenseBlock layer.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the aortic dissection boundary determination method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for determining the boundary of an aortic dissection according to any one of claims 1 to 7 is implemented.

Images