Aortic dissection detection system based on non-enhanced CT images

By using the aortic segmentation network and conditional adversarial network to generate synthetic enhanced CT images in the non-enhanced CT image detection system, the problem of poor detection of aortic dissection in the prior art is solved, and higher detection accuracy and classification effect are achieved .

CN114723726BActive Publication Date: 2024-11-08GUANGZHOU MEDICAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210422922.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-21
Publication Date
2024-11-08
Estimated Expiration
2042-04-21

AI Technical Summary

Technical Problem

The existing aortic dissection detection system based on non-enhanced CT images has poor detection effect, which can easily lead to missed diagnosis or misdiagnosis.

Method used

The preprocessing module is used to obtain the original data, and the aortic vascular lumen in the non-enhanced CT images are segmented through the aortic segmentation network, and a multi-task structure is constructed using the conditional adversarial network and classifier to generate synthetic enhanced CT images to improve detection accuracy.

Benefits of technology

It effectively improves the detection accuracy of aortic dissection, reduces misdiagnosis and misdiagnosis, and improves the sensitivity and specificity of classification results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114723726B_ABST
    Figure CN114723726B_ABST
Patent Text Reader

Abstract

The present invention discloses an aortic dissection detection system based on non-enhanced CT images, including: a preprocessing module, used to obtain raw data, preprocess the raw data to obtain preprocessed data; an aortic segmentation module, used to train an aortic segmentation network according to the preprocessed data, and segment the aortic vascular cavity of the non-enhanced CT image; a multi-task structure construction module, used to construct a multi-task structure; a conditional adversarial network training module, used to use a generator to output a synthetic image according to the aortic vascular cavity, and input the synthetic image into a discriminator for training; an aortic dissection classification module, used to input the non-enhanced CT image and the synthetic enhanced CT image into a classifier to obtain the aortic dissection classification result of the synthetic enhanced CT image, and output the final aortic dissection classification result when the multi-task generation structure converges. The embodiment of the present invention can effectively improve the accuracy of aortic detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to an aortic dissection detection system based on non-enhanced CT images. Background Art

[0002] Aortic dissection (AD) is one of the most common aortic diseases. Its symptoms are that blood enters the aortic wall through the aortic intimal tear, causing the arterial wall to separate. Any delay or misdiagnosis may have serious consequences for patients with aortic dissection, and even lead to a high mortality rate. The incidence of aortic dissection is low, and non-cardiovascular physicians and primary care clinicians lack knowledge of aortic dissection, which can easily lead to misdiagnosis and missed diagnosis. Therefore, rapid and accurate detection of aortic dissection can save patients valuable time and provide effective help in the selection of clinical treatment options.

[0003] At present, the diagnosis of aortic dissection mainly relies on imaging examinations. Commonly used screening methods in clinical practice include transthoracic echocardiography (TTE), CT angiography (CTA), and magnetic resonance imaging (MRI). CTA has fast scanning speed, high temporal and spatial resolution, and powerful post-processing functions. It is currently the preferred non-invasive examination in clinical practice, especially in emergency situations. The enhanced CT image technology used can clearly display the true and false aortic cavities, and can fully evaluate the type, range, location of the rupture, and involvement of major branch vessels or abdominal organs in the dissection. However, due to the sensitization and nephrotoxicity of contrast agents, and the lack of 24-hour service available in many emergency departments, especially in rural areas or underserved areas with lack of technical and staff support, enhanced CT image technology is greatly limited. In contrast, non-enhanced CT image technology has a wider range of use, is not only convenient and easy to use, but also has a relatively low radiation dose. However, the existing aortic dissection detection system based on non-enhanced CT images is not effective in detecting aortic dissection, which can easily lead to missed diagnosis or misdiagnosis. Summary of the invention

[0004] The present invention provides an aortic dissection detection system based on non-enhanced CT images, so as to solve the technical problem that the existing aortic dissection detection system based on non-enhanced CT images has a poor detection effect on aortic dissection.

[0005] An embodiment of the present invention provides an aortic dissection detection system based on non-enhanced CT images, comprising:

[0006] A preprocessing module, used to obtain raw data and preprocess the raw data to obtain preprocessed data; the raw data includes a non-enhanced CT image and an enhanced CT image;

[0007] an aorta segmentation module, used for training an aorta segmentation network according to the preprocessed data, and using the aorta segmentation network to segment the aorta vascular cavity of the non-enhanced CT image;

[0008] A multi-task structure construction module, used to construct a multi-task structure according to a conditional adversarial network and a classifier, wherein the conditional adversarial network includes a generator and a discriminator;

[0009] A conditional adversarial network training module, used to use the generator to output a synthetic image according to the aortic vascular cavity, and input the synthetic image into the discriminator for training; the synthetic image includes a true cavity, a false cavity and a synthetic enhanced CT image;

[0010] The aortic dissection classification module is used to input the non-enhanced CT image and the synthetic enhanced CT image into a classifier to detect aortic dissection, obtain the aortic dissection classification result of the synthetic enhanced CT image, and output the final aortic dissection classification result when the multi-task generation structure converges.

[0011] Furthermore, when the multi-task generated structure converges, outputting the final aortic dissection classification result includes:

[0012] The total objective function of the multi-task generation structure is designed by combining the loss function of the generator, the loss function of the discriminator and the loss function of the classifier. When the total objective function converges, the final aortic dissection classification result is output.

[0013] Furthermore, the overall objective function is:

[0014] G * = arg min G max D L cGAN (G,D)+λ L1 L L1 (G)+λ C L bce (G,C)

[0015] Among them, G is the generator, G is the discriminator, C is the classifier, L bce (G, C) is the binary cross entropy function of classifier C, L L1 (G) is the L1 loss term of the discriminator G, L cGAN (G, D) is the loss function of the conditional generative adversarial network, λ L1 and λ C Both are high parameters.

[0016] Furthermore, the system also includes an evaluation module, which is used to:

[0017] The Dice similarity coefficient was used to evaluate the aorta segmentation network;

[0018] The synthetic enhanced CT images were evaluated using structural similarity, peak signal-to-noise ratio, mean absolute error, and root mean square error;

[0019] The Jaccard coefficient, under-segmentation rate, and over-segmentation rate were used to evaluate the segmentation effect of the true lumen and false lumen.

[0020] Sensitivity, specificity, precision, F1 score, accuracy, and area under the ROC curve were used to evaluate the classification performance of the classifier for aortic dissection.

[0021] Furthermore, the preprocessing module is used to:

[0022] Acquire non-enhanced CT images and enhanced CT images of multiple subjects with the same scanning position, coverage and parameters;

[0023] The non-enhanced CT image and the enhanced CT image are three-dimensionally rigidly registered, the sampling values ​​of the non-enhanced CT image and the enhanced CT image are averaged in resolution using the cubic spline interpolation method, a robust statistical segmentation method is used to draw the true segmentation masks of the aortic vascular lumen, true lumen and false lumen on the enhanced CT in a semi-automatic manner, and data augmentation is performed on the data sets of the non-enhanced CT image and the enhanced CT image with the true segmentation masks to obtain preprocessed data.

[0024] Furthermore, the conditional adversarial network training module is used to:

[0025] The aortic blood vessel cavity is used as a single-channel input of the generator, a synthetic image is output through a dual-channel, the synthetic image is input into a discriminator for training, the authenticity of the synthetic image is identified by the discriminator, and the training is stopped when a preset condition is met.

[0026] The aortic segmentation module of the embodiment of the present invention uses an aortic segmentation network to segment the aortic vascular lumen in the non-enhanced CT image, reduces the influence of the non-aortic region on the segmentation result during the segmentation process, and can segment the aortic vascular lumen into a true lumen and a false lumen, thereby effectively amplifying the contrast between the blood in the vascular lumen and the intimal sheet on the synthetic enhanced CT image, thereby effectively improving the accuracy of aortic dissection detection;

[0027] Furthermore, an embodiment of the present invention adopts a multi-task generation structure composed of a generator, a discriminator and a classifier to classify and detect aortic dissection, wherein the generator and the discriminator generate an adversarial network, and the discriminator repeatedly identifies the authenticity of samples in the generator to improve the performance of the generator and the discriminator, so as to output accurate classification results of aortic dissection and no aortic dissection. In addition, an embodiment of the present invention can also use non-enhanced CT images to synthesize enhanced CT images based on subtle differences in the X-ray attenuation coefficients between blood and the endothelial sheet in the vascular cavity to further improve the classification effect of aortic detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a structural schematic diagram of an aortic dissection detection system based on non-enhanced CT images provided by an embodiment of the present invention;

[0029] Figure 2 It is a principle block diagram of realizing aortic dissection detection provided by an embodiment of the present invention;

[0030] Figure 3 is a schematic diagram of the network structure of a generator provided by an embodiment of the present invention;

[0031] Figure 4 is a schematic diagram of a network structure of an identifier provided by an embodiment of the present invention;

[0032] Figure 5 is a schematic diagram of the network structure of a classifier provided by an embodiment of the present invention;

[0033] Figure 6 This is a rendering of a synthetic enhanced CT image provided by an embodiment of the present invention;

[0034] Figure 7 This is a segmentation effect diagram of the vascular cavity and the true and false cavities provided by an embodiment of the present invention;

[0035] Figure 8 is a ROC curve diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0037] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the feature. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0038] In the description of this application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0039] See also Figure 1-8 One embodiment of the present invention provides an aortic dissection detection system based on non-enhanced CT images as shown in 1, comprising:

[0040] The preprocessing module 10 is used to obtain raw data and preprocess the raw data to obtain preprocessed data; the raw data includes non-enhanced CT images and enhanced CT images;

[0041] an aorta segmentation module 20, for obtaining an aorta segmentation network by training the preprocessed data, and using the aorta segmentation network to segment the aorta vascular cavity of the non-enhanced CT image;

[0042] In an embodiment of the present invention, by using an aortic segmentation network to segment the true lumen and false lumen of the aortic vascular lumen in a non-enhanced CT image, a region of interest can be provided for NC2C transformation (Non-contrast to contrast CT transformation), true lumen and false lumen segmentation, and aortic dissection detection in a subsequent multi-task structure, thereby effectively improving the accuracy of aortic detection.

[0043] A multi-task structure construction module 30, used to construct a multi-task structure according to a conditional adversarial network and a classifier, wherein the conditional adversarial network includes a generator and a discriminator;

[0044] The conditional adversarial network training module 40 is used to use the generator to output a synthetic image according to the aortic vascular cavity, and input the synthetic image into the discriminator for training; the synthetic image includes a true cavity, a false cavity and a synthetic enhanced CT image;

[0045] The aortic dissection classification module 50 is used to input the non-enhanced CT image and the synthetic enhanced CT image into the classifier to detect aortic dissection, obtain the aortic dissection classification result of the synthetic enhanced CT image, and output the final aortic dissection classification result when the multi-task generation structure converges.

[0046] See also Figure 2 , which is a principle block diagram of implementing aortic dissection detection provided by an embodiment of the present invention.

[0047] The aortic segmentation module of the embodiment of the present invention uses an aortic segmentation network to segment the aortic vascular lumen in the non-enhanced CT image, reduces the influence of the non-aortic region on the segmentation result during the segmentation process, and can segment the aortic vascular lumen into a true lumen and a false lumen, thereby effectively amplifying the contrast between the blood in the vascular lumen and the intimal sheet on the synthetic enhanced CT image, thereby effectively improving the accuracy of aortic dissection detection;

[0048] Furthermore, an embodiment of the present invention adopts a multi-task generation structure composed of a generator, a discriminator and a classifier to classify and detect aortic dissection, wherein the generator and the discriminator generate an adversarial network, and the discriminator repeatedly identifies the authenticity of samples in the generator to improve the performance of the generator and the discriminator, so as to output accurate classification results of aortic dissection and no aortic dissection. In addition, an embodiment of the present invention can also use non-enhanced CT images to synthesize enhanced CT images based on subtle differences in the X-ray attenuation coefficients between blood and the endothelial sheet in the vascular cavity to further improve the classification effect of aortic detection.

[0049] In one embodiment, when the multi-task generated structure converges, the final aortic dissection classification result is output, including:

[0050] Combining the loss function of the generator, the loss function of the discriminator and the loss function of the classifier, the overall objective function of the multi-task generation structure is designed. When the overall objective function converges, the final aortic dissection classification result is output.

[0051] In an embodiment of the present invention, the conditional generative adversarial network composed of the generator and the discriminator can take the cascade of the image x and the noise vector z as input and convert it to the target image y, where y is the cascade of the enhanced CT image and the segmentation mask in the channel direction, and its loss function is:

[0052] L cGAN (G,D)=E x,y [log D(x,y)]+E x,z [log(1-D(x,G(x,z)))] (1)

[0053] Among them, the goal of G is to minimize the loss function in formula (1), and the goal of D is to maximize the loss function in formula (1). * = arg min G max D L cGAN (G,D).

[0054] The embodiment of the present invention uses the L1 loss term to ensure the similarity of the values ​​between G(x,z) and y and reduce the ambiguity of G(x,z):

[0055] L L1 (G) = E x,y,z [||yG(x,z)||1] (2)

[0056] The synthetic enhanced-CT is defined as G(x,z)1, and the binary cross entropy loss function of the classifier C is fed back to the generator G:

[0057]

[0058] In formula (3), n represents the number of subjects, y is the true category of the input image x, and y i =0 means no aortic dissection, y i =1 indicates the presence of aortic dissection, and C(·) represents the predicted value of classifier C.

[0059] In one embodiment, the overall objective function is:

[0060] G * = arg min G max D L cGAN (G,D)+λ L1 L L1 (G)+λ C L bce (G,C)

[0061] Among them, G is the generator, G is the discriminator, C is the classifier, L bce (G, C) is the binary cross entropy function of classifier C, L L1 (G) is the L1 loss term of the discriminator G, L cGAN (G, D) is the loss function of the conditional generative adversarial network, λ L1 and λ C Both are high parameters.

[0062] In an embodiment of the present invention, the overall objective function enables the generator, discriminator and classifier in the multi-task generation structure to be trained simultaneously to improve the efficiency of aortic dissection additional testing. Specifically, in an embodiment of the present invention, the classifier C is placed in parallel with the discriminator D, and the classifier C takes the synthetic CE-CT and non-enhanced CT as input. The classification error is back-propagated through the generator G and the classifier C, and the classifier C is placed at the output of the generator G and minimizes the binary cross entropy (L bce ) as the classification loss, and the Jensen-Shannon Divergence (JSD) between patients without and with aortic dissection was added.

[0063] In the multi-task generation mechanism of the embodiment of the present invention: the generator G uses L L1 loss to ensure the quality of synthesized CE-CT and segmentation mask; L bce The input of classifier C is composed of the original non-enhanced CT image and the synthesized enhanced CT image, and L bce The loss generates features of no or presence of aortic dissection in synthetic enhanced CT, which can improve the classification effect of the classification task.

[0064] In a specific implementation, the algorithm for generating a multi-task structure is as follows:

[0065]

[0066]

[0067] In one embodiment, the aortic dissection detection system further includes an evaluation module for:

[0068] The Dice similarity coefficient was used to evaluate the aorta segmentation network;

[0069] In the embodiment of the present invention, the Dice similarity coefficient is:

[0070]

[0071] In formula (5), TP is the true positive, which is predicted as the target area and is also the actual target detection area; TN is the true negative, which is predicted as the non-target area and is actually the non-target detection area; FP is the false positive, which is predicted as the target area and is actually the non-target detection area; FN is the false negative, which is predicted as the non-target area and is actually the target detection area.

[0072] The synthetic enhanced CT images were evaluated using structural similarity, peak signal-to-noise ratio, mean absolute error, and root mean square error;

[0073] In the embodiment of the present invention, the expressions of structural similarity, peak signal-to-noise ratio, mean absolute error and root mean square error are as follows:

[0074]

[0075]

[0076]

[0077]

[0078] In formula (6), μ v Represents the average value of the intensity value of the real enhanced CT image, μ v Represents the average value of the intensity value of the synthesized enhanced CT image. v Represents the standard deviation of the intensity value of the real enhanced CT image, Represents the standard deviation of the intensity values ​​of the synthetic enhanced CT image. is the covariance between the real CE-CT and the synthetic enhanced CT image. C1 and C2 are both constants, which are 0.01 respectively. 2 and 0.03 2 , n is the number of voxels in the 3D enhanced CT image. In formulas (7), (8) and (9), v i represents the intensity value of the ith voxel in the real CE-CT, represents the intensity value of the i-th voxel in the synthetic CE-CT.

[0079] Please refer to Table 1, which is a performance evaluation table of the synthetic enhanced CT image provided in an embodiment of the present invention.

[0080] Table 1: Performance evaluation table of synthetic enhanced CT images

[0081]

[0082] In the embodiment of the present invention, according to the image quality rating indicators in Table 1, when the non-enhanced CT is synthesized into the enhanced CT, the SSIM of the embodiment of the present invention reaches 0.982, the PSNR reaches 30.999, and the MAE and RMSE values ​​are small, which proves that the enhanced CT synthesized by the embodiment of the present invention is close to the actual situation.

[0083] The Jaccard coefficient, under-segmentation rate, and over-segmentation rate were used to evaluate the segmentation effect of the true lumen and false lumen.

[0084] In the embodiment of the present invention, the expressions of the Jaccard coefficient, the under-segmentation rate and the over-segmentation rate are as follows:

[0085]

[0086]

[0087]

[0088] Among them, in formulas (10), (11) and (12), TP is the true positive (True Positive), which is predicted as the target area and is the actual target detection area; TN is the true negative (True Negative), which is predicted as the non-target area and is the actual non-target detection area; FP is the false positive (False Positive), which is predicted as the target area and is the actual non-target detection area; FN is the false negative (False Negative), which is predicted as the non-target area and is the actual target detection area.

[0089] See also Figure 6 , which is an effect diagram of the synthetic enhanced CT image provided by an embodiment of the present invention.

[0090] Please refer to Table 2, which is a segmentation performance evaluation table of the vascular lumen and the true and false lumens provided in an embodiment of the present invention.

[0091] Table 2: Segmentation performance evaluation table of vascular lumen and true and false lumens

[0092]

[0093] See also Figure 7 , which is a diagram showing the segmentation effect of the vascular cavity and the true and false cavities provided in an embodiment of the present invention.

[0094] Sensitivity, specificity, precision, F1 score, accuracy, and area under the ROC curve were used to evaluate the classification performance of the classifier for aortic dissection.

[0095] In the embodiment of the present invention, the expressions of sensitivity, specificity, precision, F1 score, accuracy and area under the ROC curve are as follows:

[0096]

[0097]

[0098]

[0099]

[0100]

[0101] In the above formula, TP is a true positive (True Positive), which is predicted to be aortic dissection, and in fact it is aortic dissection; TN is a true negative (True Negative), which is predicted to be no aortic dissection, and in fact it is no aortic dissection. FP is a false positive (False Positive), which is judged to be aortic dissection, but in fact it is a negative sample without aortic dissection; FN is a false negative (False Negative), which is judged to be no aortic dissection, but in fact it is aortic dissection.

[0102] Please refer to Table 3, which is a performance comparison table of radiologists provided in an embodiment of the present invention and aortic dissection detection in an embodiment of the present invention.

[0103] Table 3: Performance comparison of aortic dissection detection by radiologists and embodiments of the present invention

[0104]

[0105] See also Figure 8 , which is a ROC curve diagram provided by an embodiment of the present invention, wherein the AUC is 0.824.

[0106] In the embodiment of the present invention, it can be seen from Table 3 that the Acc, Sens, Spec, precision and F1 of the aortic dissection detection in the embodiment of the present invention are 0.844, 0.938, 0.750, 0.789 and 0.857, respectively. The above-mentioned multiple indicators reflect that the embodiment of the present invention has a good effect on the detection of aortic dissection in non-enhanced CT images, and has high sensitivity, which can effectively reduce missed diagnosis and misdiagnosis.

[0107] In one embodiment, the preprocessing module 10 is used to:

[0108] Acquire non-enhanced CT images and enhanced CT images of multiple subjects with the same scanning position, coverage and parameters;

[0109] The non-enhanced CT images and enhanced CT images were three-dimensionally rigidly registered, and the sampling values ​​of the non-enhanced CT images and enhanced CT images were averaged with the cubic spline interpolation method. The robust statistical segmentation method was used to semi-automatically draw the true segmentation masks of the aortic lumen, true lumen and false lumen on the enhanced CT. The data sets of the non-enhanced CT images and enhanced CT images with the true segmentation masks were augmented to obtain the preprocessed data.

[0110] In an embodiment of the present invention, the collected raw data comes from different medical centers. Specifically, the raw data of the embodiment of the present invention comes from 207 subjects, including 78 patients with aortic dissection and 129 patients without aortic dissection but suffering from other cardiovascular diseases. Non-enhanced CT technology and enhanced CT technology are sequentially performed on each subject with the same scanning position, coverage and parameters. In order to reduce motion artifacts and misregistration between non-enhanced CT and enhanced-CT, the embodiment of the present invention uses ECG gating triggering at the end of respiration to collect data, and the patient is asked to hold his breath. Paired non-enhanced CT images and enhanced CT images have the same image size: 512×512×(385~1224), and the same resolution, which ranges from 0.549×0.549×0.625 mm 3 To 0.977×0.977×1.250mm 3 .

[0111] In a specific embodiment, the embodiment of the present invention uses the Elastix tool to perform three-dimensional rigid registration on the non-enhanced CT image and the enhanced CT image, and then uses the cubic spline interpolation method to sample the non-enhanced CT image and the enhanced CT image to an average resolution of: 0.7444×0.7444×1.031mm 3 , and their image size was cropped to 256 × 256 × 512. Finally, the robust statistical segmentation method in the open source software 3D Slicer was used to draw the true segmentation mask of the aortic lumen, true lumen and false lumen on the enhanced CT images in a semi-automatic manner.

[0112] It should be noted that, since the CT value of the metal stent in some cases is much higher than that of the aortic tissue, the CT value threshold intervals of the non-enhanced CT image and the enhanced CT image are truncated to [0,200] and [0,800], respectively. The preprocessed data are obtained by performing data augmentation on the non-enhanced CT image and enhanced CT image datasets with the true segmentation mask by randomly rotating (-10,10) degrees along the (x,y), (y,z) and (x,z) coordinate planes. In the embodiment of the present invention, the metal stent refers to the stent graft visible around the true lumen of the patient after surgery.

[0113] The embodiment of the present invention trains the aortic segmentation network and the multi-task generation structure based on the preprocessed data. In order to train the aortic segmentation network, the embodiment of the present invention randomly selects the non-enhanced CT images and the real aortic segmentation masks of 51 subjects without aortic dissection, and then uses the trained three-dimensional aortic segmentation network (Pre-trained3D U-Net) to segment the aorta of the non-enhanced CT images of the remaining 156 subjects (78 patients with aortic dissection and 78 patients without aortic dissection). The non-enhanced CT images, enhanced CT images, and the corresponding segmentation masks of the aortic vascular lumen, true and false lumens of the remaining 156 subjects are used to train and test the multi-task generation structure (MTGA). The embodiment of the present invention adopts 5 cross-validations, each of which uses 80% as the training samples of the model and 20% as the test samples of the model.

[0114] In one embodiment, the conditional adversarial network training module 40 is used to:

[0115] The aortic lumen is used as the single-channel input of the generator, and the synthetic image is output through the dual channels. The synthetic image is input into the discriminator for training. The authenticity of the synthetic image is identified by the discriminator, and the training is stopped when the preset conditions are met.

[0116] In an embodiment of the present invention, the conditional adversarial network training module includes a generator and a discriminator, wherein the generator is used to take the aortic vascular cavity of the non-enhanced CT image as a single-channel input and synthesize a dual-channel output for the aorta, specifically, the output is a synthesized enhanced CT image: and the segmentation mask The cascade in the axial direction, i.e. Ideally, the synthesized output That is, the synthesized image is considered to be indistinguishable from the corresponding real Y:Y1⊕Y2. The discriminator D determines the difference between the real Y and the synthesized Y at the 3D patch level (70×70×70). Whether it is real or forged, all patch responses are averaged to provide the final output of the discriminator D. In particular, the aortic vessel lumen of the input non-enhanced CT image and the synthetic The cascaded Y forms a fake sample (Fakes), and the real Y forms a real sample (Reals). In the embodiment of the present invention, the real sample and the fake sample are input into the discriminator D to output the corresponding two matrices (D F and D R ) to predict the authenticity of the input sample. In the embodiment of the present invention, the condition for stopping the training of the conditional adversarial network can be when a preset number of iterations is reached, or when the result of the discriminator identification reaches a certain accuracy.

[0117] See also Figure 3 , is a schematic diagram of the network structure of the generator provided by an embodiment of the present invention. The dimension of the generator is 256×256×512, and the input data with a channel number of 1 first passes through the encoding path of the generator, using Convolution-BatchNorm-ReLU (convolution-batch normalization-ReLU activation layer). Batch normalization is not used in the first layer of the encoding path. All ReLU units in the encoder are LeakeyReLU with a slope of 0.2. All convolutions in the encoding path and deconvolutions in the decoding path are three-dimensional spatial filters of size 4×4×4, and the step size of the three directions (x, y, z) is 2. The last layer of the decoding path maps its input to a two-channel output, followed by a Tanh nonlinear set function.

[0118] Please continue reading Figure 3 , Figure 3 The architecture shown is a symmetrical structure with an encoding path and a decoding path, and the two paths include a convolution block and a deconvolution block respectively. In the embodiment of the present invention, skip connections are used between the convolution block and the deconvolution block to provide information from the encoding path for image synthesis.

[0119] In the embodiment of the present invention, high-level features of non-enhanced CT images can be extracted by downsampling, and then two upsamplings are performed, one of which outputs the segmentation results of the true and false cavities, and the other upsampling outputs the synthetic enhanced CT image. The loss function of the segmentation task in the embodiment of the present invention can replace the L1 loss with the Dice loss.

[0120] See also Figure 4 , Figure 4 The discriminator structure shown includes 4 convolution blocks and an additional convolution layer with a convolution kernel of 1×1×1. Specifically, each convolution block includes a convolution kernel of 2×2×2, a batch normalization layer, and a LeakyRelu activation function. In an embodiment of the present invention, the specific training process of the discriminator D is as follows: the cascade of the enhanced CT image and the segmentation mask generated by the generator G, and the cascade of the real enhanced CT image and the mask manually outlined by the doctor are sequentially input into four convolution blocks, and the last additional convolution layer with a convolution kernel of 1×1×1 is responsible for extracting features and comparing them, judging and synthesizing the authenticity of the cascade, and forcing the synthesized cascade to be close to the corresponding real cascade, thereby further constraining the generator G to reduce the possibility of erroneous synthesis, which is conducive to improving the accuracy of the generator synthesized image and improving the accuracy of aortic dissection detection.

[0121] See also Figure 5, is the network structure of the classifier provided in the embodiment of the present invention. In the embodiment of the present invention, the synthesized enhanced CT image is input into the classifier C, and the classifier classifies the non-enhanced CT image (X) and the synthesized enhanced CT image Cascade As input, the output is the classification result of whether there is aortic dissection, that is, AD / non-AD. It should be noted that It is the segmentation result of the true and false cavities, and is only used to assist in the generation of and are not input into the classifier.

[0122] In an embodiment of the present invention, the classifier used to detect aortic dissection is a lightweight CNN. It includes 4 convolution blocks, 4 pooling layers, 2 fully connected layers and 1 Softmax layer, which are used to classify the category subscript 1 or 0, where 1 indicates the presence of aortic dissection and 0 indicates the absence of aortic dissection. In the embodiment of the present invention, each convolution block includes a 2×2×2 convolution kernel, a batch normalization layer and a LeakyReLU activation function. In this embodiment, the specific training process of the classifier C is as follows: the enhanced CT generated by the generator G and the original non-enhanced CT are sequentially input into four convolution blocks and the following maximum pooling layer, features are extracted through two fully connected layers, and the predicted probability of each category is output through the final softmax layer, forcing the prediction result to be close to the corresponding true classification result, thereby effectively improving the accuracy of the classifier in accurately classifying aortic dissection.

[0123] The implementation of the embodiments of the present invention has the following beneficial effects:

[0124] The aortic segmentation module of the embodiment of the present invention uses an aortic segmentation network to segment the aortic vascular lumen in the non-enhanced CT image, reduces the influence of the non-aortic region on the segmentation result during the segmentation process, and can segment the aortic vascular lumen into a true lumen and a false lumen, thereby effectively amplifying the contrast between the blood in the vascular lumen and the intimal sheet on the synthetic enhanced CT image, thereby effectively improving the accuracy of aortic dissection detection;

[0125] Furthermore, an embodiment of the present invention adopts a multi-task generation structure composed of a generator, a discriminator and a classifier to classify and detect aortic dissection, wherein the generator and the discriminator generate an adversarial network, and the discriminator repeatedly identifies the authenticity of samples in the generator to improve the performance of the generator and the discriminator, so as to output accurate classification results of aortic dissection and no aortic dissection. In addition, an embodiment of the present invention can also use non-enhanced CT images to synthesize enhanced CT images based on subtle differences in the X-ray attenuation coefficients between blood and the endothelial sheet in the vascular cavity to further improve the classification effect of aortic detection.

[0126] The above are preferred embodiments of the present invention. It should be noted that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A system for detecting aortic dissection based on non-enhanced CT images. It is characterized in that include: A preprocessing module, used to obtain raw data and preprocess the raw data to obtain preprocessed data; the raw data includes a non-enhanced CT image and an enhanced CT image; an aorta segmentation module, used for training an aorta segmentation network according to the preprocessed data, and using the aorta segmentation network to segment the aorta vascular cavity of the non-enhanced CT image; A multi-task structure construction module, used to construct a multi-task structure according to a conditional adversarial network and a classifier, wherein the conditional adversarial network includes a generator and a discriminator; A conditional adversarial network training module, used to use the generator to output a synthetic image according to the aortic vascular cavity, and input the synthetic image into the discriminator for training; the synthetic image includes a true cavity, a false cavity and a synthetic enhanced CT image; an aortic dissection classification module, for inputting the non-enhanced CT image and the synthetic enhanced CT image into a classifier to detect aortic dissection, obtaining an aortic dissection classification result of the synthetic enhanced CT image, and outputting a final aortic dissection classification result when the multi-task generation structure converges; When the multi-task generation structure converges, outputting the final aortic dissection classification result includes: combining the loss function of the generator, the loss function of the discriminator and the loss function of the classifier, designing the total objective function of the multi-task generation structure, and when the total objective function converges, outputting the final aortic dissection classification result; the total objective function is: Among them, G is the generator, D is the discriminator, and C is the classifier. is the binary cross entropy function of classifier C, is the discriminator G L 1 Loss item, is the loss function for conditional generative adversarial networks, and Both are high parameters.

2. The aortic dissection detection system based on non-enhanced CT images as claimed in claim 1, It is characterized in that After outputting the final aortic dissection classification results, an evaluation module is also included for: The Dice similarity coefficient was used to evaluate the aorta segmentation network; The synthetic enhanced CT images were evaluated using structural similarity, peak signal-to-noise ratio, mean absolute error, and root mean square error; The Jaccard coefficient, under-segmentation rate, and over-segmentation rate were used to evaluate the segmentation effect of the true lumen and false lumen. Sensitivity, specificity, precision, F1 score, accuracy, and area under the ROC curve were used to evaluate the classification performance of the classifier for aortic dissection.

3. The aortic dissection detection system based on non-enhanced CT images as claimed in claim 1, It is characterized in that The preprocessing module is used to: Acquire non-enhanced CT images and enhanced CT images of multiple subjects with the same scanning position, coverage and parameters; The non-enhanced CT image and the enhanced CT image are three-dimensionally rigidly registered, the sampling values ​​of the non-enhanced CT image and the enhanced CT image are averaged in resolution using the cubic spline interpolation method, a robust statistical segmentation method is used to draw the true segmentation masks of the aortic vascular lumen, true lumen and false lumen on the enhanced CT in a semi-automatic manner, and data augmentation is performed on the data sets of the non-enhanced CT image and the enhanced CT image with the true segmentation masks to obtain preprocessed data.

4. The aortic dissection detection system based on non-enhanced CT images as claimed in claim 1, It is characterized in that The conditional adversarial network training module is used to: The aortic blood vessel cavity is used as a single-channel input of the generator, a synthetic image is output through a dual-channel, the synthetic image is input into a discriminator for training, the authenticity of the synthetic image is identified by the discriminator, and the training is stopped when a preset condition is met.