Systems and methods for segmenting ascending aorta and coronary arteries from CCTA using hybrid approach

Abstract

The present application provides a system and method for segmenting the ascending aorta and coronary arteries from coronary CT angiography (CCTA) using a hybrid approach, which relates to a technique for extracting only the shape of the coronary arteries and the ascending aorta from an input coronary CT medical image.

Description

[0001] Cross-references to related applications

[0002] This application claims priority to Korean Patent Application No. 10-2022-0027282, filed on March 3, 2022, with the Korean Intellectual Property Office, the disclosure of which is incorporated herein by reference in its entirety. Technical Field

[0003] The following disclosure relates to a system and method for segmenting the ascending aorta and coronary arteries from coronary CT angiography (CCTA) using a hybrid approach, and more particularly to a system and method for segmenting the ascending aorta and coronary arteries from CCTA using a hybrid approach, which is capable of accurately segmenting only the shape of the coronary arteries and ascending aorta for morphological and hemodynamic analysis of coronary artery diseases. Background Technology

[0004] According to a report by Statistics Korea, heart disease was the second leading cause of death in 2019. Furthermore, heart disease was reported as the leading cause of death globally in 2019 by the World Health Organization (WHO). Consequently, medical expenses related to heart disease are gradually increasing. According to the "Health Insurance Statistics" released by the Health Insurance Review and Evaluation Organization, since 2015, medical expenses in South Korea due to circulatory system diseases have increased by an average of 8.4% annually, reaching approximately 10 trillion won and 500 billion won respectively in 2019.

[0005] For this reason, the need for early and accurate diagnosis of heart disease has naturally become more apparent.

[0006] In clinical practice, medical images such as coronary angiography, coronary CT imaging, and cardiac magnetic resonance imaging (MRI) have been used for diagnosis, especially non-invasive, high-resolution coronary CT imaging, which can be scanned at low cost, and is now the most widely used.

[0007] For morphological diagnosis of blood vessels, it is necessary to quantitatively analyze the degree of stenosis or the severity of the lesion. This requires accurate extraction of lesions from medical images. Furthermore, by performing hydrodynamic modeling on the extracted vessels, diagnosis can be made while taking flow factors into account.

[0008] In particular, fractional flow reserve (FFR), which is widely used in the diagnosis of coronary artery disease, is a flow factor that represents a decrease in pressure and can be noninvasively derived through modeling.

[0009] To enable morphological diagnosis of blood vessels, numerous studies in the field have recognized the necessity of segmenting the ascending aorta and coronary arteries, and rule-based (RB) or deep-learning (DL) methods have been employed.

[0010] As their names suggest, the RB method performs algorithms according to a pre-established rule sequence, while the DL method derives results through a prediction model learned by artificial intelligence.

[0011] However, since the RB method and the DL method each have their own advantages and disadvantages, it is impossible to determine which method is better.

[0012] Specifically, since the RB method is not a learning-based method, it does not require a large amount of labeled data, but it is relatively slow to process and it is difficult to set fixed hyperparameters in the algorithm because the image quality and shape of blood vessels vary depending on the equipment used for image scanning, the scanning settings, and the patient.

[0013] Meanwhile, deep learning (DL) methods using deep learning models can make more accurate and faster predictions, but require a large amount of labeled data to create the learning model. However, for coronary artery CT images, it is difficult to obtain a large amount of labeled data, especially for coronary arteries, due to the lack of standard label databases. The complex shape of the coronary arteries makes the labeling process itself time-consuming and requires in-depth anatomical knowledge.

[0014] Given the advantages and disadvantages mentioned above, the RB method is considered more suitable for coronary artery segmentation.

[0015] Because the ascending aorta has a simpler shape and larger volume than the coronary arteries, it is relatively easy to label, and model learning does not require overfitting. Considering this, applying deep learning (DL) methods can predict results faster and more accurately in ascending aorta extraction. However, related techniques inevitably reduce accuracy by selecting either DL or redox methods simultaneously for extracting both the ascending aorta and coronary arteries, rather than considering both methods concurrently.

[0016] In this regard, Korean Patent Application No. 10-1793499 (“Aortic Extraction Method Using Geometric Information from Z-Axis Images”) discloses a technique for effectively extracting the aortic location based on geometric information from Z-axis images obtained from three-dimensional images of the heart.

[0017] [Related Technical Documents]

[0018] [Patent Documents]

[0019] Korean Patent Registration No. 10-1793499 (Registration Date: October 30, 2017) Summary of the Invention

[0020] An exemplary embodiment of the present invention aims to provide a system and method for segmenting the ascending aorta and coronary arteries from coronary CT angiography (CCTA) using a hybrid approach. This system and method, by combining the advantages of rule-based (BP) methods and deep learning (DL) methods, is able to rationally, quickly and accurately segment only the shape of the coronary arteries and ascending aorta from input coronary CT medical images.

[0021] In one general aspect, a system for segmenting the ascending aorta and coronary arteries from coronary CT angiography (CCTA) using a hybrid approach includes: an image preprocessing unit 100 that receives a 2D coronary CT image dataset obtained by coronary CT angiography (CCTA) and performs image preprocessing to extract the shapes of the ascending aorta and coronary arteries; a first processing unit 200 that receives the preprocessed image dataset from the image preprocessing unit 100 via a pre-stored deep learning model and segments the ascending aorta image region; a second processing unit 300 that receives the preprocessed image dataset from the image preprocessing unit 100 based on pre-stored rules and segments the coronary artery image region using the ascending aorta image region segmented by the first processing unit 200; and a combining unit 400 that overlaps the ascending aorta image region and the coronary artery image region to obtain a combined structural image.

[0022] The image preprocessing unit 100 may include: a first brightness adjustment unit 110, which adjusts the maximum and minimum brightness values ​​of each image data based on a preset first brightness value range; a voxel transformation unit 120, which isotropically processes the size of a unit voxel for each image data whose brightness values ​​have been adjusted by the first brightness adjustment unit 110 according to a predetermined reference; a noise processing unit 130, which removes noise regions from each image data that has been isotropically processed by the voxel transformation unit 120 by applying a morphological closure operator; and a second brightness adjustment unit 140, which adjusts the maximum and minimum brightness values ​​of each image data for which noise has been removed by the noise processing unit 130 based on a preset second brightness value range.

[0023] The noise processing unit 130 may include: a threshold processing unit 131, which analyzes the pixels contained in each image data based on brightness values ​​below a specific threshold and assigns a predetermined value; an operator application unit 132, which processes each image data by applying a morphological closure operator; a mask generation unit 133, which generates a mask by using the predetermined value provided by the threshold processing unit 131; and a mask processing unit 134, which masks the image data processed by the operator application unit 132 using the mask generated by the mask generation unit 133 to remove noise regions.

[0024] The first processing unit 200 may include: a deep learning segmentation unit 210, which inputs the image data set by the second brightness adjustment unit 140 into a deep learning model for training the ascending aorta segmentation and receives the image region of the ascending aorta; and a segmentation post-processing unit 220, which analyzes the output results from the deep learning segmentation unit 210 and performs noise removal.

[0025] The segmentation post-processing unit 220 may include: a filter processing unit 221, which applies a preset filter to the output ascending aorta image region to generate an image that emphasizes the boundaries of the included structures; and a transformation processing unit 222, which detects circular structures included in the image generated by the filter processing unit 221 by applying a preset transformation technique, wherein the circular structures detected by the transformation processing unit 222 are set as the ascending aorta image region.

[0026] The segmentation post-processing unit 220 may further include: a region-of-interest (ROI) processing unit 223, which processes the image from the filter processing unit 221 to set the ROI; and an aortic setting unit 224, which analyzes the circular structure detected by the transformation processing unit 222, the circular structure being contained within the ROI set by the ROI processing unit 223, and sets the final ascending aortic image region in the detected circular structure by means of a predetermined reference.

[0027] The second processing unit 300 may include: a blood vessel calculation unit 310, which calculates the blood vessel values ​​of each pixel constituting each image data preprocessed by the image preprocessing unit 100 by using a Hessian-based blood vessel filter; a blood vessel correction unit 320, which uses the blood vessel values ​​from the blood vessel calculation unit 310 to adjust the maximum and minimum blood vessel values ​​based on a preset blood vessel value range, and provides binarized data to each pixel using the adjusted blood vessel values; a coronary artery extraction unit 330, which uses the final ascending aorta image region set by the aorta setting unit 224 to extract a structural image region connected to the final ascending aorta image region in the structural image obtained by the binarized data provided by the blood vessel correction unit 320; and a coronary artery setting unit 340, which uses the blood vessel values ​​from the blood vessel calculation unit 310 to remove noise pixels contained in the structural image region extracted by the coronary artery extraction unit 330, and sets it as the final coronary artery image region.

[0028] The combining unit 400 can overlap the final ascending aortic image region from the aortic setting unit 224 and the final coronary artery image region from the coronary artery setting unit 340 to obtain a structural region based on the combination of pixel values ​​at the same location.

[0029] In another general aspect, a method for segmenting the ascending aorta and coronary arteries from coronary CT angiography (CCTA) using a hybrid approach, the hybrid approach using segmentation systems for the ascending aorta and coronary arteries from CCTA, wherein each operation is performed by a computational processing unit including a computer, the segmentation method comprising: an image input operation S100, wherein an image preprocessing unit receives 2D images obtained from coronary CT angiography. CCTA image dataset; preprocessing operation S200, wherein the image preprocessing unit preprocesses the image dataset based on the image input operation S100; first processing operation S300, wherein the first processing unit uses a pre-stored deep learning model to segment the ascending aorta image region from the image dataset preprocessed by the preprocessing operation S200; second processing operation S400, wherein the second processing unit analyzes the image dataset preprocessed in the preprocessing operation S200 and uses the analysis results and the ascending aorta image region segmented in the first processing operation S300 to segment the coronary artery image region; and combination operation S500, wherein the combination unit overlaps the ascending aorta image region based on the first processing operation S300 and the coronary artery image region based on the second processing operation S400.

[0030] The preprocessing operation S200 may include: a first brightness adjustment operation S210, which adjusts the maximum and minimum brightness values ​​of each image data constituting the image dataset based on a preset first brightness value range; a voxel processing operation S220, which isotropically processes the size of unit voxels for each image data whose brightness values ​​were adjusted according to a predetermined reference in the first brightness adjustment operation S210; a noise processing operation S230, which removes noise regions from each image data based on the voxel processing operation S220 by applying a morphological closure operator; and a second brightness adjustment operation S240, which adjusts the maximum and minimum brightness values ​​of each image data that has had noise removed in the noise processing operation S230 based on a preset second brightness value range.

[0031] The noise processing operation S230 may include: a thresholding operation S231, which analyzes the pixels contained in each image data based on brightness values ​​below a specific threshold and assigns a predetermined value; an operator application operation S232, which processes each image data by applying a morphological closure operator; a mask generation operation S233, which generates a mask using the predetermined value provided in the thresholding operation S231; and a mask processing operation S234, which performs masking processing on the image data processed in the operator application operation S232 using the mask in the mask generation operation S233 to remove noise regions.

[0032] The first processing operation S300 may include: a deep learning segmentation operation S310, which inputs the image dataset based on the second brightness adjustment operation S240 into a learned deep learning model for ascending aorta segmentation and receives the ascending aorta image region; and a segmentation post-processing operation S320, which performs noise removal by analyzing the output results based on the deep learning segmentation operation S310.

[0033] The post-segmentation processing operation S320 may include: a filter processing operation S321, which applies a preset filter to the output ascending aorta image region to generate an image that emphasizes (highlights) the boundary region of the contained structures; and a transformation processing operation S322, which detects circular structures contained in the image based on the filter processing operation S321 by applying a preset transformation technique, wherein the circular structures detected in the transformation processing operation S322 are set as the ascending aorta image region.

[0034] The post-segmentation processing operation S320 may further include: a region-of-interest (ROI) processing operation S323, which sets the ROI by processing the image based on the filter processing operation S321; and an aortic setting operation S324, which sets the final ascending aortic image region in the circular structure detected by the transformation processing operation S322, the circular structure being contained in the ROI set by the ROI processing operation S323, by analyzing the circular structure detected by the transformation processing operation S322.

[0035] The second processing operation S400 may include: a blood vessel calculation operation S410, which calculates the blood vessel values ​​of each pixel constituting the image data preprocessed by the preprocessing operation S200 by using a Hessian-based blood vessel filter; a blood vessel correction operation S420, which uses the blood vessel values ​​based on the blood vessel calculation operation S410 to adjust the maximum and minimum blood vessel values ​​based on a preset blood vessel value range, and provides binarized data to each pixel using the adjusted blood vessel values; and a coronary artery extraction operation S430. Operation S430 uses the final ascending aorta image region set by the aorta setting operation S324 to extract a structural image region connected to the final ascending aorta image region in the structural image obtained from the binarized data given by the vascular correction operation S420; and coronary artery setting operation S440 sets the final coronary artery image region after removing noise pixels contained in the structural image region extracted by the coronary artery extraction operation S430 by using the vascular values ​​based on the vascular calculation operation S410.

[0036] The combination operation S500 can overlap the final ascending aortic image region based on the aortic setting operation S324 and the final coronary artery image region from the coronary artery setting operation S440 to obtain a structural region based on the combination of pixel values ​​at the same location.

[0037] Other features and aspects will become apparent from the following detailed description, accompanying drawings, and scope of protection. Attached Figure Description

[0038] Figure 1 An exemplary configuration diagram is provided to illustrate a system for segmenting the ascending aorta and coronary arteries from a CCTA using a hybrid method according to an exemplary embodiment of the present invention;

[0039] Figure 2-9 The diagram illustrates the image processing of a system and method for segmenting the ascending aorta and coronary arteries from CCTA using a hybrid approach according to exemplary embodiments of the present invention, in various configurations.

[0040] Figure 10 A flowchart illustrating a method for segmenting the ascending aorta and coronary arteries from a CCTA using a hybrid approach according to an exemplary embodiment of the present invention. Detailed Implementation

[0041] In the following, the system and method of the present invention for segmenting the ascending aorta and coronary arteries from CCTA using a hybrid approach will be described in detail with reference to the accompanying drawings. Exemplary embodiments of the invention described below are provided by way of example so that the spirit of the invention is fully conveyed to those skilled in the art to which this invention pertains. Therefore, the scope of the invention is not limited to the following description and drawings, and may be implemented in other forms. Furthermore, throughout the specification, the same reference numerals denote the same parts.

[0042] Unless otherwise stated, the terms used in this specification (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains, and detailed descriptions of known functions and constructions that may obscure the essential points of the invention will be omitted.

[0043] In addition, a system refers to a set of components, including equipment, mechanisms, units, etc., which are organized and interact regularly to perform the required functions.

[0044] In a broader sense, according to exemplary embodiments of the present invention, the system and method for segmenting the ascending aorta and coronary arteries from CCTA using a hybrid approach aims to extract only the shape of the ascending aorta and coronary arteries from coronary CT medical images. Using the extracted shapes of the coronary arteries and ascending aorta, the degree of cardiovascular stenosis can be quantified, and subsequently, coronary artery disease can be diagnosed using 3D reconstructed vessel shapes through hemodynamic modeling. That is, extracting the shape of the ascending aorta and coronary arteries is a technique prior to morphological and hemodynamic analysis of coronary artery disease.

[0045] In existing technologies, either the RB method or the DL method is selected for extraction, or the operator performs the extraction manually, which is time-consuming and has a limitation: the operator's level of anatomical knowledge and subjectivity may inevitably be reflected.

[0046] Therefore, according to an exemplary embodiment of the present invention, the method for segmenting the ascending aorta and coronary artery system from CCTA using a hybrid approach is a hybrid method that utilizes the advantages of both the RB method and the DL method, wherein, in terms of morphological features, the RB method is applied to coronary artery extraction and the DL method is applied to ascending aorta extraction, so that results (extraction of the ascending aorta and coronary arteries) can be obtained reasonably, quickly and more accurately.

[0047] Figure 1An exemplary configuration diagram is provided to illustrate a system for segmenting the ascending aorta and coronary arteries from a CCTA using a hybrid method according to an exemplary embodiment of the present invention. Reference will be made to... Figure 1 A system for segmenting the ascending aorta and coronary arteries from CCTA using a hybrid method according to an exemplary embodiment of the present invention is described in detail.

[0048] like Figure 1 As shown, a system for segmenting the ascending aorta and coronary arteries from CCTA using a hybrid method according to an exemplary embodiment of the present invention may include an image preprocessing unit 100, a first processing unit 200, a second processing unit 300, and a combination unit 400, and each component may be included in an arithmetic processing unit or multiple arithmetic processing units including a computer.

[0049] Each component will be described in detail.

[0050] The image preprocessing unit 100 preferably receives coronary CT medical images obtained by coronary CT angiography (CCTA) and performs image preprocessing to extract the shape of the ascending aorta and coronary arteries.

[0051] The coronary CT medical images input from the image preprocessing unit 100 can refer to the two-dimensional (2D) coronary CT image dataset obtained through coronary CT angiography, and three-dimensional (3D) coronary CT images can be obtained by overlaying the images.

[0052] like Figure 1 As shown, the image preprocessing process of the image preprocessing unit 100 is executed by the first brightness adjustment unit 110, the voxel transformation unit 120, the noise processing unit 130, and the second brightness adjustment unit 140.

[0053] Figure 2 An example diagram illustrating the image preprocessing process through the image preprocessing unit 100 is provided. Specifically, Figure 2 a) represents an image data selected from the 2D coronary CT image dataset input to the image preprocessing unit 100.

[0054] The first brightness adjustment unit 110 performs brightness cropping of the image data, such as... Figure 2 As shown in b).

[0055] Specifically, many patients have calcified areas in their coronary arteries. Since the brightness of calcified areas is much higher than that of the vascular area of ​​the coronary artery, extracting this area solely based on brightness without considering this may cause problems such as coronary artery rupture.

[0056] To prevent this problem in advance, the first brightness adjustment unit 110 performs brightness adjustment processing on the calcified region. Preferably, the maximum and minimum brightness values ​​of each pixel in each image data of the 2D coronary CT image dataset input to the image preprocessing unit 100 are adjusted based on a first brightness value range.

[0057] The preferred reference range for the first brightness value set experimentally is -350 to 550 HU. Reflecting this, for example, if the brightness value of each pixel is less than -350 HU, the brightness value is adjusted to -350 HU; if the brightness value is greater than 550 HU, the brightness value is adjusted to 550 HU. Therefore, it is desirable to reduce the brightness difference between the coronary artery region and the calcified region by adjusting the brightness value of each pixel constituting each image data to -350 HU to 550 HU. This can improve the segmentation accuracy of subsequent analysis.

[0058] Preferably, the voxel transformation unit 120 performs voxel isotropic processing of the image data, such as... Figure 2 As shown in c).

[0059] Specifically, the voxel transformation unit 120 preferably isotropically processes the size of each image data unit voxel, and each image data has a brightness value adjusted by the first brightness adjustment unit 110 according to a predetermined size.

[0060] In other words, since the width, length, and height of pixels differ among the various image data constituting the image dataset, it is preferable to convert the size of each voxel to have the same width, length, and height, such as... Figure 3 As shown. This is performed with the voxels needing to be preprocessed to be isotropic, because the blood vessel filter applied later is essentially a 3D-based filter. For example, the size of the voxel units constituting the image dataset is converted to 0.6 mm in width, length, and height.

[0061] like Figure 2 As shown in d), the noise processing unit 130 preferably removes the noise region of each image data that has been isotropically processed by the voxel transformation unit 120 by applying a morphological closure operator.

[0062] like Figure 1 As shown, the noise processing unit 130 performs operations through the threshold processing unit 131, the operator application unit 132, the mask generation unit 133, and the mask processing unit 134.

[0063] Specifically, tracheas and bronchi are located in the lungs and their shape resembles blood vessels. If these tracheas and bronchi are not removed, they are likely to be incorrectly detected as vascular regions when a vascular filter is applied later. Therefore, to improve segmentation accuracy, they are removed beforehand through preprocessing. Preferably, points in the lungs where the brightness of the removed objects (tracheas, bronchi, etc.) differs from that of blood vessels are used.

[0064] Figure 4 a) is an example diagram showing image data processed isotropically by the voxel transformation unit 120. For example... Figure 4 As shown in b), the threshold processing unit 131 preferably analyzes the pixels contained in each image data based on brightness values ​​below a specific threshold. That is, pixels with brightness values ​​below the specific threshold (set to -280HU in the experiments of this invention) are assigned a value of 1, and otherwise the pixels are assigned a value of 0.

[0065] Preferably, the operator application unit 132 processes each image data by applying a morphological closure operator. That is, as... Figure 4 As shown in c), pores corresponding to the internal regions of the lung are filled by applying a morphological closure operator.

[0066] Preferably, the mask generation unit 133 generates a mask by using a predetermined value assigned by the threshold processing unit 131. That is, as... Figure 4 As shown in d), the pixel generation mask is assigned a value of 0 when the threshold processing unit assigns a value of 1, and the pixel generation mask is assigned a value of 1 when the threshold processing unit assigns a value of 0.

[0067] The masking unit 134 performs masking processing on the image data processed by the operator application unit 132 using the mask generated by the masking unit 133, thereby removing noise regions. In other words, as... Figure 4 As shown in e), when the mask generated by the mask generation unit 133 is applied (multiplied) to the image data processed by the operator application unit 132, the result of removing the vascular structure in the internal region of the lung can be obtained.

[0068] Subsequently, the second brightness adjustment unit 140 preferably adjusts the maximum and minimum brightness values ​​of each image data whose noise has been removed by the noise processing unit 130 based on a preset second brightness value range. Specifically, as shown... Figure 2 e) and Figure 4 As shown in e), the noise is removed by the masking processing unit 134.

[0069] The preferred reference range for the second brightness value, set experimentally, is 0 to 255. To reflect this, for example, when the brightness value of each pixel is less than 0 HU, it is adjusted to 0 HU, and when the brightness value is greater than 255 HU, it is adjusted to 255 HU. Thus, the brightness values ​​of all pixels constituting each image data are adjusted to correspond to 0 HU to 255 HU, thereby improving the segmentation accuracy of subsequent analysis.

[0070] Preferably, the first processing unit 200 receives the image dataset preprocessed by the image preprocessing unit 100 and segments the ascending aorta image region by using a deep learning model pre-stored in the DL method.

[0071] like Figure 1 As shown, the first processing unit 200 is configured to include a deep learning segmentation unit 210 and a segmentation post-processing unit 220.

[0072] The deep learning segmentation unit 210 uses a deep learning model trained for ascending aorta segmentation, inputs image data set by the second brightness adjustment unit 140, and outputs the ascending aorta image region.

[0073] Because the ascending aorta is a structure with a simple shape and relatively large volume compared to the coronary arteries, results can be obtained quickly, while maintaining sufficient segmentation performance through deep learning models.

[0074] like Figure 5 As shown, the deep learning segmentation unit 210 performs learning processing by adding a dropout layer and a batch normalization layer based on 2D U-Net, as... Figure 5 The deep learning algorithms shown improve the stability and performance of deep learning models. Of course, besides the modified 2D U-Net, various other deep learning algorithms can be used to generate deep learning models, such as the unmodified 2D U-Net, 3D U-Net, 2D Fully Convolutional Networks (FCN) models, and 3DFCN models.

[0075] The segmentation post-processing unit 220 can analyze the output results of the deep learning segmentation unit 210 (segmentation results of the ascending aorta) and remove noise through post-processing, thereby further improving the segmentation accuracy.

[0076] In the post-segmentation processing unit 220, operations are performed in the filter processing unit 221 and the transformation processing unit 222 based on the fact that the ascending aorta has a substantially circular cross-section.

[0077] Filter processing unit 221 applies a preset filter to the ascending aorta image region output by deep learning segmentation unit 210 (see [link]). Figure 6 a)) thus generating an image that emphasizes the boundaries of the contained structures.

[0078] That is to say, such as Figure 6 As shown in b), the filter processing unit 221 applies a Gaussian gradient filter to the output of the deep learning segmentation unit 210 (the ascending aortic image region) to obtain an image that includes the boundaries of the emphasized object (structure).

[0079] like Figure 6 As shown in c), the transformation processing unit 222 applies the Hough circle transform to detect circular structures in the first image data based on the filter processing unit 221. This Hough circle transform is a preset transformation technique. The post-segmentation processing unit 220 can then set the circular structures detected by the transformation processing unit 222 as the ascending aorta image region.

[0080] However, as Figure 7 As shown in a), in some cases, multiple circular structures can be detected in the image data. Among these circular structures, additional filtering is preferably performed to extract only the circular structures corresponding to the ascending aorta.

[0081] like Figure 1 As shown, for additional filtering processing, the segmentation post-processing unit 220 is preferably configured to further include a region-of-interest (ROI) processing unit (ROI) 223 and an aortic setting unit 224.

[0082] Preferably, the ROI processing unit 223 sets the ROI by processing the image based on the filter processing unit 221. For example, preferably, the interior region is designated as the ROI based on 20% of the horizontal and vertical lengths of the top, bottom, left, and right corners. Here, the 20% limit is derived experimentally but is only an example.

[0083] like Figure 7 As shown in b), preferably, the aortic setting unit 224 analyzes the circular structure detected by the transformation processing unit 222 based on the ROI set by the ROI processing unit 223, so as to set the final ascending aortic image region in the circular structure detected by the predetermined reference.

[0084] In other words, the aortic setting unit 224 analyzes only the circular structures included in the ROI set by the ROI processing unit 223 among the circular structures detected by the transformation processing unit 222, and sets the closest circle as the final ascending aortic image based on the center point of the image (center point of the ROI), excluding circular structures with a radius of less than 10 mm or more than 20 mm.

[0085] At this point, apart from some circular structures based on radius, setting the final ascending aorta imaging region based on the center point is a standard set by experimentally considering the position of the ascending aorta, but this is also an exemplary embodiment.

[0086] The second processing unit 300 preferably receives the image dataset preprocessed by the image preprocessing unit 100 according to pre-stored rules, and uses the ascending aorta image segmented by the first processing unit 200 to segment the coronary artery image region as the RB method.

[0087] Therefore, such as Figure 1 As shown, the second processing unit 300 includes a blood vessel calculation unit 310, a blood vessel correction unit 320, a coronary artery extraction unit 330, and a coronary artery setting unit 340.

[0088] The blood vessel calculation unit 310 calculates the blood vessel value of each pixel constituting each image data preprocessed by the image preprocessing unit 100 by using a Hessian-based blood vessel filter.

[0089] Specifically, in order to apply the Hessian matrix-based blood vessel filter, the blood vessel calculation unit 310 first superimposes the image data preprocessed by the image preprocessing unit 100 and applies it as 3D image data. In this case, a Gaussian smoothing filter is applied to make the brightness values ​​between adjacent pixels have mantissa and distribution.

[0090] Next, the Hessian matrix and its eigenvalues ​​for each pixel are calculated. In this case, the Hessian matrix is ​​replaced by Equation 1 below, and the eigenvalues ​​are defined by Equation 2 below.

[0091] [Formula 1]

[0092]

[0093] [Formula 2]

[0094] 0≈|λ1|<<|λ2|≈|λ3|

[0095] Since the Hessian matrix on the 3D image data is calculated, each pixel has three feature values. The blood vessel value of each pixel is calculated by substituting these three feature values ​​into the blood vessel calculation formula, for example, Formula 3 proposed by Frangi (Frangi et al., 1998).

[0096] [Formula 3]

[0097]

[0098] (Here, R) A Indicates sensitivity to plate-like structures.

[0099] R B Indicates sensitivity to speckled structures, and

[0100] S represents the sensitivity to image contrast.

[0101] R A R B And S can be defined by the following formula 4.

[0102] [Formula 4]

[0103]

[0104] R A R B S is defined by eigenvalues ​​and is preferably optimized by inputting three hyperparameters (α, β, and γ) for adjustment. Therefore, the blood vessel value ranges from 0 to 1, and when the blood vessel value is close to 1, it indicates that the pixel is close to the center line of the blood vessel.

[0105] The vascular correction unit 320 adjusts the maximum and minimum vascular values ​​based on the set vascular range so that the coronary artery region set by the aortic setting unit 224 and the final ascending aortic region overlap each other by using the vascular values ​​calculated by the vascular calculation unit 310, and assigns binarized data to each pixel using the adjusted vascular values.

[0106] Specifically, Figure 8 a) consists of image data preprocessed by the image preprocessing unit 100, and as follows Figure 8 As shown in b), the calculated values ​​for blood vessels are corrected to a range of 0 to 255, and as... Figure 8 As shown in c), pixels with a corrected blood vessel value of 60 or greater are assigned a 1, and the remaining pixels are assigned a 0, thus obtaining binarized data. Therefore, 1 represents the blood vessel region, and 0 represents the remaining region.

[0107] The coronary artery extraction unit 330 preferably extracts a structural image region connected to the final ascending aorta image region in the structural image obtained from the binarized data provided by the vessel correction unit 320, using the final ascending aorta image region defined by the aorta setting unit 224. That is, since the coronary artery is a vessel extending from and connecting to the ascending aorta, the structure connected to the final ascending aorta image region defined by the aorta setting unit 224 is extracted only from the vessel shape structure segmented by the vessel correction unit 320. In this case, the extracted structure represents the coronary artery.

[0108] The coronary artery setting unit 340 preferably sets the final coronary artery image region by using the vascular values ​​from the vascular calculation unit 310 to remove noisy pixels contained in the structural image region extracted by the coronary artery extraction unit 330.

[0109] like Figure 9 As shown, when the pixels constituting each image data preprocessed by the image preprocessing unit 100 have a brightness of 100 or greater, the coronary artery setting unit 340 removes noise in the blood vessel segmentation result by assigning a value of 1 to the pixels identified as structural image regions by the coronary artery extraction unit 330 and assigning a value of 0 to the remaining pixels.

[0110] The combining unit 400 overlaps the ascending aortic image region and the coronary artery image region to obtain a combined structural image.

[0111] Specifically, the combining unit 400 overlaps the final ascending aortic image region set by the aortic setting unit 224 and the final coronary artery image region set by the coronary artery setting unit 340 to obtain a coupled structure image.

[0112] The final ascending aorta region set by the aorta setting unit 224 and the final coronary artery image region set by the coronary artery setting unit 340 are 2D image datasets in binary form (1 for blood vessels, 0 for the remaining areas). With this in mind, if pixel values ​​at the same location are added after overlapping these two structures, 1 or 2 represents the region of the coronary artery or ascending aorta, and 0 corresponds to the rest of the background. Thus, pixels with a value of 1 or greater are assigned a value of 1, and the remaining pixels are assigned a value of 0, thereby ultimately yielding the structure of the combined ascending aorta and coronary artery.

[0113] Figure 10 A flowchart illustrating a method for segmenting the ascending aorta and coronary arteries from a CCTA using a hybrid approach according to an exemplary embodiment of the present invention will be referenced. Figure 10 A detailed description is provided of a method for segmenting the ascending aorta and coronary arteries from a CCTA using a hybrid approach, according to exemplary embodiments of the present invention.

[0114] According to an exemplary embodiment of the present invention, the method for segmenting the ascending aorta and coronary arteries from CCTA using a hybrid method is a segmentation method that uses a segmentation system to segment the ascending aorta and coronary arteries from CCTA using a hybrid method, wherein each step is performed by a computing processing unit including a computer.

[0115] like Figure 10 As shown, the method for segmenting the ascending aorta and coronary arteries from CCTA using a hybrid method according to an exemplary embodiment of the present invention includes an image input operation S100, a preprocessing operation S200, a first processing operation S300, a second processing operation S400, and a combination operation S500.

[0116] Each operation will be described in detail.

[0117] In the image input operation S100, the image preprocessing unit 100 obtains coronary CT medical images through coronary CT angiography (CCTA).

[0118] Coronary CT medical imaging refers to 2D coronary CT image datasets obtained through coronary CT angiography, as well as 3D coronary CT images obtained by overlaying 2D coronary CT image datasets.

[0119] In the preprocessing operation S200, the image preprocessing unit 100 preprocesses the image dataset set by the image input operation S100 to extract the shape of the ascending aorta and coronary arteries.

[0120] Specifically, the preprocessing operation S200 includes a first brightness adjustment operation S210, a voxel processing operation S220, a noise processing operation S230, and a second brightness adjustment operation S240, such as... Figure 10 As shown.

[0121] Figure 2 a) indicates selecting an image data from the 2D coronary CT image dataset obtained through the image input operation S100, and in the first brightness adjustment operation S210, as shown... Figure 2 As shown in b), the image data brightness is cropped.

[0122] Specifically, many patients have calcified areas in their coronary arteries. Since the brightness of calcified areas is much higher than that of the vascular area of ​​the coronary artery, extracting this area solely based on brightness without considering this may cause coronary artery rupture.

[0123] To prevent this problem in advance, brightness adjustment processing of the calcified area is performed through the first brightness adjustment operation S210.

[0124] In the first brightness adjustment operation S210, the maximum and minimum brightness values ​​of each pixel in each image dataset constituting the 2D coronary CT image dataset are adjusted based on a preset first brightness value range.

[0125] The preferred reference range for the first brightness value set experimentally is -350 to 550 HU. Reflecting this, for example, if the brightness value of each pixel is less than -350 HU, the brightness is adjusted to -350 HU; if the brightness value is greater than 550 HU, the brightness value is adjusted to 550 HU. Therefore, it is desirable to reduce the brightness difference between the coronary artery region and the calcified region by adjusting the brightness value of each pixel constituting each image data to -350 HU to 550 HU. This can improve the segmentation accuracy of subsequent analysis. While the reference range for the first brightness value set experimentally is the optimal range set experimentally, the present invention is not limited thereto.

[0126] In voxel processing operation S220, such as Figure 2 As shown in c), the size of each image data unit voxel having a brightness value adjusted by the first brightness adjustment operation S210 is isotropically processed according to a predetermined size.

[0127] In other words, since the width, length, and height of pixels differ among the various image data constituting the image dataset, it is preferable to convert the size of each voxel to have the same width, length, and height, such as... Figure 3 As shown. This is performed when voxels need to be preprocessed to become isotropic, because the blood vessel filter applied later is essentially a 3D-based filter. For example, the size of the voxel units constituting the image dataset is converted to 0.6 mm in width, length, and height. In this case, the size of the unit voxel is also the optimal size set experimentally, but is not limited to this.

[0128] In noise processing operation S230, such as Figure 2 As shown in d), noise regions in each image data that are isotropically processed by the voxel transformation unit 120 are removed by applying a morphological closure operator.

[0129] like Figure 10 As shown, the noise processing operation S230 includes a threshold processing operation S231, an operator application operation S232, a mask generation operation S233, and a mask processing operation S234.

[0130] Specifically, tracheas and bronchi are present in the lungs and resemble blood vessels in shape. If these tracheas and bronchi are not removed, they are likely to be incorrectly detected as vascular regions when a vascular filter is applied later. Therefore, to improve segmentation accuracy, they are removed beforehand through preprocessing. Preferably, points within the lungs where the brightness of the removed objects (tracheas, bronchi, etc.) differs from that of blood vessels are used.

[0131] Figure 4 a) is an example diagram showing image data isotropically processed by the voxel transformation unit 120. Figure 4 In the threshold processing operation (S231), such as Figure 4 As shown in b), preferably, the pixels contained in each image data can be analyzed based on brightness values ​​below a certain threshold, and these pixels can be assigned predetermined values. That is, pixels with brightness values ​​below a certain threshold (set to -280HU in the experiments of this invention) are assigned a value of 1, otherwise the pixels are assigned a value of 0.

[0132] In operator application operation S232, image data are processed by applying a morphological closure operator. That is to say, as... Figure 4 As shown in c), pores corresponding to the internal regions of the lung are filled by applying a morphological closure operator.

[0133] In mask generation operation S233, a mask is generated using the predetermined value assigned in threshold processing operation S231. That is to say, as... Figure 4 As shown in d), the pixel generation mask is assigned a value of 0 when the threshold processing unit 131 assigns a value of 1, and the pixel generation mask is assigned a value of 1 when the threshold processing unit 131 assigns a value of 0.

[0134] In the masking operation S234, the image data processed by the operator application operation S232 is masked using the mask generated by the mask generation operation S233, thereby removing noisy regions. In other words, as... Figure 4 As shown in e), when the mask generated by the mask generation operation S233 is applied (multiplied) to the image data processed by the operator application operation S232, the result of removing the vascular structures contained in the internal region of the lung can be obtained.

[0135] In the second brightness adjustment operation S240, the maximum and minimum brightness values ​​of each image data that has had noise removed by the noise processing operation S230 can be adjusted based on the second brightness value range, such as... Figure 2 e) and Figure 4 As shown in e).

[0136] The second brightness value range reference set experimentally is preferably 0 to 255. For example, when the brightness value of each pixel is less than 0 HU, it is adjusted to 0 HU, and when the brightness value is greater than 255 HU, it is adjusted to 255 HU. Thus, the brightness values ​​of all pixels constituting each image data are adjusted to correspond to 0 HU to 255 HU, thereby improving the segmentation accuracy of subsequent analysis. In this case, although the reference for the second brightness value range set experimentally is the optimal range set experimentally, the present invention is not limited thereto.

[0137] In the first processing operation S300, the first processing unit 200 receives the preprocessed image dataset from the image preprocessing operation S200 by using a deep learning model pre-stored in the DL method, and segments the ascending aorta image region.

[0138] In the first processing operation S300, such as Figure 10 As shown, deep learning segmentation operation S310 and post-segmentation processing operation S320 are performed.

[0139] In the deep learning segmentation operation S310, the image data set by the second brightness adjustment operation S240 is used as input to the deep learning model trained for ascending aorta segmentation, and the ascending aorta image region is output.

[0140] Because the ascending aorta is a structure with a simple shape and relatively large volume compared to the coronary arteries, results can be obtained quickly while maintaining sufficient segmentation performance through deep learning models.

[0141] like Figure 5 As shown, learning processing is performed by adding an exit layer and a batch normalization layer based on 2D U-Net as a deep learning algorithm to generate a deep learning model, thereby improving the stability and performance of the deep learning model. Of course, in addition to the modified 2D U-Net, various deep learning algorithms can be used to generate deep learning models, such as the unmodified 2DU-Net, 3D U-Net, 2DFCN model, and 3DFCN model.

[0142] In the post-segmentation processing operation S320, the output results of the deep learning segmentation operation S310 (segmentation results of the ascending aorta) can be analyzed, and noise removal can be performed through post-processing to further improve the segmentation accuracy.

[0143] In the post-segmentation processing operation S320, based on the fact that the ascending aorta has a circular cross-section, filter processing operation S321 and transformation processing operation S322 are performed.

[0144] In filter processing operation S321, a preset filter is applied to the output of deep learning-based segmentation operation S310 (ascending aorta segmentation result) to generate an image that emphasizes the boundaries of the contained structures.

[0145] That is, Figure 6 As shown in b), by applying a Gaussian gradient filter to the output of the deep learning segmentation operation S310 (the ascending aorta image region), an image containing the boundaries of emphasized objects (structures) is obtained.

[0146] In the transformation processing operation S322, such as Figure 6 As shown in c), circular structures in the first image data based on filter processing operation S321 are detected by applying the Hough circular transform, which is a preset transform technique. In this case, the detected circular structure can be set as the ascending aorta image region.

[0147] However, as Figure 7 As shown in a), in some cases, multiple circular structures can be detected in the image data. Among these circular structures, additional filtering is preferably performed to extract only the circular structures corresponding to the ascending aorta.

[0148] For additional filtering processing, such as Figure 10 As shown, ROI processing operation S323 and aortic setting operation S324 are performed.

[0149] In ROI processing operation S323, the ROI is set by processing the image based on filter processing operation S321. For example, the interior region is designated as ROI based on 20% of the horizontal and vertical lengths of the top, bottom, left, and right corners. Here, the 20% limit is derived experimentally and is only an example.

[0150] In the aortic setting operation S324, such as Figure 7 As shown in b), the circular structure detected by the transformation processing operation S322 is analyzed based on the ROI set by the ROI processing operation S323, so as to set the final ascending aortic region in the circular structure detected by the predetermined reference.

[0151] In other words, in the aortic setting operation S324, among the circular structures detected by the transformation processing operation S322, only the circular structures included in the ROI set by the ROI processing operation S323 are analyzed, and the closest circle is set as the final ascending aortic image region based on the center point of the image (the center point of the ROI), excluding circular structures with a radius of less than 10 mm or more than 20 mm.

[0152] At this point, apart from some radius-based circular structures, setting the final ascending aorta imaging region based on the center point is a standard set by experimentally considering the location of the ascending aorta, but this is also an exemplary embodiment.

[0153] In the second processing operation S400, the second processing unit 300 analyzes the preprocessed image dataset in the preprocessing operation S200 based on the rules pre-stored in the RB method, and segments the coronary artery image region by using the analysis results and the segmented ascending aortic image region in the first processing operation S300.

[0154] In the second processing operation S400, such as Figure 10 As shown, the following operations are performed: blood vessel calculation operation S410, blood vessel correction operation S420, coronary artery extraction operation S430, and coronary artery setting operation S440.

[0155] In the blood vessel calculation operation S410, the blood vessel value of each pixel of each image data constituting the preprocessing in the image preprocessing operation S200 is calculated by using a Hessian-based blood vessel filter.

[0156] Specifically, firstly, in order to apply a blood vessel filter based on the Hessian matrix, the preprocessed image data are superimposed and used as 3D image data. In this case, a Gaussian smoothing filter is applied to make the brightness values ​​between adjacent pixels have a tail number and distribution.

[0157] Next, the Hessian matrix and its eigenvalues ​​for each pixel are calculated. In this case, the Hessian matrix is ​​replaced by Equation 1 above, and the eigenvalues ​​are defined by Equation 2 above.

[0158] Since the Hessian matrix on the 3D image data is calculated, each pixel has three feature values. The blood vessel value of each pixel is calculated by substituting these three feature values ​​into the blood vessel calculation formula, for example, Formula 3 proposed by Frangi (Frangi et al., 1998).

[0159] R A R B S is defined by eigenvalues ​​and is preferably optimized for the blood vessel filter by inputting three hyperparameters (α, β, and γ) for adjustment. Thus, the blood vessel value ranges from 0 to 1, and a blood vessel value close to 1 indicates that the pixel is close to the centerline of the blood vessel.

[0160] In the blood vessel correction operation S420, the blood vessel values ​​obtained in the blood vessel calculation operation S410 are used to adjust the maximum and minimum blood vessel values ​​based on a preset blood vessel value range, and the adjusted blood vessel values ​​are used to assign binarized data to each pixel.

[0161] Specifically, Figure 8a) represents the pre-processed image data, and as shown in the image data... Figure 8 As shown in b), the calculated values ​​for blood vessels are corrected to a range of 0 to 255, and as... Figure 8 As shown in c), pixels with a corrected blood vessel value of 60 or greater are assigned a 1, and the remaining pixels are assigned a 0, thus obtaining binarized data. Therefore, 1 represents the blood vessel region, and 0 represents the remaining region.

[0162] In the coronary artery extraction operation S430, the structural image region connected to the final ascending aortic image region in the structural image is extracted by using the final ascending aortic image region set in the aortic setting operation S324 and the structural image obtained by the binarized data given by the vascular correction operation S420.

[0163] In other words, since the coronary arteries are vessels that extend to and connect to the ascending aorta, in the vessel shape structure segmented by the vessel correction operation S420, only the structure connecting to the final ascending aorta image region set by the aorta setting operation S324 is extracted based on this. At this time, the extracted structure represents the coronary artery.

[0164] In the coronary artery setting operation S440, after removing the noisy pixels in the structural image region extracted in the coronary artery extraction operation S430 by using the blood vessel value obtained in the blood vessel calculation operation S410, it is set as the final coronary artery image region.

[0165] That is to say, such as Figure 10 As shown, when each pixel of each image data constituting the preprocessing has a brightness of 100 or greater, in the coronary artery extraction operation S430, a value of 1 is assigned to the pixel identified as a structural image region, and a value of 0 is assigned to the remaining pixels, thereby removing noise in the blood vessel segmentation result.

[0166] In the combined operation S500, a combined structural image is obtained by overlapping the image regions of the ascending aorta and the coronary arteries.

[0167] Specifically, in the combination operation S500, the final ascending aorta image region set in the aorta setting operation S324 and the final coronary artery image region set in the coronary artery setting operation S440 overlap to obtain a coupled structure image.

[0168] The final ascending aorta and coronary artery image regions are 2D image datasets in binary form (1 for vessels, 0 for the rest of the region). With this in mind, if pixel values ​​at the same location are added after overlapping these two structures, 1 or 2 represents the region of the coronary artery or ascending aorta, and 0 corresponds to the rest of the background. Thus, pixels with values ​​of 1 or greater are assigned a value of 1, and the remaining pixels are assigned a value of 0, ultimately yielding the combined structure of the ascending aorta and coronary arteries.

[0169] In the system and method of the present invention based on the above configuration, which uses a hybrid method to segment the ascending aorta and coronary arteries from CCTA, a hybrid method combining the RB method and the DL method is used according to the characteristics of the coronary arteries and the ascending aorta, thereby extracting the results (extracting the coronary artery and ascending aorta regions) reasonably, quickly and accurately.

[0170] In the foregoing, although the present invention has been described in detail by way of specific matters such as detailed components, exemplary embodiments, and accompanying drawings, these are provided only to aid in a thorough understanding of the invention. Therefore, the invention is not limited to the exemplary embodiments. Various modifications and changes can be made by those skilled in the art based on this specification.

[0171] Therefore, the spirit of the present invention should not be limited to these exemplary embodiments, but the claims and all modifications equivalent to or related to the claims are intended to fall within the scope and spirit of the present invention.

[0172] [Detailed Description of Key Elements]

[0173] 100: Image Preprocessing Unit

[0174] 110: First brightness adjustment unit; 120: Voxel transformation unit

[0175] 130: Noise processing unit; 140: Second brightness adjustment unit

[0176] 131: Threshold processing unit; 132: Operator application unit

[0177] 133: Mask generation unit; 134: Mask processing unit

[0178] 200: First Processing Unit

[0179] 210: Deep learning segmentation unit; 220: Post-segmentation processing unit.

[0180] 221: Filter processing unit; 222: Transformation processing unit

[0181] 223: ROI processing unit; 224: Aortic setting unit

[0182] 300: Second Processing Unit

[0183] 310: Vascular Calculation Unit; 320: Vascular Correction Unit

[0184] 330: Coronary artery extraction unit; 240: Coronary artery setting unit

[0185] 400: Combination Unit

Claims

1. A segmentation system for segmenting the ascending aorta and coronary arteries from coronary CT angiography using a hybrid approach, the segmentation system comprising: The image preprocessing unit receives a 2D coronary CT image dataset obtained by coronary CT angiography and performs image preprocessing to extract the shape of the coronary arteries and ascending aorta. The first processing unit receives the preprocessed image dataset from the image preprocessing unit through a pre-stored deep learning model and segments the ascending aorta image region. A second processing unit receives a preprocessed image dataset from the image preprocessing unit based on pre-stored rules, and segments the coronary artery image region using the ascending aorta image region segmented by the first processing unit; and A combining unit overlaps the ascending aorta image region and the coronary artery image region to obtain a combined structural image. The second processing unit includes: A blood vessel calculation unit calculates the blood vessel values ​​of each pixel constituting each image data preprocessed by the image preprocessing unit by using a Hessian-based blood vessel filter. A vascular correction unit uses vascular values ​​from the vascular calculation unit to adjust the maximum and minimum vascular values ​​based on a preset vascular value range, and uses the adjusted vascular values ​​to provide binarized data to each pixel. A coronary artery extraction unit uses the ascending aorta image region segmented and processed by the first processing unit to extract a structural image region connected to the ascending aorta image region in the structural image obtained from the binarized data. as well as The coronary artery setting unit sets the extracted structural image region as the final coronary artery image region by using the vascular values ​​from the vascular calculation unit to remove noisy pixels contained in the extracted structural image region.

2. The segmentation system according to claim 1, wherein, The image preprocessing unit includes: A first brightness adjustment unit adjusts the maximum and minimum brightness values ​​of each image data based on a preset first brightness value range. The voxel transformation unit is a unit that isotropically processes the size of a unit voxel by adjusting the brightness value of each image data according to a predetermined reference by the first brightness adjustment unit. A noise processing unit removes noise regions from image data that have undergone isotropic processing by the voxel transformation unit by applying a morphological closure operator; and The second brightness adjustment unit adjusts the maximum and minimum brightness values ​​of each image data that has had noise removed by the noise processing unit based on a preset second brightness value range.

3. The segmentation system according to claim 2, wherein, The noise processing unit includes: A threshold processing unit analyzes the pixels contained in each image data based on brightness values ​​below a specific threshold and assigns them predetermined values. The operator application unit processes image data by applying morphological closure operators; A mask generation unit that generates a mask using a predetermined value provided by the threshold processing unit; and A masking unit, which uses the mask generated by the masking unit to mask the image data processed by the operator application unit to remove noise regions.

4. The segmentation system according to claim 2, wherein, The first processing unit includes: A deep learning segmentation unit, wherein the deep learning segmentation unit inputs the image data set by the second brightness adjustment unit into a deep learning model trained for ascending aorta segmentation, and receives the image region of the ascending aorta; and A segmentation post-processing unit analyzes the output from the deep learning segmentation unit and performs noise removal.

5. The segmentation system according to claim 4, wherein, The post-segmentation processing unit includes: A filter processing unit applies a preset filter to the output ascending aorta image region to generate an image that emphasizes the boundaries of the included structures; and The transformation processing unit detects circular structures contained in the image generated by the filter processing unit by applying a preset transformation technique. The circular structure detected by the transformation processing unit is set as the ascending aorta imaging region.

6. The segmentation system according to claim 5, wherein, The post-segmentation processing unit further includes: A Region of Interest (ROI) processing unit processes the image from the filter processing unit to set the ROI; and An aortic setting unit analyzes a circular structure detected by the transformation processing unit, the circular structure being contained within an ROI set by the ROI processing unit, and sets a final ascending aortic image region within the detected circular structure using a predetermined reference.

7. The segmentation system according to claim 6, wherein, The coronary artery extraction unit uses the final ascending aortic image region set by the aortic setting unit to extract a structural image region connected to the final ascending aortic image region in the structural image obtained from the binarized data provided by the vascular correction unit.

8. The segmentation system according to claim 7, wherein, The combining unit overlaps the final ascending aortic image region from the aortic setting unit and the final coronary artery image region from the coronary artery setting unit to obtain a structural region based on the combination of pixel values ​​at the same location.

9. A segmentation method for segmenting the ascending aorta and coronary arteries from coronary CT angiography (CCTA) using a hybrid approach, said hybrid approach using segmentation systems for the ascending aorta and coronary arteries from CCTA, wherein, Each operation is performed by a computing processing unit including a computer, and the segmentation method includes: Image input operation, wherein the image preprocessing unit receives a 2D CCTA image dataset obtained through coronary CT angiography; Preprocessing operation, wherein the image preprocessing unit preprocesses the image dataset based on the image input operation; A first processing operation, wherein a first processing unit uses a pre-stored deep learning model to segment the ascending aorta image region from the image dataset preprocessed by the preprocessing operation; The second processing operation involves a second processing unit analyzing the preprocessed image dataset from the preprocessing operation and using the analysis results and the segmented ascending aorta image region from the first processing operation to segment the coronary artery image region; and The combination operation, wherein the combination unit overlaps the ascending aorta image region based on the first processing operation and the coronary artery image region based on the second processing operation. The second processing operation includes: A blood vessel calculation operation is performed by using a Hessian-based blood vessel filter to calculate the blood vessel values ​​of each pixel that constitutes each image data preprocessed by the preprocessing operation. The vascular correction operation uses vascular values ​​based on the vascular calculation operation to adjust the maximum and minimum vascular values ​​based on a preset vascular value range, and uses the adjusted vascular values ​​to provide binarized data to each pixel. A coronary artery extraction operation, which uses the ascending aorta image region segmented by the first processing operation, extracts a structural image region connected to the ascending aorta image region in the structural image derived from the binarized data provided by the vascular correction operation; and The coronary artery setting operation removes noisy pixels from the structural image region extracted by the coronary artery extraction operation by using the vascular values ​​based on the vascular calculation operation, and then sets it as the final coronary artery image region.

10. The segmentation method according to claim 9, wherein, The preprocessing operations include: The first brightness adjustment operation adjusts the maximum and minimum brightness values ​​of each image data constituting the image dataset based on a preset first brightness value range. The voxel processing operation is a process in which the image data whose brightness values ​​are adjusted according to a predetermined reference in the first brightness adjustment operation are isotropically processed to adjust the size of a unit voxel. A noise processing operation, wherein the noise processing operation removes noise regions from each image data based on the voxel processing operation by applying a morphological closure operator; and The second brightness adjustment operation adjusts the maximum and minimum brightness values ​​of each image data that has had noise removed in the noise processing operation, based on a preset second brightness value range.

11. The segmentation method according to claim 10, wherein The noise processing operations include: A thresholding operation is performed, which analyzes the pixels contained in each image data based on brightness values ​​below a certain threshold and assigns a predetermined value. Operator application operations, wherein the operator application operations process each image data by applying morphological closure operators; A mask generation operation, wherein the mask generation operation uses a predetermined value provided in the threshold processing operation to generate a mask; as well as A masking operation is performed on the image data processed in the operator application operation by using the mask generated in the masking operation to remove noise regions.

12. The segmentation method according to claim 10, wherein, The first processing operation includes: A deep learning segmentation operation, wherein the deep learning segmentation operation inputs the image dataset based on the second brightness adjustment operation into a learned deep learning model for ascending aorta segmentation, and receives the ascending aorta image region; and The post-segmentation processing operation removes noise by analyzing the output of the deep learning segmentation operation.

13. The segmentation method according to claim 12, wherein, The post-segmentation processing operations include: A filter processing operation, wherein the filter processing operation applies a preset filter to the output ascending aorta image region to generate an image emphasizing the boundary regions of the included structures; and The transformation processing operation, which applies a preset transformation technique, detects circular structures contained in the image processed by the filter operation. Specifically, the circular structure detected during the transformation processing operation is set as the ascending aorta imaging region.

14. The segmentation method according to claim 13, wherein, The post-segmentation processing operation also includes: Region of Interest (ROI) processing operations, wherein the ROI processing operations set the ROI by processing the image based on the filter processing operations; and The aortic setting operation analyzes the circular structure detected by the transformation processing operation and sets the final ascending aortic image region in the circular structure detected by a predetermined reference, wherein the circular structure is contained in the ROI set by the ROI processing operation.

15. The segmentation method according to claim 14, wherein, The coronary artery extraction operation uses the final ascending aortic image region set by the aortic setting operation to extract the structural image region connected to the final ascending aortic image region in the structural image obtained from the binarized data given by the vascular correction operation.

16. The segmentation method according to claim 15, wherein, The combined operation overlaps the final ascending aortic image region based on the aortic setup operation and the final coronary artery image region from the coronary artery setup operation to obtain a structural region based on a combination of pixel values ​​at the same location.

Images