WO2018155765A1 - Method and device analyzing plaque from computed tomography image - Google Patents

Abstract

Disclosed is a method for analyzing a plaque from a medical image. The method comprises the steps of: receiving a medical image including a computed tomography (CT) image; generating n-channel data by adjusting the window width (WW) and the window level (WL) of the CT image, n being a natural number equal to or larger than 2; orthogonally reconfiguring images from each of the n-channel data such that the same are reconfigured into an axial image, a sagittal image, and a coronal image; machine-learning the reconfigured images on the basis of a convolutional neural network (CNN); generating a mask image from an image including at least one of the machine-learned, orthogonally reconfigured images and acquiring a sectional image of a coronary artery vessel on the basis of the generated mask images. Above steps are performed in an individual or integrated manner with regard to the coronary artery inner wall and the coronary artery outer wall.

Description

Method and apparatus for analyzing plaque in computed tomography images

The present invention relates to a method and apparatus for analyzing plaque in a computed tomography image. More specifically, plaque analysis using deep learning techniques to automatically generate masks of the inner and outer walls of coronary arteries from computed tomography images, thereby enabling easier and more accurate analysis of plaques in coronary arteries. The present invention relates to a method and an image processing apparatus.

Coronary artery disease can produce coronary plaque in blood vessels that provide blood to the heart, such as stenosis (a disease in which blood vessels become abnormally narrowed). Coronary plaques may limit blood flow to the heart, and patients suffering from coronary artery disease may experience chest pain such as unstable angina when at rest or chronic invariant angina when vigorous physical exercise.

Patients suffering from pain or exhibiting symptoms of coronary artery disease may receive one or more tests that can provide some indirect evidence of coronary plaque. For example, non-invasive tests may include biomarker evaluation from electrocardiograms, blood tests, treadmill tests, electrocardiogram recording, single positron emission computed tomography (SPECT) and positron emission tomography (PET). ) May be included. Anatomical data can be obtained non-invasively using coronary computed tomography angiography (CCTA). CCTA can be used for imaging of patients with chest pain and involves the use of computed tomography (CT) technology to image the heart and coronary arteries following intravenous infusion of contrast agents.

As such, quantitative analysis of atherosclerotic plaques in coronary artery is very important in the treatment of coronary artery disease. CCTA has become a reliable test method for the diagnosis of anatomically obstructed coronary artery disease, and CCTA can use analytical software to automatically or semi-automatically analyze atherosclerotic plaques in CCTA images.

However, in the process of analyzing coronary arteries using analysis software in CCTA, sophisticated modifications by skilled specialists are necessary because the analysis software cannot accurately analyze the boundaries of the inner and outer walls of the coronary arteries. Therefore, there is a problem that requires a lot of manual work to produce clinically useful results.

In addition, a study comparing how plaques measured in CCTA and actual plaques found that the volume of plaques measured in CCTA was overestimated by reference standard IVUS (intravascular ultrasound).

Therefore, there is an increasing need for a new plaque analysis method and apparatus for analyzing the plaque in the coronary artery to derive accurate results to be utilized in the actual clinical environment.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and the present invention uses a new CNN structure for image learning and a new medical image format suitable for such a CNN structure, and automatically masks the inner and outer walls of the coronary artery. It is an object of the present invention to provide a method and an image processing apparatus which can generate and thereby analyze the plaque in the coronary artery more easily and accurately.

According to an aspect of the present invention, there is provided a method for analyzing a plaque in a medical image, the method including: receiving a medical image including a computed tomography (CT) image; Generating n-channel data by adjusting a window width (WW) and a window level (WL) of the CT image, wherein n is a natural number of two or more; Reconstructing an image into an orthogonal image in each of the n-channel data and reconstructing the image into a horizontal image, a sagittal image, and a coronal image; Machine learning the reconstructed images based on a convolutional neural network (CNN); And generating a mask image from an image including at least one of a machine-learned orthogonally reconstructed image, and acquiring a cross-sectional image of the coronary vessel based on the generated mask images. It is characterized in that the inner wall and the coronary outer wall are performed individually or integrally, respectively.

The method may further include analyzing a plaque in the coronary artery using the obtained cross-sectional image of the coronary inner wall and the acquired cross-sectional image of the coronary outer wall.

In addition, generating the n-channel data by adjusting the window width (WW) and window level (WL) of the CT image, WW ₁ and WL ₁ for observation of coronary lumen, WW for calcium analysis Generating the n-channel data by setting WW ₃ and WL ₃ for ₂ and WL ₂ and lipid plaques, where n is 3.

In addition, the composite product neural network (CNN) may be configured by successively stacking two brief convolutional networks (BCN).

In addition, prior learning may be performed by using an auto-encoder in the preceding first BCN among the two simple synthetic product networks BCN.

In addition, the image processing apparatus according to an embodiment of the present invention for solving the above technical problem, the image receiving unit configured to receive a CT image, and the heart image received from the image receiving unit And an image processor configured to display at least one of a coronary inner wall image, an outer wall image, and a plaque image output from the image processor, and to control the image receiver, the image processor, and the display unit. An n-channel data generation unit configured to generate n-channel data by adjusting a window width (WW) and a window level (WL) of the CT image, wherein n is two or more; Natural mission; An image reconstruction unit configured to reconstruct an image at right angles from each of the n-channel data to reconstruct it into a horizontal plane image, a sagittal plane image, and a coronal plane image; A machine learning unit configured to machine learn based on a composite product neural network (CNN) for the reconstructed images; And an image acquisition unit configured to generate a mask image from each of the images including any one of the machine-reversed right-angle reconstructed images, and obtain a cross-sectional image of the coronary vessel based on the generated mask images. have.

The image processor may further include a plaque analysis unit configured to analyze plaque in the coronary artery using the obtained cross-sectional image of the coronary inner wall and the acquired cross-sectional image of the coronary outer wall.

In addition, the n-channel data generating unit is configured by setting WW ₁ and WL ₁ for coronary lumen observation, WW ₂ and WL ₂ for calcium analysis, and WW ₃ and WL ₃ for lipid plaques. Generate channel data, where n is three.

In addition, the composite product neural network (CNN) may be configured by successively stacking two brief convolutional networks (BCN).

In addition, prior learning may be performed by using a self-encoder in the preceding first BCN of the two simple synthetic product networks BCN.

According to a method for analyzing plaque in a computed tomography (CT) image and an image processing apparatus according to an embodiment of the present invention, analysis of coronary plaque in a computed tomography (CT) image using deep learning is performed. It can be done more easily and accurately.

In addition, according to the method and the image processing apparatus for analyzing the plaque in the computed tomography (CT) image according to an embodiment of the present invention, in analyzing the plaque of the coronary artery, accurate results to be utilized in the actual clinical environment It is possible to derive

1 is an exemplary diagram for describing a format of a general DICOM standard medical image.

2 is a flowchart of a method S200 for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention.

3 is an exemplary diagram for describing a process of generating 3-channel data according to an embodiment of the present invention.

FIG. 4 is an exemplary diagram for describing a process of reconstructing three-channel data generated in FIG. 3 into horizontal, coronal, and sagittal images according to an embodiment of the present invention.

5A is a schematic block diagram of a simplified composite product network (BCN) according to an embodiment of the present invention, and FIG. 5B is a schematic block diagram for explaining machine learning logic according to an embodiment of the present invention.

FIG. 6 is a flowchart schematically illustrating the overall flow of a method for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention.

7 is a schematic block diagram of an image processing apparatus 700 configured to perform a method for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention.

DETAILED DESCRIPTION A detailed description of preferred embodiments of the present invention will now be described with reference to the accompanying drawings. It should be noted that the same components in the figures represent the same numerals wherever possible. Specific details are set forth in the following description, which is provided to help a more general understanding of the present invention. In describing the present invention, when it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted.

The terminology used herein is to select general terms that are widely used as possible in consideration of the functions of the present invention, but may vary according to the intention or precedent of the person skilled in the art to which the present invention belongs, the emergence of new technologies, etc. have. In addition, in certain cases, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the corresponding description of the present invention. Therefore, the terms used in the present specification should be defined based on the meanings of the terms and the contents throughout the present specification, rather than simple names of the terms.

When a part of the present specification is said to "include" any component, this means that it may further include other components, without excluding the other components unless otherwise stated. In addition, the terms "... unit", "module", etc. described in the specification mean a unit for processing at least one function or operation, which may be implemented in hardware or software or a combination of hardware and software. .

1 is an exemplary diagram for describing a format of a general DICOM standard medical image.

DICOM (Digital Imaging and Communications in Medicine) standard is a term used to refer to various standards used for digital image representation and communication in medical devices. The American College of Radiology (ACR) and the US Electric Presented by a coalition committee formed by the National Electrical Manufacturers Association (NEMA).

FIG. 1 schematically illustrates the format of such DICOM standard medical images, as shown, DICOM images are 2 bytes, i.e. 16 bits in size, for which the user directly measures the window width (see FIG. We are using the window width (WW) and window level (WL) controls.

CT images can represent different stages of black and white using CT numbers, which are the relative linear attenuation coefficients of pixels that are set based on water, bone, and air. Here, the window width (WW) means a range of CT numbers that can be expressed in gray-scale, which is various levels of black and white, and the window level (WL) means the median value of the gray scale.

For example, if the window width WW of the abdominal image is set to +300 and the window level WL is set to 0, the range of hounsfield units (HU) appearing in the image is -150 to +150. Therefore, a material with an absorption lower than -150 appears black, a material with an absorption higher than +150 appears bright, and a material having a HU between -149 and +149 can be represented by a level between black and white.

However, when analyzing coronary arteries, especially the inner wall / outer wall / plaque of coronary arteries using these common DICOM format images, most of them require sophisticated modifications by experienced experts and due to the performance limitations of the analysis algorithms, There is an inconvenience of not having to analyze the outer wall precisely and go through a lot of manual work.

2 is a flowchart of a method S200 for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention.

For reference, in the following specification of the present invention, coronary computed tomography angiography (CCTA) images are exemplified for easy understanding of the present invention, but these are only examples and are limited to CCTA images. It could be extended to other medical images such as MRI image and X-ray image.

Method (S200) for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention, receiving a medical image including a computed tomography (CT) image (S210); Generating n-channel data by adjusting a window width (WW) and a window level (WL) of the CT image (S220), where n is a natural number of two or more; Reconstructing an image from each of the n-channel data at right angles and reconstructing the image into a horizontal image, a sagittal image, and a coronal image (S230); Machine learning on the reconstructed images based on a convolutional neural network (CNN) (S240); Generating a mask image from each of the images including any one of the machine-learned right-angle reconstructed images, and obtaining a cross-sectional image of the coronary blood vessel based on the generated mask images (S250); And analyzing the plaque in the coronary artery using the obtained cross-sectional image of the coronary inner wall and the obtained cross-sectional image of the coronary outer wall (S260).

For reference, each step S210 to S260 shown in FIG. 2 corresponds to an example for easy understanding of the present invention, and thus it will be apparent that additional steps not shown in FIG. 2 may be further performed. will be.

S210 corresponds to receiving a CT image, such as coronary computed tomography angiography (CCTA) image. Computed tomography (CT) is one of the medical image processing methods using tomography produced by computer processing, wherein the CCTA image is any coronal image that can be obtained by any computed tomography (CT) equipment / device / instrument Arterial CT imaging.

S220 corresponds to a step of generating n-channel data by adjusting a window width WW and a window level WL of the received CT image. Here, n is a natural number of two or more, and in the following description, three-channel data will be described as an example for easy understanding of the present invention, but it will be apparent that other plurality of channel data may additionally or alternatively be implemented. will be.

It is an object of the present invention to provide a method and apparatus for analyzing coronary plaques more easily and accurately compared to conventional methods, and to implement machine learning (eg, CNN) of images in implementing such methods / apparatus. It entails and proposes three-channel data with a new medical image format optimized for such machine learning, which is a clear contrast to conventional DICOM medical images with a single channel of 2 byte size.

More specifically, the three-channel data according to an embodiment of the present invention is composed of three data, each of which adjusts the window width (WW) and the window level (WL) of the CT image, each of which is coronary lumen Data for observation, data for calcium analysis, and data for lipid plaques.

For example, the image for coronary lumen observation can be set to window width WW ₁ and window level WL ₁ , and the image for calcium analysis can be set to window width WW ₂ and window level WL ₂ , for lipid plaques. The image can be set to window width WW ₃ and window level WL ₃ . For example, it is possible to set the window width / window level to values of 740/220 for coronary lumen observation, 1500/550 for calcium analysis and 500/100 for lipid plaque, Specific values of may correspond to experimentally derived values as optimal values for coronary lumen observation, calcium analysis and lipid plaque, respectively, in the art, but additional values other than those illustrated Or alternatively, it may be used.

For reference, the window width / window level of the first channel for coronary lumen observation is described in Eur Radiol. 2016 Sep; 26 (9): 3190-8, doi: 10.1007 / s00330-015-4121-5. Epub 2015 Dec 2. "Optimal boundary detection method and window settings for coronary atherosclerotic plaque volume analysis in coronary computed tomography angiography: comparison with intravascular ultrasound."

3 is an exemplary diagram for describing a process of generating 3-channel data according to an embodiment of the present invention. For reference, FIG. 3 exemplarily shows an image of a coronary lumen.

As shown in Figure 3, according to one embodiment of the present invention, to adjust the window width / window level for the same coronary artery lumen images in each _{WW 1 / WL 1, WW 2} / WL 2 and WW ₃ / WL ₃ Three channel data ch1, ch2 and ch3 can be generated. According to a further embodiment of the invention, the three-channel data ch1, ch2 and ch3 thus generated can be reconstructed on a scale of [0, 255], for this [-150HU, 590HU], [-200HU] , 1300 HU] and [-150 HU, 350 HU] different crop windows can be used.

The reason for constructing such three-channel data is to simulate the technique used in the conventional CT image analysis method. In fact, the above-described WW ₁ / WL ₁ , WW ₂ / WL ₂ and WW ₃ / WL _{3 are} presented. By setting the window width / window level to analyze coronary plaques, the present invention therefore has an important meaning that the deep learning model simulates the actual working environment to allow it to learn.

In other words, the purpose of the new three-channel data according to the present invention is to simulate the process of adjusting the width and level of the data display window to suit its characteristics when a real expert analyzes plaque by hand.

After the three-channel data is generated in S220 through the above-described process, S230 reconstructs the image at right angles from each of the three-channel data generated in such a manner so that the horizontal image, the coronal image, and Corresponds to the reconstruction of sagittal images.

For reference, because calcified plaques are three-dimensionally located in the coronary arteries, there is a limit in learning the overall characteristics of the plaques only by learning horizontal images. Therefore, the present invention is characterized by additionally generating a coronal image and a sagittal image by rotating the horizontal image as well as the horizontal image and reconstructing the image at right angles, and using them for image machine learning.

FIG. 4 is an exemplary diagram for describing a process of reconstructing three-channel data generated in FIG. 3 into horizontal, coronal, and sagittal images according to an embodiment of the present invention. As shown in Fig. 4, each of the three-channel data ch1, ch2 and ch3 is reconstructed into a horizontal plane image, a coronal plane image and a sagittal plane image.

Here, the size of the horizontal plane image is fixed to 64 × 64, but since the blood vessel length is different according to the degree of disease of a person and a person, a 64 × 64 × 64 size cube can be defined as a sampling unit. The part may be excluded from the learning object.

For reference, FIG. 4 exemplarily shows a 'horizontal plane image', 'coronal plane image' and 'sagittal plane image' for easy understanding of various embodiments of the present invention. It will be apparent in the art that it can be used.

As such, 3-channel data is generated by adjusting the window width (WW) and the window level (WL) of the CT image (S220), and reconstructing the horizontal / sagittal and coronal images for each of the generated 3-channel data. After S230, S240 corresponds to machine learning a reconstructed image using a composite product neural network (CNN).

For reference, a convolutional neural network (CNN) corresponds to a neural network having a structure for performing pretreatment in a compound multiplication layer consisting of one or several compound multiplication layers and a general artificial neural network layer mounted thereon. It may also be referred to as a convolutional neural network, a circuit neural network, a brain network, and the like.

For reference, the general composite neural network (CNN) structure is not suitable for detailed and sophisticated segmentation work such as coronary arteries. Therefore, unlike the conventional CNN structure that performs the segmentation in the conventional three-channel color image, in the present invention, a so-called " compression layer and decompression layer " Use the structure of a brief convolutional network (BCN).

FIG. 5A shows a schematic block diagram of this simplified composite product network (BCN). As shown in FIG. 5A, a simple synthetic product network (BCN) according to an embodiment of the present invention includes two compressed layers consisting of three composite product layers and one max-pooling layer, and three composite products. It may consist of two decompression layers consisting of a product layer and one up-sampling layer. Here, the max-pooling layer and the up-sampling layer may use 2 pixel striding.

For reference, all of the composite product layers were applied with a normalization penalty of 0.0005 (L2-weight), and activated using ReLU (Rectified Linear Unit). In addition, batch normalization was applied prior to activation of the composite product network to prevent gradients from weakening or converging, and a dropout technique at the end of each composite product network to prevent overfitting of the neural network. Was used and the ratio was set to 0.5.

Therefore, if we assume that an input image of size n × m is X and the output image of the same size is Y, then the probability distribution for x ∈ X and y 하여 Y is proposed using the proposed simple synthetic product network (BCN). It is possible to perform labeling on the output image using p (y | x).

5B is a schematic block diagram for explaining machine learning logic according to an embodiment of the present invention. As shown in FIG. 5B, machine learning for coronary inner wall / outer wall / plaque areaization according to an embodiment of the present invention is performed using a network of simple stacked composite product networks (BCNs). It is done.

That is, the composite product neural network (CNN) used in the image machine learning according to the present invention may be configured by successively stacking two simple composite product networks (BCN). Here, pre-training is performed by using an auto-encoder in the preceding first BCN (left BCN in FIG. 5B) of the two simplified composite product networks BCN.

More specifically, pre-learning is carried out using a simple synthetic product network (BCN) structure in order to obtain better initial values prior to coronary inner and outer wall prediction, where the output computed from the input is the input itself. It is characterized by using an auto-encoder to be similar to.

Therefore, through the self-encoder process by the first BCN, not only can the image noise be reduced, but also the boundary between the inner wall and the outer wall of the coronary artery is cleared, and the characteristics of the calcified plaques are improved to facilitate the learning.

As shown in FIG. 5B, when image learning is performed in a second BCN structure subsequent to the first BCN for pre-learning, a predictive model of the coronary inner and outer walls may be completed. For reference, the prediction model is individually learned for the horizontal plane image, the sagittal plane image, and the coronal plane image, and the learning and prediction for the coronary inner wall and the coronary outer wall may be separately or collectively performed.

As such, when machine learning is performed on the reconstructed horizontal plane, sagittal plane, and coronal plane image based on the double simple compound network (BCN) (S240), S250 is respectively obtained from the machine-learned horizontal plane, sagittal plane, and coronal plane images. Generating a mask image and obtaining a cross-sectional image of a coronary vessel based on the generated mask images.

In other words, since the input of the first BCN is an image and the output is also an image, one learns himself, and a plurality of composite products in the BCN structure learn features of the image in the learning process. That is, meaningless things (e.g. noise) will naturally disappear in the learning process, and only meaningful things (e.g. plaques) will remain.

In this way, if the image is input to the layers learned by the self-encoding technique, the same image may appear to be output, but in reality, it is possible to acquire an image in which noise is reduced and the plaque remains as it is.

The pre-learned BCN and the unlearned BCN are then stacked, and in the stacked structure, the input is an image and the output is a mask and the stacked layers are learned in this state. When the learning is completed through this process, a partial weight-update occurs in the pre-learned first BCN, where the weights of all the composite product layers in the second BCN are all the weights learned from the random-weighted values. Is updated. Therefore, when inputting a new image not used for learning, a mask may be output through all learned weights.

FIG. 6 is a flowchart schematically illustrating the overall flow of a method for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention.

When machine learning is performed on the reconstructed horizontal plane / sagittal plane / coronal plane image based on the double simple compound network (BCN) 50, the mask image for each channel (ch1, ch2 and ch3) as a result of the machine learning. The mask images 61, 62, and 63 may be generated, and a cross-sectional image 70 of the coronary vessel may be obtained based on the mask image 60 generated as described above. For reference, the entire flow chart of FIG. 6 illustrates the regionalization of the coronary lumen.

In this case, the probability of each pixel of the output image for each pixel of the input image may be obtained through the trained simplified composite product network (BCN) structure, and thus the probability maps f _a = p _a (for horizontal, sagittal and coronal images). y | x), f _c = p _c (y | x), and f _s = p _s (y | x) can be obtained. In addition, to obtain a more robust prediction result, an amplified feature that amplifies the probability for each pixel is defined as follows.

f _m : = exp (p _a (y | x)) + exp (p _c (y | x)) + exp (p _s (y | x))

As such, a feature vector {f _a , f _c , f _s , f _m } for determining the label for each pixel is defined, and label determination for each pixel of the output image uses a gradient boosting technique. The gradient boosting model is trained on the extracted feature vectors and label true values.

When the mask image 60 is generated as a result of the machine learning, and based on the acquired cross-sectional image 70 of the blood vessel (S250), S260 is a cross-sectional image of the acquired coronary inner wall and the obtained cross-sectional image of the coronary outer wall. Corresponds to the step (S260) of analyzing the plaque in the coronary artery using. In general, except for the region of the inner wall of the coronary artery from the region of the outer wall of the coronary artery, the remaining region may correspond to the plaque region.

For reference, the illustrated series of steps S210 to S260 of FIG. 2 may be performed separately or integrally with respect to the coronary inner wall and the coronary outer wall, respectively.

7 is a schematic block diagram of an image processing apparatus 700 configured to perform a method for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention.

As shown in FIG. 7, an image processing apparatus 700 configured to execute a method for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention is a computed tomography (CT) image. An image receiver 720 configured to receive a medical image including an image, an image processor 730 configured to process a cardiac image received by the image receiver 720, and a coronal tube output from the image processor 730. A controller configured to display at least one of an artery inner wall image, an outer wall image, and a plaque image, and a controller configured to control the image receiver 720, the image processor 730, and the display 740. 710 may be included.

For reference, each element of the image processing apparatus 700 of the block diagram illustrated in FIG. 7 is merely an example for easy understanding of the present disclosure, and elements other than the element illustrated in FIG. 7 may be used as the image processing apparatus 700. Will be included in the

Here, the image processing unit 730 is configured to adjust the window width (WW) and the window level (WL) of the CT image to generate n-channel data, where n-channel data generation unit 731, where n Is at least two natural numbers; An image reconstruction unit 732, configured to reconstruct an image at right angles from each of the n-channel data to reconstruct it into a horizontal plane image, a sagittal plane image, and a coronal plane image; A machine learning unit (733) configured to machine learn based on a composite product neural network (CNN) for the reconstructed images; A mask image is generated from a horizontal plane image, a sagittal plane image, and a coronal plane image, respectively, from an image including at least one of a machine-learned orthogonally reconstructed image, and a cross-sectional image of a coronary vessel is obtained based on the generated mask images. An image acquisition unit 734, configured to be configured to be; And a plaque analysis unit 735 configured to analyze plaque in the coronary artery using the obtained cross-sectional image of the coronary inner wall and the acquired cross-sectional image of the coronary outer wall.

Since specific functions and operations of the units 731 to 735 of the image processor 730 have been described above, the description thereof will be omitted.

In addition, the controller 710 may be configured to collectively control the image receiver 720, the image processor 730, and the display 740. For example, the controller 710 may be implemented as a single controller, or may be implemented as a plurality of micro-controllers.

Experiment Results and Analysis

For reference, the present inventors experimented with the performance of the machine-learned model using 10 coronary artery data, the results obtained with 1 channel image and 3 channel image to confirm the effect on the 3-channel data Was compared. In addition, as in the model training step, the experiment was performed by reconstructing the input data into three-channel data and reconstructing the image into horizontal, sagittal, and coronal images. To evaluate the performance of the model, we compared Dice Similarity Coefficient (DSC) with the true value and the results are shown in the table below.

In order to verify the clinical validity of the proposed model, two modalities are used to match the plaque volume obtained from the simple synthetic product network (BCN) model with the plaque region in the IVUS, which is the clinical true value, and the plaque region in the CT image. Plaques were positioned in the image. The table below shows the results of comparing the plaque volume measured by simple synthetic product network (BCN) with the plaque volume measured by clinical true value IVUS.

As can be seen through the above descriptions and techniques in connection with the present invention, according to the method and an image processing apparatus for analyzing plaque in a computed tomography (CT) image according to an embodiment of the present invention, Running can be used to more easily and accurately analyze coronary plaques in computed tomography (CT) images.

In addition, according to the method and the image processing apparatus for analyzing the plaque in the computed tomography (CT) image according to an embodiment of the present invention, in analyzing the plaque of the coronary artery, accurate results to be utilized in the actual clinical environment It is possible to derive

The above-described embodiments of the present invention can be written as a program that can be executed in a computer, and can be implemented in a general-purpose digital computer that operates the program using a computer-readable recording medium.

Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, computer readable media may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Communication media typically includes computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave, or other transmission mechanism, and includes any information delivery media.

The foregoing description of the present invention is intended for illustration, and it will be understood by those skilled in the art that the present invention may be easily modified in other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present invention is shown by the following claims rather than the above description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present invention. something to do.

According to the present invention, a deep learning technique is used to automatically generate a mask of an inner wall and an outer wall of a coronary artery from a computed tomography image, thereby enabling easier and accurate analysis of plaques in a coronary artery. It is available.

Claims (10)

1. As a method for analyzing plaque in medical images,

   Receiving a medical image including a computed tomography (CT) image;

   Generating n-channel data by adjusting a window width (WW) and a window level (WL) of the CT image, wherein n is a natural number of two or more;

   Reconstructing an image into an orthogonal image in each of the n-channel data and reconstructing the image into a horizontal image, a sagittal image, and a coronal image;

   Machine learning the reconstructed images based on a convolutional neural network (CNN); And

   Generating a mask image from an image including at least one of the machine-learned orthogonally reconstructed images, and obtaining a cross-sectional image of the coronary vessel based on the generated mask images

   Including,

   Characterized in that the steps are performed separately or integrally on the coronary inner wall and coronary outer wall, respectively,

   Method for analyzing plaque in medical images.
2. The method of claim 1,

   Analyzing the plaque in the coronary artery using the obtained cross-sectional image of the coronary inner wall and the acquired cross-sectional image of the coronary outer wall,

   Method for analyzing plaque in medical images.
3. The method of claim 1,

   Generating n-channel data by adjusting the window width (WW) and the window level (WL) of the CT image,

   Generating the n-channel data by setting WW ₁ and WL ₁ for coronary lumen observation, WW ₂ and WL ₂ for calcium analysis and WW ₃ and WL ₃ for lipid plaques, Where n is 3,

   Method for analyzing plaque in medical images.
4. The method of claim 1,

   The composite product neural network (CNN) is configured by successively stacking two brief convolutional networks (BCN),

   Method for analyzing plaque in medical images.
5. The method of claim 4, wherein

   Pre-learning is performed using an auto-encoder in a preceding first BCN of the two simple synthetic product networks (BCN),

   Method for analyzing plaque in medical images.
6. An image processing apparatus configured to perform the method according to any one of claims 1 to 5,

   An image receiver configured to receive a computed tomography (CT) image;

   An image processor configured to process a heart image received by the image receiver;

   A display unit configured to display at least one of a coronary inner wall image, an outer wall image, and a plaque image output from the image processor;

   A controller configured to control the image receiver, the image processor, and the display unit

   Including,

   The image processor,

   An n-channel data generation unit, configured to adjust the window width WW and window level WL of the CT image to generate n-channel data, where n is a natural number of two or more;

   An image reconstruction unit configured to reconstruct an image at right angles from each of the n-channel data to reconstruct it into a horizontal plane image, a sagittal plane image, and a coronal plane image;

   A machine learning unit configured to machine learn based on a composite product neural network (CNN) for the reconstructed images; And

   An image acquisition unit, configured to generate a mask image from an image including at least one of the machine-learned right-angle reconstructed images, and obtain a cross-sectional image of the coronary vessel based on the generated mask images

   Including,

   Image processing device.
7. The method of claim 6,

   The image processing unit may further include a plaque analysis unit configured to analyze plaque in the coronary artery using the obtained cross-sectional image of the coronary inner wall and the obtained cross-sectional image of the coronary outer wall.

   Image processing device.
8. The method of claim 6,

   The n-channel data generation unit sets the n-channel by setting WW ₁ and WL ₁ for coronary lumen observation, WW ₂ and WL ₂ for calcium analysis, and WW ₃ and WL ₃ for lipid plaques. Generate data, where n is 3,

   Image processing device.
9. The method of claim 6,

   The composite product neural network (CNN) is configured by successively stacking two brief convolutional networks (BCN),

   Image processing device.
10. The method of claim 9,

    Pre-learning is performed using a self-encoder in the preceding first BCN of the two simple synthetic product networks (BCN),

    Image processing device.

Images