Spectral clustering for detecting atypical cardiac coronary arteries

Abstract

A method of spectral clustering includes obtaining a coronary artery image dataset comprising a plurality of voxels of a volume of interest around a heart; processing the plurality of voxels to order the plurality of voxels for likelihood of characterizing a blood vessel; identifying similarities between pairs of voxels to quantify link strength between voxels in each pair of voxels; spectrally clustering the plurality of voxels based on the likelihood of characterizing the blood vessel and a link strength between the pairs of voxels; selecting at least one voxel subtree based on the spectral clustering; classifying each sub-tree based on the coronary artery image data set; and reconstructing a representation of at least one blood vessel in the volume of interest as a first reconstruction to include at least one subtree classified based on the coronary artery image dataset.

Description

Background Art

[0001] The assessment and quantification of coronary arteries in diagnostic and / or preoperative computed tomography (CT) image volumes are performed regularly and require at least semi - automatic segmentation of the coronary arteries. Most anatomical constellations of coronary arteries are characterized by type, and the current assessment and quantification of such coronary arteries are performed by type. Standard machine - learning models can be trained on conventional computed tomography scans of coronary artery constellations of known types in large image and annotation databases. The inability to identify, segment, and develop atypical anatomical constellations may lead to the inability to diagnose and treat certain cardiac events.

[0002] However, the assessment of atypical coronary artery configurations is not widespread in current image and annotation databases, let alone well - curated and annotated. Additionally, spectral and photon - counting computed tomography scans with improved spectral and / or spatial resolution have not been widely used in such databases, let alone curated or annotated for the assessment of atypical coronary artery configurations.

[0003] There is a need to segment atypical configurations of coronary arteries. Such atypical configurations of coronary arteries include, for example, collateral vessels, bypasses, and coronary artery bypass grafts (CABGs). Summary of the Invention

[0004] According to one aspect of the present disclosure, a method of spectral clustering includes: obtaining a coronary artery image data set including a plurality of voxels of an interest volume around a heart; processing the plurality of voxels to rank the plurality of voxels for the likelihood of characterizing a blood vessel; identifying similarities between pairs of voxels to quantify the link strength between the voxels in each pair; performing spectral clustering on the plurality of voxels based on the likelihood of characterizing a blood vessel and the link strength between the pairs of voxels; selecting at least one voxel subtree based on the spectral clustering; classifying each subtree based on the coronary artery image data set; and reconstructing a representation of at least one blood vessel in the interest volume as a first reconstruction to include at least one subtree classified based on the coronary artery image data set.

[0005] In another aspect of the present disclosure, a system for spectral clustering includes: a memory storing instructions; and a processor executing the instructions. When executed by the processor, the instructions cause the system to: obtain a coronary artery image dataset including a plurality of voxels in a volume of interest around the heart; process the plurality of voxels to rank the plurality of voxels according to the likelihood of representing a blood vessel; identify similarities between pairs of voxels to quantify the link strength between the voxels in each pair; perform spectral clustering on the plurality of voxels based on the likelihood of representing a blood vessel and the link strength between the pairs of voxels; select at least one voxel subtree based on the spectral clustering; classify each subtree based on the coronary artery image dataset; and reconstruct a representation of at least one blood vessel in the volume of interest as a first reconstruction to include at least one subtree classified based on the coronary artery image dataset.

[0006] In another aspect of the present disclosure, a tangible non-transitory computer-readable storage medium stores a computer program. The computer program, when executed by a processor, causes a system: A system for spectral clustering includes a memory storing instructions; and a processor executing the instructions. When executed by the processor, the instructions cause the system to: obtain a coronary artery image dataset including a plurality of voxels in a volume of interest around the heart; process the plurality of voxels to rank the plurality of voxels according to the likelihood of representing a blood vessel; identify similarities between pairs of voxels to quantify the link strength between the voxels in each pair; perform spectral clustering on the plurality of voxels based on the likelihood of representing a blood vessel and the link strength between the pairs of voxels; select at least one voxel subtree based on the spectral clustering; classify each subtree based on the coronary artery image dataset; and reconstruct a representation of at least one blood vessel in the volume of interest as a first reconstruction to include at least one subtree classified based on the coronary artery image dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Example embodiments are best understood from the following detailed description when read in conjunction with the accompanying drawings. It should be emphasized that the various features are not necessarily drawn to scale. In fact, the dimensions may be arbitrarily increased or decreased for the sake of discussion. Where applicable and practical, like reference numerals refer to like elements.

[0008] Figure 1 A system for spectral clustering for detecting atypical coronary arteries of the heart according to a representative embodiment is shown.

[0009] Figure 2 A method for spectral clustering for detecting atypical coronary arteries of the heart according to a representative embodiment is shown.

[0010] Figure 3Another method of spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0011] Figure 4 An example of power iteration for spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0012] Figure 5 An example of projection for spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0013] Figure 6A A graph of power iteration for spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0014] Figure 6B It is shown according to a representative embodiment Figure 6A A graph of the maximum power size within an iteration of the power iteration in

[0015] Figure 7 An exemplary application on the pulmonary vasculature tree in spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0016] Figure 8 An example diffusion process of network ranking for evolving candidate voxels in spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0017] Figure 9 A computer system according to another representative embodiment is shown, on which a method of spectral clustering for detecting atypical cardiac coronary arteries is implemented. Detailed Description

[0018] In the following detailed description, for purposes of explanation and not limitation, representative embodiments disclosing specific details are set forth in order to provide a thorough understanding of embodiments according to the present teachings. However, other embodiments consistent with the present disclosure that depart from the specific details disclosed herein are still within the scope of the claims. Descriptions of known systems, devices, materials, operating methods, and manufacturing methods may be omitted so as not to obscure the description of the representative embodiments. Nevertheless, systems, devices, materials, and methods within the knowledge of those of ordinary skill in the art are within the scope of the present teachings and may be used in accordance with the representative embodiments. It should be understood that the terms used herein are for the purpose of describing particular embodiments only and are not intended to be limiting. The definitions and interpretations of the terms herein supplement the technical and scientific meanings of the terms commonly understood and accepted in the technical field of the present teachings.

[0019] It should be understood that although the terms first, second, third, etc. may be used herein to describe various elements or components, these elements or components should not be limited by these terms. These terms are only used to distinguish one element or component from another. Thus, without departing from the teachings of the inventive concept, the first element or component discussed below may be referred to as the second element or component.

[0020] As used in the specification and claims, the singular forms of the terms "a" and "the" are intended to include both the singular and the plural forms, unless the context clearly dictates otherwise. Additionally, the terms "comprising", "including" and / or "containing" and / or similar terms used in this specification specify the presence of the stated features, elements and / or components, but do not preclude the presence or addition of one or more other features, elements, components and / or groups thereof. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0021] Unless otherwise specified, when an element or component is referred to as "connected to", "coupled to" or "adjacent to" another element or component, it should be understood that the element or component can be directly connected or coupled to the other element or component, or there may be intermediate elements or components. That is, these terms and similar terms cover situations where one or more intermediate elements or components can be employed to connect two elements or components. However, when an element or component is referred to as "directly connected" to another element or component, this only covers the situation where the two elements or components are connected to each other without any intermediate or intervening elements or components.

[0022] Accordingly, through one or more of the various aspects, embodiments and / or specific features or sub-components of the present disclosure, the present disclosure aims to provide one or more advantages as specifically pointed out below.

[0023] As described herein, atypical anatomical configurations can be segmented and developed. Additionally, the advanced spectral and spatial resolution voxel features of emerging imaging modalities (such as spectral and photon counting computed tomography) can be utilized. Examples of atypical anatomical configurations that can be segmented and developed using emerging imaging modalities include, for example, collateral vessels, bypasses, and coronary artery bypass grafts. The identification, segmentation, and development of atypical anatomical configurations may result in the ability to diagnose and treat cardiac events that would otherwise be undiagnosable and untreatable.

[0024] Figure 1 System 100 for spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0025] Figure 1System 100 therein is a system for spectral clustering of atypical cardiac coronary arteries and includes components that may be provided together or may be distributed. System 100 includes an imaging system 110, a computer 140, and a display 180. Computer 140 includes a controller 150, and controller 150 includes a memory 151 and a processor 152. In some embodiments, imaging system 110 is equipped with computer 140 and display 180, such as in a hospital complex or other medical environment. In other embodiments, imaging system 110 is provided remotely from computer 140 and display 180, such as when computer 140 represents a cloud implementation of the functions described herein and display 180 is provided in a separate facility.

[0026] Imaging system 110 represents an imaging system for performing cardiac coronary artery imaging. In the description herein, imaging system 110 is primarily referred to as a computed tomography imaging system. However, the teachings herein apply to all three-dimensional imaging modalities, including computed tomography enterography (CTE), magnetic resonance imaging (MRI), and ultrasound.

[0027] Computer 140 represents a desktop or server implemented in a facility or in the cloud. Figure 9 The computers that may be used to implement computer 140 are depicted, but computer 140 may include more or fewer elements than Figure 1 or Figure 9 shown therein.

[0028] The controller 150 includes at least a memory 151 that stores instructions and a processor 152 that runs the instructions. The controller 150 runs the instructions to perform a method based on coronary artery image data such as from computed tomography images. The method includes obtaining a coronary artery image data set including a plurality of voxels of an interested volume around the heart. That is, the image data is three-dimensional image data including a plurality of voxels of an interested volume around the heart. The method also includes processing the voxels to rank the voxels according to the likelihood of characterizing a blood vessel. The likelihood of characterizing a blood vessel may be referred to as vascularity because the basic determination is whether the voxel represents a part of a blood vessel included in the interested volume. The method also includes identifying the similarity between paired voxels to quantify the link strength between the voxels in each pair of voxels. The method implemented using the controller 150 searches for paired voxels, each pair of voxels representing a part of a blood vessel included in the interested volume. The method also includes performing spectral clustering on the plurality of voxels based on the likelihood of characterizing a blood vessel and the link strength between paired voxels. Voxels that may characterize a blood vessel may be clustered. Clustering may mean dividing the voxels into two or more groups. The method implemented using the controller 150 also includes selecting at least one voxel subtree based on spectral clustering. The subtree may include a set of adjacent voxels that are considered likely to represent a blood vessel. The method also includes classifying each subtree based on the coronary artery image data set. The method also includes reconstructing a representation of at least one blood vessel in the interested volume into a first reconstruction to include at least one subtree classified based on the coronary artery image data set.

[0029] The display 180 may be local to the computer 140 or may be remotely connected to the computer 140 (e.g., via a local area network or via a wide area network such as the Internet). When locally connected, the display 180 may be connected to the computer 140 via a local wired interface such as an Ethernet cable or via a local wireless interface such as a Wi-Fi connection. The display 180 may interface with other user input devices (including a mouse, a keyboard, a thumb wheel, etc.) through which a user can input instructions. The display 180 may be a monitor such as a computer monitor, a display on a mobile device, an augmented reality display, a television, an electronic whiteboard, or another screen configured to display an electronic image. The display 180 may also include one or more input interfaces, such as those mentioned above that may be connected to other elements or components, and an interactive touch screen configured to display prompts to the user and collect touch inputs from the user.

[0030] The controller 150 may also include interfaces, such as a first interface, a second interface, a third interface, and a fourth interface. One or more interfaces may include ports, disk drives, wireless antennas, or other types of receiver circuits that connect the controller 150 to other electronic components. One or more interfaces may also include a user interface, such as buttons, keys, a mouse, a microphone, a speaker, a display separate from the display 180, or other elements that a user may use to interact with the controller 150 (e.g., input instructions and receive outputs).

[0031] The controller 150 may directly perform some of the operations described herein and may indirectly implement other operations described herein. For example, the controller 150 may indirectly control operations, such as by generating and transmitting content to be displayed on the display 180. The controller 150 may directly control other operations based on inputs received from electronic components and / or the user via the interfaces, such as logical operations performed by the processor 152 running instructions from the memory 151. Thus, the processes implemented by the controller 150 when the processor 152 runs instructions from the memory 151 may include steps not directly performed by the controller 150.

[0032] Figure 2 A method for spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0033] Figure 2 The method may be executed by a system 100 including a computer 140 having a controller 150. At S210, the method begins with obtaining a coronary artery image. The coronary artery image is composed of coronary artery image data including a volume of interest around the heart. The coronary artery image may be received directly or indirectly by the computer 140 from the imaging system 110. For example, the coronary artery image may be transferred in the same facility (such as a hospital complex). Alternatively, the coronary artery image may be uploaded to the cloud and received at a server in a data center, in which case the server in the data center is the computer 140 that executes Figure 2 most of the methods in.

[0034] At S220, the voxels are processed. The method of Figure 2 is executed to identify and connect vascular structures in the volume of interest around the heart without prior anatomical training. The coronary artery configurations identified and connected in Figure 2 may include typical and atypical coronary artery configurations. At S220, multiple voxels are processed to rank the voxels according to the likelihood of representing a blood vessel. The processing at S220 may include filtering the multiple voxels based on expected image attributes of the blood vessel. The processing of the multiple voxels uses power iteration. The likelihood of a voxel representing a blood vessel may be determined based on a vasculature filter applied to the voxel.

[0035] At S230, the similarity between paired voxels is identified. Identifying the similarity quantifies the link strength between the voxels in each pair of voxels. The similarity is identified based on the expected image attributes of blood vessels. The similarity can be identified based on the proximity of the coordinates of each voxel in a pair of voxels and the directionality of each voxel in a pair of voxels. Proximity reflects how close a pair of voxels are to each other, including whether the voxels are adjacent. Directionality reflects whether the voxels belong to blood vessels flowing in the same direction. For example, the similarity can be identified for paired adjacent voxels rather than for paired distant voxels.

[0036] As an example similarity measure, consider s_ij = n||x i - x j || / ρ·k(v_i, v_j)·d(f_i, f_j), where x_i is the integer coordinate of voxel i, ρ is the spatial influence range in the voxel, for example, ρ = 1, n is a stationary covariance with compact support, used to quantify the local neighborhood. v_i is the vascularity of voxel i. v_j is the vascularity of voxel j, k is the similarity between the vascularities of voxels i and j. d is the similarity between other image features f_i, such as the similarity in image voxel intensity, the similarity in various spectral computed tomography channels, the vector alignment between the local Hessian eigenvectors, the vector alignment between the local Hessin eigenvector and the node-connecting spatial vector direction, the similarity between the vascularity and the local Hessian eigenvalue, etc.

[0037] At S240, the method includes performing spectral clustering. Spectral clustering is performed based on the likelihood of representing blood vessels and the connection strength between pairs of blood vessels. Spectral clustering can be performed based on a matrix of multiple voxels in the volume of interest and using power iteration clustering. As an example, the more likely two adjacent voxels are to represent blood vessels, the more likely these two adjacent voxels are to be clustered together. However, voxel clustering is not limited to multiple voxels adjacent to each other.

[0038] As an example of power iteration clustering at S240, the maximum eigenvector can be iteratively found by power iteration, making it tractable even for very large matrices. When all voxels have equal ranking, power iteration can be calculated starting from an initial unit vector, and then the vector containing the ranking of all voxels is updated using the graph links and the current weights of all local neighborhood voxels. The voxels that are part of the flow path appear as the ranking scores increase. The voxels at the root of the subtree have the highest scores because it is the confluence location of multiple paths.

[0039] In some embodiments, during power iteration towards the random walk convergence state, the ranking score s of each voxel can correspond to its entry in the eigenvector, and is updated from its current state s^k to state s^k+1 using matrix-vector multiplication. For a sparse matrix L, this can be reduced to a low-computation update of the node scores based on the current neighbor scores. For numerical stability, the state vector can be normalized after each iteration.

[0040] In some embodiments, at each iteration state of the clustering, the current eigenvector elements can be interpreted as graph node scores, and the graph can be clustered into subgraphs, such as using k-means clustering on the scores, graph pattern search, or finding connected components above a certain score threshold.

[0041] The processing at S220, the similarity identification at S230, and the spectral clustering at S240 can be performed on a one-to-one basis by the kernels of a graphics processing unit. For example, the processor 152 can be or include a graphics processing unit with dozens, hundreds, or thousands of cores for processing image data of multiple pixels and multiple voxels. Figure 2 The iterative nature of parts of the method benefits from parallel implementations, such as using a graphics processing unit or a server central processing unit (CPU) supporting SIMD, or being mapped to hardware-optimized sparse matrix operations.

[0042] In some embodiments, the parameters for identifying similarity can be optimized. If an annotated training set of the image and the vascular clusters is available through a fidelity criterion such as mean squared error (MSE) or mean absolute error (MAE), the parameters of the similarity metric can be optimized.

[0043] In some embodiments, the user can be provided with the ability to interactively seed specific blood vessels as a guide for the approximate calculation of spectral clustering. For example, the opening of the left anterior descending branch on the aorta can be specified by the user, and the approximate calculation of spectral clustering can be guided by starting from a non-uniform vector instead of a uniform or random initialization in the power iteration.

[0044] At S250, one or more subtrees are selected. The voxel subtrees are selected based on spectral clustering. The selection of the subtrees can involve depicting the voxels including the subtrees and implicitly identifying the boundaries of the subtrees in three dimensions.

[0045] At S260, each selected subtree is classified. The classification at S260 is based on a coronary artery image dataset. The classification of the subtree can be a coronary vessel or a pulmonary vessel. Possible classifications of the subtree can include typical major coronary artery segments such as LAD, LCD, RCD, Ramus, etc.; atypical coronary arteries as anatomical variants; coronary veins; pulmonary veins or arteries; artificial bypass segments; artificial wires such as pacemaker wires; and / or possible image artifacts. The classified subtrees can be labeled. Of course, the classification is not limited to the types listed above, and the classification can involve more or fewer differentiable types than those listed above.

[0046] At S270, Figure 2 the method includes reconstructing a blood vessel as a first reconstruction. Specifically, at S270, a representation of at least one blood vessel in the volume of interest is reconstructed to include at least one subtree classified based on the coronary artery image dataset. The reconstruction includes segmenting atypical cardiac coronary arteries identified by Figure 2 spectral voxel map clustering in the method. The reconstruction may also include segmenting typical cardiac coronary arteries identified by spectral voxel map clustering. Figure 2 The method of the method allows for the detection and segmentation of coronary artery configurations that are not easily learned by standard machine learning techniques. Examples of such coronary artery configurations include those of patients after surgical remodeling or after artificial bypass.

[0047] At S280, a second reconstruction is generated. The second reconstruction can be performed independently of the execution of S220 to S270, but is performed based on performing a conventional analysis on typical cardiac coronary arteries. That is, the second reconstruction in the volume of interest can be performed without performing spectral clustering and selection. The conventional analysis can include applying a trained artificial intelligence model to identify a predetermined type of blood vessel.

[0048] At S290, a comparison is made between the first reconstruction and the second reconstruction. The comparison at S290 can be performed to identify any atypical coronary artery configurations from the voxel subtrees selected at S260 based on spectral clustering. In the case where the second reconstruction should result in a conventional segmentation of the coronary vessels in the volume of interest, the first reconstruction can relatively show atypical coronary vessels. In addition to the machine learning algorithm, an analysis algorithm described by Figure 2 the method can also be run to compare the results and mark uncertainties when differences occur.

[0049] Figure 2 The method can be used for preprocessing before coronary artery intervention. Figure 2 The method allows for the separation of coronary artery and pulmonary vessel subtrees surrounding the heart.

[0050] Figure 3Another method for spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0051] Figure 3 The method begins at S310 by detecting a region of interest that may be the volume of interest. Identifying the region of interest involves segmenting the cardiac chamber into a wide volume of images of interest around the heart. The segmentation for detecting the region of interest can be performed using model-based segmentation (MBS) or other machine learning-based semantic segmentation. For example, machine learning-based semantic segmentation can be performed by a deep convolutional neural network.

[0052] Figure 3 The method treats all voxels of the volume of images of interest as graph nodes. Each voxel is considered to be fully connected to its local neighbors, but each neighbor has a continuous (rather than binary) link weight. The link weight depends on vascularity features, multispectral similarity, and radial and directional affinity.

[0053] At S320, the attributes of each voxel are determined. The attributes can include characteristics of each voxel that can reflect the likelihood that the voxel represents a blood vessel in the coronary artery configuration. The determination at S320 can be performed as a vascularity filter response. Considering all spectral channels, a single-channel image containing the vascularity filter response for each voxel in the volume of interest can be generated. The vascularity filter response can include magnitude, radius, and direction estimates. The response can be generated as a function / eigenvalue of the Hessian matrix of the second derivative, optionally with additional provisions for image noise.

[0054] In some embodiments, a trained artificial intelligence model can be used to implement the vascularity filter. The vascularity filter can be implemented as a machine learning component, for example, by using a convolutional neural network, rather than being implemented as an analytical filter function. The vascularity filter can be trained according to an available dataset (e.g., according to annotated coronary arteries).

[0055] At S330, neighborhood links are identified. The neighborhood link is a similarity measure calculation in the local neighborhood of voxels in the volume of interest. For each pair of voxels within a finite neighborhood, a symmetric similarity measure s_ij between voxels i and j is calculated to quantify the link strength between the two voxels. The symmetric similarity measure can be defined as positive definite and semi-definite.

[0056] Approximate spectral clustering involves implicitly requiring the Laplacian matrix L or variants such as "normalized" or "random walk" between all voxels in the volume of interest. The matrix can be given by L = D - S, where S = (s_ij) is the similarity matrix and D = diag(S1) is the (diagonal) degree matrix. Spectral clustering proceeds by computing the first few eigenvectors of the (very sparse) Laplacian matrix. In practice, these algorithms rely on matrix-vector multiplications (MVMs) with the Laplacian matrix, such as power iteration, Lanczos, etc.

[0057] At S340, iterative eigenvector approximation is performed. Iterative eigenvector approximation is the iterative identification of significant through-flow voxels and root tree voxels. These voxels are prominent due to their most "influential" role in the voxel graph.

[0058] In some embodiments, once identified, low-ranked voxels can be excluded from the iteration. Excluding low-ranked voxels can accelerate Figure 3 the processing in the method.

[0059] At S350, significant components are selected. S350 involves the selection of subtrees. Separated local ranking score peak positions can be selected and used as seeds for constructing subtrees from all "upstream" voxels.

[0060] At S360, the selected components are classified. Rule-based subtree classification can use heuristics, i.e., pulmonary vessels or coronary arteries. Classification as pulmonary vessels can be based on being embedded in lung tissue and can use semantic segmentation (labeling). Classification as coronary arteries can be based on having endpoints appear somewhere in the aorta and myocardium.

[0061] Figure 3 The method is used for preprocessing / normalization and vessel segment clustering. For preprocessing / normalization, standard vessel filters produce uncalibrated responses based on varying image characteristics such as contrast, noise, and resolution. In contrast, the eigenvectors of the Laplacian performed at S340 are normalized vectors such that the voxel-by-voxel response is determined by topology, flow, confluence, stenosis, etc. For vessel segment clustering, the scattered vessel filter responses are aggregated into natural clusters representing vessel segments, which can then be classified and further processed based on their global characteristics. The natural clusters are connected by flow-determined affinities. The size of the natural clusters increases with iteration.

[0062] Figure 2 The method and Figure 3 The method describe global analysis algorithms with few parameters. A large amount of machine learning training data is not required, which helps avoid potential requirements that may accompany machine learning training data such as annotation, sampling, imaging protocol coverage, and regulatory work. In fact, by Figure 2 The method andFigure 3 The analysis algorithm provided by the method can be used to generate and / or accelerate the semi-automatic management of "ground truth" for the training of artificial intelligence algorithms.

[0063] Figure 2 and Figure 3 Embodiments of are mainly described with respect to spectral computed tomography. However, other imaging modalities can use the spectral clustering described herein. Other imaging modalities include magnetic resonance imaging, ultrasound, single photon emission computed tomography, positron emission tomography, etc. Additionally, the teachings herein are not limited to the cardiac coronary vasculature as the volume of interest. Instead, the teachings herein apply to other anatomical trees, such as the pulmonary vasculature tree of veins and arteries and the lobar, segmental bronchial airway tree, hepatic vasculature tree, etc.

[0064] Figure 4 Shows an example of the power iteration of spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment.

[0065] Figure 4 Shows an example of the power iteration on the Laplacian matrix in the cardiac volume of interest, shown as the maximum projection perpendicular to the long axis of the left ventricle. The exemplary figure consists of 6,677,315 nodes (= the length of the eigenvector), which have links in a 7×7×7 neighborhood, with a maximum of 171 links and an average of 33 neighbor links per node. The top left panel shows all the vascular filter responses. The top middle panel shows the component connected to the ascending aorta (segmented by model-based segmentation (MBS)). The top right panel and all three bottom panels show the iterations, where the connected components are formed by fractions above a threshold of 2× the average fraction and are connected to the ascending aorta.

[0066] Figure 5 Shows an example of the projection of spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment.

[0067] Figure 5 Shows the exemplary axial, coronal, and sagittal maximum projections with respect to the left ventricle - long axis fraction after 8 power iterations of the cluster / component connected to the ascending aorta.

[0068] Figure 6A Shows a graph of the power iteration of spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment. Figure 6B Shows according to a representative embodiment Figure 6A a graph of the maximum power magnitude during the iteration process of the power iteration in.

[0069] In Figure 6A the power iteration on the Laplacian matrix in the cardiac volume of interest shows a reduction in the connected components with fractions above 2× the average fraction. InFigure 6B In it, the maximum component size also decreases with iteration because more graph nodes have vanishing scores and only significant nodes remain significantly higher than (2x) the average score.

[0070] Figure 7 An example application of the pulmonary vascular tree in spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0071] In Figure 7 on the left side, a dense voxel candidate input is shown, while on the right side, after iteration, the network ranking of the voxels is shown as brightness, showing the most influential stump in the subtree. In Figure 7 on the right side of, the most influential stump in the subtree corresponds to the sagittal projection.

[0072] Figure 8 An example diffusion process of the network ranking of evolving candidate voxels in spectral clustering for detecting atypical cardiac coronary arteries according to a representative embodiment is shown.

[0073] Figure 8 The diffusion process in is an example iteration of the evolving network ranking of all candidate voxels and is based on the sagittal projection as in Figure 7 on the right side.

[0074] Figure 9 A computer system according to another representative embodiment is shown, on which a method for spectral clustering for detecting atypical cardiac coronary arteries is implemented.

[0075] Referring to Figure 9 , computer system 900 includes a set of software instructions that can be run to cause computer system 900 to perform any method or computer-based function disclosed herein. Computer system 900 can operate as a stand-alone device or can be connected to other computer systems or peripheral devices, for example, using network 901. In an embodiment, computer system 900 performs logical processing based on digital signals received via an analog-to-digital converter.

[0076] In a networked deployment, computer system 900 operates in the capacity of a server or as a client user computer in a server-client user network environment, or as a peer computer system in a peer-to-peer (or distributed) network environment. Computer system 900 can also be implemented as or incorporated into various devices, such as a workstation including a controller, a fixed computer, a mobile computer, a personal computer (PC), a laptop computer, a tablet computer, or any other machine capable of running a set of software instructions (sequential or otherwise) that specify actions to be taken by the machine. Computer system 900 can be incorporated as a device or within a device that is in turn within an integrated system that includes additional devices. In an embodiment, computer system 900 can be implemented using an electronic device that provides voice, video, or data communication. Additionally, although computer system 900 is shown in the singular, the term "system" should also be considered to include any collection of systems or subsystems that operate individually or jointly to execute one or more sets of software instructions to perform one or more computer functions.

[0077] As Figure 9 shown, computer system 900 includes a processor 910. Processor 910 can be considered a representative example of a processor of a controller and runs instructions to implement some or all aspects of the methods and processes described herein. Processor 910 is tangible and non-transitory. As used herein, the term "non-transitory" should not be construed as a permanent property of a state, but rather as a property of a state that will persist for a period of time. The term "non-transitory" expressly negates transient properties, such as those of a carrier wave or signal, or other forms that exist only temporarily at any given time and place. Processor 910 is an article of manufacture and / or a machine component. Processor 910 is configured to run software instructions to perform functions as described in various embodiments herein. Processor 910 can be a general-purpose processor or can be part of an application specific integrated circuit (ASIC). Processor 910 can also be a microprocessor, a microcomputer, a processor chip, a controller, a microcontroller, a digital signal processor (DSP), a state machine, or a programmable logic device. Processor 910 can also be a logic circuit, including a programmable gate array (PGA), such as a field programmable gate array (FPGA), or another type of circuit including discrete gates and / or transistor logic. Processor 910 can be a central processing unit (CPU), a graphics processing unit (GPU), or both. Additionally, any processor described herein can include multiple processors, parallel processors, or both. Multiple processors can be included within a single device or multiple devices, or coupled to a single device or multiple devices.

[0078] As used herein, the term "processor" encompasses an electronic component capable of running a program or machine-executable instructions. A reference to a computing device that includes a "processor" should be construed to include multiple processors or processing cores, such as in a multi-core processor. A processor may also refer to a collection of processors within a single computer system or distributed among multiple computer systems. The term computing device should also be construed to include a collection or network of computing devices, each including one or more processors. A program has software instructions that are run by one or more processors, which may be within the same computing device or may be distributed across multiple computing devices.

[0079] Computer system 900 also includes a main memory 920 and a static memory 930, where the memories in computer system 900 communicate with each other and with processor 910 via bus 908. Either or both of main memory 920 and static memory 930 may be considered representative examples of the memory of the controller, and store instructions for implementing some or all aspects of the methods and processes described herein. The memories described herein are tangible storage media for storing data and executable software instructions, and are non-transitory during the period when the software instructions are stored therein. As used herein, the term "non-transitory" should not be construed as a permanent property of a state, but rather as a property of a state that will persist for a period of time. The term "non-transitory" expressly negates transient properties, such as the properties of a carrier wave or signal or other forms that exist only temporarily at any given time and place. Main memory 920 and static memory 930 are articles of manufacture and / or machine components. Main memory 920 and static memory 930 are computer-readable media from which a computer (e.g., processor 910) can read data and executable software instructions. Each of main memory 920 and static memory 930 may be implemented as one or more of a random access memory (RAM), read-only memory (ROM), flash memory, electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disk, removable disk, magnetic tape, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), floppy disk, Blu-ray disc, or any other form of storage medium known in the art. The memory may be volatile or non-volatile, secure and / or encrypted, insecure and / or unencrypted.

[0080] "Memory" is an example of a computer-readable storage medium. Computer memory is any memory directly accessible by a processor. Examples of computer memory include, but are not limited to, RAM memory, registers, and register files. A reference to "computer memory" or "memory" should be construed to potentially be multiple memories. The memory may be, for example, multiple memories within the same computer system. The memory may also be multiple memories distributed among multiple computer systems or computing devices.

[0081] As shown, computer system 900 also includes a video display unit 950, such as a liquid crystal display (LCD), an organic light emitting diode (OLED), a flat panel display, a solid state display, or a cathode ray tube (CRT). Additionally, computer system 900 includes an input device 960, such as a keyboard / virtual keyboard or a touch-sensitive input screen or voice input with voice recognition, and a cursor control device 970, such as a mouse or a touch-sensitive input screen or a keyboard. Computer system 900 also optionally includes a disk drive unit 980, a signal generation device 990 (such as a speaker or a remote control), and / or a network interface device 940.

[0082] In one embodiment, as Figure 9 shown, disk drive unit 980 includes a computer-readable medium 982 in which a set or multiple sets of software instructions 984 (software) are embedded. The software instruction set 984 is read from the computer-readable medium 982 for execution by the processor 910. Further, the software instructions 984, when executed by the processor 910, perform one or more steps of the methods and processes described herein. In an embodiment, the software instructions 984 reside, in whole or in part, within the main memory 920, the static memory 930, and / or the processor 910 during execution by the computer system 900. Additionally, the computer-readable medium 982 may include the software instructions 984 or receive and execute the software instructions 984 in response to a propagated signal such that a device connected to the network 901 transmits voice, video, or data over the network 901. The software instructions 984 may be sent or received over the network 901 via the network interface device 940.

[0083] In an embodiment, a dedicated hardware implementation such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a programmable logic array, and other hardware components are constructed to implement one or more of the methods described herein. One or more embodiments described herein may use two or more specific interconnected hardware modules or devices to implement functions, and these modules or devices have associated control and data signals that can be transmitted between and through the modules. Thus, the present disclosure encompasses software, firmware, and hardware implementations. Nothing in this application should be construed as implementing or being implementable using only software rather than hardware such as tangible non-transitory processors and / or memories.

[0084] In accordance with various embodiments of the present disclosure, the methods described herein may be implemented using a hardware computer system running a software program. Additionally, in an exemplary non-limiting embodiment, the implementation may include distributed processing, component / object distributed processing, and parallel processing. Virtual computer system processing may implement one or more of the methods or functions described herein, and the processors described herein may be used to support a virtual processing environment.

[0085] Accordingly, spectral clustering for detecting atypical cardiac coronary arteries enables the identification, segmentation, and development of cardiac coronary arteries that might otherwise be missed by machine learning models seeking typical cardiac coronary arteries. The teachings herein enable the ability to diagnose and treat cardiac events that would otherwise go undiagnosed and untreated.

[0086] Although spectral clustering for detecting atypical cardiac coronary arteries has been described with reference to several exemplary embodiments, it should be understood that the words used are words of description and illustration, not of limitation. Changes may be made within the scope of the claims, as presently set forth and modified, without departing from the scope and spirit of spectral clustering for detecting atypical cardiac coronary arteries in its aspects. Although spectral clustering for detecting atypical cardiac coronary arteries has been described with reference to specific means, materials, and embodiments, spectral clustering for detecting atypical cardiac coronary arteries is not intended to be limited to the details disclosed; rather, spectral clustering for detecting atypical cardiac coronary arteries extends to all functionally equivalent structures, methods, and uses within the scope of the claims.

[0087] The illustrations of the embodiments described herein are intended to provide a general understanding of the structure of the various embodiments. These illustrations are not intended to be a complete description of all elements and features of the present disclosure described herein. Many other embodiments may be apparent to those of ordinary skill in the art upon reading the present disclosure. Other embodiments may be utilized and derived from the present disclosure, such that structural and logical substitutions and changes may be made without departing from the scope of the present disclosure. Additionally, the illustrations are merely representative and may not be drawn to scale. Some ratios within the illustrations may be exaggerated while others may be minimized. Accordingly, the present disclosure and the drawings are to be considered illustrative, not restrictive.

[0088] One or more embodiments of the present disclosure may be referred to herein individually and / or collectively as the "invention" merely for convenience and are not intended to voluntarily limit the scope of this application to any particular invention or inventive concept. Moreover, although specific embodiments have been shown and described herein, it should be understood that any subsequent arrangement designed to achieve the same or similar purpose may be substituted for the specific embodiments shown. The present disclosure is intended to cover any and all subsequent adaptations or variations of the various embodiments. Combinations of the above-described embodiments and other embodiments not specifically described herein will be apparent to those of ordinary skill in the art upon reading the specification.

[0089] The abstract of the disclosure is provided to comply with 37 C.F.R. § 1.72(b), and it is understood that it will not be used to interpret or limit the scope or meaning of the claims at the time of filing. Additionally, in the foregoing detailed description, for the purpose of simplifying the disclosure, various features may be combined together or described in a single embodiment. The disclosure should not be construed as reflecting an intention that the claimed embodiments require more features than are expressly recited in each claim. Rather, as reflected in the following claims, the subject matter of the invention may involve less than all of the features of any of the disclosed embodiments. Accordingly, the following claims are incorporated into the detailed description, where each claim independently defines a separately claimed subject matter.

[0090] The foregoing description of the disclosed embodiments is provided to enable a person skilled in the art to practice the concepts described in the disclosure. Accordingly, the subject matter disclosed above is to be considered illustrative, not restrictive, and the claims are intended to cover all such modifications, enhancements, and other embodiments that fall within the true spirit and scope of the disclosure. Thus, to the maximum extent permitted by law, the scope of the disclosure will be determined by the broadest permissible interpretation of the claims and their equivalents, and should not be limited or constrained by the foregoing detailed description.

Claims

1. A method of spectral clustering, comprising: Obtaining a coronary artery image dataset of a plurality of voxels including an interested volume around the heart; Processing the plurality of voxels to rank the plurality of voxels for the likelihood of characterizing blood vessels; Identifying similarities between paired voxels to quantify the link strength between voxels in each pair of voxels; Performing spectral clustering on the plurality of voxels based on the likelihood of characterizing blood vessels and the link strength between the paired voxels; Selecting at least one voxel subtree based on the spectral clustering; Classifying each subtree based on the coronary artery image dataset; and Reconstructing a representation of at least one blood vessel in the interested volume as a first reconstruction to include at least one subtree classified based on the coronary artery image dataset.

2. The method according to claim 1, further comprising: Generating a second reconstruction in the interested volume without performing the spectral clustering and selection; And Comparing the first reconstruction and the second reconstruction to identify an atypical coronary artery configuration based on the at least one voxel subtree selected based on the spectral clustering.

3. The method according to claim 1, wherein, The processing of the plurality of voxels includes filtering the plurality of voxels based on expected blood vessel image attributes.

4. The method according to claim 3, wherein, The similarity is identified based on the expected blood vessel image attributes.

5. The method according to claim 1, wherein, The processing of the plurality of voxels, the identification of the similarity, and the spectral clustering are performed by a core of a graphics processing unit on a one-to-one basis for the plurality of voxels in a power iteration manner.

6. The method according to claim 1, wherein, The similarity is identified based on the proximity of the coordinates of each voxel in a pair of voxels and the directionality of each voxel in a pair of voxels.

7. The method according to claim 1, wherein, The likelihood of a voxel characterizing a blood vessel is determined based on a vesselness filter applied to the voxel.

8. The method according to claim 1, wherein The similarity is identified for paired adjacent voxels rather than for paired distant voxels.

9. The method according to claim 1, wherein, The spectral clustering is performed based on a matrix of the plurality of voxels in the interested volume and using power iteration clustering.

10. A system for spectral clustering, comprising: A memory that stores instructions; And A processor that runs the instructions, wherein when run by the processor, the instructions cause the system to: Obtain a coronary artery image dataset of a plurality of voxels including an interested volume around the heart; Process the plurality of voxels to rank the plurality of voxels for the likelihood of characterizing blood vessels; Identify similarities between paired voxels to quantify the link strength between voxels in each pair of voxels; Perform spectral clustering on the plurality of voxels based on the likelihood of characterizing blood vessels and the link strength between the paired voxels; Select at least one voxel subtree based on the spectral clustering; Classify each subtree based on the coronary artery image dataset; and Reconstruct a representation of at least one blood vessel in the interested volume as a first reconstruction to include at least one subtree classified based on the coronary artery image dataset.

11. The system according to claim 10, wherein, The processing of the plurality of voxels includes filtering the plurality of voxels based on expected blood vessel image attributes.

12. The system according to claim 11, wherein The similarity is identified based on the expected blood vessel image attributes.

13. The system according to claim 10, wherein, The processing of the plurality of voxels, the recognition of the similarity, and the spectral clustering are performed by a core of a graphics processing unit on a one-to-one basis for the plurality of voxels in a power iteration manner.

14. The system according to claim 10, wherein, The similarity is recognized based on the proximity of the coordinates of each voxel in a pair of voxels and the directionality of each voxel in the pair of voxels.

15. The system according to claim 10, wherein The likelihood that a voxel represents a blood vessel is determined based on a vesselness filter applied to the voxel.

16. The system according to claim 10, wherein, The similarity is recognized for pairs of neighboring voxels, rather than for pairs of distant voxels.

17. The system according to claim 10, wherein The spectral clustering is performed based on a matrix of the plurality of voxels in the volume of interest and using power iteration clustering.

18. A tangible non-transitory computer-readable storage medium storing a computer program, wherein, The computer program, when run by a processor, causes the system to: obtain a coronary artery image data set of a plurality of voxels including a volume of interest around the heart; process the plurality of voxels to rank the plurality of voxels according to the likelihood of representing a blood vessel; recognize a similarity between pairs of voxels to quantify a link strength between the voxels in each pair of voxels; perform spectral clustering on the plurality of voxels based on the likelihood of representing a blood vessel and the link strength between the pairs of voxels; select at least one voxel subtree based on the spectral clustering; classify each subtree based on the coronary artery image data set; and reconstruct a representation of at least one blood vessel in the volume of interest as a first reconstruction to include at least one subtree classified based on the coronary artery image data set.

Images