Evaluation of plaque load in the peripheral blood vessels of the subject.

Abstract

A system and method are provided for evaluating plaque load in the peripheral blood vessels of a subject. This involves acquiring 3D ultrasound images of blood vessels using non-invasive ultrasound imaging, performing 3D vascular segmentation of blood vessels in the ultrasound images using a vascular segmentation machine learning model to identify vessel boundaries, performing 3D plaque segmentation of plaque in the ultrasound images using a plaque segmentation machine learning model to identify plaque within vessel boundaries, detecting the vessel's midline based on 3D vascular segmentation, determining the volume of plaque within vessel boundaries based on 3D vascular segmentation (S517), and determining the local thickness of plaque in at least one cross-section of the vessel relative to the detected midline based on 3D vascular segmentation and 3D plaque segmentation.

Description

Technical Field

[0001] The present invention relates to a method, a system, and a computer program product for evaluating plaque burden in the peripheral blood vessels of a subject.

Background Art

[0002] Cardiovascular disease (CVD), the most common cause of death in the United States, is caused by atherosclerosis, and early detection is optimal. The gold standard methods for CVD evaluation and risk stratification, such as CT angiography, are expensive and not readily available. Also, CT angiography exposes clinicians and patients to ionizing radiation and exposes patients to nephrotoxic contrast agents.

[0003] In comparison, ultrasound imaging provides a non-invasive, inexpensive, easily accessible, and non-ionizing approach for CVD evaluation and risk stratification. However, direct imaging of the coronary arteries using, for example, non-invasive ultrasound can be difficult and is dependent on the movement of the heart and the skills of the sonographer.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Research has demonstrated that, as an alternative to direct ultrasound imaging of the coronary arteries, it can be indirectly evaluated by quantifying plaque burden in peripheral arteries such as the carotid artery. However, the evaluation of peripheral artery plaque burden by current software approaches is manual, intensive, and time-consuming. Therefore, what is needed is an automated technique for rapidly, accurately, and consistently measuring the plaque burden (volume, thickness, etc.) of peripheral arteries using three-dimensional (3D) ultrasound images.

Means for Solving the Problems

[0005] The present invention is defined by the independent claims. The dependent claims define advantageous embodiments.

[0006] In a typical embodiment, a system is provided for evaluating plaque load in the peripheral blood vessels of a subject. The system includes a display configured to display three-dimensional (3D) ultrasound images of peripheral blood vessels acquired using non-invasive ultrasound imaging, a user interface, and a processor that communicates with the display and user interface, and includes non-temporary memory for storing instructions. The processor is configured to receive 3D ultrasound images of peripheral blood vessels. Vascular segmentation A machine learning model is used to perform 3D vascular segmentation of blood vessels in the 3D ultrasound images to identify peripheral blood vessel boundaries. Plaque segmentation A machine learning model is used to perform 3D plaque segmentation of plaques in the ultrasound images to identify plaque within the peripheral blood vessel boundaries. The midline of the blood vessels is detected based on the vascular segmentation. The volume of plaque within the peripheral blood vessel boundaries is determined based on the 3D plaque segmentation. Based on 3D vascular segmentation and 3D plaque segmentation, the local thickness of plaque in at least one cross-section of a peripheral vessel is determined relative to the detected midline of the vessel. Both 3D vascular segmentation and 3D plaque segmentation are performed on the entire image volume, in contrast to the need to perform 2D vascular segmentation and 2D plaque segmentation on individual 2D cross-sections (frames) of the vessel.

[0007] Another representative embodiment provides a method for evaluating plaque load in the peripheral blood vessels of a subject. This method involves acquiring 3D ultrasound images of the peripheral blood vessels using non-invasive ultrasound imaging. Vascular segmentation 3D vascular segmentation of the peripheral blood vessels in the ultrasound images is performed using a machine learning model to identify the boundaries of the peripheral blood vessels. Plaque segmentation 3D plaque segmentation of the plaques in the ultrasound images is performed using a machine learning model to identify plaques within the boundaries of the peripheral blood vessels. Based on the 3D vascular segmentation of the peripheral blood vessels, the midline of the vessels is detected. Based on the 3D plaque segmentation, the volume of plaque within the vessel boundaries is determined. Based on the 3D vascular segmentation and 3D plaque segmentation, the local thickness of the plaque in at least one cross-section of the peripheral blood vessel is determined relative to the detected midline. Both 3D vascular segmentation and 3D plaque segmentation are performed on the entire image volume, in contrast to the need to perform 2D vascular segmentation and 2D plaque segmentation on individual 2D cross-sections (frames) of blood vessels.

[0008] In another typical embodiment, a computer program product (e.g., software that can be downloaded from a server via the Internet or stored on a non-temporary computer-readable medium) includes instructions for evaluating the plaque load in the blood vessels of a subject. When executed by a processor, the instructions cause the processor to perform the method described above.

[0009] These and other aspects of the present invention will become apparent from and be explained with reference to the embodiments described below.

[0010] Exemplary embodiments are best understood from the following detailed description when read in conjunction with the accompanying drawings. It should be emphasized that various features are not necessarily depicted to a constant scale. In fact, dimensions may be arbitrarily increased or decreased for the sake of clarity in the discussion. Similar reference numbers refer to similar elements, insofar as they are applicable and practical. [Brief explanation of the drawing]

[0011] [Figure 1] This is a simplified block diagram of a system for evaluating plaque load in the blood vessels of a subject, representing a typical embodiment. [Figure 2] This is an image of a branched blood vessel with a detected central line, according to a typical embodiment. [Figure 3A] This is an ultrasound image showing the initial vascular boundary and cross-section of a vessel with plaque, obtained by 3D vascular segmentation and 3D plaque segmentation according to a typical embodiment. [Figure 3B] This is an ultrasound image showing a cross-section of a blood vessel with a corrected (modified) vascular boundary, according to a typical embodiment. [Figure 4A] These are ultrasound images showing cross-sections of blood vessels and plaques with initial plaque boundaries, obtained by 3D vascular segmentation and 3D plaque segmentation according to a typical embodiment. [Figure 4B] This is an ultrasound image showing a cross-section of a blood vessel with a corrected (modified) plaque boundary, according to a typical embodiment. [Figure 5] This is a flowchart illustrating a method for evaluating plaque load in a subject's blood vessels according to a typical embodiment. [Figure 6] This shows an example of training data for training a full 3D artificial intelligence algorithm. [Modes for carrying out the invention]

[0012] In the following detailed descriptions, representative embodiments disclosing specific details are provided for illustrative purposes only, not limitation, to provide a complete understanding of the embodiments described herein. Descriptions of known systems, apparatus, materials, methods of operation, and methods of manufacture may be omitted to avoid obscuring the description of the representative embodiments. Nevertheless, systems, apparatus, materials, and methods within the scope of those skilled in the art may be used in accordance with the representative embodiments within the scope of this teaching. It should be understood that the terms used herein are intended solely to describe specific embodiments and are not intended to be limiting. Defined terms are in addition to the technical and scientific meanings of the defined terms as generally understood and accepted in the art of this teaching.

[0013] In this specification, terms such as first, second, third, etc., may be used to describe various elements or components, but it should be understood that these elements or components should not be limited by these terms. These terms are used solely to distinguish one element or component from another. Accordingly, the first element or component discussed below may be referred to as the second element or component without departing from the teaching of the concept of the present invention.

[0014] The terms used herein are for the sole purpose of describing specific embodiments and are not intended to limit them. As used in the specification and the claims, the singular forms of the terms “a,” “an,” and “the” are intended to include both singular and plural forms unless the context explicitly indicates otherwise. Furthermore, the terms “contains,” “consist of,” and / or similar terms specify the presence of the described features, elements, and / or components, but do not exclude the presence or addition of one or more other features, elements, components, and / or groups thereof. As used herein, the terms “and / or” include any combination of one or more of the related listed items. As used in the specification and the claims of the appendix, and in addition to their usual meanings, the terms “substantial” or “substantial” mean within an acceptable limit or degree.

[0015] Unless otherwise specified, when an element or component is said to be “connected” or “joined” to another element or component, it is understood that the element or component may be directly connected or joined to the other element or component, or that there may be an intermediary element or component. That is, these terms and similar terms encompass cases where one or more intermediate elements or components may be used to connect two elements or components. However, when an element or component is said to be “directly connected” to another element or component, this only encompasses cases where two elements or components are connected to each other without any intermediate or intermediary elements or components.

[0016] Accordingly, this disclosure is intended to elicit one or more of the advantages described below specifically, through its various aspects, embodiments, and / or one or more of its particular features or sub-components. For illustrative purposes, not limiting, exemplary embodiments disclosing specific details are described to provide a complete understanding of the embodiments described herein. However, other embodiments consistent with this disclosure that deviate from the specific details disclosed herein are within the scope of the appended claims. Furthermore, descriptions of known apparatus and methods may be omitted so as not to obscure the descriptions of exemplary embodiments. Such methods and apparatus are within the scope of this disclosure.

[0017] In general, the various embodiments described herein provide systems and methods for providing fully automated assessment of peripheral artery atherosclerotic plaque load using non-invasive 3D ultrasound imaging, as opposed to 2D ultrasound imaging. The embodiments provide end-to-end deep learning models for fully automated 3D (peripheral artery) vascular segmentation and 3D plaque segmentation, and branch-compatible 3D vascular centerline detection to facilitate plaque thickness measurement. The vascular and plaque segmentation results can also be edited. 3D quantification of plaque echogenicity and 3D visualization of results by color-coded plaque thickness are also supported.

[0018] In various embodiments, in contrast to the need to perform 2D vascular segmentation and 2D plaque segmentation on individual 2D cross-sections (frames) of blood vessels, both 3D vascular segmentation and 3D plaque segmentation are performed on the entire image volume, thus reducing the need for manual intervention by the user. Therefore, the segmentation is spatially consistent and coherent throughout the vessel, for both vascular boundaries and plaque. Furthermore, the plaque segmentation correction process is accelerated by enabling 3D editing of the plaque segmentation results by the user, not just editing on a per-2D frame basis. Branch-compatible vascular centerline detection enables automatic detection of the main vessel and its branches, reducing the amount of manual intervention required to determine the plaque load of different vascular branches that may have different clinical significance. Various embodiments enable automatic analysis of 3D plaque echogenicity measurements, in contrast to 2D plaque echogenicity assessment. Furthermore, the 3D color map may correlate with the severity of plaque load (lumen narrowing, plaque area, and plaque echogenicity), allowing for rapid identification of vascular regions with the most severe disease. In addition, according to the present invention, plaque is segmented by identification in image volume, rather than being inferred by identifying the vascular wall and residual lumen, as disclosed, for example, in Zhou et al.'s "Deep learning-based carotid media-adventitia and lumen-intima boundary segmentation from three-dimensional ultrasound images" (2019). Therefore, the method according to the present invention does not require assumptions about the thickness of the vascular wall, resulting in a more accurate determination of plaque thickness. Moreover, local plaque density can be distinguished from local wall deformation. Thus, the present invention provides a method for determining the volume of significant plaque (TPV).

[0019] FIG. 1 is a simplified block diagram of a system for evaluating atherosclerotic plaque burden in a subject's blood vessel belonging to a representative embodiment. The blood vessel may be a peripheral artery such as the carotid artery or femoral artery, which is relatively large, superficial, and stationary, and thus is easier to accurately image with ultrasound compared to the much smaller coronary artery on the moving surface of the beating heart.

[0020] Referring to FIG. 1, system 100 includes a workstation 105 for implementing and / or managing the processes described herein with respect to evaluating plaque burden in a blood vessel of interest of a subject (patient) 150 using 3D ultrasound images from an ultrasound imaging device 140. Workstation 105 includes a processor 120, a memory 130, a user interface 122, and a display 124. Processor 120 communicates with ultrasound imaging device 140 via an imaging interface (not shown). Ultrasound imaging device 140 includes an ultrasound transducer probe 145 operable by a user to acquire 3D ultrasound images of plaques within a blood vessel of subject 150, such as the carotid artery, which has been shown to correlate with a patient's cardiovascular risk.

[0021] Memory 130 stores instructions executable by processor 120. When the instructions are executed, processor 120 implements one or more processes for performing a plaque burden assessment of subject 150 using the ultrasound images acquired by ultrasound imaging device 140. The ultrasound images may be provided from ultrasound imaging device 140 in real-time or near real-time during the scan procedure, or may be acquired from storage (e.g., database 112) after the scan procedure. For purposes of illustration, memory 130 is shown as including software modules, each module including instructions executable by processor 120 corresponding to related functions of system 100.

[0022] Processor 120 represents one or more processing units and is used by a general-purpose computer, a central processing unit (CPU), a digital signal processing unit (DSP), a graphical processing unit (GPU), a computer processor, a microprocessor, a state machine, a programmable logic device, a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or a combination thereof, hardware, software, firmware, wired logic circuits, or a combination thereof. A processor or processing unit as used herein may include multiple processors, parallel processors, or both. Multiple processors may be contained in a single device or multiple devices, or may be combined. As used herein, the term “processor” encompasses an electronic component capable of executing a program or machine-executable instructions. A processor may also refer to a collection of processors within a single computer system or distributed across multiple computer systems, such as a cloud-based or other multi-site application. A program has software instructions that are executed by one or more processors, which may be within the same computing device or distributed across multiple computing devices.

[0023] Memory 130 represents one or more memories, which may include main memory and / or static memory, and such memories may communicate with the processor 120 via one or more buses. Memory 130 can be implemented by any number, types, and combinations of random access memory (RAM) and read-only memory (ROM), and may store various kinds of information such as software algorithms, artificial intelligence (AI) machine learning models, and computer programs, all of which are executable by the processor 120. Various types of ROM and RAM include disk drives, flash memory, electrically programmable read-only memory (EPROM), electrically erasable and programmable read-only memory (EEPROM), registers, hard disks, removable disks, tapes, compact disc read-only memory (CD-ROM), digital general-purpose multi-purpose discs (DVDs), floppy disks, Blu-ray discs, universal serial bus (USB) drives, or other forms of storage media. Memory 130 is a tangible storage medium for storing data and executable software instructions, and is non-transient for the time the software instructions are stored. The term “non-transient” as used herein is interpreted as a characteristic of a state that lasts for a period of time, rather than as a permanent characteristic of the state. The term “non-transient” specifically negates fleeting characteristics, such as carrier waves or signals or other forms of characteristics that exist only temporarily at any given time and place. Memory 130 may store software instructions and / or computer-readable code that enable the performance of various functions. Memory 130 may be secure and / or encrypted, or non-secure and / or unencrypted.

[0024] System 100 may also include a database 112 for storing information that can be used by various software modules of memory 130. For example, database 112 may include image data from previously obtained ultrasound images of object 150 and / or other objects at similar locations. Stored image data may be used to train AI machine learning models, such as 3D UNet models and neural network models, as described below. Database 112 may be implemented by any number, types, and combinations of RAM and ROM. Various types of ROM and RAM may include any number, types, and combinations of computer-readable storage media, such as disk drives, flash memory, EPROM, EEPROM, registers, hard disks, removable disks, tapes, CD-ROMs, DVDs, floppy disks, Blu-ray disks, USB drives, or any form of storage media known to those skilled in the art. Database 112 includes tangible storage media for storing data and executable software instructions, which may be non-temporary while the data and software instructions are stored therein. Database 112 may be secure and / or encrypted, or it may be insecure and / or not encrypted. For illustrative purposes, database 112 is shown as a separate storage medium, but it will be understood that it may be combined with and / or contained within memory 130 without departing from the scope of this teaching.

[0025] The processor 120 may include, or have access to, an artificial intelligence (AI) engine, which may be implemented as software that provides artificial intelligence (e.g., UNet models, neural network models) and applies the machine learning described herein. The AI ​​engine may reside in addition to the processor 120, or in any of the various components other than the processor 120, such as memory 130, an external server, and / or a cloud. If the AI ​​engine is implemented in a cloud (not shown), such as a data center, the AI ​​engine may be connected to the processor 120 via the Internet or other communication network using one or more wired and / or wireless connections.

[0026] The user interface 122 is configured to provide the user with information and data output by the processor 120, memory 130, and / or ultrasound imaging device 140, and / or to receive information and data input by the user. That is, the user interface 122 enables the user to input data and control or operate aspects of the processes described herein, and enables the processor 120 to indicate the effect of user input, including control or operation of the ultrasound transducer probe 145. All or part of the user interface 122 may be implemented by a graphical user interface (GUI), such as a GUI 128 that can be displayed on the display 124, as described below. The user interface 122 may include, for example, one or more interface devices such as a mouse, keyboard, trackball, joystick, microphone, video camera, touchpad, touchscreen, and voice or gesture recognition captured by the microphone or video camera.

[0027] The display 124 may be, for example, a computer monitor, television, liquid crystal display (LCD), organic light-emitting diode (OLED) display, flat panel display, solid-state display, cathode ray tube (CRT) display, or electronic whiteboard. The display 124 includes a screen 126 for displaying an ultrasound image of the subject 150, for example, with various features described herein for communicating to the user the volume of the plaque shown in the image, and a GUI 128 that allows the user to interact with the displayed image and features. In embodiments, the ultrasound imaging apparatus 140 may include another dedicated display for acquiring ultrasound images, where the dedicated display is also represented by the display 124.

[0028] Referring to memory 130, various modules within it store sets of data and instructions that can be executed by processor 120 to perform plaque load assessment as described above. The ultrasound imaging module 131 is configured to receive and process 3D ultrasound images of blood vessels of interest in subject 150. The 3D ultrasound images may be acquired non-invasively by ultrasound imaging device 140 via ultrasound transducer probe 145 and received from ultrasound imaging device 140 in real time or near real time during a simultaneous imaging session of subject 150. Alternatively, the 3D ultrasound images may be previously acquired images retrieved from storage (e.g., database 112) during a previous imaging session. The 3D ultrasound images may be displayed on display 124. In particular, the display of real-time images allows the user to visualize the anatomical structure of subject 150 while operating the ultrasound transducer probe 145.

[0029] The vascular segmentation module 132 is configured to receive a 3D ultrasound image of a target vessel from the ultrasound imaging module 131 and to perform 3D vascular segmentation of the vessel to identify the vessel boundaries. 3D vascular segmentation may be performed using a vascular segmentation machine learning model, which may include, for example, various known vascular segmentation machine learning models. 3D vascular segmentation can segment the entire range of a vessel more efficiently than conventional 2D segmentation techniques. The vascular segmentation machine learning model may be defined by a deep learning neural network such as a convolutional neural network (CNN), e.g., UNet, artificial neural network (ANN), or recurrent neural network (RNN). In embodiments, at least a portion of the 3D vascular segmentation may be performed manually by the user via the user interface 122, as described below, and / or the results of the 3D vascular segmentation may be edited by the user via the user interface 122.

[0030] In the embodiment, as a preprocessing step, 3D vascular segmentation involves performing normalization of 3D ultrasound images (e.g., z-score normalization) and increasing the resolution of the 3D ultrasound images (e.g., approximately 0.2 × 0.2 × 0.2 mm). 3 This may include resampling. The vascular segmentation machine learning model takes normalized and resampled 3D ultrasound images as input and outputs a 3D vascular mask (image) as a 3D model. For example, the 3D vascular mask is a binary vascular mask where "1" corresponds to the vessel (including the vessel boundary and the inside of the vessel) and "0" corresponds to the background (including the outside of the vessel). The vascular mask fits the vessel boundary and includes only the internal volume of the vessel defined by the identified vessel boundary.

[0031] As a post-processing step for 3D models, 3D vascular segmentation involves performing connected component analysis to facilitate vascular analysis. Generally, connected component analysis involves uniquely labeling binary masks to individual components based on component connectivity. In 3D, component isolation can be defined by 6 connectivity (faces connected), 18 connectivity (edges connected), or 26 connectivity (nodes connected). Connected component analysis may include, for example, extracting the largest component from a resampled 3D ultrasound image to remove potential sparse false-positive segmentation. Furthermore, image intensity-guided ballooning of vascular masks may be performed as a post-processing step to further fit the vascular boundaries. Ballooning refers to a user interface technique for editing or modifying the position of vascular (and / or plaque) boundaries. User-defined ballooning "pushes" the boundary, forcing it to move to a different position. The boundary may respond by "pushing back" depending on the brightness of the surrounding pixels. The balloon method is generally based on the idea that the vessel wall generates a strong echo and is written as a bright pixel in the ultrasound image, while the blood-filled interior of the vessel (and some types of acoustic noise) generally generates a weak echo and is a black or dark pixel in the image. The centerline of the vessel may be detected using a vessel mask, as discussed below with reference to the centerline detection module 134. In other words, the method involves calculating transverse cuts along the longitudinal axis of the vessel, where each transverse cut contains multiple pixels. Classifying the multiple pixels of each transverse cut, the classification includes identifying the pixels as lumen, plaque, vessel wall, etc. The classification may be binary or non-binary. For each transverse cut, the centroid of at least one classification is determined. The determined centroids are connected to provide an estimate of the centerline fragment.

[0032] The vascular segmentation module 132 is further configured to derive a polygonal mesh containing vertices and edges from the vascular mask. The mesh can be derived during post-processing, for example, using a known marching cube algorithm. The boundaries of the vessels are represented by the polygonal mesh derived by the vascular segmentation module 132.

[0033] The plaque segmentation module 133 is configured to receive 3D ultrasound images of blood vessels from the ultrasound imaging module 131, receive a vascular mask from the vascular segmentation module 132, perform 3D plaque segmentation of plaques in the ultrasound images, and identify plaques within the vascular boundaries. 3D plaque segmentation can be performed, for example, using a plaque segmentation machine learning model, which may include various known plaque segmentation machine learning models. 3D plaque segmentation can segment the entire volume of plaque within blood vessels more efficiently than conventional 2D segmentation techniques. Plaque segmentation models can be defined, for example, by deep learning neural networks such as CNNs (e.g., UNet), ANNs, or RNNs. It is understood that vascular segmentation machine learning models and plaque segmentation machine learning models may be different machine learning models (as illustrated), or they may be implemented by a single machine learning model trained for both types of segmentation, without departing from the scope of this teaching. In the embodiments, at least a portion of the 3D plaque segmentation may be performed manually by the user via the user interface 122, as described below, and / or the results of the 3D plaque segmentation may be edited by the user via the user interface 122. Furthermore, according to the present invention, plaque is segmented by identification in the image volume, rather than being inferred by identifying the vessel wall and residual lumen, as disclosed, for example, in Zhou et al.'s "Deep learning-based carotid media-adventitia and lumen-intima boundary segmentation from three-dimensional ultrasound images" (2019). Therefore, the method according to the present invention does not require the assumption of vessel wall thickness, resulting in a more accurate plaque thickness. Moreover, local plaque density can be distinguished from local wall deformation.

[0034] In the embodiment, as a preprocessing step, 3D plaque segmentation involves performing normalization of 3D ultrasound images (e.g., z-score normalization) and raising the resolution of the 3D ultrasound images to a higher resolution (e.g., approximately 0.18 × 0.18 × 0.18 mm). 3 This includes resampling. The plaque segmentation machine learning model takes normalized and resampled 3D ultrasound images and a 3D vascular mask as input and outputs a 3D plaque mask (image) as a 3D model. As a post-processing step, 3D plaque segmentation includes performing connected component analysis to facilitate plaque analysis. For example, connected component analysis may include color-coding different components within the segmented plaque regions. The plaque segmentation module 133 is further configured to derive a polygonal mesh containing vertices and edges from the plaque mask. The mesh can be derived during post-processing, for example, using a known marching cube algorithm. The plaque boundaries are represented by the polygonal mesh derived by the plaque segmentation module 133.

[0035] Machine learning models for vascular and plaque segmentation can be applied sequentially. For example, with respect to resampling, 3D ultrasound images are first resampled for 3D vascular segmentation and then resampled again for 3D plaque segmentation. This approach allows for interaction (and editing) with the vascular mask by the vascular segmentation machine learning model before applying the plaque segmentation machine learning model, thereby ensuring correct 3D vascular segmentation and consistent 3D plaque segmentation that matches the vascular boundaries determined by the 3D vascular segmentation. Interaction with the vascular mask can be performed using a 3D deformable model.

[0036] In various embodiments, all or part of the processes provided by the vascular and plaque segmentation machine learning model may be implemented, for example, by one or more of the AI ​​engines described above. The vascular and plaque segmentation machine learning model can be trained using retrospective training ultrasound images of peripheral vessels from multiple patients and ground truth data indicating vascular and plaque boundaries associated with the training ultrasound images. Segmented vessels and segmented plaques are associated with the corresponding vascular and plaque volumes. The training data may be stored in one or more databases, such as database 112, and annotated by a clinical expert.

[0037] The centerline detection module 134 is configured to detect the centerlines of blood vessels based on the blood vessel boundaries determined by the blood vessel segmentation module 132 (e.g., the output of a blood vessel segmentation machine learning model). In embodiments, the centerlines of blood vessels may be detected using image processing techniques, such as a fast marching algorithm (e.g., the Eikonal equation). Generally, a fast marching algorithm solves a wave propagation equation in which the refractive index of the wave propagation equation is derived from a distance map to the blood vessel boundary. The fast marching algorithm takes pairs of points in a segmented blood vessel as input and outputs the least geodesic path (i.e., Fermat's principle) between the pairs of points. For example, in the case of carotid arteries, input pairs of points include (center of the proximal common carotid artery, center of the distal internal carotid artery) and (center of the proximal common carotid artery, center of the distal external carotid artery).

[0038] A pair of points may be the endpoints of a fragment of the central line within a blood vessel. For example, a pair of points can be determined by calculating a transverse cut along the longitudinal axis of the vessel (e.g., volume z-axis or volume x-axis depending on the orientation of the ultrasound transducer probe 145), tracking the two-dimensional connected components of the transverse cut, and identifying the location of any branch and the direction of the ultrasound scan to provide a 3D ultrasound image for the division of the main vessel into multiple vascular branches. Determining a pair of points involves determining the centroid of each connected component to provide an estimate of the central line portion (including the vascular branches). The endpoints of the estimated central line piece are input as pairs of points into a fast marching algorithm of a 3D vascular potential map derived from 3D vascular segmentation. With a suitable 3D vascular potential map, the fast marching algorithm generates a central line that converges deep within the carotid artery. However, in the carotid artery, the clinical focus is set on the internal common branch, and the external carotid artery is considered a clinically less important secondary branch. Therefore, after applying a high-speed marching algorithm, the midline of the external carotid artery can be corrected by forcing an early merger of the midline of the external carotid artery and the midline of the internal common carotid artery branch when the midline of the external carotid artery is sufficiently close. This results in an asymmetrical result that better represents the anatomical structure. In other words, the method comprises the steps of: calculating transverse cuts along the longitudinal axis of a vessel, each transverse cut comprising a plurality of pixels; classifying the plurality of pixels of each transverse cut, the classification comprising identifying the pixels as lumen, plaque, vessel wall, etc., and the classification may be binary or non-binary; determining the centroid of at least one classification for each transverse cut; and connecting the determined centroids to provide an estimate of the midline fragment. The midline, or fragment thereof, may then be used to determine the vessel branches and to identify the internal carotid artery, external carotid artery, and common carotid artery.

[0039] The process of detecting the midline of blood vessels automatically detects the midline of the main vessel and the midlines of its branches. Generally, the midlines of vascular branches need to provide a natural coordinate system, such as a curvilinear coordinate system or a local 2D axis system perpendicular to the midline. A natural coordinate system is necessary for 2D quantification of plaque, such as measuring local plaque thickness, as described below.

[0040] Figure 2 shows an image of a branched vessel with a detected centerline, according to a typical embodiment. As shown in Figure 2, vessel 200 includes a main vessel 210 and vessel branches 221 and 222, which are followed by a branch 220. The detected centerline 215 follows the main vessel 210 longitudinally, and the detected centerlines 223 and 224 follow the vessel branches 221 and 222 longitudinally, respectively. Plaque 230 shown in vessel 200 is measured relative to the detected centerlines 215, 223 and 224, as described below.

[0041] Referring again to Figure 1, the plaque measurement module 135 is configured to determine the 3D volume and 2D thickness of plaque within the vascular boundary. The 3D volume and thickness of the plaque are used for CVD assessment and risk stratification of subject 150. Generally, the larger the 3D volume and / or thickness of the plaque determined by the plaque measurement module 135, the more advanced the CVD in subject 150. The 3D volume and 2D thickness of the plaque can be evaluated based on the total value of the plaque and / or the ratio of the plaque to the 3D volume and / or 2D cross-sectional area of ​​the vessel.

[0042] The plaque measurement module 135 is configured to determine the 3D volume of the plaque based on 3D plaque segmentation. In particular, the 3D volume of the plaque is mathematically calculated from the coordinates of the vertices and edges of the mesh derived by the plaque segmentation module 133 described above.

[0043] The plaque measurement module 135 is further configured to determine the 2D local thickness of plaque in a cross-section (slice) of a vessel relative to the detected centerline of the vessel. The plaque measurement module 135 can optionally also determine the 2D plaque region in these cross-sections. The detected centerline is used to define a local 2D plane perpendicular to the centerline, and the local thickness of the plaque is determined as the maximum thickness of the plaque within the local 2D plane. Multiple local thicknesses of plaque can be determined in multiple local 2D planes defined by the detected centerline. As mentioned above, the detected centerline of the vessel provides a natural coordinate system, such as a curvilinear coordinate system, for identifying the local 2D plane, and a local 2D axial system perpendicular to the detected centerline. The local thickness provides 2D quantification of plaque and can also be used to determine plaque region, stenosis, and other measurements that implicitly require local orthogonality to the vessel boundary. Stenosis, in particular, is the reduction in the diameter or area of ​​the vessel lumen in any cross-section by the thickness of the plaque in the same cross-section, and thus obstructs blood flow.

[0044] In various embodiments, the system 100 allows the user to edit options for 3D vascular segmentation and 3D plaque segmentation via the user interface 122. In some cases, image artifacts such as shadows due to calcification and clutter noise degrade the performance of automatic segmentation of 3D vascular and plaque boundaries, reducing the level of manual editing by the user to a desirable level for obtaining accurate quantitative results. Generally, conventional methods for correcting 2D planar contours are very cumbersome and inherently unsuitable for correcting 3D segmentation models, as they may require hundreds of 2D planes or slices to accurately represent the 3D volume containing the vessels. Therefore, an essential 3D editing scheme for correcting 3D vascular and plaque segmentation, from which the vascular and plaque boundaries are derived, can be applied to 3D ultrasound images of vessels. Corrections to 3D vascular and plaque segmentation can be input by the user in any planar direction and extended to the 3D neighborhood at a scale selected by the user. For 3D plaque segmentation, the correction is also applied to the 3D view. In other words, a plaque can consist of multiple separate (unconnected) components, and the fix involves removing one of the spurious components (or all but one).

[0045] In system 100, the vascular editing module 136 enables editing of 3D vascular segmentation, and the plaque editing module 137 enables editing of 3D plaque segmentation. Specifically, the vascular editing module 136 is configured to implement editing commands received from the user via the user interface 122 to modify 3D vascular segmentation and provide modified vascular segmentation of the vessels, and the plaque editing module 137 is configured to implement editing commands received from the user via the user interface 122 to provide modified plaque segmentation of the vessels. In this case, centerline detection by the centerline detection module 134 is performed based on the modified vascular segmentation, and plaque thickness determination by the plaque measurement module 135 is performed on the modified plaque segmentation within the boundaries of the modified vascular boundary relative to the centerline of the modified vessel.

[0046] The editing commands that the vascular editing module 136 receives to edit 3D vascular segmentation may include input from the user interface 122 indicating points where the surface of the vascular boundary is to be repositioned. Each point can be placed inside or outside the initial vascular boundary to indicate that a smaller or larger cross-sectional area is desired. For example, the input may include touching a touchscreen or point-clicking from a mouse or other input device on a 3D ultrasound image using the GUI 128. The vascular editing module 136 then implements the editing commands to modify the 3D vascular segmentation by repositioning (moving) the surface of the vascular boundary to the commanded points and providing the repositioned surface.

[0047] Figure 3A is an ultrasound image showing a cross-section of a vessel with an initial vascular boundary and plaque, obtained by 3D vascular segmentation and 3D plaque segmentation according to a typical embodiment. As shown in Figure 3A, vessel 310 is followed by 3D vascular and plaque segmentation having an initial vascular boundary 314, an initial centerline 316 in the lumen of vessel 310, and an initial plaque boundary 318 along a portion of the initial vascular boundary 314.

[0048] Figure 3B is an ultrasound image showing a cross-section of a vessel with a corrected (modified) vessel boundary, according to a typical embodiment. Referring to Figure 3B, the modified vessel 310A is shown in response to an editing command that sets an anchor point 325 inward from the initial vessel boundary 314 to establish the modified vessel boundary 324. The modified vessel boundary 324 also results in a modified centerline 326. The shape of the initial plaque boundary 318 is also slightly shifted in response to the adjustment represented by the modified vessel boundary 324, resulting in the modified plaque boundary 328.

[0049] The amount of change to the initial vessel boundary 314 gradually decreases as the distance from the anchor point 325 along the line of the vessel boundary 314 (e.g., user clicks) increases until the original surface of the initial vessel boundary 314 remains unchanged. For example, in Figures 3A and 3B, the left side of the modified vessel boundary 324 remains unchanged compared to the left side of the initial vessel boundary 314 because the distance from the anchor point 325 is relatively large. Since the shape of the vessel 310 is smooth, the modification works on a scale that preserves its smooth nature, and the degree of change is determined by the distance from the anchor point 325 to the surface. In embodiments, the degree of change may be previewed on the display 124, for example, by a vessel mesh rendered by colorization.

[0050] Edit commands received by the plaque editing module 137 to edit 3D plaque segmentation may include input from the user interface 122 indicating a selected radius of a virtual balloon within the plaque region provided by the 3D plaque segmentation in the ultrasound image. For example, the input may include a touch on a touchscreen or a point click from a mouse or other input device on the 3D ultrasound image using the GUI 128. Accordingly, the plaque editing module 137 implements edit commands to modify the 3D plaque segmentation by repositioning (moving) the plaque boundary surface to the indicated point (to match the selected radius of the virtual balloon). Alternatively, or in addition, edit commands received by the plaque editing module 137 may include input from the user interface 122 indicating that one or more individual connected components (determined by connected component analysis in post-processing) are to be removed from between connected components.

[0051] Figure 4A is an ultrasound image showing a cross-section of a vessel and plaque with an initial plaque boundary obtained by 3D vessel segmentation and 3D plaque segmentation according to a typical embodiment. As shown in Figure 4A, the vessel 410 is shown following 3D vessel and plaque segmentation with an initial vessel boundary 414, an initial centerline 416 in the lumen of the vessel 410, and an initial plaque boundary 418 along a portion of the initial vessel boundary 414.

[0052] Figure 4B is an ultrasound image showing a cross-section of a vessel with a corrected (modified) plaque boundary, according to a typical embodiment. Referring to Figure 4B, the modified vessel 410A is shown in response to an editing command in which a virtual balloon 425 is established within the boundary of the initial plaque boundary 418 to establish the modified plaque boundary 428. Meanwhile, the shape of the initial vessel boundary 414 and the position of the initial centerline 416 remain substantially unchanged by the modification indicated by the modified plaque boundary 428. The plaque segmentation modification is set by the virtual balloon 425, the radius of which is interactively determined by the distance between an initial input (e.g., the first screen touch or click) and a subsequent input in any planar cut. The user can enlarge the size of the virtual balloon 425 to match the initial plaque boundary 418, and then push the virtual balloon 425 beyond the initial plaque boundary 418, which deforms accordingly to accommodate the shape of the expanding virtual balloon 425 in order to provide the modified plaque boundary 428. This process also works in reverse, and the size of the virtual balloon 425 may be interactively reduced. Because the modification is interactive, the user can see the results of the changes to the virtual balloon 425 and their corresponding effects on the shape of the modified plaque boundary 428 virtually in real time and immediately on the display 114.

[0053] The user can repeat modifications to the virtual balloon 425, and therefore repeat modifications to the modified plaque boundary 428 as needed, either starting from the beginning or from the latest shape. The user may also add plaque that may have been missing due to segmentation using the lumen representation of the plaque surface in the initial plaque boundary 418 or the modified plaque boundary 428 (i.e., the complement surface of the plaque within the vessel 410). The interior of the vessel 410 consists of the lumen and plaque separated by the initial plaque boundary 418. Therefore, plaque modification may be performed in a "plaque view" or "lumen view" that complements the plaque view, knowing that the size of the vessel 410 remains unchanged. Plaque modification can be achieved by pushing the lumen boundary of the internal vessel (which is not possible in the plaque representation). For example, moving the lumen away from the vessel wall will also effectively move the plaque boundary 418, thereby increasing the amount of plaque.

[0054] The virtual balloon 425 may have a spherical shape driven by image features (e.g., a distance map, an implicit function of which the sphere is a level set). The edges or ridges of the sphere define geodesic distances instead of Euclidean distances, and the shape of the sphere conforms to these distances. The user can select different settings in the user interface 122 to adjust the degree of geodesic adaptation that best suits the situation. In the absence of such adaptation, the spherical shape of the virtual balloon 425 is a Euclidean sphere.

[0055] Under certain circumstances, users may want to extend the vascular branches identified during the plaque measurement process described above. In carotid artery ultrasound scans, image quality deteriorates quasi-systematically as the ultrasound transducer probe 145 moves distally. This is primarily due to increased depth and obstruction by the subject's jaw. Therefore, such ultrasound images are not clinically grade and cannot accurately confirm the presence of plaque. Furthermore, these areas are typically excluded from training vascular and plaque segmentation machine learning (AI) models.

[0056] Figure 5 is a flowchart of a method for evaluating plaque load in a subject's blood vessels according to a typical embodiment. This method may be implemented, for example, using instructions stored in memory 130 and executable by the processor 120 of system 100, as described above.

[0057] Referring to Figure 5, a 3D ultrasound image of the blood vessel of interest to the subject is acquired in block S511. The 3D ultrasound image is acquired using non-invasive ultrasound imaging and can be received from an ultrasound imaging device or a database of previously acquired 3D ultrasound images.

[0058] In block S512, a vascular segmentation machine learning model is used to perform 3D vascular segmentation of blood vessels in ultrasound images and identify the boundaries of blood vessels. In block S513, a plaque segmentation machine learning model is used to perform 3D plaque segmentation of plaques in ultrasound images and identify the boundaries of plaques within the boundaries of blood vessels. The vascular segmentation model and the plaque segmentation model may each be a neural network such as UNet, ANN, RNN, or CNN. The neural network may be a deep learning neural network. Furthermore, the vascular segmentation model and the plaque segmentation model can be implemented by a single unified machine learning model or by separate independent machine learning models without departing from the scope of this instruction.

[0059] In block S514, it is determined whether editing is necessary to modify the 3D vascular segmentation and / or 3D plaque segmentation. This determination may be made by the user based on a visual inspection of the segmented ultrasound images. If it is determined that editing is necessary for either or both of the 3D vascular segmentation or 3D plaque segmentation (block S514: yes), the process proceeds to block S515. In block S515, an editing command is received from the user interface and implemented to modify the 3D vascular segmentation to provide a modified vascular segmentation of the vessels and / or to modify the 3D plaque segmentation to provide a modified plaque segmentation of the plaques within the vessels. Generally, an editing command to modify 3D vascular segmentation may include the location of anchor points in the ultrasound image where a portion of the vessel boundary is moved. Editing commands for correcting 3D plaque segmentation may include adjusting the position and selected radius of a virtual balloon placed within the plaque region in the ultrasound image, and adjusting the surface of the plaque region boundary to match the position of the selected radius of the virtual balloon. Editing commands for correcting 3D vascular segmentation and 3D plaque segmentation may be received from the user, for example, through a GUI or a user interface that may be included by a GUI.

[0060] If it is determined in block S514 that no editing is required for 3D vascular segmentation or 3D plaque segmentation (block S514: No), the process proceeds to block S516. In block S516, the centerlines of the vessels are detected based on the 3D vascular segmentation of the vessels. This refers to either the initial 3D vascular segmentation provided by block S513, or, if determined, the modified 3D vascular segmentation provided by block S515. Once the centerlines of the vessels are detected, the centerlines of the main vessels and the branches of vessels from the main vessels are automatically detected, thereby identifying the vessel branches. While the branches of vessels can be automatically identified, the identification of branches (e.g., internal and external carotid arteries) is checked by the user. The main centerline must enter the internal carotid artery from the common carotid artery (before branching). If the centerline runs from the common carotid artery to the external carotid artery, the user may, in optional block S517, for example, use a swap button on the user interface to swap the external and internal identifications so that the initially identified external carotid artery is identified as the internal carotid artery, and vice versa.

[0061] In block S518, the volume of plaque within the vessel boundary is determined based on 3D plaque segmentation (dependent on 3D vessel segmentation). As described above, the volume of plaque can be determined by obtaining the plaque mesh and determining the volume of the plaque from the coordinates of the mesh vertices and edges.

[0062] In block S519, the local thickness of plaque within the vessel boundary is determined relative to the detected centerline based on 3D vessel segmentation and 3D plaque segmentation. As described above, determining the local thickness involves identifying one or more cross-sections (slices) of the vessel relative to the detected vessel centerline (or multiple centerlines if the vessel contains branches), and measuring the maximum thickness of the plaque in the 2D plane defined by the cross-section.

[0063] Plaque volume and local thickness are used to assess CVD and stratify risk in subjects. Generally, the greater the plaque volume and / or thickness, the more advanced the subject's CVD, and any prescribed treatments will be noted. Plaque volume and thickness can be assessed based on the overall value of the plaque and / or the ratio of plaque to the volume and / or cross-sectional area of ​​the vessel.

[0064] The segmentation algorithms disclosed herein, namely 3D vascular segmentation and 3D plaque segmentation algorithms, may include trained AI configured to perform complete 3D segmentation of blood vessels and / or plaques. These trained AI may be further trained neural networks (NNs).

[0065] The neural network (NN) was trained on a dataset annotated to represent the boundaries between blood vessels and plaque. For example, the annotated dataset was generated using Philips Vascular Plaque Quantification (VPQ) software, as reported by Skillessen et al. in "Carotid Plaque Burden as a Measure of Subclinical Atherosclerosis: Comparison with Other Tests for Subclinical Arterial Disease in the High Risk Plaque BioImage Study" (2012), to generate the boundaries between blood vessels and plaque. In other words, the software generates 2D slices from a 3D image volume, as shown in Figure 6, and segments each 2D slice to identify blood vessel boundaries and plaque. Each segmented 2D image slice was further validated by an expert leader, who was free to manually adjust the segmentation. Skillessen et al. describe in detail how they used VPQ software to create training data and analyze 3D carotid ultrasound images from 6,101 subjects (2012). All ultrasound images were read using Philips QLAB-VPQ plaque analysis software in the core laboratory of the Department of Vascular Surgery at Riggs Hospitalet, University of Copenhagen (Copenhagen, Denmark). All core lab reads were performed by blinded readers using individual reading stations, without being aware of the results of other imaging modalities. Each 2D slice was segmented, and after review by expert readers, the 2D boundaries of each slice were interpolated into a 3D mesh representing the carotid artery wall and plaque, as shown in Figure 2.Examples of training data obtained using this method are described in Skillessen et al. (2012), Fernandez-Friela et al. (2015) "Prevalence, Vascular Distribution, and Multiterritorial Extent of Subclinical Atherosclerosis in a Middle-Aged Cohort: The PESA (Progression of Early Subclinical Atheroscleros) Study," and Buxton et al. (2023) "The pregnancy research on inflammation, nutrition, and city environment: systematic analyses study (PRINCESA) cohort, 2009–2015."

[0066] Figure 6 shows, at the bottom, 2D ultrasound image slices from a 3D ultrasound imaging volume. These slices are segmented by a 2D segmentation algorithm as described herein. For each slice, the segmentation algorithm may identify the boundaries of the tunica media-to-tunica media boundary 622, the intima-to-tunica media boundary 624, the residual lumen 626, and the plaque 628. In other examples, additional, fewer, or different boundaries may be identified by the segmentation algorithm. Each boundary may be reviewed by a specialist reader and modified as necessary. As shown at the top of Figure 6, the segmented boundaries of the 2D image slices of the 3D volume can be further combined, and data between the 2D image slices can be interpolated to generate a 3D mesh of vessels and plaque, as shown in Figure 2.

[0067] In another example, training data was acquired using a 3D software tool based on a deformable model. This tool was used to initialize a 3D vascular mesh and deform it to fit the boundaries of the vessel walls. Similarly, a 3D plaque mesh was initialized within the vessels and deformed to fit the boundaries of the plaque. Finally, the 3D vascular and plaque meshes were reviewed by expert readers and adjusted using an interactive 3D modification tool. An example of a 3D software tool based on a deformable model is shown in Mory et al., "Real-Time 3D Image Segmentation by User-Constrained Template Deformation" (2012).

[0068] For example, the AI ​​model is trained using a symmetric encoder-decoder architecture similar to U-Net (Ronneberger et al., "U-net: Convolutional networks for biomedical image segmentation" (2015)) adapted for 3D patch segmentation. This architecture consists of seven hierarchical levels, each containing two convolutional blocks. Each of these blocks consists of a 3x3x3 convolutional layer paired with a normalization layer and a LeakyReLu activation function. To improve robustness and performance, intermediate outputs across multiple network stages are integrated, with half of the total loss weights attributed to the final stage and the other half to these intermediate stages (deep supervision). This loss was calculated as an equally weighted combination of binary cross-entropy loss and soft dice loss. The model was trained over 2000 epochs using an Adam optimizer with a decay coefficient of 0.9, reducing the learning rate which was previously initialized to 10⁻².

[0069] The input data, i.e., the training data, is normalized to the range of 0–1, and linear interpolation is used to obtain values ​​from 0.17 to 0.23 mm. 3The data was resampled to an isotropic resolution. Data augmentation techniques such as Gaussian noise, brightness, contrast, rotation, and cropping were randomly applied during the training phase to introduce variability and enhance robustness.

[0070] Two models were trained according to the described methodology: one for evaluating blood vessels and the other for acquiring plaque. The only difference is that the plaque model uses a vascular mask as input in addition to ultrasound images.

[0071] In accordance with various embodiments of this disclosure, the methods described herein may be implemented using a hardware computer system that executes a software program stored on a (non-temporary) storage medium. Furthermore, in exemplary, non-limiting embodiments, 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.

[0072] The assessment of intravascular plaque load is described with reference to exemplary embodiments, but the language used is understood to be descriptive and illustrative, not restrictive. Modifications can be made within the scope and spirit of the embodiments, as described and modified, within the scope and spirit of the embodiments. The assessment of plaque load is described with reference to specific means, materials and embodiments, but is not intended to be limited to the disclosed details. Rather, the assessment of plaque load extends to all functionally equivalent structures, methods, and applications, such as those within the scope of the appended claims.

[0073] The figures of the embodiments described herein are intended to provide a general understanding of the structure of various embodiments. The figures are not intended to serve as a complete description of all elements and features of the disclosure described herein. Many other embodiments may be apparent to those skilled in the art by examining this disclosure. Other embodiments may be used and derived from this disclosure so as to be structural and logical substitutions and modifications without departing from the scope of this disclosure. In addition, the figures are merely representational and do not have to be drawn to scale. Certain proportions in the figures may be exaggerated, and other proportions may be minimized. Accordingly, the disclosure and figures should be considered illustrative and not restrictive.

[0074] One or more embodiments of this disclosure may be referenced individually and / or collectively by the term “invention” for convenience only, without any intention to spontaneously limit the scope of this application to any particular invention or inventive concept. Furthermore, it should be understood that while certain embodiments are illustrated and described herein, subsequent configurations designed to achieve the same or similar objectives may substitute for the specific embodiments shown. This disclosure is intended to cover any and all subsequent adaptations or variations of various embodiments. Combinations of the embodiments described herein, and other embodiments not specifically described herein, will become apparent to those skilled in the art upon reviewing this specification.

[0075] The abstract is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. Furthermore, the aforementioned detailed description may, for the purpose of streamlining the disclosure, group various features together or describe them in a single embodiment. This disclosure should not be interpreted as reflecting an intention that the claimed embodiments require more features than those explicitly enumerated in each claim. Rather, as reflected in the following claims, the subject matter of the invention may be directed toward fewer than all of the features of any of the disclosed embodiments. Therefore, the following claims are incorporated into the detailed description, and each claim exists independently as defining the individually claimed subject matter.

[0076] The foregoing description of the disclosed embodiments is provided to enable those skilled in the art to practice the concepts described herein. Accordingly, the subject matter disclosed above should be considered illustrative rather than restrictive, and the appended claims are intended to cover all such modifications, extensions, and other embodiments that fall within the true scope of this disclosure. Therefore, to the maximum extent permitted by law, the scope of this disclosure shall be determined by the most broadly permissible interpretation of the following claims and their equivalents, and shall not be limited or restricted by the foregoing detailed description.

[0077] No reference symbol placed in parentheses in a claim should be construed as limiting that claim. Advantageously, means described in different dependent claims may be used in combination.

Claims

1. A computer-implemented method for evaluating plaque load in the peripheral blood vessels of a subject, wherein the method is: The steps include receiving a 3D ultrasound image of the peripheral blood vessel using non-invasive ultrasound imaging, The steps include: using a vascular segmentation machine learning model to perform 3D vascular segmentation of blood vessels in the ultrasound image to identify the boundaries of the blood vessels; A plaque segmentation machine performs 3D plaque segmentation of plaque in the ultrasound image using a learning model to identify plaque within the boundaries of the blood vessel, wherein, in addition to performing 2D vascular segmentation and 2D plaque segmentation on individual 2D cross-sectional frames of the blood vessel, both the 3D vascular segmentation and the 3D plaque segmentation are performed on the entire image volume. The steps include detecting the centerline of the blood vessel based on 3D vascular segmentation of the blood vessel, The steps include determining the volume of plaque within the vessel boundary based on the 3D plaque segmentation, Based on the 3D vascular segmentation and the 3D plaque segmentation, the steps include determining the local thickness of the plaque in at least one cross-section of the blood vessel with respect to the detected centerline. A computer implementation method having

2. The method according to claim 1, further comprising the step of identifying the main and secondary branches of the blood vessel based on the detected centerline, wherein the volume of the plaque and the local thickness are determined with respect to the main branch, the main branch being preferably the common carotid artery.

3. A step of calculating a horizontal cut along the longitudinal axis of the blood vessel, The steps include tracking the 2D connection components of the aforementioned lateral cut, The steps include determining the centroid of each of the 2D connected components and providing an estimate of the portion of the center line, The steps include: applying a high-speed marching algorithm to the estimated portion of the centerline to detect the centerline of the blood vessel; The method according to claim 1, further comprising:

4. Steps to measure the narrowing of the blood vessel with respect to the center line of the detected blood vessel. The method according to claim 1, further comprising:

5. The method according to claim 1, wherein each of the vascular segmentation machine learning model and the plaque segmentation machine learning model has a 3DUNet model.

6. The method according to claim 1, wherein the step of performing 3D plaque segmentation of the blood vessel comprises using a vascular mask that fits the boundaries of the blood vessel and the ultrasound image, the vascular mask comprising only the internal volume of the blood vessel as defined by the boundaries of the blood vessel identified by performing the 3D vascular segmentation.

7. The steps include: executing an editing command to edit the 3D vascular segmentation to provide a modified 3D vascular segmentation of the vessels; and A step of automatically updating the 3D plaque segmentation in response to the modified 3D vascular segmentation of the blood vessels, and / or The step of executing an editing command to edit the 3D plaque segmentation to provide a modified 3D plaque segmentation of the plaque. It has, The volume of plaque within the boundary of the blood vessel is determined based on the modified 3D plaque segmentation. The method according to claim 1.

8. The editing command for correcting the 3D vascular segmentation has multiple point clicks in the ultrasound image received from the user interface, and the surface of the vascular boundary is repositioned inside or outside the original surface of the vascular boundary. The step of executing an editing command to modify the 3D blood vessel segmentation includes the step of moving the boundaries of the blood vessels to the relocated surface. The method according to claim 7.

9. The editing command for editing the 3D plaque segmentation has a selected radius of a virtual balloon received from the user interface within the region of the plaque provided by the 3D plaque segmentation in the ultrasound image, The step of executing an editing command to modify the 3D plaque segmentation includes moving the surface of the plaque region to match the selected radius of the virtual balloon, Optionally, the above method further, The steps include executing an editing command to correct the 3D plaque segmentation by displaying the movement of the surface area of ​​the plaque region on a display in real time so as to match the selected radius of the virtual balloon, and The method according to claim 7, having the following characteristics.

10. The local thickness of the plaque is determined relative to the detected centerline of the blood vessel using the curved coordinates of the detected centerline and the local two-dimensional axis perpendicular to the detected centerline, or The method according to claim 1, wherein the local thickness of the plaque in at least one cross-section of the blood vessel is determined as the maximum thickness of the plaque in the local 2D plane defined by the at least one cross-section.

11. The steps include: resampling the ultrasound image to provide a first resampled 3D ultrasound image; and performing 3D vascular segmentation of the blood vessels in the first resampled 3D ultrasound image. A step of resampling the ultrasound image to provide a second resampled 3D ultrasound image, and performing 3D plaque segmentation of the plaque in the second resampled 3D ultrasound image, wherein the portion of the ultrasound image showing the plaque is resampled at a higher resolution than the portion of the ultrasound image showing the blood vessels. The method according to claim 1, further comprising:

12. The method according to claim 1, further comprising the step of determining a 2D plaque region of at least one cross-section of the blood vessel with respect to the detected centerline, based on the 3D blood vessel segmentation and the 3D plaque segmentation.

13. The blood vessel has a carotid artery including the common carotid artery prior to the branching of the carotid artery, and the method further, The steps include confirming that the midline of the blood vessel enters the internal carotid artery from the common carotid artery, rather than from the external carotid artery, The step of switching the identification of the internal carotid artery and the external carotid artery when the midline of the blood vessel enters the external carotid artery from the common carotid artery, In the step of determining the local thickness of plaque in at least one cross-section of the vessel with respect to the detected centerline, based on the 3D vessel segmentation and the 3D plaque segmentation, the detected centerline is the detected centerline of the common carotid artery branching into the internal carotid artery. The method according to any one of claims 1 to 12, comprising

14. A system for evaluating the plaque load in the blood vessels of a subject, wherein the system is A display configured to display a 3D ultrasound image of the blood vessel acquired using non-invasive ultrasound imaging, User interface and A processor that communicates with the display and the user interface, configured to perform the method described in any one of claims 1 to 13, and A system that has

15. A computer program product having instructions for evaluating the plaque load in the blood vessels of a subject, wherein, when executed by a processor, the computer program product is configured to cause the processor to execute the method according to any one of claims 1 to 13.

Images