Container shape

Abstract

In one embodiment, a method (100) is described. The method includes identifying (102) a set of vascular landmarks within a subject's body using a machine learning ML model configured to identify adjacent vascular landmarks from radiological imaging data of the subject's body. The method further includes determining (104) a path having the identified adjacent landmarks in the set. A shape of the vessel between the identified adjacent landmarks is determined based on the determined path and imaging conditions indicative of a wall of the vessel.

Description

[Technical field]

[0001] The present invention relates to a method, non-transitory machine-readable medium and apparatus for determining blood vessel geometry, for example from computed tomography (CT) imaging data. [Background technology]

[0002] A subject's body may be imaged using a CT imaging device for various purposes, such as diagnosis, medical intervention, etc. Three-dimensional (volume) reconstructed images and / or associated multiplanar reconstruction (MPR) images obtained from imaging data, such as a set of CT imaging slices, may show various anatomical structures within the subject's body. However, certain structures, such as blood vessels (e.g., blood vessels such as arteries and veins), may not be readily identifiable in the reconstructed images.

[0003] For example, during analysis of head and neck vessels in certain post-processing applications, a user, such as a clinician or other imaging device operator, may define vessels by extracting (e.g., identifying) and manually naming the vessels on the image. This extraction / naming procedure may be performed in conjunction with observing and / or measuring possible lesions. It may take a user several minutes to perform this extraction and naming procedure, especially if the vessels are long and / or have complex curvatures. Summary of the Invention [Problem to be solved by the invention]

[0004] The time a user spends during vessel extraction and the accuracy of the vessel extraction may depend heavily on the experience of the user. [Means for solving the problem]

[0005] Aspects or embodiments described herein may relate to vessel shape determination or segmentation based on certain radiological image data. Such vessel shape determination / segmentation may be performed automatically with minimal or no user input. Aspects or embodiments described herein may avoid one or more problems associated with user (i.e., human) identification of vessel shapes from radiological image data.

[0006] In a first aspect, a method is described. The method is a computer-implemented method. The method includes identifying a set of vascular landmarks in a subject's body using a machine learning ML model configured to identify adjacent vascular landmarks from radiation image data of the subject's body. The method further includes determining a path including the identified adjacent landmarks in the set. The method further includes determining a shape of the vessel between the identified adjacent landmarks based on the determined path and imaging conditions indicative of a vessel wall.

[0007] Several embodiments relating to the first aspect and other aspects are described below.

[0008] In some embodiments, the ML model includes a Deep Q Network, DQN model.

[0009] In some embodiments, the ML model is trained to recognize a set of landmarks based on a training dataset that includes radiological image data obtained from a plurality of training subjects, each of which includes a target vessel that is annotated with the set of landmarks to be recognized.

[0010] In some embodiments, determining the path includes identifying a pixel intensity threshold within the radiation image data, the pixel intensity threshold may be indicative of identified landmarks that are connected to form the path.

[0011] In some embodiments, the pixel intensity threshold is determined based on a minimum pixel intensity value and / or an average pixel intensity value associated with each of the identified landmarks.

[0012] In some embodiments, the method includes determining a route by identifying coordinates of corresponding vascular structures that indicate a vascular segment between adjacent landmarks, and estimating a profile of the route based on the coordinates.

[0013] In some embodiments, estimating the path profile includes determining a shortest path between adjacent landmarks by setting a first landmark as a seed point and identifying, within the iteratively growing volumetric region, a second landmark that corresponds to a closest landmark to the first landmark, such that coordinates associated with the first landmark define a path between the first landmark and the second landmark.

[0014] In some embodiments, determining the path includes identifying a vascular structure corresponding to at least a portion of a cross-section of an ellipse and determining a center point of the ellipse. The center point may define coordinates of a centerline of the blood vessel that defines the path.

[0015] In some embodiments, the imaging conditions indicative of a wall of a vessel are based on gradient conditions indicative of the presence of a wall.

[0016] In some embodiments, the method includes determining coordinates corresponding to a center point of the vessel defining an origin according to a local coordinate system, and determining a shape of the vessel in a vessel cross section perpendicular to the path at the origin by simulating a number of rays extending radially from the origin until a gradient condition is satisfied for each radial ray at a specified distance from the origin. The specified distance of each radial ray from the origin may be used to estimate a boundary in the vessel cross section corresponding to at least a portion of the shape of the vessel.

[0017] In some embodiments, the method includes determining a vessel segment that corresponds to the shape of the vessel between adjacent landmarks by connecting estimated boundaries from adjacent vessel cross-sections.

[0018] In some embodiments, the gradient condition is determined based on a gradient model of pixel intensity values ​​in the vessel cross-section.

[0019] In some embodiments, the blood vessel includes an internal carotid artery, ICA, and / or a vertebral artery, VA. The set of landmarks may include a set of identified landmarks LX from at least one of the internal carotid artery and the vertebral artery of the subject identified by the ML model. Each landmark LX in the identified set may have an associated index X and a corresponding anatomical feature location as follows:

[0020] L0 is shared by the internal carotid and vertebral arteries and is located in the brachiocephalic trunk of the aortic arch of the subject.

[0021] L1 is located at the common carotid artery of the aortic arch.

[0022] L2 is located in the subclavian artery of the aortic arch.

[0023] L3 and L4 are located at the bifurcation of the internal carotid artery and the subject's external carotid artery.

[0024] L5 and L8 are the locations of the internal carotid arteries around the subject's C6 cervical vertebra.

[0025] L9 and L14 are the locations of the vertebral arteries surrounding the subject's C6 cervical vertebra.

[0026] L6 and L7 are the locations of the internal carotid arteries around the subject's C2 cervical vertebra.

[0027] L10 and L13 are the locations of the vertebral arteries surrounding the subject's C2 cervical vertebra.

[0028] L11 and L12 are the locations of the internal carotid arteries around the subject's C1 cervical vertebra.

[0029] L15 and L18 are the locations of the vertebral arteries surrounding the subject's C1 cervical vertebra.

[0030] L16 and L17 are the locations of the internal carotid arteries through which blood is transferred to structures inside and outside the skull.

[0031] L19 and L22 are the locations of the vertebral arteries on the subject's posterior fossa where the vertebral arteries enter the subject's skull.

[0032] L20, L21, L23 and L24 are the locations of the intracranial internal carotid artery, and L25 is a location shared by both the left and right vertebral arteries of the subject.

[0033] In some embodiments, the method includes causing a user interface for displaying a simulation of the subject's body to display a shape of the blood vessel via the user interface.

[0034] In a second aspect, a non-transitory machine-readable medium is described comprising instructions that, when executed by at least one processor, cause the at least one processor to implement a method according to the first aspect or any of the related embodiments.

[0035] In a third aspect, a device is described for extracting vascular information from radiation image data of a subject. The device comprises at least one processor communicatively coupled to an interface for receiving the radiation image data. The device further comprises a non-transitory machine-readable medium storing instructions readable and executable by the at least one processor to perform a method according to the first aspect or any of the related embodiments.

[0036] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

[0037] Exemplary embodiments of the present invention will now be described, by way of example only, with reference to the following drawings, in which: [Brief description of the drawings]

[0038] [Figure 1] Reference is now made to a method for determining vessel shape according to one embodiment. [Diagram 2] 1 is a schematic diagram of a system for determining vascular geometry according to one embodiment. [Diagram 3] Reference is now made to a method for determining vessel shape according to one embodiment. [Figure 4] 2 shows various views of a three-dimensional reconstruction of a subject's body from radiation image data. [Diagram 5] 2 shows various views of a three-dimensional reconstruction of a subject's body from radiation image data. [Figure 6] 1 shows a representation of landmarks detected on a subject's body, according to one embodiment. [Figure 7] 1 is a schematic diagram of a model of a blood vessel cross-section according to one embodiment. [Figure 8] Reference is now made to a method for extracting paths between connected landmarks according to one embodiment. [Figure 9] Reference is now made to a method for determining a boundary of a vessel according to an embodiment. [Figure 10] 4 illustrates the results of a vessel shape determination procedure according to an embodiment. [Figure 11] 1 is a schematic diagram of a machine-readable medium for determining container shape according to one embodiment. [Figure 12] 1 is a schematic diagram of an apparatus for determining container shape according to one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0039] 1 illustrates a method 100 (e.g., a computer-implemented method) for determining vascular geometry. Method 100 may be implemented by a computer, such as a user interface, or a user computer communicatively coupled to a server or cloud-based service (e.g., communicatively coupled to the user computer and / or the user interface).

[0040] The method 100 includes, at block 102, identifying a set of vascular landmarks within the subject's body using a machine learning ML model configured to identify adjacent vascular landmarks from radiation image data of the subject's body.

[0041] As depicted by Figures 5 and 6 (described in more detail below), a "landmark" refers to a location within a subject's body relative to a blood vessel. A blood vessel within a body can be considered relative to the body's anatomy. The body's anatomy can refer to the arrangement of bones and tissues within the body. Although there may be some differences between each subject, the anatomy can be expected to be similar for each subject. Thus, a particular blood vessel within a subject's body can have a profile and / or shape that is expected relative to the subject's anatomical features.

[0042] The ML model may be trained to identify a set of landmarks that may be pre-determined based on anatomical structure. For example, the ML model described in more detail below (including, in some embodiments, a Deep Q Network, DQN, model) may be trained to identify landmarks associated with at least one of the subject's internal carotid artery (ICA) and / or vertebral artery (VA) in the head, neck, and / or thoracic regions of the subject's body (although a similar procedure may be used to identify landmarks associated with other blood vessels in the subject's body). From the user's perspective, these landmarks may be distinguishable from each other based on certain factors, such as distance from each other, proximity to anatomical features such as bones, organs, etc. Such factors may or may not be taken into account by the ML model when identifying the set of landmarks. For example, once trained, the ML model may appropriately weight any information derived from the radiological image data to facilitate landmark identification using the ML model. By appropriate selection of the set of landmarks to be used for training (e.g., by a human expert based on anatomical structure), the ML model may enable repeatable and accurate identification of the set of landmarks in each subject.

[0043] Radiographic data may refer to raw data obtained from a radiographic imaging device such as a CT scanner, and / or may refer to processed data derived from the raw data, such as three-dimensional reconstructed images and / or associated MPR images.

[0044] The method 100 further includes, at block 104, determining a path that includes the identified adjacent landmarks in the set.

[0045] In some embodiments described below, a path may be determined based on various information sources, such as a pixel intensity threshold indicating a path connecting adjacent landmarks in a set, and / or identifying the nearest landmark within an enhancement region in the radiation image data.

[0046] Thus, identifying connected landmarks of a vessel may enable identifying a path in the radiological image data, such a path may correspond to a profile (e.g., a centerline) of the vessel.

[0047] Block 104 further includes determining a shape of the vessel between the identified adjacent landmarks based on the determined path and imaging conditions indicative of the vessel walls.

[0048] An imaging condition indicative of a wall of a blood vessel may refer to a portion of the radiation image data (e.g., at least one pixel in the radiation image data) that appears to correspond to a structure that appears to represent a wall of a blood vessel. For example, if a blood vessel has an elliptical (e.g., including a circular) cross-section as determined based on the imaging condition, the portion of the radiation image data in the vicinity of the identified landmark that satisfies the imaging condition may be considered a wall of the blood vessel. In this manner, a combination of information about the path (e.g., centerline) of the blood vessel and information about the wall of the blood vessel at each relevant point along the path (e.g., center point of the blood vessel at each of the identified landmarks) may be used to determine the shape of the blood vessel between the connected landmarks. This determination of the blood vessel shape may be referred to as "vessel segmentation." By determining the blood vessel shape, a user interface may highlight the shape of the blood vessel in a radiographic image (e.g., a three-dimensional reconstruction image and / or an MPR image) of the subject's body. For example, the blood vessel may be highlighted in a different color relative to the remaining structures in the image, annotated with text, or the like, thereby allowing a user to visualize and / or more easily or accurately identify or distinguish the blood vessel in the image.

[0049] Accordingly, certain methods described herein (e.g., method 100 and other methods described below) may be capable of automatically identifying certain blood vessels within a subject's body with minimal or no human input. This identification may include extracting and / or naming the blood vessels such that a user interface for visualizing the processed radiological image data of the subject's body may display and / or highlight the blood vessel shape. This may save a user time and / or help the user reduce blood vessel identification errors. Automated extraction of blood vessels from radiological image data may aid in efficiently and accurately analyzing such blood vessels, which may be useful in certain procedures, including identifying and / or measuring lesions.

[0050] FIG. 2 illustrates a system 200 with certain components for implementing certain methods described herein (e.g., method 100 and other methods described below). The system 200 includes a radiological imaging device (e.g., a CT scanner or other type of radiological imaging device, e.g., a magnetic resonance imaging (MRI) scanner) 202, a treatment couch 204, and a subject 206 (e.g., a patient) supported by the treatment couch 204. The radiological imaging device 202 is communicatively coupled to a computer 208 configured to receive radiological image data from the radiological imaging device 202. The computer 208 includes a memory 210 (e.g., a non-transitory machine-readable medium) that stores instructions for implementing certain methods described herein, and at least one processor 212 for executing instructions. The computer 208 is communicatively coupled to a user interface 214 (e.g., a display) that provides certain information derived from the radiological image data. For example, a three-dimensional reconstruction image and / or an MPR image of the subject's body may be displayed on the user interface 214. Additionally, output from certain methods described herein (eg, method 100 ) may be displayed on the user interface 214 .

[0051] The computer 208 and / or the user interface 214 may be implemented locally or remotely. For example, the computer 208 may be a local machine (e.g., a user computer located near the radiography device 202 or at another location) or a remote machine or computing system (e.g., a server and / or the cloud). The user interface 214 may be implemented locally (e.g., near the radiography device 202, e.g., in a hospital where the radiography device 202 is located) or remotely (e.g., where analysis is performed collaboratively with the involvement of a remote expert).

[0052] 3 illustrates a method 300 (e.g., a computer-implemented method) for determining vascular geometry. Method 300 may be implemented by a computer, such as a user interface, or a user computer communicatively coupled to a server or cloud-based service (e.g., communicatively coupled to the user computer and / or the user interface). Method 300 includes method 100 and refers to certain methods described below. Some blocks of method 300 may be performed in a different order than illustrated by FIG. 3 or may be omitted.

[0053] After start 302 of method 300, in block 304, method 300 includes calculating a starting point for performing landmark detection. The starting point may be any landmark within the subject's body that the ML model is configured to initially recognize as part of the landmark identification procedure according to block 102 of method 100. Exemplary starting points are provided in the description below.

[0054] At block 306 of method 300, landmark detection is performed (eg, according to block 102 of method 100) to detect landmarks on the target vessels (eg, the ICA and / or the VA).

[0055] In some embodiments, the ML model is trained to recognize a set of landmarks based on a training dataset obtained from multiple training subjects. Radiographic image data obtained from each training subject may include a target vessel annotated with a set of landmarks to be recognized. For example, radiographic image data associated with each training subject may be annotated by a human trainer to indicate a set of landmarks associated with a target vessel within the body of the training subject. The set of landmarks may be selected based on anatomical structures (e.g., bones or other tissues, etc.) associated with the vessel. For example, a particular landmark on a vessel may be relatively easy for the ML model to identify if a portion of the vessel is adjacent to a particular anatomical structure that is relatively easy to distinguish from other anatomical structures. Thus, portions of the vessel that can be easily identified may be candidate landmarks to be used for the set of landmarks.

[0056] In one example, the training dataset includes radiological image data acquired from 139 training subjects (i.e., 139 series). The radiological image data can be obtained from the same or different types of radiological imaging devices (e.g., from different machines and / or from different countries), and 129 of the 139 series are used for training, where most of the data includes 26 annotated landmarks per series (not all series have 26 landmarks). The remaining 10 series of the 139 series were used for validation of the training dataset. There were a total of about 3614 landmarks (139 x 26) in the training dataset. A coordinate transformation method was used on the radiological image data to prepare the data for the training procedure. In this example, a landmark detection accuracy rate of 96.680% was obtained from the 138 test dataset (e.g., from the original series of the training data).

[0057] According to the above example, once the head and neck enhanced data is loaded, the trained ML model may be able to detect a total of 26 landmarks (discussed below) on the internal carotid and vertebral arteries.

[0058] As described below, the detected landmarks may be optimized to improve the accuracy of the detected location of each landmark. Thus, once the ML model has determined the "approximate" locations of a set of landmarks in the radiological image data, a different method is used for optimization (e.g., a non-ML model called a "gradient model" is used, although it should be understood that ML methods may be used instead).

[0059] In block 308 of method 300, the detected landmarks are optimized (e.g., by locating their positions more accurately) based on a gradient model of a predicted profile of pixel intensity values ​​(e.g., a "pixel intensity value" may refer to the measured intensity at each spatial location (pixel) in the radiation image data across a cross-section of the vessel). For example, a vessel may have an elliptical (e.g., including circular) cross-section perpendicular to a path defined along the vessel. The initial landmark detection performed in block 306 may not necessarily identify a point that is at the center of the vessel. However, by assuming that the pixel intensity values ​​vary in a certain way across the cross-section (e.g., have an "expected profile"), the center point of the vessel at the landmark can be calculated. In the gradient model described below, the center point of the vessel corresponds to the highest pixel intensity value, and the pixel intensity values ​​decrease radially from the center point. Thus, the optimization in block 308 includes identifying a region in the radiation image data that corresponds to a predicted profile of the cross-section of the vessel, fitting the radiation image data in the region according to the gradient model, and calculating the center point based on the fitted data. As used herein, a cross section may refer to any cross section of image data (e.g., in an axial, coronal, or sagittal plane, or in an oblique plane). Although the examples shown herein refer to axial cross sections, some examples refer to cross sections in other planes.

[0060] In block 310 of method 300, which may correspond to an embodiment of block 104 of method 100, a pixel intensity threshold (e.g., a "CT value" threshold in the case of CT image derived pixel intensity values) is estimated based on pixel intensity values ​​obtained in the vicinity of a set of detected landmarks. The pixel intensity threshold may indicate a path connecting adjacent landmarks in the set based on the radiological image data. For example, the pixel intensity threshold may refer to similar pixel intensity values ​​(e.g., within a range) obtained in the vicinity of each of the set of landmarks. Since each of these pixel intensity values ​​corresponds to a vasculature, it may be assumed that other portions of the radiological image data (e.g., in other CT slices) having similar pixel intensity values ​​may also correspond to a vasculature. To prevent or avoid other anatomical features that are not vasculature being assumed to be part of the vasculature, the region for searching (e.g., using the pixel intensity threshold) for other portions of the vasculature is the path of adjacent landmarks that are connected.

[0061] Radiation image data may include a series of measured pixel intensity values ​​acquired by a radiation imaging device. In the absence of human input, the pixel intensity values ​​are relatively complex for a computer (e.g., a computer program executable by a processor) to interpret in terms of recognizing whether a pixel or set of pixels corresponds to a tissue. For example, a human can visually inspect an image and identify anatomical structures (including the blood vessels themselves), but a computer may not necessarily be able to automatically classify different anatomical structures when the measurements have similar values. For example, certain anatomical structures may have similar measurements that are completely different but cannot be understood by a computer.

[0062] By determining a pixel intensity threshold indicative of a path connecting adjacent landmarks, method 300 can enable a computer to distinguish between related and non-related measurements, for example, for purposes of automatically identifying portions of the radiological image data related to a blood vessel (e.g., via a set of identified landmarks). In other words, with appropriate identification of a pixel intensity threshold (e.g., based on the radiological image data related to a set of identified landmarks), method 300 can filter the radiological image data based on the pixel intensity threshold, such that method 300 can determine whether or how adjacent landmarks are connected. In cases where there are multiple blood vessels, certain landmarks in the set may not be connected.

[0063] At block 312 of method 300, which may or may not be performed in conjunction with block 310, a vascular path, such as a centerline between adjacent landmarks that are connected, is extracted. In one implementation, this may be based on the shortest (straight line) path between adjacent landmarks (e.g., based on the coordinates of the center points of the adjacent landmarks), while in other implementations the path may be modeled on a different profile, such as a curved profile.

[0064] At block 314 of method 300, the boundary of the vessel is calculated. The boundary may refer to the wall of the vessel. The boundary calculation may be based on the imaging conditions referenced at block 104 of method 100. For example, a gradient model may assume that pixel intensity values ​​may vary in a characteristic way across the boundary (e.g., due to structural changes). By identifying parts in the radiological image data that exhibit such characteristic variations, the boundary of the vessel may be calculated. Using information from the boundary calculation at each point along the path determined by block 312, it may be possible to determine the vessel shape between adjacent landmarks. This procedure may be extended along the vessel to extract or segment the vessel between the aortic arch and / or the connecting portions of the vessel associated with the brain (i.e., the ICA and / or the VA).

[0065] At block 316 of the method 300, the ICA and / or VA associated with the aortic arch segment may be extracted using the same procedure described above.

[0066] At block 318 of the method 300, the same procedure as described above can be used to extract the ICA and / or VA associated with the cerebrovascular segment.

[0067] Block 320 of the method 300 includes causing a user interface (eg, user interface 214) to display a simulation of the subject's body to indicate a shape of the blood vessel via the user interface.

[0068] At block 322, the method 300 ends.

[0069] Further details of some of the specific blocks of method 300 are described below.

[0070] The internal carotid and vertebral arteries are two commonly studied vessels of the head and neck in X-ray image data.

[0071] Figures 4 and 5 show exemplary views (side and front views) of 3D reconstruction images of various anatomical features in the head and neck of a subject. In Figure 5, landmark detection as described herein is performed and shows the location of each of the detected landmarks 502 (see angled arrows) based on the image shown in Figure 4. As shown by Figure 5, some of the detected landmarks are "adjacent" to each other.

[0072] Figure 6 shows a simpler representation of the landmarks depicted by the chevron arrows in Figure 5. The landmarks are labeled LX in the order of detection, from index X = 0 to X = 25. Thus, L0 refers to the first landmark detected, L25 refers to the last landmark detected (i.e., a total of 26 landmarks), etc. Thus, the ML model is trained to recognize landmarks LX in the radiological image data, and each landmark LX is associated with an anatomical feature of the subject.

[0073] The number and accuracy of the landmarks can determine the completeness and accuracy of the vessel extraction. The selection of the landmarks for detection can be set according to the size of the target vessel(s), since the range of the landmarks can determine the range of the vessel extraction. For example, more landmarks may be needed for a long vessel, and the location of the landmarks is appropriately selected to cover the entire length of the vessel. If the detection subject is not within the data range, the detection accuracy may be reduced. Considering that the bottom of the head and neck CT scan generally includes the aortic arch, in the example described herein, three landmarks are selected on the three branches of the aortic arch (i.e., the brachiocephalic trunk, the left common carotid artery, and the left subclavian artery). The detection targets can have certain characteristics to make the detection more accurate, such as the proximity of certain anatomical features. However, landmarks close to each other tend to converge, which means redundancy does not help in vessel extraction. Therefore, the distribution of the landmarks over the detection range can be appropriately selected to avoid the concentration of too many landmarks close to each other, while ensuring a reasonably uniform distribution of the landmarks over the entire vessel in the window. An example of the selection of the set of selected landmarks for the ICA and VA is shown in FIG. 6.

[0074] Thus, in some embodiments where the vessel includes an ICA and / or a VA, the set of landmarks includes a set of identified landmarks "LX" from at least one of the ICA and VA of the subject identified by the ML model. These are shown as the left ICA and VA in Figures 4 and 5, but the method can be extended to other ICAs and / or VAs and / or other vessels within the subject's body. Each landmark "LX" in the identified set has an associated index "X" and a corresponding anatomical feature location as follows:

[0075] L0 is shared by the ICA and VA and is located in the brachiocephalic trunk of the subject's aortic arch.

[0076] L1 is located in the (eg, left) common carotid artery of the aortic arch.

[0077] L2 is located in the (eg, left) subclavian artery of the aortic arch.

[0078] L3 and L4 are located at the bifurcation of the ICA and the subject's external carotid artery.

[0079] L5 and L8 are the locations of the ICA around the subject's C6 cervical vertebra.

[0080] L9 and L14 are the locations of VA around the subject's C6 cervical vertebra.

[0081] L6 and L7 are the locations of the ICA around the subject's C2 cervical vertebra.

[0082] L10 and L13 are the locations of VA around the subject's C2 cervical vertebra.

[0083] L11 and L12 are the locations of the ICA around the C1 cervical vertebra of the subject.

[0084] L15 and L18 are the locations of VA around the subject's C1 cervical vertebra.

[0085] L16 and L17 are the locations of the ICA where blood is transferred to structures on the inside and outside of the skull.

[0086] L19 and L22 are the locations of the VAs on the subject's posterior fossa where the VAs enter the subject's skull.

[0087] L20, L21, L23 and L24 are the locations of the ICA within the skull.

[0088] L25 is a location shared by both the left and right VA of the subject.

[0089] The ML model may be configured to identify a specified set of landmarks LX in order from X=0 to X=25. Thus, the model may identify L0 first, then identify the rest of the landmarks LX to proceed to L25 at the end of the detection process. In FIG. 6, the arrows represent the order of detection at each of the landmarks "LX" (i.e., L1 is detected after L0, L3 is detected after L4, L5 and L6 are detected after L7 and L8, etc.). The order is based on the proximity of the landmarks from each other. The landmark closest to L0 is L1, the landmark closest to L2 is L1, and so on.

[0090] Because different image pixel intervals may represent different characteristics, the training image and the detection image may have a unified image pixel interval. If the image pixel interval is too large, the detection accuracy may decrease (features near the landmarks are not obvious), and if the pixel interval is too small, the calculation efficiency may be affected. Therefore, to improve the detection accuracy and efficiency, the pixel interval as large as possible is selected. According to the 138 radiographic image dataset used in the test, the pixel interval of the calculation is selected to be 0.8 mm.

[0091] In some embodiments, landmark detection is performed by an ML model (such as the DQN model in some embodiments) using a Markov decision process (MDP). The ML model may be based on a convolutional neural network (CNN) architecture. As part of the MDP process, the artificial agent learns to make certain decisions aimed at finding a target landmark. An environment may be defined in an input image. The agent navigates the environment using a set of actions (i.e., moving up, down, left, right, forward, or backward). The agent aims to find an anatomical landmark based on the results obtained from a set of actions (e.g., the agent's movement direction) and states (e.g., positions corresponding to potential landmark positions), along with a reward function (e.g., to indicate whether the agent is close to a ground truth target position). The agent may be considered to have found the target landmark when the agent oscillates around the location when performing its actions, at which point the target landmark is considered to be located. In other words, the Q-value associated with a state is lower the closer it is to the target landmark during training, such that the agent may be rewarded for finding the lowest Q-value at a location. During testing, the agent behaves according to the ML model it is trained on and does not receive any rewards and / or update the model.

[0092] After landmark detection, a method is performed (eg, see block 306 of method 300) to verify that the detected points are located on the vessel (ie, verify that the landmark is at the center of the vessel).

[0093] FIG. 7 shows an exemplary model of a cross-section of a blood vessel perpendicular to the path of the blood vessel. In this exemplary model, which may be referred to as a gradient model, the highest pixel intensity value (represented by dark shading) is obtained at the center of the blood vessel, and radially outer locations have lower pixel intensity values ​​(represented by lighter shading). Thus, if the radiation image data shows the exemplary model, the center point of the blood vessel can be located by finding the center of an ellipse (such as a circle). Thus, the location of the landmark can be optimized to the center of the blood vessel based on the gradient model.

[0094] In an exemplary implementation using CT scan data, an iteration around the detection result is implemented. The center of the blood vessel is higher than the surrounding CT value. The evaluation function of the blood vessel cross section is designed to find the center of the closest circular region. The gradient model can be based on the CT image values ​​of the blood vessel belonging to a normal distribution defined as follows: TIFF2024504025000002.tif22117

[0095] where ct is the CT image value of the point being tested, M is the predicted CT value, and ε is a constant that is a function of the distance between the test point and the target point.

[0096] To prevent the improved / optimized point position from being too far from the origin, the number of iterations is set to 3, with each iteration corresponding to a movement of 1 pixel.

[0097] Although the landmarks L19, L22, and L25 can determine the region of VA, they are not always accurate enough. Therefore, a shape-based method such as the one described above can improve the detection results.

[0098] The target landmarks may be displayed on a viewing window in the user interface. The viewing window may be set such that the independent region of the target landmark (i.e., the vessel cross-section) is not connected to the window boundary. Thus, the window size may be predefined to be, for example, slightly larger than the vessel cross-section.

[0099] In the case of L25 (and / or any other landmark), the vascular region may be connected to a bone region (or other anatomical feature that results in a vessel in the slice not in an independent region in the window), and therefore other portions of the radiological image data may be used to find an independent region containing a vessel associated with the landmark. For example, the (axial) slice may be checked first. If there is no independent region in the slice, the coronal slice may be checked next to identify whether there is an independent region, followed by the sagittal slice (or in any other suitable order).

[0100] Once the detected landmarks are determined, a pixel intensity threshold for the target vessel and a boundary of the vessel region can be determined (e.g., according to blocks 310 and 314 of method 300). The pixel intensity threshold may be low enough to connect adjacent landmarks, while being high enough to avoid significant leakage into bones connecting the vessel. In this example, the pixel intensity threshold is estimated by using pixel intensity values ​​from 26 detected landmarks.

[0101] A path (e.g., corresponding to a centerline) can be found by finding the shortest connected path between adjacent landmarks on the vessel.

[0102] In some embodiments, determining the path includes identifying a pixel intensity threshold within the radiation image data, the pixel intensity threshold may be indicative of identified landmarks that are connected to form the path.

[0103] In some embodiments, a pixel intensity threshold is determined based on the minimum pixel intensity value for connecting adjacent identified landmarks.

[0104] In some embodiments, a pixel intensity threshold is determined based on the average pixel intensity value associated with each of the identified landmarks, A range may be defined around the pixel intensity threshold, and only pixel intensity values ​​within this range may be selected for use in certain methods described herein.

[0105] 8 illustrates a method 800 for extracting (e.g., identifying) paths between connected landmarks. Method 800 may be implemented as part of methods 100 or 300 (e.g., according to block 312).

[0106] Method 800 begins at block 802. At block 804 of method 800, a first landmark is set as a seed point in the radiological image data. At block 806 of method 800, a region around the seed point is grown such that at block 808 of method 800, a determination is made as to whether a second (adjacent) landmark is within the region. If not, method 800 returns to block 806. If so, method 800 ends at block 810.

[0107] In some embodiments, the path is determined (e.g., according to block 312 and / or via method 800) by identifying coordinates of corresponding vasculature that indicate vascular segments between adjacent landmarks and estimating a profile of the path based on the coordinates. In some embodiments, the identification of the coordinates is based on a pixel intensity threshold.

[0108] In some embodiments, estimating the path profile includes determining a shortest path between adjacent landmarks by setting a first landmark as a seed point and identifying, within the iteratively growing volumetric region, a second landmark that corresponds to a closest landmark to the first landmark, such that coordinates associated with the first landmark define a path between the first landmark and the second landmark.

[0109] A corresponding contour / profile of the path can be calculated based on each center point identified according to method 800. For example, the contour / profile may be obtained by taking the center point as the coordinate origin, establishing a local coordinate system, and calculating, for example, X and Y direction vectors of the coordinate system passing through adjacent center points.

[0110] To determine the coordinates of the vessel wall (e.g., according to block 314 of method 300), rays are emitted from each of the center points to find sampling points that satisfy a local gradient value condition (e.g., "imaging condition") as a boundary. As discussed below, the calculation of the gradient condition indicating the boundary can be based on the selection of a strong or weak boundary model. Once the gradient condition is determined based on the selected model (e.g., "gradient model"), the boundary coordinates around the center point can be determined according to the pixel intensity values ​​in the section that corresponds (fits) the gradient condition. Depending on the concentration of the radial rays, a smoothing algorithm (e.g., to generate an elliptical cross section corresponding to a vessel lumen-like shape) can be used to make the boundary more realistic (e.g., to avoid sharp transitions in distance from the center point around the origin).

[0111] The "local gradient value condition" may refer to the "maximum gradient value" that corresponds to the edge of the vessel (ie, when the gradient is maximum, it corresponds to the edge of the vessel).

[0112] Thus, in some embodiments, determining the route includes identifying a vascular structure corresponding to at least a portion of a cross-section of an ellipse and determining a center point of the ellipse, the center point defining the coordinates of a centerline of the vessel that defines the route.

[0113] In some embodiments, the imaging conditions indicative of a wall of a vessel are based on gradient conditions indicative of the presence of a wall.

[0114] In some embodiments, a particular method (e.g., block 314 of method 300) includes determining coordinates corresponding to a center point of a vessel that defines an origin according to a local coordinate system, and determining a shape (e.g., a set of coordinates corresponding to a shape) of the vessel in a vessel cross-section perpendicular to the path at the origin by simulating a number of rays extending radially from the origin until a gradient condition is satisfied for each radial ray at a specified distance from the origin. The specified distance of each radial ray from the origin may be used to estimate a boundary in the vessel cross-section that corresponds to at least a portion of the shape of the vessel.

[0115] In some embodiments, a particular method (e.g., part of block 104 of method 100) includes determining a vessel segment that corresponds to the shape of the vessel between adjacent landmarks by connecting estimated boundaries from adjacent vessel cross sections.

[0116] In some embodiments, the gradient condition is determined based on a gradient model of pixel intensity values ​​in a vessel cross section. The problem of identifying the gradient condition can be based on a model including a strong boundary model and a weak boundary model of the boundary. Either or both of these models can be used to calculate the boundary. Thus, the gradient condition can be determined by calculating a corresponding energy function for each model and selecting whether to use the strong boundary model or the weak boundary model based on the minimum energy obtained by calculating the energy function based on the pixel strong intensity values.

[0117] 9 illustrates a method 900 for determining a boundary (e.g., a wall) of a blood vessel. Method 900 may be implemented as part of methods 100 or 300 (e.g., according to block 314).

[0118] The method 900 has two parts, the first part refers to vessel boundary detection and the second part refers to vessel smoothing. In FIG. 9, "i" refers to point identification in the path (e.g., identification of each of the center points in the centerline), and "N" refers to the number of points in the path. In some cases (depending on the slice number of the data), there may be about 100 sections, each of which may include a center point.

[0119] The method 900 begins at block 902 .

[0120] At block 904 of the method 900, the index "i" of the center point is input.

[0121] At block 906 of method 900, if "i" is less than "N", then a local coordinate system is established at block 908, then a boundary is calculated based on a gradient condition at block 910, and an energy function is calculated to optimize the boundary (e.g., find the most likely coordinates of the boundary around center point "i") at block 912. The index is incremented according to "i++" at block 914, after which method 900 returns to block 906 and repeats blocks 908-912 for each center point identification.

[0122] Once the boundary of each center point is found (i.e., "i=N"), the method 900 proceeds to block 916, where smoothing is performed within the plane of the center point (i.e., its cross section). The method 900 then proceeds to block 918, where smoothing is performed across the plane (e.g., between adjacent boundaries around the center point). The method 900 then proceeds to block 920, where boundaries around the paths between adjacent center points are filled to ensure that a smooth shape is defined for the vessel segment. In other words, after calculation of the vessel boundary on each cross section, a smoothing procedure is performed in which the boundaries that are respectively smoothed on each cross section and the boundaries defined by the segments between adjacent cross sections are smoothed (i.e., along the centerline). The method 900 ends at block 922.

[0123] In some embodiments, the method 900 segments the aortic arch and cerebral vessels separately, and thus can automatically extract four vessels (e.g., at least one ICA and / or at least one VA, e.g., left VA and / or right VA), the aortic arch and / or cerebral vessels, potentially with a single button press.

[0124] 10 illustrates an example result of segmentation according to some methods described herein in which the above-referenced vessels are extracted and highlighted for display in a user interface (e.g., user interface 214) (e.g., according to block 320 of method 300). The methods described herein can be used to extract other vessels within a subject's body and are not limited to extracting the vessels shown in FIG.

[0125] 11 illustrates a non-transitory machine-readable medium 1100. The non-transitory machine-readable medium 1100 comprises instructions 1102 that, when executed on at least one processor 1104, cause the at least one processor 1104 to implement certain methods described herein. In this embodiment, the instructions 1102 include instructions 1106 configured to implement block 102 of the method 100. The instructions 1102 further comprise instructions 1108 configured to implement block 104 of the method 100. Any of the methods described herein may be implemented by the non-transitory machine-readable medium 1100 that causes the at least one processor 1104 to implement such methods.

[0126] 12 shows an apparatus 1200 for determining a shape of a blood vessel. The apparatus 1200 comprises at least one processor 1202 (which may correspond to the at least one processor 212 referred to in FIG. 2). The at least one processor 1202 is communicatively coupled to an interface 1204 (e.g., a "receive" interface) for receiving radiation image data of a subject acquired by a radiation imaging device (e.g., the radiation imaging device 202). The interface 1204 may be part of the computer 208 referred to in FIG. 2.

[0127] The apparatus 1200 further comprises a non-transitory machine-readable medium 1206 (corresponding to memory 210 in FIG. 2 ) that stores instructions 1208 readable and executable by the at least one processor 1202 to perform a method corresponding to some of the methods described herein (e.g., any of methods 100, 300, 800, and / or 900).

[0128] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered as illustrative or exemplary and not restrictive, and the invention is not limited to the disclosed embodiments.

[0129] One or more features described in one embodiment may be combined with or substituted for features described in another embodiment.

[0130] Embodiments of the present disclosure may be provided as a method, a system, or a combination of machine-readable instructions and processing circuitry. Such machine-readable instructions may be contained on a non-transitory machine (e.g., computer) readable storage medium (including, but not limited to, disk storage devices, CD-ROMs, optical storage devices, etc.) having computer-readable program code therein or thereon.

[0131] The present disclosure will be described with reference to flowcharts and block diagrams of methods, apparatuses, and systems according to embodiments of the present disclosure. Although the above flowcharts show a specific instruction execution order, the instruction execution order may differ from that shown. Blocks described in connection with one flowchart may be combined with blocks of another flowchart. It will be understood that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented by machine-readable instructions.

[0132] The machine-readable instructions can be executed, for example, by a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other field programmable gate array data processing device to realize the functions described in the description and figures. In particular, a processor or processing circuit, or modules thereof, can execute machine-readable instructions. Thus, the functional modules of the apparatus and other devices described herein can be implemented by a processor that executes machine-readable instructions stored in a memory, or a processor that operates according to instructions embedded in a logic circuit. The term "processor" should be interpreted broadly to include a CPU, a processing unit, an ASIC, a logic unit, or a programmable gate array, etc. All the methods and functional modules can be executed by a single processor or divided among several processors.

[0133] Such machine-readable instructions may also be stored in a computer-readable storage device that can direct a computer or other programmable data processing device to operate in a particular mode.

[0134] Such machine-readable instructions may also be loaded onto a computer or other programmable data processing apparatus such that the computer or other programmable data processing apparatus performs a sequence of operations to generate a computer-implemented process, such that the instructions executed on the computer or other programmable device implement the functions identified by the blocks in the flowcharts and / or block diagrams.

[0135] Furthermore, the teachings herein may be implemented in the form of a computer program product, the computer program product being stored on a storage medium and comprising a plurality of instructions for causing a computing device to implement the methods recited in the embodiments of the present disclosure.

[0136] An element or step described in connection with one embodiment may be combined with or substituted by an element or step described in connection with another embodiment. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain means are recited in mutually different dependent claims does not indicate that a combination of these means cannot be used to advantage. A computer program may be stored or distributed on a suitable medium, such as an optical storage medium or a solid-state medium provided together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. Any reference signs in the claims should not be interpreted as limiting the scope.

Claims

1. 1. A computer-implemented method comprising: identifying a set of vascular landmarks within a subject's body using a machine learning model configured to identify vascular adjacent landmarks from radiological imaging data of the subject's body; determining a path having adjacent landmarks identified in the set by identifying coordinates of corresponding vasculature indicative of a vascular segment between adjacent landmarks and estimating a profile of the path based on the coordinates; determining a shape of a blood vessel between the identified adjacent landmarks based on the determined path and imaging conditions indicative of a wall of the blood vessel, the imaging conditions referencing a portion of the radiological imaging data corresponding to a structure indicative of a wall of the blood vessel; The method comprising:

2. The method of claim 1 , wherein the machine learning model comprises a deep Q-network model.

3. 3. The method of claim 1, wherein the machine learning model is trained to recognize the set of landmarks based on a training dataset comprising radiological imaging data obtained from a plurality of training subjects, the radiological imaging data obtained from each training subject having a target vessel annotated with the set of landmarks to be recognized.

4. 4. The method of claim 1, wherein determining the path comprises identifying pixel intensity thresholds in the radiographic data, the pixel intensity thresholds being indicative of identified landmarks that are connected to form the path.

5. The method of claim 4 , wherein the pixel intensity threshold is determined based on a minimum pixel intensity value and / or an average pixel intensity value associated with each of the identified landmarks.

6. A method according to any one of claims 1 to 5, wherein the step of estimating the path profile comprises determining the shortest path between adjacent landmarks by setting a first landmark as a seed point and identifying a second landmark corresponding to a landmark closest to the first landmark within an iteratively growing volumetric region, such that coordinates associated with the first and second landmarks define the path between the first landmark and the second landmark.

7. A method according to any one of claims 1 to 6, wherein the imaging conditions indicative of the walls of the blood vessel are based on gradient conditions indicative of the presence of the walls.

8. A method as described in claim 7, comprising the steps of determining coordinates corresponding to a center point of a blood vessel defining an origin according to a local coordinate system, and determining a shape of the blood vessel in a blood vessel cross section perpendicular to the path at the origin by simulating a number of rays extending radially from the origin until the gradient condition is satisfied for each radial ray at a specified distance from the origin, wherein the specified distance of each radial ray from the origin is used to estimate a boundary in the blood vessel cross section corresponding to at least a portion of the shape of the blood vessel.

9. The method of claim 8, further comprising determining a vascular segment corresponding to the shape of the vessel between the adjacent landmarks by connecting estimated boundaries from adjacent vascular cross sections.

10. A method according to claim 7, 8 or 9, wherein the gradient condition is determined based on a gradient model of pixel intensity values ​​in a vascular cross section.

11. The blood vessel comprises an internal carotid artery and / or a vertebral artery, the set of landmarks comprising an identified set LX of landmarks from at least one of the internal carotid artery and the vertebral artery of the subject identified by the machine learning model, each landmark LX of the identified set being: L0 is shared by the internal carotid artery and the vertebral artery and is located in the brachiocephalic trunk of the aortic arch of the subject; L1 is located at the common carotid artery of the aortic arch; L2 is located in the subclavian artery of the aortic arch; L3 and L4 are located at the bifurcation of the subject's external carotid artery and the internal carotid artery; L5 and L8 are the positions of the internal carotid artery around the C6 cervical vertebra of the subject; L9 and L14 are the locations of the vertebral arteries around the C6 cervical vertebra of the subject; L6 and L7 are the positions of the internal carotid artery around the C2 cervical vertebra of the subject; L10 and L13 are the positions of the vertebral arteries around the C2 cervical vertebra of the subject; L11 and L12 are the positions of the internal carotid artery around the C1 cervical vertebra of the subject; L15 and L18 are the positions of the vertebral arteries around the C1 cervical vertebra of the subject; L16 and L17 are the locations of the internal carotid arteries, through which blood is transferred to structures inside and outside the skull. L19 and L22 are the locations of the vertebral artery on the subject's posterior fossa where the vertebral artery enters the subject's skull; L20, L21, L23 and L24 are the locations of the internal carotid arteries in the skull; L25 is a location shared by both the left and right vertebral arteries of the subject 11. The method of claim 1, further comprising:

12. A non-transitory machine-readable medium having instructions that, when executed by at least one processor, cause the at least one processor to perform a method according to any one of claims 1 to 11.

13. An apparatus for extracting vascular information from radiological imaging data of a subject, the apparatus comprising: at least one processor communicatively coupled to an interface for receiving the radiological imaging data; identifying a set of vascular landmarks within the subject's body using a machine learning model configured to identify vascular adjacent landmarks from radiological imaging data of the subject's body; determining a path having adjacent landmarks identified in the set by identifying coordinates of corresponding vasculature indicative of a vascular segment between adjacent landmarks and estimating a profile of the path based on the coordinates; determining a shape of a blood vessel between the identified adjacent landmarks based on the determined path and imaging conditions indicative of a wall of the blood vessel, the imaging conditions referencing a portion of the radiological imaging data corresponding to a structure indicative of a wall of the blood vessel; a non-transitory machine-readable medium storing instructions readable and executable by the at least one processor to perform a method having the steps of: The system has:

Images