Computer-implemented method, computer program product and imaging system for calculating center line from 3D model of tubular structure

Abstract

A computer-implemented method according to the invention comprises receiving (100) a patient-specific volumetric image comprising structural information of a patient's blood vessel represented in voxels, comprising extracting and / or presenting patient-specific data using centerlines, the method comprising a correction (150) step comprising an iterative process, the iterative process includes at least the operations of centering (552) the points defining the centerline by positioning the points defining the centerline at a position further from a nearest surface defined in the geometry, and equalizing (553) the points defining the centerline by bringing each point closer to an average position of adjacent points. The invention also relates to a computer program product and an imaging system.

Description

Technical Field

[0001] The present invention relates to a computer-implemented method, a computer program product, and an imaging system for calculating a centerline from a 3D model of a tubular structure. The present invention is particularly suitable for simulating blood flow using computational fluid dynamics, in particular in the extraction of vascular geometry from a medical 3D volume of an artery filled with a contrast agent, and for enhancing the visualization of blood vessels in the medical 3D volume. Background Art

[0002] Cardiovascular disease is one of the leading causes of death worldwide. With the development of computed tomography (CT), accurate diagnosis of patients with suspected ischemic heart disease using non-invasive methods has become possible.

[0003] CT scans are typically presented as 3D volumetric images. These 3D volumetric images represent physical quantities as a function of three spatial coordinates. In a digital volumetric image, each sample (voxel) represents that quantity measured at a specific location. The image consists of a spatial sequence of 2D slices of the object of interest.

[0004] Typically, a slice is represented as an image matrix of pixels (X and Y coordinates). The slice number represents the Z coordinate.

[0005] Interpreting CT images and diagnosing cardiovascular disease are complex tasks requiring highly trained physicians. A common problem is that interpreting the projections visible on CT images requires highly skilled radiologists, while diagnosing cardiovascular disease requires experienced cardiologists. Skilled individuals in both areas are in short supply. Therefore, computer-assisted technology is needed.

[0006] One method for diagnosing cardiac pathology with the aid of a computer is to generate a diagnostic-aided 3D model of the blood vessels. In particular, in the field of cardiac pathology, detailed 3D models of the coronary arteries are required. Detailed models allow for accurate assessment of patient-specific anatomical and functional characteristics, such as hemodynamic parameters. In the prior art, there are many techniques that utilize computational fluid dynamics (CFD) for blood flow simulation. The geometry of the blood vessels presented in a 3D volume image can be represented as a subset of voxels, a surface mesh, a volume mesh, or a point cloud of the 3D image. An example of CFD technology is disclosed in European Patent EP3820357, which is used to calculate the fractional flow reserve (FFR) based on a 3D model of the coronary arteries from coronary computed tomography angiography (CCTA).

[0007] A centerline is defined as a polygonal chain of points and is calculated at subvoxel resolution. A centerline represents the geometric location of the center of a vessel's cross section along the length of the vessel. Centerlines are used to model and generate image representations of vessels. Methods for determining centerlines are widely developed and known in the art, for example, in US20170018116A1, US9129417B2, CN108133478B, and US11069078B2.

[0008] When representing 3D curved objects in CT images of limited resolution, numerous artifacts are prone to occur, distorting the shape of the centerline. One typical artifact that adversely affects known algorithms for determining the centerline is the so-called blooming artifact, which is caused by the high density of calcified atherosclerotic plaques and the density of the contrast agent applied to the patient for imaging. The blooming artifact is most severe when the contrast agent is concentrated in a certain area. Another typical artifact is the so-called motion artifact, which frequently occurs in imaging, including in cardiac imaging. Motion artifacts are caused by patient movement during imaging. Misalignment artifacts appear as blurring, streaking, or shadowing. Image blurring can also occur due to patient movement during radiological examinations, and the blurring leads to wave-like deformations in blood vessels.

[0009] The key to accurate simulation and diagnosis of cardiovascular disease is an accurate 3D model of the coronary arteries. In order to be able to create accurate 3D models and perform precise simulations, a method for extracting centerlines from geometric shapes is required. The centerline of a tubular object is defined as a chain of polygons, that is, a curve, specified by a sequence of points (or vertices) connected by edges. Since we are only considering 3D models, these points always lie in 3D space. Both the geometric shape and the centerline are represented with sub-voxel accuracy. The first and last points on the centerline need to be located at the start and end points of the geometric shape, respectively. All points and corresponding edges need to be located inside and at the center of the cross-section of the geometric shape. Since the coronary artery consists of several branches, so does its 3D model. Therefore, for each coronary artery, several centerlines must be extracted, one for each branch.

[0010] Centerline extraction is a crucial part of the 3D model creation process. The centerlines of 3D vascular models are used to generate appropriate vascular views, known as curved surface reconstructions (CPRs), which physicians use for segmentation tasks, topology analysis, and diagnostic procedures. Therefore, constructing appropriate coronary artery centerlines is crucial for diagnosing vascular disease, estimating the size of coronary artery stenosis, and simulating blood flow.

[0011] One of the most common methods for calculating the centerline of a given geometric shape is to use the Voronoi diagram of a 3D geometric model, as disclosed in "Patient-specific modeling of geometry and blood flow in large arteries" by Antiga and Luca (Politecnico di Milano (2002)), and provided as a working computer program library document that can be accessed through http: / / www.vmtk.org / tutorials / Centerlines.html Available online. The main advantage of this method is that it produces mathematically well-defined, high-quality centerlines at the expense of providing a geometric model of the vessel. However, this method often fails when the vessel is thin and curved.

[0012] Another approach is to extract the centerline directly from the provided image. This allows the geometry extraction to be skipped in the process at the expense of centerline quality. An example approach is to feed a local block of the image to a neural network and try to determine the next point on the centerline, the bifurcation point, and the location of the vessel end in an iterative process, as disclosed in "Deep Reinforcement Learning for Vessel Centerline Tracing in Multi-modality 3D Volumes" by Zhang, Pengyue, Fusheng Wang, and Yefeng Zheng (MICCAI (2018)). Another example approach disclosed in US9129417B2 is to extract the entire centerline by applying a filter to the image to determine the location of the vessel, and find the centerline by applying a refinement process (skeletonization). In addition, other applicable methods have been published, such as Xiong G, Sun P, Zhou H, Ha S, Hartaigh BO, Truong QA, and Min JK, "Comprehensive Modeling and Visualization of Cardiac Anatomy and Physiology from CT Imaging and Computer Simulations," IEEE Trans Vis Comput Graph, February 2017; 23(2):1014-1028. doi:10.1109 / TVCG.2016.2520946. Epub 2016 Feb 5. PMID: 26863663; PMCID: PMC4975682. The main disadvantage of these methods is that they are highly sensitive to anatomical differences between patients and image acquisitions.YIN WANG et al., "3-D Quantitative Vascular Shape Analysis for Arterial Bifurcations via Dynamic Tube Fitting" (IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, IEEE, USA, Vol. 59, No. 7, July 2012, pp. 1850-1860, XP011447498, ISSN: 0018-9294, DOI: 10.1 109 / TBME.2011.2179654) discloses a computer-implemented method for determining a centerline based on a cubic model, comprising the following steps: receiving a patient-specific volumetric image comprising structural information of the patient's blood vessels represented in voxels; retrieving a three-dimensional geometry of at least one blood vessel; and extracting a skeleton graph from the geometry to determine the centerline. The method is performed directly on a mesh domain, allegedly eliminating the need for image upsampling. "Skeleton extraction by mesh contraction" by OSCAR KIN-CHUNG AU et al. (ACMTRANSACTIONS ON GRAPHICS, ACM, NY, USA, Vol. 27, No. 3, August 2008, pp. 1-10, XP058355353, ISSN: 0730-0301, DOI: 10.1145 / 1360612.1360643) discloses a geometry-based framework for extracting a curved skeleton from a mesh surface. The extraction is applied directly to the mesh domain, without the need for voxelization. This method works well as long as the vessel fits well within the elliptical model. If this is not the case, model inadequacies can lead to significant errors in the centerline trajectory.

[0013] Other related methods include European Patent EP2074585B1, which discloses a system for registering a blood vessel model with an image dataset based on a connection model. The system includes a reference object model and a blood vessel model having a centerline described by a plurality of control points.

[0014] The specification of European patent EP3245632B1 discloses an imaging system comprising a sub-resolution luminal stenosis detector that detects sub-resolution stenosis of a vessel lumen in an image volume by centerline profile analysis and calculates a sub-resolution determined diameter by modifying an approximate visible luminal diameter with the detected sub-resolution stenosis.

[0015] The publication of European patent application EP3723041 discloses an image processing system configured to generate a model of a tubular object from a volume image, the system comprising a volume image acquisition unit, a geometry extraction unit and a centerline generator.

[0016] US20130064435 discloses a system and method for determining a centerline of a tubular structure based on volumetric data of the blood vessel, wherein a contrast agent is injected into the bloodstream to enhance imaging of the centerline.

[0017] An active contour model is disclosed that is adapted to determine a contour of a blood vessel defined in image data by normalizing the image data around a starting point.

[0018] Problem to be solved:

[0019] Known methods of generating and correcting centerlines tend to fail near bends, narrowings, or other dimensional variations or other anomalies that cause the vessel to differ significantly from an assumed model (eg, an elliptical cross section).

[0020] It is an object of the present invention to enhance known methods and improve the centerline generation mechanism in order to reduce the need for manual corrections. Summary of the Invention

[0021] According to the present invention, a computer-implemented method comprises the following steps: receiving a patient-specific volumetric image comprising structural information of a patient's blood vessels represented in voxels; retrieving a three-dimensional geometry of at least one blood vessel; extracting a skeleton from the geometry; determining endpoints of at least one blood vessel to identify an inlet and at least one outlet; calculating at least one initial centerline for the at least one blood vessel, the at least one initial centerline being defined by points and calculated with subvoxel resolution; correcting the at least one initial centerline with subvoxel accuracy to obtain and output a corrected centerline; and extracting and / or presenting patient-specific data using the corrected centerline. According to the present invention, the step of correcting the at least one initial centerline comprises an iterative process comprising at least the operations of centering and equalizing the points defining the centerline. Centering the points defining the centerline involves positioning them further away from the nearest surface defined in the geometry. Equalizing the points defining the centerline involves moving each point closer to the average position of its neighboring points. Centering and equalizing are performed sequentially until a predefined stopping condition is satisfied. Centering is performed such that the centerline points remain equidistant from the vessel wall, i.e., located at the center of the vessel. Equalization makes the centerline correspond better to the direction of blood flow and prevents it from taking on a shape that would lead to erroneous simulations. Performing the centering and equalization operations sequentially in an iterative manner can yield a centerline that is very useful for performing simulations in a fully or at least mostly automatic manner.

[0022] The steps of calculating at least one initial centerline include verifying the continuity of the skeleton image and, if necessary, correcting the skeleton image by connecting disjointed segments. The initial centerline in the skeleton image is determined as the shortest path between the inlet and outlet. This series of steps can eliminate certain types of errors caused by inadequate imaging. This method does not rely on any vessel model and is therefore unaffected by unusual or pathological vessel shapes. In particular, the cross-section of the vessel needs to be modeled as an ellipse.

[0023] Advantageously, the iterative process of the correction step includes subjecting the points of the centerline to image forces. Centering includes subjecting the points of the centerline to image retraction forces, thereby pushing these points away from the nearest surface of the geometric shape. Equalization includes subjecting the points on the centerline to image tension forces toward adjacent points, thereby pushing the points toward the average position of the two adjacent points.

[0024] Advantageously, the stopping condition is the reaching of a predetermined number of iterations. When testing such a stopping condition, the required computing power is low. An alternative stopping condition involves an analysis of the convergence of the points to a stable position.

[0025] Advantageously, the step of retrieving the three-dimensional geometry includes generating a surface mesh representation of the blood vessel from the volume image. This makes the calculations required for centering easier and makes it easier to present to the physician when manual enhancement is required. The alternative volume representation has the advantage that it can be used to simulate blood flow.

[0026] Advantageously, extracting a skeleton graph from the geometric shape includes the following steps: voxelizing the geometric shape to obtain a volumetric representation of the vessel of interest; skeletonizing the volume to obtain a volumetric skeleton representation of the vessel; and creating a skeleton graph based on the volumetric skeleton representation. Testing has shown that this method is faster and more stable than the Voronoi method commonly used in the art. It is particularly suitable for vascular curvature and narrow areas.

[0027] Advantageously, the determination of endpoints includes an automatic breadth-first search of geometric nodes and uses thickness gradients as criteria to determine the inlet and outlet of the blood vessels. The method can reliably, automatically or semi-automatically assign endpoints and orient the skeleton graph.

[0028] Advantageously, prompting the user to input the approximate location of the portal and the determination of the endpoint includes selecting the portal based on the approximate location given by the user with the assistance of filtering the endpoint candidates using a filter, thereby speeding up the operation with less effort on the part of the user.

[0029] Advantageously, the corrected centerline points are post-processed, the post-processing comprising at least one operation selected from the group consisting of: resampling the centerline, smoothing the positions of the centerline points in space, calculating centerline point metadata, and pruning. Pruning reduces the memory consumption required to perform the method, while the metadata is useful for visualizing the 3D geometry.

[0030] The computer program product according to the invention comprises a set of instructions which are loaded into a memory of a computer such that said computer performs the method according to the invention.

[0031] An image processing system is configured to generate a model of a tubular object from a volume image. The system comprises a volume image acquisition unit, a geometry extraction unit, a centerline generator, and a diagnosis assistance module according to the present invention. The image processing system is suitable for implementing the method according to the present invention.

[0032] Advantageously, the volume image acquisition unit is a computer module adapted to receive images via a telecommunication network.

[0033] Advantageously, the volume image acquisition unit is an imaging device. Advantageously, the imaging device is selected from the group consisting of a computed tomography scanner, a photon counting computed tomography scanner, a magnetic resonance scanner, a positron emission tomography scanner, a single photon emission computed tomography scanner, and an ultrasound imaging device. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention has been described in detail below with reference to the following drawings.

[0035] Figure 1 A flowchart showing the centerline generation method.

[0036] Figure 2 A flow chart of step 120 of extracting a skeleton graph from a geometric shape is shown.

[0037] Figure 3 A flow chart is shown for step 130 of determining the inlet endpoint and the outlet endpoint.

[0038] Figure 4 A flow chart is shown for step 140 of calculating an initial centerline.

[0039] Figure 5 A flow chart of the initial centerline correction step 150 is shown.

[0040] Figure 6 A flow chart of the centerline post-processing step 160 is shown.

[0041] Figures 7a to 7d The surface mesh of the coronary arteries is shown, where Figures 7b to 7dVoxelized by different image spacing values: Figure 7b d is 1.0 mm, Figure 7c d is 0.5 mm and Figure 7d The d is 0.15mm.

[0042] Figure 8 Shows the Figure 7b The skeleton obtained from the voxelized mesh shown.

[0043] Figure 9a shows the effect of correcting the initial centerline, Figure 9b The center line corrected according to the present invention is shown.

[0044] Figure 10a Shows the corrected centerline before post-processing. Figure 10b The same centerline is shown after being smoothed in a post-processing step.

[0045] Figures 11a to 11d An example of a coronary artery surface mesh with corresponding centerlines (white) determined according to the present invention is shown.

[0046] Figure 12 The mean and standard deviation of the voxelization time Tv [s] calculated based on a set of 160 example data are shown as a function of the image spacing value d [mm] in voxel size.

[0047] Figure 13 The mean and standard deviation of the skeletonization time Ts [s] calculated based on a set of 160 example data are shown as a function of the image spacing value d [mm] in voxel size.

[0048] Figure 14 The mean and standard deviation of the correction time Tc[s]] calculated based on a set of 160 example data are shown as a function of the number of correction steps N[1 / 1].

[0049] Figures 15a to 15d shows the correction effect after multiple iteration steps, where Figure 15a 5 times, Figure 15b 20 times, Figure 15c For 50 times and Figure 15d is 100 times.

[0050] Figure 16a shows the effect of the correction in the absence of tension, and Figure 16b The effect of the correction is shown without retraction force. DETAILED DESCRIPTION

[0051] Figure 1A high-level flow chart of a method according to the present invention is shown. Solid arrows represent the flow of the method, and double arrows represent information input from other steps. According to the present invention, the method begins at step 100: receiving a patient-specific volume image, which includes structural information about the patient's blood vessels. The volume image can be the result of any known imaging technology for vascular imaging. This includes computed tomography (CT) scans, photon counting computed tomography (PCCT) scans, computed tomography angiography (CTA) scans, coronary computed tomography angiography (CCTA) scans, magnetic resonance imaging (MRI) scans, positron emission tomography (PET) scans, ultrasound scans, or any other medical imaging modality that includes at least one blood vessel. The method has been tested and works very well with CCTA image sequences.

[0052] After receiving the volumetric image, in step 110, the three-dimensional (3D) geometry of at least one blood vessel is retrieved from the image. The 3D geometry is a representation of the at least one blood vessel. The 3D geometry is typically represented in the form of a graph, such as a surface mesh, but can also be represented in the form of a volume mesh or point cloud. A 3D representation is required to adequately model and accurately simulate blood flow.

[0053] The 3D geometry can be retrieved manually by, for example, a radiologist or trained engineer using computer segmentation software tools. In this embodiment, software tools that enhance the creation of 3D geometry include, but are not limited to, active contours, active meshes, and machine learning. Active contours or active meshes allow the creation and correction of 3D geometry by moving its edges toward visible contrast in the image. Machine learning methods are applied to generate an initial 3D geometry that can be further refined or accepted by an expert. Use of the software tools results in an accelerated geometry retrieval step 110. Alternatively, fully automated tools can be applied using machine learning or artificial intelligence models trained on geometry created by radiologists to predict the presence of blood vessels in the 3D volume and convert them into a mesh.

[0054] The 3D geometry outputted from step 110 is represented in the form of a surface mesh. The resolution of the surface mesh is a hyperparameter of the method. Generally, setting a lower resolution will increase the speed of the operation at the expense of the quality of the centerline. In step 120, the 3D geometry is used to extract an undirected skeleton graph. In this embodiment, the undirected skeleton graph is created in three steps, such as Figure 2 As shown:

[0055] voxelization 221, which involves converting the geometry of the blood vessels into voxels;

[0056] Skeletonization 222 , which involves thinning to create a skeleton from voxels;

[0057] Create 223 an undirected skeleton graph from the skeleton.

[0058] Voxelization 221 involves a repartitioning of the 3D space. It can result in a uniform partitioning of the 3D space into a grid of voxels or an octree of voxels. An octree is a tree-like data structure in which each internal node has exactly eight children. Octrees are most commonly used to partition a 3D space by recursively subdividing the 3D space into eight octaves. An example of an efficient algorithm for voxelization for this purpose is the TriangleMashToBinaryImageFilter method implemented in the ITK library. This method converts a triangular mesh into a predefined volume representation by checking whether each voxel lies within the mesh of the geometric shape. However, a person skilled in the art can easily find alternative algorithms for voxelization that provide almost as good results.

[0059] In this embodiment, the skeletonization step 222 is performed using a medial axis transform (MAT). The MAT determines the closest boundary points for each point in the vessel. An interior point belongs to the skeleton if it has at least two closest boundary points. The formal definition of a skeleton is the locus of the centers of all maximum inscribed spheres. An example thinning algorithm that can be used in this step is the algorithm of Zhang and Lee.

[0060] In step 223, an undirected skeleton graph is created from the skeleton volume by converting voxels into graph nodes and adding edges between connected voxels. For each node, its original position in the image is stored. Then, potentially disconnected graph parts are connected together based on proximity. In step 130, the endpoints of a vessel are determined so that the inlet and outlet can be located. This step is repeated for each vessel in the 3D geometry. In this embodiment, the following reference is applied Figure 3 Described method.

[0061] In step 331, a breadth-first search (BFS) algorithm is used to traverse the vertices of the geometric shape to create a BFS graph. The algorithm starts from an arbitrary vertex. Because blood vessels are often quasi-tubular structures, connecting vertices at the same distance from the starting vertex should form a cross-sectional ring. In successive iterations of the BFS algorithm, the ring is advanced along the vessel.

[0062] For each ring, the centroid of the vertices contained in it is calculated. An undirected graph is constructed with the centroids as nodes, and connected by edges based on the vertex connectivity of each ring.

[0063] The leaf nodes of this graph are expected to be located where the BFS algorithm cannot proceed further along the vessel, ie, potential ends of the vessel geometry. These leaf nodes are also used to determine the centerline endpoints in step 332.

[0064] The algorithm works particularly well for quasi-tubular structures and tends to misorient rings in structures of other shapes.

[0065] In step 332, the centerline endpoints are found using the BFS graph, the skeleton graph, and the leaf nodes of the geometry. The leaf node in the skeleton graph closest to the BFS graph is selected as the endpoint and projected onto the geometry surface.

[0066] In step 333, the inlet is selected under the assumption that the vessel is thickest at the inlet and becomes increasingly thinner with each further branch.

[0067] Calculate the diameter of the artery mesh at each node of the skeleton graph. Then, starting at each endpoint, traverse the skeleton graph completely and calculate the difference in diameter between adjacent nodes. If the adjacent node corresponds to a smaller diameter, the result is given a positive sign; otherwise, it is given a negative sign. The differences are then summed together. The maximum sum corresponds to the starting endpoint where the drop (gradient) in diameter is the largest. This is the best candidate for the vessel entrance.

[0068] Optionally, the portal is selected with the help of a user who indicates the region in which the portal is located. The portal is then selected from the endpoints located near the designated region using the above method.

[0069] The remaining endpoints are considered egress.

[0070] At least one initial centerline of at least one blood vessel is generated in step 140. The centerline is defined by points and is calculated with subvoxel resolution relative to the volume image.

[0071] Although there are vascular diagnostic methods based on 2D images and two-dimensional centerlines (e.g. EP3061015), accurate blood flow simulation requires a 3D representation.

[0072] The 3D representation requires the acquisition of 3D centerlines, which are necessary for obtaining digitally straightened curved planar reconstruction (CPR) and multiplanar reconstruction (MPR) views, as well as for the simulation itself, in particular for calculating cross-sections for modeling blood flow.

[0073] Although the 2D approach simplifies the determination of the centerline and reduces the required computing power, it does not provide sufficient modeling accuracy and is not consistent with the present invention.

[0074] In the step 140 of calculating the initial centerline, the skeleton graph 120 obtained in the points is used to initialize the centerline. First, in step 441, disjoint skeleton graph parts are found and connected to obtain a continuous graph.

[0075] Then, step 442 is executed to find the shortest path between the inlet and the outlet in the undirected skeleton graph. The shortest path is calculated by breadth-first traversal of the graph nodes.

[0076] The points of the shortest path are resampled to ensure the same distance between points. These points are used together with the information of the start (entry point) and end (exit point) of the vessel geometry to create a directed graph consisting of multiple directed chains of points in volume coordinates. This only gives information about the position of each node of the centerline in space. Each centerline C m,n It consists of m paths and n points. The paths are described in the form of directed chains, where each path consists of a path from the entry point p inlet To the exit point p m The points are shared between these paths, and except for the entrance or exit points in the path, each point has two neighbors. Along the path, there is one in front of the point and one behind the point. At the end of the point, the coordinates are converted to real coordinates to obtain the initial centerline. In step 445, the points are resampled to the desired spacing. In the step of preprocessing the centerline, some centerline paths that are shorter than a predefined threshold are removed.

[0077] The step 150 of correcting at least one initial centerline with subvoxel accuracy begins at step 551 by receiving the initial centerline and the 3D geometry. Correcting the at least one initial centerline then comprises an iterative process comprising at least the following operations: centering 552 the points defining the centerline by positioning them further from the nearest surface defined in the geometry; and equalizing 553 the points defining the centerline by moving each point closer to the average position of its neighboring points. These operations are performed sequentially until a stopping condition 554 is met. The stopping condition can be a certain number of iterations or a predefined measure of difference between subsequent steps. The result is output at step 555.

[0078] The calculation is convenient if the 3D geometry is a surface mesh, however, the vessel surface in a voxel-represented geometry can be identified as the outer layer of the voxels. If the geometry is represented as a point cloud covering the vessel surface, the distance to the closest surface is the distance to the closest point.

[0079] Centering 552 includes subjecting a point of the centerline to a retractive force that pushes the point away from the nearest surface of the geometric shape.

[0080] Equalization 553 involves subjecting the points of the centerline to a tensile force toward the adjacent points, thereby pushing each point toward the average position of the two adjacent points.

[0081] This process can be implemented using the following pseudo code:

[0082]

[0083]

[0084] Good results were obtained using the following parameter values:

[0085] n: int=100; n∈[0, 1000];

[0086] dt: float=0.5; dt∈[0,1];

[0087] dr: float = 0.02; dr∈[0,1]—represents the fraction of the vessel diameter at a given point on the centerline.

[0088] The corrected initial centerline is subjected to optional post-processing 160. This includes resampling 641 and 3D smoothing 642 of each path in the centerline.

[0089] In a post-processing step 160, optional features of the vessel are calculated for the points of the centerline and added 643 to the optional metadata. These features may include, but are not limited to: a perpendicular vector to the centerline, a tangent to the centerline, a cumulative distance from the starting point, a diameter of a circle inscribed in the geometric surface shape, a diameter of a disk inscribed in the geometric surface shape, a semi-minor axis and a semi-major axis of an ellipse inscribed in the geometric surface shape, a minimum distance to the nearest vertex of the vessel geometry, and a radius of a disk having the same area as the shape determined by the vessel geometry on a plane perpendicular to the centerline at a given point.

[0090] Since all centerline paths start from the same entrance, they overlap in some parts and share multiple points. In order to merge them into a single centerline structure represented as a directed graph, the branches need to be pruned 644, the overlapping parts are removed, and the now non-overlapping branches are connected with edges.

[0091] The centerline is smooth, continuous, and located at the center of the geometry. These properties are necessary for accurate CFD simulations and good MPR and CPR views.

[0092] Finally, the centerline is output in step 170 and used in conjunction with conventional blood flow simulations (eg as disclosed in EP3820357) or simply for visualization to allow a physician to extract and / or present patient-specific data.

[0093] According to the invention, a computer program product can be run on a conventional computer that receives a volume image and is adapted to simulate blood flow. The computer program product comprises a set of instructions that, when loaded into a memory of the computer, result in the execution of the method according to the invention.

[0094] According to the present invention, the program can be run on a stand-alone computer or on a processing unit of a medical imaging device with an integrated image processing device. In addition, the computer program product can be provided in a "software as a service" model, wherein the volume image is uploaded by the user to a server, and the centerline and simulation are calculated on the server side, and the results are transmitted to the user via a communication network.

[0095] According to the present invention, an image processing system is configured to generate a model of a tubular object from a volumetric image. The system comprises a volumetric image acquisition unit, a geometry extraction unit, a centerline generator, and a diagnostic assistance module adapted to implement the method according to the present invention. The diagnostic assistance module preferably includes a blood flow simulator and a display module for presenting a view of the blood vessels to a physician.

[0096] The volumetric image acquisition unit can be a computer module adapted to receive images via a telecommunications network. Alternatively, when the system is implemented in an imaging device, the device itself can be a volumetric image acquisition device. For example, European Patent Publication No. EP3820357B1 discusses methods for geometry extraction, centerline generation, and blood flow simulation suitable for implementation in a computer to obtain a geometry extraction unit, a centerline generator, and a blood flow simulator. According to the present invention, implementation of the aforementioned centerline generation and correction module allows for the realization of this system.

[0097] The imaging device may be any medical imaging device having a display and a processing unit. This includes, among others, a computed tomography scanner, a PCCT (photon counting computed tomography) scanner, an MRI (magnetic resonance) scanner, a PET (positron emission tomography) scanner, a SPECT (single photon emission computed tomography) scanner, an ultrasound scanner, or any other relevant medical imaging device.

[0098] The method proves useful over a wide range of parameters. Examples of parameters where a trade-off between computational power and accuracy is necessary are discussed below.

[0099] Figure 7a The geometric surface mesh of the coronary artery image obtained in the step of retrieving the geometry 110 is shown. Extracting 120 the skeleton from the geometry requires voxelization 221. The image size in the voxelization affects the accuracy and computation time of each step. Figures 7b to 7d Shows Figure 7a The voxelization 221 results of the grid shown are for different values of image spacing (voxel size) d. Figure 7b The voxelization result when d=0.15mm is shown. Figure 7c The voxelization result when d=0.5mm is shown. Figure 7d The voxelization results when d=1 mm are shown.

[0100] Figure 8 An example of the skeleton obtained in step 222 is shown, which is used to create 223 a skeleton graph, which is then used to generate initial centerlines.

[0101] Figure 9a An example of a portion of an initial centerline generated in step 140 of the method according to the present invention is shown. Figure 9b The center line corrected using the method according to the invention is shown.

[0102] Figure 10a and Figure 10b The post-processing results of the vessel ends are shown. Figure 10a The centerline of the vessel end before smoothing 642 is shown. Figure 10b The centerline is shown after smoothing 642. Smoothing may include applying a relatively weak low-pass filter to the coordinates of the points of the centerline. Figure 10b In the example shown, a Butterworth filter is applied. Smoothing can remove numerical artifacts, especially those related to specific geometries that were not previously corrected.

[0103] Figures 11a to 11d An example of a coronary artery surface mesh with corresponding centerlines (white) determined according to the present invention is shown.

[0104] Figure 12 This figure shows how the image spacing (voxel size) d [mm] affects the voxelization time Tv [s]. The data are from 160 images from different patients. Because the number of voxels increases cubically, a balance must be found between voxelization accuracy and duration. For the tested data, the value of d that achieved the best quality and speed was d = 0.15 mm. Using this value, voxelization took approximately 5 seconds on the test device.

[0105] Figure 13 The image shows the skeletonization time Ts [s] versus the image spacing d [mm]. The results were obtained using the scikit-learn image skeletonization algorithm, based on the method developed by Lee et al. Skeletonization iteratively removes border voxels until only a single pixel remains. Higher resolutions require more algorithm iterations to complete their work. For d = 0.15 mm, the skeletonization time was approximately 6 seconds.

[0106] Figure 14The linear relationship between the number of correction iterations, N[1 / 1], and the correction time, Tc[s], is shown. This relationship is approximately linear. Increasing the number of iterations introduces small differences in the centerline over time. On more powerful computers, more corrections can be forced by default. On slower machines, particularly when integrating with older imaging devices, a lower number of iterations can be defaulted, with the user being given the option to manually override and add additional corrections if desired.

[0107] Figures 15a to 15d The correction effect is shown after 5, 20, 50 and 100 correction iterations. It has been observed that 100 iterations provides the best balance between accuracy and processing time.

[0108] By comparison Figure 15d and Figure 16a In the images shown, the effect of applying equalization 553 (eg by tension) can be observed. Figure 16a The centerline is shown after the correction procedure without equalization 553. This form of centerline is not useful because at least some properties calculated for each point (such as the tangent) will be severely disturbed.

[0109] By comparison Figure 15d and Figure 16b In the images shown, the effect of applying centering 552 (eg, by a retraction force) can be observed. Figure 16b The centerline is shown after the correction procedure without centering 552. The centerlines are smoother, but they tend to be too close to the vessel wall. Furthermore, it is unclear how many iterations should be used, as typically the forces will balance each other, leading to convergence. Without retraction forces, the centerline will not converge, and increasing the number of iterations will result in a continuously straightening line.

[0110] The exemplary embodiments of the present invention described above are intended to illustrate and not to limit the present invention. The present invention includes methods and systems for calculating blood vessel centerlines, particularly coronary artery centerlines, based on a surface mesh. According to embodiments of the present invention, the methods can be easily implemented and can produce robust and accurate segmentations with very limited time and storage costs (typically less than 20 seconds and 3 GB on average). Combining the above-described methods for determining centerlines with state-of-the-art coronary artery segmentation methods and state-of-the-art methods for calculating hemodynamic factors produces a simulation of blood flow, thereby allowing diagnosis with little to no intervention from a skilled operator.

[0111] It is worth noting that, based on the representation of objects in volumetric imaging, the well-defined centerline can also be applied to the modeling and simulation of basic tubular objects. Therefore, the present invention is also applicable to the modeling of blood vessels other than coronary arteries, the digestive system, or other tubular objects, especially the modeling of objects that guide certain fluids.

[0112] A computer in the sense described above is understood to be a hardware device, including computers, microcontrollers, signal processing, programmable gate arrays, other processing digital devices considered as hardware computers, graphics cards, application specific integrated circuits or for image processing, as well as distributed solutions including cloud computing environments.

[0113] Based on the teachings of the above description, a person skilled in the art can routinely come up with various hardware and software solutions for the device according to the present invention, for the computer-implemented execution method according to the present invention, and for the method for obtaining training information.

[0114] It should be emphasized that the above description is only illustrative of the present invention and that those skilled in the art will be able to propose many alternative embodiments encompassed by the scope of protection defined in the appended claims.

[0115] In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The use of the verb "comprise" does not mean that there are no elements or steps other than those stated in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements.

Claims

1. A computer-implemented method comprising the following steps: receiving (100) a patient-specific volumetric image, the patient-specific volumetric image including structural information of a patient's blood vessels represented in voxels; retrieving (110) a three-dimensional geometry of at least one blood vessel; extracting (120) a skeleton graph from the geometric shape; determining (130) endpoints of the at least one blood vessel to determine an inlet and at least one outlet; calculating (140) at least one initial centerline of at least one blood vessel, the at least one initial centerline being defined by points and calculated at a subvoxel resolution; Correcting (150) the at least one initial centerline with subvoxel accuracy to obtain a corrected centerline, including an iterative process involving equalizing (553) the points by bringing each point closer to an average position of neighboring points defining the centerline; output (170) and use the corrected centerline to extract and / or present patient-specific data; It is characterized in that The step of calculating (140) at least one initial centerline comprises: verifying the continuity of the skeleton graph and, if necessary, correcting the skeleton graph by connecting disjoint parts, determining an initial centerline from the skeleton graph as the shortest path between the inlet and the outlet, and The step of correcting (150) the at least one initial centerline includes an iterative process, the iterative process further comprising at least the following operations: Centering the points defining the centerline by positioning them further from the nearest surface defined in the geometric shape (552), and The centering (552) and the equalization (553) are performed sequentially until a predetermined stop condition (554) is met.

2. The computer-implemented method of claim 1 , wherein: The iterative process of the correction (150) step comprises subjecting the points of the center line to image forces, wherein The centering (552) includes subjecting a point of the centerline to an image retraction force that pushes the point away from the nearest point of the 3D geometric shape, and The equalization (553) involves subjecting a point of the centerline to image tension toward an adjacent point, thereby pushing the point toward the average position of the two adjacent points.

3. The computer-implemented method of claim 1 , wherein: The stopping condition (554) is to run the simulation a predetermined number of times.

4. The computer-implemented method according to claim 1 or 2, wherein: The step of retrieving (110) the three-dimensional geometry includes generating a surface mesh representation of the blood vessel from the volumetric image.

5. The computer-implemented method of claim 1, 2, 3, or 4, wherein: The step of extracting (120) the skeleton graph from the geometric shape comprises: subjecting the geometric shape to receptor pixelation (221) to obtain a volumetric representation of the vessel of interest; skeletonizing (222) the volume to obtain a volumetric skeleton representation of the vessel; and The skeleton graph is created (223) from the volumetric skeleton representation.

6. The computer-implemented method according to any one of claims 1 to 5, wherein: The endpoint determination includes an automatic breadth-first search of geometric nodes and uses thickness gradient as a criterion to determine the entrance and exit of the blood vessel.

7. The computer-implemented method according to any one of claims 1 to 6, wherein: Prompting a user to input an approximate location of the portal and determining the endpoint includes selecting the portal with the assistance of filtering endpoint candidates using a filter based on the approximate location given by the user.

8. The computer-implemented method according to any one of claims 1 to 7, wherein: The points of the corrected centerline are post-processed, the post-processing comprising at least one operation selected from the group consisting of: resampling (641) the centerline, smoothing (642) the positions of the centerline points in space, calculating centerline point metadata (643), and pruning (644).

9. A computer program product comprising a set of instructions, which are loaded into a memory of a computer so as to cause the computer to perform the method according to any one of claims 1 to 8.

10. An image processing system configured to generate a model of a tubular object from a volume image, the system comprising a volume image acquisition unit, a geometry extraction unit, a centerline generator, and a diagnosis assistance module, wherein: The system is suitable for implementing the method according to any one of claims 1 to 8.

11. The image processing system according to claim 10, wherein: The volume image acquisition unit is a computer module adapted to receive the image via a telecommunication network.

12. The image processing system according to claim 10, wherein: The volume image acquisition unit is an imaging device.

13. The image processing system according to claim 12, wherein: The imaging device is selected from the group consisting of a computed tomography scanner, a photon counting computed tomography scanner, a magnetic resonance scanner, a positron emission tomography scanner, a single photon emission computed tomography scanner, and an ultrasound imaging device.

Images