Image Analysis-Based Simulation Method for Intravascular Contrast Agent Diffusion

Abstract

This invention relates to the field of medical image processing technology, specifically to a method for simulating intravascular contrast agent diffusion based on image analysis. This method receives time-series image data from angiography equipment, establishes a spatiotemporal correlation model constrained by the vessel wall topology, and forms a spatiotemporal concentration field. Within this model, the contrast agent flow front is identified, and a physical process model describing diffuse filling is constructed based on its dynamic evolution. Key spatiotemporal nodes in the model are extracted and input into a recurrent neural network based on a temporal memory mechanism. This network generates diffusion path prediction trajectories based on the geometry of the vessel bifurcation structure. The predicted trajectories and the spatiotemporal concentration field are integrated to generate a contrast agent perfusion probability map containing multi-level vessel branches, and spatial probability density calibration is performed based on the local curvature of the vessel and the blood flow velocity distribution. The calibrated map is then fused with the original time-series image data at the pixel level to output a dynamic simulation image sequence of the intravascular contrast agent diffusion process.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method for simulating intravascular contrast agent diffusion based on image analysis. Background Technology

[0002] Accurate simulation of contrast agent diffusion behavior within vascular networks is valuable in clinical diagnosis and interventional treatment planning. Current techniques typically rely on visual observation of continuous angiographic images or simple post-processing analysis. A common approach is to coarsely estimate the arrival time or perfusion area of ​​the contrast agent based on image intensity thresholding. This method treats each pixel as an independent time series and determines the perfusion status by setting a fixed threshold. Another approach is to apply simplified computational fluid dynamics models to simulate contrast agent flow in idealized tubular geometries. These existing methods have significant drawbacks. Thresholding methods completely ignore the topological connectivity of the vascular system and the physical continuity of contrast agent transport, leading to judgments of perfusion sequence in complex bifurcation regions that often do not conform to physiological reality, and simulation results exhibiting discrete, discontinuous artifacts. While simplified fluid models take into account flow continuity, they typically treat vascular networks as regular and separate flow tubes. They fail to effectively integrate anatomical structures with complex geometric variations (such as local stenosis, tortuosity, and bifurcation angles) extracted from real images as boundary conditions for simulation. This results in simulated diffusion front morphology that is too idealized and cannot reflect the individualized vascular morphological effects.

[0003] Meanwhile, most existing methods treat contrast agent filling as a homogeneous, uniform process, lacking explicit modeling of the physical process of secondary diffusion filling of the contrast agent in the "perfused" area due to mechanisms such as leakage and mixing. This limits the realism of simulated images in assessing microscopic perfusion. Furthermore, the prediction of diffusion paths often employs heuristic algorithms based on local gradients or shortest paths. These algorithms cannot learn and remember the multiple possible patterns that the contrast agent exhibits when traversing complex bifurcation structures over time. The prediction results lack adaptability to the specific geometry of the vascular tree, leading to path selection at bifurcation points that may violate hydrodynamic principles. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an image analysis-based method for simulating intravascular contrast agent diffusion.

[0005] To achieve the above objectives, the present invention employs the following technical solution: an image analysis-based method for simulating intravascular contrast agent diffusion, comprising: Continuous temporal image data is received from an angiography scanning device, wherein each frame of the temporal image data contains contrast agent concentration signals of multiple pixels; A spatiotemporal correlation model is established for the time-series image data. The spatiotemporal correlation model uses the topological structure of the blood vessel wall as a constraint condition to correlate the contrast agent concentration signals between pixels into a continuously changing spatiotemporal concentration field. In the spatiotemporal correlation model, a contrast agent flow front is identified, which divides the spatiotemporal concentration field into perfused and unperfused regions. Using the dynamic evolution trajectory of the contrast agent flow front, a physical process model describing the diffuse filling of the contrast agent in the perfused area is constructed. Multiple key spatiotemporal nodes are extracted from the physical process model, and the information of the key spatiotemporal nodes is fed into a recurrent neural network based on a temporal memory mechanism. The recurrent neural network generates a predicted trajectory of the contrast agent diffusion path based on the geometric morphology of the blood vessel bifurcation structure. By integrating the predicted trajectory with the spatiotemporal concentration field, a contrast agent perfusion probability map containing multi-level vascular branches is generated. Spatial probability density calibration is performed on the map, and the probability values ​​are adjusted according to the local curvature of the blood vessels and the blood flow velocity distribution. The calibrated contrast agent perfusion probability map is fused with the original time-series image data at the pixel level to generate a dynamic simulated image sequence of the intravascular contrast agent diffusion process.

[0006] As a further aspect of the present invention, the step of establishing a spatiotemporal correlation model for the time-series image data includes: Spatiotemporal registration is performed on continuous temporal image data to unify the spatial coordinate system and timestamps of all image frames; Based on the time-varying curve of the contrast agent concentration signal of each pixel, the correlation coefficient matrix of the concentration change of adjacent pixels is calculated. By combining the topological structure of the blood vessel wall obtained from the angiography image, and using the blood vessel wall as the boundary constraint, pixels with correlation coefficients exceeding a preset threshold are clustered into the same spatiotemporal concentration field unit. The clustered spatiotemporal concentration field units are hierarchically organized according to the connectivity of vascular branches to form a spatiotemporal correlation model describing the continuous transmission of contrast agent concentration in the vascular network.

[0007] As a further aspect of the present invention, identifying the contrast agent flow front in the spatiotemporal correlation model includes: In the spatiotemporal correlation model, gradient abrupt boundary of spatiotemporal concentration field is detected along the blood flow direction; Morphological closure and refinement operations are performed on the detected gradient abrupt boundary to form a continuous front. The advancing speed and direction of the moving forward are calculated based on the displacement vector of the front line in consecutive image frames; Based on the difference in advance speed, the mobile forwards are further divided into fast-advancing main forwards and slow-advancing lateral forwards.

[0008] As a further aspect of the present invention, the step of constructing a physical process model describing the diffuse filling of the contrast agent within the perfused area using the dynamic evolution trajectory of the contrast agent flow front includes: Record the entire time from the first arrival of the flow front to the stable concentration of contrast agent in each perfused area; Using the arrival time of the flow front as the time origin, a spatial concentration profile function describing the decay of contrast agent concentration from the center of the blood vessel to the vessel wall is established. The transverse diffusion coefficient of contrast agent molecules in plasma is defined by combining the vessel radius and local blood flow shear force. Based on the spatial concentration profile function and the lateral diffusion coefficient, a physical process model with the convection diffusion equation as its core is constructed to simulate the homogenization process of contrast agent concentration in the perfused area.

[0009] As a further aspect of the present invention, the extraction of multiple key spatiotemporal nodes from the physical process model includes: Each key spatiotemporal node corresponds to a significant topological change in the contrast agent concentration field; Monitor the contrast agent concentration homogenization process described in the physical process model and record the moment when the concentration change rate changes abruptly; At the moment when the abrupt change occurs, a snapshot of the contrast agent concentration distribution on the cross-section of the blood vessel is taken; Analyzing the distribution snapshots identifies significant changes in the spatial distribution pattern of contrast agent concentration due to vascular bifurcation, stenosis, or dilation. The specific time and spatial location corresponding to each significant change event are defined as a key spatiotemporal node, and the concentration distribution pattern type of the key spatiotemporal node is recorded.

[0010] As a further aspect of the present invention, feeding the information of key spatiotemporal nodes into a recurrent neural network based on a temporal memory mechanism includes: For each key spatiotemporal node, its corresponding spatial location, occurrence time, and concentration distribution pattern type are encoded to form a node feature vector; Key spatiotemporal nodes are arranged in chronological order to form a sequence of node feature vectors, which serve as the input to the recurrent neural network. An attention mechanism is introduced into the hidden layer of the recurrent neural network, enabling the network to dynamically adjust the memory weights of historical node information based on the geometric morphological features of the blood vessel bifurcation structure. The recurrent neural network processes the node feature vector sequence and outputs the predicted trajectory of the contrast agent diffusion path at future time steps.

[0011] As a further aspect of the present invention, the integration of the predicted trajectory and the spatiotemporal concentration field to generate a contrast agent perfusion probability map containing multi-level vascular branches includes: The predicted trajectory output by the recurrent neural network is transformed into an initial field of probability density distribution in three-dimensional space; The initial field of the probability density distribution is superimposed on the spatiotemporal concentration field in the spatiotemporal correlation model; Based on the results of the superposition operation, different probability diffusion prior weights are assigned to the trunk vessels, main branches and terminal branches according to the pre-divided vascular branch hierarchy. By applying the graph propagation algorithm, the probability values ​​with prior weights are smoothly propagated along the vascular network topology, ultimately generating a contrast agent perfusion probability map that covers the entire region of interest and can distinguish between different levels of blood vessels.

[0012] As a further aspect of the present invention, the step of performing spatial probability density calibration and adjusting the probability value according to the local curvature of the blood vessel and the blood flow velocity distribution includes: The vessel centerline is segmented from the angiography image, and the local curvature of each point on the vessel centerline is calculated. Acquire Doppler blood flow imaging data, or estimate the average blood flow velocity distribution in different segments of the blood vessel by contrast agent transit time; Establish a calibration function, the input of which is the local curvature of the blood vessel and the blood flow velocity, and the output is an adjustment coefficient for the baseline probability value; The calibration function is applied to each spatial location of the contrast agent perfusion probability map, and the probability values ​​in the map are calibrated point by point using adjustment coefficients, so that the probability values ​​are reduced in regions with large curvature and slow flow rate, and increased in regions with small curvature and fast flow rate.

[0013] As a further aspect of the present invention, the step of performing pixel-level probabilistic fusion of the calibrated contrast agent perfusion probability map with the original time-series image data includes: For each frame of the original time-series image data, its pixel intensity value is extracted and normalized to the range of zero to one, which is used as the observation probability. The calibrated contrast agent perfusion probability map is interpolated in the time dimension to generate a dynamic probability map sequence that corresponds one-to-one with the time points of the original image frames. For each pixel at each time point, the observed probability and the corresponding predicted probability in the dynamic probability map are weighted and averaged according to the preset fusion weights. The fusion probability values ​​of all pixels obtained after weighted averaging are subjected to spatiotemporal smoothing filtering, and finally reconstructed into a set of dynamic simulated images of intravascular contrast agent diffusion that are continuous in time and coherent in space.

[0014] As a further aspect of the present invention, the application of the graph propagation algorithm smoothly propagates the probability values, after being assigned prior weights, along the vascular network topology, including: The vascular network topology is abstracted as a graph structure, where the intersection of vascular branches is defined as a graph node, and vascular branch segments are defined as graph edges. Based on the hierarchical relationship of blood vessel branches, a propagation weight coefficient is assigned to each graph edge, and the propagation weight coefficient is inversely proportional to the hierarchy of the blood vessel branches; The probability values ​​after assigning prior weights are used as the initial probability values ​​of the graph nodes; A graph propagation algorithm based on random walk is adopted to iteratively calculate the probability value of each graph node. In each iteration, the probability value of each graph node is updated to the average of its own probability value and the probability values ​​of its neighboring graph nodes multiplied by the corresponding propagation weight coefficient. When the change in the probability values ​​of all graph nodes in two consecutive iterations is less than a preset threshold, the iteration stops, and the probability values ​​of the graph nodes obtained in the final iteration are mapped back to the original spatial location of the vascular network, thus completing the smooth propagation of the probability values.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: A spatiotemporal correlation model is established using the vascular wall topology as a constraint, correlating the concentration signals of discrete pixels into a continuous spatiotemporal concentration field. This approach confines the evolution of contrast agent concentration to the connected domains defined by the anatomical structure, eliminating non-physiological spatial jumps. The constraints force the evolution of the concentration field within the vascular lumen to maintain spatial continuity and strictly follow the topological connections of vascular branches. This process naturally distinguishes the clear interface between perfused and non-perfused areas, i.e., the flow front, providing accurate spatiotemporal boundaries for modeling subsequent physical processes. The vascular topology is incorporated into the model as prior knowledge, ensuring that the simulated concentration field is morphologically consistent with the patient's own vascular structure. This overcomes the problem of discontinuous perfusion areas caused by traditional thresholding methods and avoids the drawback of simplified flow tube models deviating from real anatomical geometry.

[0016] A recurrent neural network (RNN) based on temporal memory mechanism is constructed to generate predicted trajectories of contrast agent diffusion paths based on the geometry of vascular bifurcation structures. The temporal memory mechanism of the RNN enables it to learn and encode the dynamic patterns of the contrast agent front's propagation along the vessel in historical frames, capturing its movement trends. The network infers from the input key spatiotemporal node information and the specific geometry of the current vessel bifurcation (such as branch angle and diameter ratio) to generate predicted trajectories that conform to the flow behavior at bifurcation points. This approach transforms path prediction from a simple geometric extrapolation or application of fixed rules into a data-driven reasoning process that adapts to different bifurcation morphologies. It can handle the problem of multiple possible path selection for contrast agents in the face of complex bifurcation networks, generating more reasonable prediction results in terms of both anatomical and hemodynamic significance. Compared to traditional algorithms based on local image gradients or shortest paths, it significantly improves the physiological rationality and individual consistency of the predicted paths. Attached Figure Description

[0017] Figure 1 This is a flowchart of the image analysis-based intravascular contrast agent diffusion simulation method described in this invention; Figure 2 A flowchart for identifying the contrast agent flow front; Figure 3 The iterative convergence curve of the graph propagation algorithm in the calculation of contrast agent perfusion probability is shown. Figure 4 A bar chart comparing different vascular segments before and after contrast agent perfusion probability calibration. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the description of this invention, it should be understood that terms such as "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] The core of this invention lies in providing an image analysis-based method for simulating intravascular contrast agent diffusion. Please refer to [link to relevant documentation]. Figure 1The overall implementation scheme is as follows: Continuous temporal image data is received from an angiography scanning device. Each frame of the temporal image data contains contrast agent concentration signals of multiple pixels. A spatiotemporal correlation model is established for the temporal image data. This model uses the topology of the blood vessel wall as a constraint to correlate the contrast agent concentration signals between pixels into a continuously changing spatiotemporal concentration field. A contrast agent flow front is identified in the spatiotemporal correlation model, which divides the spatiotemporal concentration field into perfused and unperfused regions. Using the dynamic evolution trajectory of the flow front, a physical process model describing the diffuse filling of contrast agent within the perfused region is constructed. Multiple key spatiotemporal nodes are extracted from the physical process model, and the information of these key spatiotemporal nodes is fed into a recurrent neural network based on a temporal memory mechanism. This network generates a predicted trajectory of the contrast agent diffusion path based on the geometry of the blood vessel bifurcation structure. The predicted trajectory is integrated with the spatiotemporal concentration field to generate a contrast agent perfusion probability map containing multi-level blood vessel branches. Spatial probability density calibration is performed on this map, adjusting the probability values ​​according to the local curvature of the blood vessel and the blood flow velocity distribution. The calibrated contrast agent perfusion probability map is fused with the original time-series image data at the pixel level to generate a dynamic simulated image sequence of the intravascular contrast agent diffusion process.

[0020] In one embodiment of the present invention, see [reference] Figure 2 This embodiment details the establishment of the spatiotemporal correlation model and the identification process of the contrast agent flow front. Spatiotemporal registration is performed on continuous time-series image data to unify the spatial coordinate system and timestamps of all image frames. Based on the time-varying curves of contrast agent concentration signals at pixel points, the correlation coefficient matrix of concentration changes between adjacent pixels is calculated. Combined with the vessel wall topology segmented from the contrast images, and using the vessel wall as a boundary constraint, pixels with correlation coefficients exceeding a preset threshold are clustered into the same spatiotemporal concentration field unit. These clustered spatiotemporal concentration field units are then hierarchically organized according to the connectivity of vessel branches, forming a spatiotemporal correlation model describing the continuous transmission of contrast agent concentration in the vascular network. In this model, gradient abrupt boundaries of the spatiotemporal concentration field are detected along the blood flow direction. Morphological closure and refinement operations are performed on the detected gradient abrupt boundaries to form continuous fronts. Based on the displacement vector of the fronts in continuous image frames, the propulsion speed and direction of the flow front are calculated. Based on the difference in propulsion speed, the flow front is further distinguished into a rapidly advancing main front and a slowly advancing lateral front.

[0021] In practice, a time-series image of the coronary arteries is received from an angiography scanning device. This data contains 100 consecutive images, each 512×512 pixels in size, with a time interval of 0.1 seconds between each frame. The pixel intensity value of each frame represents the contrast agent concentration signal at that point. Spatiotemporal registration is performed on the continuous time-series image data. An image registration algorithm based on maximizing mutual information is used to unify the spatial coordinate system of all 100 frames, and the time axis is aligned according to the timestamps recorded by the device. Based on the time-varying curve of the contrast agent concentration signal of each pixel, i.e., the intensity sequence of each pixel at 100 time points, the correlation coefficient matrix of the concentration changes of adjacent pixels is calculated. Each element in the correlation coefficient matrix represents the Pearson correlation coefficient between the concentration time series of two spatially adjacent pixels. The formula for calculating the Pearson correlation coefficient is expressed as:

[0022] in: The Pearson correlation coefficient, representing the time series of contrast agent concentrations between two adjacent pixels X and Y, is a dimensionless scalar between -1 and 1. and These represent the two adjacent pixels at time points. The contrast agent concentration signal, and These are the means of two concentration time series, respectively. This is the total number of time frames, in this example. By combining the vessel wall topology obtained from contrast-enhanced images using the level set method, which identifies the boundary between the vessel lumen and surrounding tissues, and using the vessel wall as a boundary constraint, pixels with correlation coefficients exceeding a preset threshold of 0.85 in the calculated matrix are clustered into the same spatiotemporal concentration field unit. These clustered spatiotemporal concentration field units are then hierarchically organized based on the connectivity of vessel branches. The spatiotemporal concentration field units of the main trunk are defined as the first level, those of the main branches as the second level, and those of the terminal branches as the third level, forming a spatiotemporal correlation model describing the continuous transmission of contrast agent concentration in the vascular network.

[0023] In some embodiments, a contrast agent flow front is identified in a spatiotemporal correlation model. Gradient abrupt boundary changes in the spatiotemporal concentration field are detected along the blood flow direction. A three-dimensional Sobel operator is used to calculate the gradients of the spatiotemporal concentration field in three spatial directions and the time dimension. Voxels with gradient moduli greater than a set threshold are marked as gradient abrupt boundary changes. Morphological closure and thinning operations are performed on the detected gradient abrupt boundary changes. Spherical structuring elements are used for expansion and erosion to fill small holes and breaks in the boundary, forming a continuous, single-pixel-wide front. Based on the displacement vector of the front in consecutive image frames, the propulsion velocity and direction of the flow front are calculated. The centerline position of the front in frame t and frame t+1 is extracted. The average displacement vector between the two centerlines is calculated using an iterative nearest-point algorithm. The average displacement vector divided by the time interval of 0.1 seconds yields the propulsion velocity of the flow front, and the direction of the average displacement vector is the propulsion direction of the flow front. Based on the difference in advance speed, the forward movement is further divided into fast-advancing main forwards and slow-advancing lateral forwards. The speed threshold is set at 10 mm / s. Forward segments with a calculated advance speed greater than 10 mm / s are classified as main forwards, and forward segments with a calculated advance speed less than or equal to 10 mm / s are classified as lateral forwards.

[0024] Understandably, the spatiotemporal registration process eliminates image positional deviations caused by slight patient movement or equipment vibration, and the calculation of the correlation coefficient matrix quantifies the synchronicity of contrast agent concentration changes among pixels. The vascular wall topology serves as a rigid constraint, ensuring that spatiotemporal concentration field units do not exceed the anatomical boundaries of the vessel. Clustering operations group pixel regions with highly correlated concentration dynamics into the same physiological diffusion unit. Hierarchical organization reflects the tree-like structure of the vessel, providing a topological framework for subsequent analysis. Gradient detection directly locates the leading edge of contrast agent filling, morphological operations ensure the geometric integrity of the leading edge, displacement and velocity calculations transform the dynamic changes of the leading edge into quantitative parameters, and velocity classification, based on fluid dynamics principles, distinguishes between the main stream and branch flows.

[0025] Optionally, when calculating the correlation coefficient matrix, the definition of adjacent pixels can be either an eight-neighbor or twenty-six-neighbor approach, depending on the spatial resolution and isotropy of the image. The preset correlation coefficient threshold can be adjusted between 0.8 and 0.9 to suit different image qualities or vascular regions. During morphological closure and thinning operations, the size of the structuring element can be selected based on the scale of the blood vessels in the image. For example, a spherical structuring element with a radius of 2 pixels can be used for large blood vessels, while a spherical structuring element with a radius of 1 pixel can be used for small blood vessels. The velocity threshold for distinguishing between the main arterial leader and the collateral leader can also be set based on the specific blood vessel type and prior knowledge of hemodynamics; for example, a lower threshold may be used in the intracranial arterial system.

[0026] In some embodiments, the segmentation of the vessel wall topology is accomplished using a semi-automatic algorithm. An operator delineates seed points on the first image frame, and then a region growing algorithm combined with edge detection is used to segment the vessel lumen throughout the entire temporal image sequence, extracting the continuous surface model of the lumen as the vessel wall topology. Spatiotemporal registration employs an algorithm based on maximizing mutual information. Its optimization process uses gradient descent, with the first image frame serving as a fixed reference image, and the remaining 99 images are sequentially registered to the spatial coordinate system of this reference image. When calculating the advance velocity of the flow front, for vessel segments with high curvature, a non-rigid registration algorithm is used to match the centerline of the front to more accurately calculate the local displacement vector.

[0027] In one embodiment of the present invention, this embodiment elaborates on the construction of the physical process model and the extraction of key spatiotemporal nodes. The entire time process from the first arrival of the flow front to the stable concentration of the contrast agent within each perfused region is recorded. Taking the arrival time of the flow front as the time origin, a spatial concentration profile function describing the decay of the contrast agent concentration from the vessel center to the vessel wall is established. Combining the vessel radius and local blood flow shear force, the lateral diffusion coefficient of the contrast agent molecules in the plasma is defined. Based on the spatial concentration profile function and the lateral diffusion coefficient, a physical process model with the convection diffusion equation as its core is constructed to simulate the homogenization process of the contrast agent concentration within the perfused region. Multiple key spatiotemporal nodes are extracted from the physical process model, where each key spatiotemporal node corresponds to a significant topological change in the contrast agent concentration field. The homogenization process of contrast agent concentration described in the physical process model is monitored, and the moment when the concentration change rate changes abruptly is recorded. At the moment when the abrupt change occurs, a snapshot of the contrast agent concentration distribution on the cross-section of the blood vessel is captured. The distribution snapshot is analyzed to identify significant change events in the spatial distribution pattern of contrast agent concentration caused by blood vessel bifurcation, stenosis, or dilation. The specific time and spatial location corresponding to each significant change event are defined as a key spatiotemporal node, and the concentration distribution pattern type of the key spatiotemporal node is recorded.

[0028] In practice, the entire time from the first arrival of the flow front to the stabilization of the contrast agent concentration within each perfused region is recorded. For a perfused region of a cylindrical vessel segment with a radius of 0.5 mm, the contrast agent concentration signal is continuously monitored using three virtual sensors placed at the inlet, midpoint, and outlet of the vessel segment. The standard for stable contrast agent concentration is defined as the fluctuation amplitude of the concentration signal at that point in five consecutive frames of images being less than twice the baseline noise level. This determines the total duration from the arrival of the flow front to the point of concentration stabilization. Using the arrival of the flow front as the time origin, a spatial concentration profile function describing the decay of contrast agent concentration from the vessel center along the axial direction towards the vessel wall is established. This function is based on the physical assumptions of axial convection and radial diffusion, and uses a modified Gaussian distribution to describe the concentration distribution on the cross-section at a specific moment. The formula is expressed as:

[0029] in: Indicates radial distance Location, relative time Contrast agent concentration, Indicates time The contrast agent concentration at the central axis of the blood vessel is a time-dependent characteristic. A monotonically increasing function. This represents the radial distance with the central axis of the blood vessel as the origin. Indicates time The standard deviation of the concentration distribution over time is related to the diffusion process and varies with time. Increase. Combining vessel radius and local blood flow shear force, a transverse diffusion coefficient of contrast agent molecules in plasma is defined. The vessel radius is directly measured through vessel segmentation results, and the local blood flow shear force is calculated using the Poiseuille flow approximation from instantaneous blood flow velocity and vessel radius. The transverse diffusion coefficient is positively correlated with the blood flow shear force. Based on the spatial concentration profile function and the transverse diffusion coefficient, a physical process model with the convection diffusion equation as its core is constructed to simulate the homogenization process of contrast agent concentration in the perfused area. The convection term in the convection diffusion equation is driven by the propulsion velocity field of the flow front, while the diffusion term is determined by the calculated spatially varying transverse diffusion coefficient.

[0030] In some embodiments, multiple key spatiotemporal nodes are extracted from the physical process model, each corresponding to a significant topological change in the contrast agent concentration field. The contrast agent concentration homogenization process described in the physical process model is monitored. The rate of concentration change is tracked by calculating the first derivative of the contrast agent concentration with respect to time within each spatial unit. The moment when the rate of concentration change undergoes a step change is recorded; a step change is defined as an event where the absolute value of the derivative increases by more than 100% within three consecutive time steps. At the moment of the step change, a snapshot of the contrast agent concentration distribution on the vessel cross-section is captured. The distribution snapshot is stored in matrix form, with each element of the matrix corresponding to the contrast agent concentration value of a small region on the cross-section. The distribution snapshot is analyzed to identify significant transition events in the spatial distribution pattern of contrast agent concentration caused by vessel bifurcation, stenosis, or dilation. Transition events caused by vessel bifurcation are characterized by a single high-concentration core splitting into two or more cores; transition events caused by vessel stenosis are characterized by a sharp contraction in the width of the high-concentration region accompanied by an increase in the central concentration; and transition events caused by vessel dilation are characterized by an increase in the width of the high-concentration region accompanied by a dilution of the central concentration. Each significant transition event is defined as a key spatiotemporal node by its specific time and spatial location, and the concentration distribution pattern type of the key spatiotemporal node is recorded. The pattern type includes three labels: "bifurcation", "narrowing", and "expansion".

[0031] It is understandable that recording the entire time process establishes a time baseline for contrast agent filling kinetics, and the unified definition of the time origin ensures the temporal comparability of concentration evolution processes at different spatial locations. The spatial concentration profile function provides a mathematical description of the non-uniform distribution of concentration within the vessel cross-section, and its parameters are correlated with time and space. The definition of the lateral diffusion coefficient introduces the influence of hemodynamic factors on molecular diffusion, enabling the physical process model to transcend the simple assumption of uniform mixing. The convection diffusion equation, as a recognized governing equation for the transport process, provides a solid physical foundation for simulating concentration homogenization. Real-time monitoring of the concentration change rate can capture abrupt changes in the kinetic process, which often correspond to changes in vessel geometry or flow state. Capturing and analyzing distribution snapshots is a key step in discretizing the continuous physical process into characteristic events, and pattern recognition directly correlates geometric changes with concentration distribution characteristics. The definition and annotation of key spatiotemporal nodes provide structured input information for subsequent path prediction based on sequence data.

[0032] Optionally, the criteria for achieving stable contrast agent concentration can be adjusted based on different image acquisition protocols. For example, for high frame rate acquisition, fluctuations in a number of consecutive frames may need to be below a threshold. The spatial concentration profile function can also employ other mathematical models, such as parabolic distributions or multilevel exponential decay models, to accommodate different blood flow regime assumptions. The calculation of the lateral diffusion coefficient can incorporate more complex non-Newtonian blood fluid models, rather than a simple Poisson's flow approximation. When monitoring the rate of concentration change, the threshold for determining a step change can be set to other values ​​between 50% and 150% to balance sensitivity to minute changes with noise immunity. When analyzing distribution snapshots to identify transition events, in addition to "bifurcation," "narrowing," and "dilation," other pattern type labels such as "eddy currents" or "stagnation" can be defined.

[0033] In some embodiments, the physical process model is solved using the finite difference method on a three-dimensional mesh. The mesh is consistent with the geometry of the blood vessel, and the time step is dynamically determined based on the velocity of the flow front and the mesh size using the Courant number condition. The concentration change rate is calculated using a central difference scheme to improve computational accuracy. When capturing a snapshot of the contrast agent concentration distribution on a cross-section of the blood vessel, the orientation of the cross-section is determined based on the local tangent direction of the vessel's centerline to ensure that the cross-section is perpendicular to the blood flow direction. Significant transition events in the concentration distribution pattern are identified using a method based on principal component analysis and clustering. Feature extraction and classification are performed on the distribution snapshot matrix, and pattern type labels are automatically assigned. The spatial location of key spatiotemporal nodes is recorded as the three-dimensional coordinates corresponding to the centerline of the blood vessel segment where the transition event occurred, and the specific time is recorded as the absolute time offset calculated from the time origin.

[0034] In one embodiment of the present invention, the predicted trajectory generation process based on a recurrent neural network is described in detail. For each key spatiotemporal node, its corresponding spatial location, occurrence time, and concentration distribution pattern type are encoded to form a node feature vector. The key spatiotemporal nodes are arranged in chronological order to form a sequence of node feature vectors, which serves as the input to the recurrent neural network. An attention mechanism is introduced into the hidden layer of the recurrent neural network, enabling the network to dynamically adjust the memory weights of historical node information based on the geometric morphological characteristics of the vascular bifurcation structure. The recurrent neural network processes the node feature vector sequence to output the predicted trajectory of the contrast agent diffusion path at future time steps.

[0035] In specific implementation, each key spatiotemporal node is encoded with its corresponding spatial location, occurrence time, and concentration distribution pattern type to form a node feature vector. For an example of a cerebral vascular network containing 8 key spatiotemporal nodes, the spatial location of the first key spatiotemporal node is recorded as (12.7, 45.3, 98.1) mm in three-dimensional Cartesian coordinates through the center line of its vascular segment. The occurrence time is recorded as 2.3 seconds after the contrast agent is injected. The concentration distribution pattern type is "bifurcation". The pattern type is represented by a one-hot encoded vector of length 3. "Bifurcation" is encoded as [1,0,0], "stenosis" is encoded as [0,1,0], and "dilation" is encoded as [0,0,1]. The spatial location coordinates, occurrence time scalar, and pattern type one-hot encoded vector are concatenated into a node feature vector of length 7. For the example node, its node feature vector is [12.7, 45.3, 98.1, 2.3, 1, 0, 0]. The key spatiotemporal nodes are arranged in chronological order to form a node feature vector sequence, which is used as the input of the recurrent neural network. The nodes are sorted in ascending order according to their occurrence time to form a sequence containing 8 time steps. The input data of each time step is a node feature vector of length 7.

[0036] In some embodiments, an attention mechanism is introduced into the hidden layer of a recurrent neural network. The recurrent neural network adopts a long short-term memory network architecture, and the dimension of its hidden state vector is set to 128. The attention mechanism calculates the correlation score between the hidden state at the current time step and the hidden states at all historical time steps through a learnable scoring function. The scoring function is designed so that the network can dynamically adjust the memory weights of historical node information according to the geometric features of the vascular bifurcation structure. The geometric features of the vascular bifurcation structure are extracted from the vascular segmentation model, including the bifurcation angle and the diameter ratio of the sub-branches. These geometric features are encoded into a fixed-length vector and concatenated with the historical hidden states to participate in the scoring calculation. The attention weights are obtained by normalizing the correlation score through the softmax function. The weighted sum of the historical context vector is combined with the current hidden state to update the network memory. Recurrent neural networks (RNNs) process the sequence of node feature vectors to output the predicted trajectory of the contrast agent diffusion path at future time steps. At the last time step of the sequence, the output layer of the RNN maps a hidden state vector of dimension 128 to the spatial location probability distribution of three consecutive time points in the future. The output of each time point is a three-dimensional heat map. The value of each voxel in the heat map represents the probability that the contrast agent front will appear at this position at this time in the future. Connecting the voxel position with the highest probability forms a predicted trajectory.

[0037] It is understandable that constructing node feature vectors compresses the multivariate attributes of key spatiotemporal nodes into a standardized, numerical form, facilitating processing by recurrent neural networks. Arranging the node feature vector sequence in chronological order aligns with the inherent temporal causality of the contrast agent diffusion process, laying the foundation for time-dependent modeling. The inherent gating mechanism of the Long Short-Term Memory (LSTM) network architecture enables the model to learn and remember long-distance temporal dependencies, preventing gradient vanishing or exploding problems. The introduction of the attention mechanism gives the model the ability to dynamically focus on information from different historical moments, rather than treating all historical inputs uniformly. Explicitly incorporating the geometric features of vascular bifurcation into the attention scoring function allows the model to adjust its information utilization strategy based on the anatomical structure of the blood vessels; for example, when faced with an acute-angle bifurcation, the model pays more attention to diffusion patterns under similar geometric structures in the past. The method of generating a spatial probability heatmap in the output layer provides a quantitative description of the uncertainty of future diffusion paths, rather than just a single deterministic path.

[0038] Optionally, the spatial coordinates in the node feature vector can use relative coordinates with a certain anatomical landmark as the origin, rather than absolute Cartesian coordinates. Pattern type encoding can also use integer labels combined with embedding layers, instead of fixed one-hot encoding. Recurrent neural networks can use gated recurrent units or other variants; the gating structure and hidden layer dimensions of long short-term memory networks can be adjusted according to sequence length and feature complexity. The attention scoring function can use different forms of additive attention or dot-product attention, expressed as follows:

[0039] in: Indicates at time step For historical time steps Unnormalized attention score It is a step in historical time The hidden state vector, It is the current time step The corresponding vascular geometric feature vector, , It is a learnable weight matrix. It is a learnable bias vector. It is a learnable weight vector, representing the transpose of a vector. The number of future time points generated by the output layer can be set according to the clinically relevant prediction duration, and is not limited to three time points.

[0040] In some embodiments, the training data used to train the recurrent neural network comes from a large number of historical angiography cases. Each case provides a set of key spatiotemporal node sequences arranged chronologically as input, and the subsequently observed contrast agent forward trajectory as supervision labels. The geometric features of the vascular bifurcation structure are extracted from the corresponding 3D vascular model for each case through an automated image analysis process. The learnable parameters in the attention mechanism are jointly optimized with other parameters of the recurrent neural network using a backpropagation algorithm. The loss function is defined as the cross-entropy loss between the predicted probability distribution of the network output and the actual forward position. The predicted trajectory is generated using a greedy search or bundle search algorithm, selecting the path with the highest probability from the output sequence probability distribution.

[0041] In one embodiment of the present invention, the generation and propagation process of the contrast agent perfusion probability map are described in detail. The predicted trajectory output by the recurrent neural network is transformed into an initial field of probability density distribution in three-dimensional space. The initial field of probability density distribution is superimposed with the spatiotemporal concentration field in the spatiotemporal correlation model. Based on the result of the superposition operation, different probability diffusion prior weights are assigned to the main vessels, major branches, and terminal branches according to the pre-divided vascular branch hierarchy. A graph propagation algorithm is applied to smoothly propagate probability values ​​with prior weights along the vascular network topology. Specifically, the vascular network topology is abstracted as a graph structure, where the intersections of vascular branches are defined as graph nodes, and vascular branch segments are defined as graph edges. Based on the hierarchical relationship of the vascular branches, a propagation weight coefficient is assigned to each graph edge, which is inversely proportional to the hierarchical level of the vascular branch. The probability value with prior weights is used as the initial probability value of the graph node. A graph propagation algorithm based on random walks is used to iteratively calculate the probability value of each graph node. In each iteration, the probability value of each graph node is updated to the average of its own probability value and the sum of the probability values ​​of its neighboring graph nodes multiplied by their corresponding propagation weight coefficients. When the change in the probability values ​​of all graph nodes in two consecutive iterations is less than a preset threshold, the iteration stops. The final iteratively obtained graph node probability values ​​are mapped back to the original spatial location of the vascular network, completing the smooth propagation of the probability values. Finally, a contrast agent perfusion probability map covering the entire region of interest and capable of distinguishing different levels of blood vessels is generated.

[0042] In practice, the predicted trajectory output by the recurrent neural network is transformed into an initial field of probability density distribution in three-dimensional space. The predicted trajectory consists of a sequence of three-dimensional coordinate points at future time points. Each coordinate point is associated with a probability value output by the recurrent neural network. A Gaussian kernel function is used to spatially diffuse each coordinate point. The standard deviation of the Gaussian kernel is dynamically set according to the radius of the current blood vessel segment. For example, for a blood vessel segment with a radius of 0.6 mm, the standard deviation is set to 0.3 mm. The probability density values ​​of all coordinate points after being diffused by the Gaussian kernel are accumulated in the three-dimensional grid space to form a continuous probability density distribution initial field covering the entire region of interest. The spatial resolution of this three-dimensional grid is consistent with the resolution of the original time-series image data, which is 0.1 mm per pixel. The initial probability density distribution field is superimposed with the spatiotemporal concentration field in the spatiotemporal correlation model. The spatiotemporal concentration field is a continuous concentration representation after spatiotemporal registration and clustering. The superposition operation is performed on the same three-dimensional spatial grid. For each grid point, the value of the initial probability density distribution field is multiplied by a scaling factor α, and the value of the spatiotemporal concentration field is multiplied by a scaling factor β. Then, the two scaled values ​​are added point by point. The scaling factors α and β are used to balance the relative contributions of prediction information and observation information. In one example, α is set to 0.7 and β is set to 0.3. Based on the results of the superposition operation, according to the pre-divided vascular branch hierarchy, different probability diffusion prior weights are assigned to the trunk vessels, major branches, and terminal branches. The division of vascular branch hierarchy is based on the topological analysis of the vascular segmentation results, and is classified according to the vessel diameter and branch generation. The trunk vessels usually refer to first-level vessels with a diameter greater than 2 mm, the major branches refer to second-level vessels with a diameter between 1 mm and 2 mm, and the terminal branches refer to third-level and lower-level vessels with a diameter less than 1 mm. The assignment of probability diffusion prior weights follows the principle that the higher the level, the greater the weight. The specific weight values ​​are defined in Table 1.

[0043] Table 1: Prior Weights for Probability Diffusion

[0044] A graph propagation algorithm is applied to smoothly propagate probability values ​​with prior weights along the vascular network topology. The vascular network topology is abstracted as a graph structure, where the intersections of vascular branches are defined as graph nodes, and vascular branch segments are defined as graph edges. The spatial location of a graph node is taken as the three-dimensional coordinates of the intersection of vascular branches, and the connectivity of graph edges is determined by the connectivity of the vascular network. Based on the hierarchical relationship of vascular branches, a propagation weight coefficient is assigned to each graph edge. The propagation weight coefficient is inversely proportional to the hierarchy of the vascular branch. Specifically, for a vascular branch segment connecting two graph nodes, its propagation weight coefficient... The calculation formula is:

[0045] in: Represents the connection graph nodes Graph Nodes The propagation weight coefficients corresponding to the graph edges, It is a global constant, set to 1.2 in one example. This represents the hierarchical value to which the vascular branch belongs: 1 for the main trunk, 2 for major branches, and 3 for terminal branches. The initial probability value of each graph node is the probability value after being assigned prior weights. Each node's initial probability value is taken from the value at its corresponding spatial location in the superimposed and weighted 3D probability field. A graph propagation algorithm based on random walks is used to iteratively calculate the probability value of each node. In each iteration, the probability value of each node is updated to the average of its own probability value and the sum of the probabilities of its neighboring nodes multiplied by their corresponding propagation weight coefficients. The update formula is expressed as:

[0046] in: This represents the probability value of graph node i in the (n+1)th iteration after one update. Indicates the first After the second iteration, the graph nodes The probability value, Represents graph nodes The set of adjacent graph nodes, From graph nodes To graph nodes The propagation weight coefficient is used. When the change in probability values ​​of all graph nodes in two consecutive iterations is less than a preset threshold (set to 0.001), the iteration stops. The final graph node probability values ​​are mapped back to their original spatial locations in the vascular network using trilinear interpolation, completing the smooth propagation of probability values. This ultimately generates a contrast agent perfusion probability map covering the entire region of interest and capable of distinguishing between different vascular levels. The preset threshold is determined based on the vascular network topology characteristics, the convergence requirements of smooth probability propagation, and the spatial accuracy requirements of the contrast agent perfusion probability map.

[0047] In some embodiments, the initial field of the probability density distribution is generated using an adaptive bandwidth Gaussian kernel function. The bandwidth of the Gaussian kernel is adjusted according to the local geometric features of the blood vessel at each coordinate point in the predicted trajectory, including curvature and cross-sectional area. Before superimposing the spatiotemporal concentration field with the initial field of the probability density distribution, histogram matching is performed on both to eliminate systematic biases in the probability distribution caused by different data sources. The hierarchical division of blood vessel branches is performed using an automated graph theory method. Based on the vascular skeleton diagram, the Strah's number of each branch point is calculated, and labels for trunk, major branches, and terminal branches are automatically assigned according to the range of the Strah's number.

[0048] It is understandable that transforming the predicted trajectory into an initial field of probability density distribution realizes the mapping from discrete path points to a continuous probability space, and Gaussian kernel diffusion simulates the uncertainty of the predicted position. The superposition operation with the spatiotemporal concentration field integrates prediction information based on deep learning and observation information based on physical models, and the scaling factor adjusts the confidence of the two information sources. Assigning prior weights introduces prior knowledge of vascular anatomy, making probability diffusion more consistent with physiological laws; the high weight of main vessels guides probability flow towards the main path. The abstraction of the graph structure transforms the complex vascular network into a mathematically easy-to-process graph model, facilitating the application of graph theory algorithms. The design of the propagation weight coefficient being inversely proportional to the level ensures that the probability propagation resistance is low between low-level main vessels and high between high-level terminal branches. The iterative update formula based on random walk achieves local smoothing and global consistency of probability values ​​in the vascular network, the iteration termination condition ensures the convergence of the algorithm, and the interpolation mapping restores the probability values ​​on the graph nodes to a continuous spatial field, forming the final probability map.

[0049] Optionally, the Gaussian kernel function used to generate the initial field of the probability density distribution can be replaced with other kernel functions, such as the Epanechnikov kernel or a uniform kernel. The scaling factors α and β in the superposition operation can be dynamically adjusted according to the vascular region or time point; for example, giving higher weight to predicted information in the early stages of angiography and higher weight to observed information in the later stages. The hierarchical division of vascular branches can incorporate more anatomical features, such as branch angles or fractional flow reserve. The global constant κ in the propagation weighting coefficient formula can be adjusted according to the overall scale of the vascular network; for large vascular networks, the κ value can be increased to promote long-range propagation.

[0050] In some embodiments, the iterative updates of the graph propagation algorithm can employ an asynchronous update strategy, where the update order of graph nodes in each iteration follows the topological order of the vascular tree from the root node to the leaf node, thereby improving convergence speed. A preset threshold for stopping iterations can be set according to the computational accuracy and efficiency requirements of the actual application; for example, it can be set to 0.005 to accelerate computation. When mapping back to the original spatial location, in addition to trilinear interpolation, a radial basis function-based interpolation method can also be used to better preserve the distribution characteristics of probability values ​​along the vessel centerline. The generated contrast agent perfusion probability map is stored in a three-dimensional scalar field data format, where each voxel value represents the final probability of contrast agent perfusion at that location.

[0051] See Figure 3This figure shows the iterative convergence curve of the graph propagation algorithm in calculating the probability of contrast agent perfusion. The red dashed line represents the convergence threshold (0.001), indicating that the algorithm stops iterating when the probability value of all graph nodes changes by less than 0.001 in two consecutive iterations. The slope of the curve shows that the probability values ​​of all vascular levels have entered a plateau phase after 15 iterations, satisfying the convergence condition. This result directly reflects the perfusion pattern of contrast agent in different vascular levels: the main vessels have the fastest perfusion rate and the highest probability, while the peripheral vessels have the slowest perfusion rate and the lowest probability. The visualization of this convergence process verifies the effectiveness of the graph propagation algorithm, ensuring that the final generated perfusion probability map achieves a balance between computational efficiency and accuracy.

[0052] In one embodiment of the present invention, the spatial probability density calibration and pixel-level probability fusion process are described in detail. Spatial probability density calibration is performed on the contrast agent perfusion probability map, adjusting the probability values ​​based on the local curvature of the blood vessel and the blood flow velocity distribution. Specifically, this includes: segmenting the blood vessel centerline from the angiography image and calculating the local curvature at each point on the blood vessel centerline; acquiring Doppler blood flow imaging data; or estimating the average blood flow velocity distribution in each segment of the blood vessel through the contrast agent transit time; establishing a calibration function; the input of the calibration function is the local curvature of the blood vessel and the blood flow velocity, and the output is an adjustment coefficient for the base probability value; applying the calibration function to each spatial location of the contrast agent perfusion probability map; and using the adjustment coefficient to calibrate the probability values ​​in the map point by point, so that the probability value decreases accordingly in areas with high curvature and slow flow velocity, and increases accordingly in areas with low curvature and fast flow velocity. The calibrated contrast agent perfusion probability map is fused with the original time-series image data at the pixel level. Specifically, this involves: extracting the pixel intensity value of each frame in the original time-series image data and normalizing it to the range of zero to one as the observation probability; interpolating the calibrated contrast agent perfusion probability map in the time dimension to generate a dynamic probability map sequence that corresponds one-to-one with the time points of the original image frames; for each pixel at each time point, weighting the observation probability and the corresponding predicted probability in the dynamic probability map according to the preset fusion weights; and performing spatiotemporal smoothing filtering on the fusion probability values ​​of all pixels obtained after weighted averaging. Finally, a set of dynamic simulated images of intravascular contrast agent diffusion that is continuous in time and coherent in space is reconstructed.

[0053] In the specific implementation, spatial probability density calibration is performed on the contrast agent perfusion probability map. The probability values ​​are adjusted according to the local curvature of the blood vessel and the blood flow velocity distribution. The centerline of the blood vessel is segmented from the angiography image, and the local curvature of each point on the centerline is calculated. A three-dimensional thinning algorithm is used to extract a single-pixel-wide skeleton of the blood vessel centerline from the segmented binary mask. The skeleton consists of a series of continuous three-dimensional coordinate points. For each centerline point, a local ring is fitted using its preceding and following points, and the curvature is calculated. The curvature value is defined as the reciprocal of the radius of the fitted circle. Doppler blood flow imaging data is acquired, or the average blood flow velocity distribution in each segment of the blood vessel is estimated by the contrast agent transit time. In one example, the contrast agent transit time method is used. The inlet and outlet regions of the blood vessel segment are manually or automatically selected on the time-series image data. The time difference between the peak value of the contrast agent concentration-time curve is calculated as the transit time. The length of the blood vessel segment is obtained through the centerline, and the average blood flow velocity is the length divided by the transit time. A calibration function is established. Its inputs are the local curvature of the blood vessel and the blood flow velocity, and its output is an adjustment coefficient for the baseline probability value. The calibration function is designed to reduce the probability value in regions with high curvature and slow flow velocity, and increase the probability value in regions with low curvature and fast flow velocity. The specific form of the calibration function is expressed as follows:

[0054] in: Indicates the adjustment factor. This represents the local curvature of a blood vessel, expressed in millimeters to the power of negative one. This indicates blood flow velocity, measured in millimeters per second. , , This is an adjustable positive parameter, set to 1.0, 0.5, and 0.01 in one example. The calibration function is applied to each spatial location of the contrast agent perfusion probability map. An adjustment factor is used to calibrate the probability values ​​in the map point-by-point. For each voxel in the probability map, the nearest vessel centerline point is found based on its spatial coordinates. The local curvature value and the average blood flow velocity value of the corresponding vessel segment at that centerline point are obtained and substituted into the calibration function to calculate the adjustment factor. Then multiply the original probability value of the voxel by the adjustment factor. The new probability values ​​are obtained after calibration. The calibration operation ensures that the probability values ​​of regions with large curvature and slow flow velocity are reduced by multiplying by a coefficient less than 1, while the probability values ​​of regions with small curvature and fast flow velocity are increased by multiplying by a coefficient greater than 1.

[0055] In some embodiments, the calibrated contrast agent perfusion probability map is fused with the original time-series image data at the pixel level. For each frame of the original time-series image data, the pixel intensity value is extracted and normalized to a range of zero to one as the observation probability. The original time-series image data is a sequence of 100 frames (512×512 pixels). For each pixel in each frame, its grayscale value is subtracted from the average background noise value, and then divided by the maximum enhancement intensity value to obtain the normalized observation probability value. The average background noise value is taken from the image frame before the contrast agent arrives. The calibrated contrast agent perfusion probability map is interpolated in the time dimension to generate a dynamic probability map sequence corresponding one-to-one with the time points of the original image frames. The calibrated contrast agent perfusion probability map is a static three-dimensional probability field, which is extended into a four-dimensional spatiotemporal field by defining a time evolution function. The time evolution function is modeled based on the total time of contrast agent injection and the total capacity of the vascular network. A corresponding three-dimensional probability map is generated for each timestamp of the original image frame using a linear interpolation method. For each pixel at each time point, the observed probability and the corresponding predicted probability in the dynamic probability map are weighted and averaged according to preset fusion weights. The fusion weights are set to fixed values: the weight for the observed probability is set to 0.4, and the weight for the predicted probability is set to 0.6. The weighted average is calculated independently at each pixel location, and the formula is: Fusion Probability = 0.4 × Observed Probability + 0.6 × Predicted Probability. The fusion probability values ​​of all pixels obtained after weighted averaging are then subjected to spatiotemporal smoothing filtering. The spatiotemporal smoothing filtering uses a three-dimensional Gaussian filter convolved in the spatial dimension and a one-dimensional Gaussian filter convolved in the temporal dimension. The spatial Gaussian kernel size is 3×3×3 pixels, and the temporal Gaussian kernel size is 5 frames. Finally, a set of temporally continuous and spatially coherent simulated images of intravascular contrast agent diffusion is reconstructed. The reconstructed image sequence maintains the same frame rate and spatial resolution as the original time-series image data.

[0056] Understandably, segmentation of the vessel centerline and curvature calculation provide a quantitative indicator describing the degree of vascular geometric tortuosity; high curvature regions typically correspond to increased blood flow resistance and delayed contrast agent passage. Obtaining the blood flow velocity distribution introduces hemodynamic information, as flow velocity directly affects the contrast agent's transmission rate. The calibration function combines geometric and hemodynamic factors into a unified adjustment coefficient, achieving physiological correction of the baseline probability values. Point-by-point calibration makes the spatial distribution of the probability map more consistent with actual blood flow characteristics. Normalization of observed probabilities transforms the original image intensity into a range comparable to the probability values. Temporal interpolation dynamizes the static probability field to match the temporal characteristics of the angiography process. Weighted average fusion combines real-time information from direct observation with prior information from model prediction, balancing the advantages of data-driven and model-driven approaches. Spatiotemporal smoothing filtering eliminates local noise and discontinuities introduced during fusion, ensuring the visual coherence and physical plausibility of the generated simulated image sequence.

[0057] Optionally, the calculation of local vascular curvature can employ a geometric differential method based on a sequence of centerline points, such as using a curvature formula determined by three adjacent points. The estimation of blood flow velocity distribution can be combined with color Doppler ultrasound data; when Doppler data is available, velocity values ​​are directly read and registered to the angiographic image space. Parameters in the calibration function... , , Optimization can be achieved using machine learning methods on historical data, rather than fixed settings. During the normalization of observation probabilities, the maximum enhancement intensity value can be taken as the maximum grayscale value of that pixel throughout the entire time series, or as the average maximum enhancement value of a reference vascular region. The temporal evolution function can employ a more complex model, such as an exponential decay function based on the blood flow front propagation velocity. The fusion weights can be adaptively adjusted based on image quality or vascular region; for example, the weight of the observation probability can be reduced in areas with high image noise. The kernel size of the spatiotemporal smoothing filter can be adjusted based on the image's spatial resolution and temporal frame rate.

[0058] In some embodiments, the extraction of observation probabilities employs a local background correction strategy. For each pixel, its background noise value is taken from the average grayscale value of multiple frames of images at the same anatomical location before the arrival of the contrast agent, rather than the global average. The generation of the dynamic probability map sequence uses a physics-based transport model for temporal interpolation, rather than simple linear interpolation. This transport model considers the convection-diffusion process of the contrast agent in the vascular network. A weighted average fusion step introduces confidence estimation. The confidence of the observation probability is calculated based on the signal-to-noise ratio of the pixel in different frames, while the confidence of the predicted probability is calculated based on the probability distribution entropy output by the recurrent neural network. The fusion weights are dynamically allocated according to the confidence levels. Spatiotemporal smoothing filtering employs an anisotropic diffusion filter, with the filtering intensity adjusted based on vascular edge information to smooth noise while protecting vascular boundaries. The final reconstructed dynamic analog image sequence is output in a digital imaging and communications medicine standard format, containing three-dimensional data from all time frames.

[0059] See Figure 4This is a bar chart comparing the contrast agent perfusion probability before and after calibration for different vessel segments. The calibrated probabilities for all segments are slightly lower than the initial probabilities, which aligns with the calibration rule you mentioned earlier—the algorithm adjusts the initial probabilities based on local vessel curvature and blood flow velocity, reducing probability values ​​in areas with high curvature and slow flow. This result directly verifies the effectiveness of the calibration process, combining the initial probability based on the image with the physiological characteristics of the vessel, making the final perfusion probability map more consistent with real hemodynamics. The probability differences between different vessel segments can help physicians identify potential perfusion risk areas; for example, the significantly decreased probability in the right posterior descending artery suggests stenosis or hemodynamic abnormalities in that area. From the aorta to the terminal branches, both the initial and calibrated probabilities show a decreasing trend, perfectly consistent with the physiological principle that the higher the level of the vessel branch, the lower the perfusion probability.

[0060] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for simulating intravascular contrast agent diffusion based on image analysis, characterized in that, The method includes: Continuous temporal image data is received from an angiography scanning device, wherein each frame of the temporal image data contains contrast agent concentration signals of multiple pixels; A spatiotemporal correlation model is established for the time-series image data. The spatiotemporal correlation model uses the topological structure of the blood vessel wall as a constraint condition to correlate the contrast agent concentration signals between pixels into a continuously changing spatiotemporal concentration field. In the spatiotemporal correlation model, a contrast agent flow front is identified, which divides the spatiotemporal concentration field into perfused and unperfused regions. Using the dynamic evolution trajectory of the contrast agent flow front, a physical process model describing the diffuse filling of the contrast agent in the perfused area is constructed. Multiple key spatiotemporal nodes are extracted from the physical process model, and the information of the key spatiotemporal nodes is fed into a recurrent neural network based on a temporal memory mechanism. The recurrent neural network generates a predicted trajectory of the contrast agent diffusion path based on the geometric morphology of the blood vessel bifurcation structure. By integrating the predicted trajectory with the spatiotemporal concentration field, a contrast agent perfusion probability map containing multi-level vascular branches is generated. Spatial probability density calibration is performed on the map, and the probability values ​​are adjusted according to the local curvature of the blood vessels and the blood flow velocity distribution. The calibrated contrast agent perfusion probability map is fused with the original time-series image data at the pixel level to generate a dynamic simulated image sequence of the intravascular contrast agent diffusion process.

2. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 1, characterized in that, The establishment of a spatiotemporal correlation model for the time-series image data includes: Spatiotemporal registration is performed on continuous temporal image data to unify the spatial coordinate system and timestamps of all image frames; Based on the time-varying curve of the contrast agent concentration signal of each pixel, the correlation coefficient matrix of the concentration change of adjacent pixels is calculated. By combining the topological structure of the blood vessel wall obtained from the angiography image, and using the blood vessel wall as the boundary constraint, pixels with correlation coefficients exceeding a preset threshold are clustered into the same spatiotemporal concentration field unit. The clustered spatiotemporal concentration field units are hierarchically organized according to the connectivity of vascular branches to form a spatiotemporal correlation model describing the continuous transmission of contrast agent concentration in the vascular network.

3. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 1, characterized in that, The identification of the contrast agent flow front in the spatiotemporal correlation model includes: In the spatiotemporal correlation model, gradient abrupt boundary of spatiotemporal concentration field is detected along the blood flow direction; Morphological closure and refinement operations are performed on the detected gradient abrupt boundary to form a continuous front. The advancing speed and direction of the moving forward are calculated based on the displacement vector of the front line in consecutive image frames; Based on the difference in advance speed, the mobile forwards are further divided into fast-advancing main forwards and slow-advancing lateral forwards.

4. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 1, characterized in that, The method of constructing a physical process model describing the diffuse filling of the contrast agent within the perfused area using the dynamic evolution trajectory of the contrast agent flow front includes: Record the entire time from the first arrival of the flow front to the stable concentration of contrast agent in each perfused area; Using the arrival time of the flow front as the time origin, a spatial concentration profile function describing the decay of contrast agent concentration from the center of the blood vessel to the vessel wall is established. The transverse diffusion coefficient of contrast agent molecules in plasma is defined by combining the vessel radius and local blood flow shear force. Based on the spatial concentration profile function and the lateral diffusion coefficient, a physical process model with the convection diffusion equation as its core is constructed to simulate the homogenization process of contrast agent concentration in the perfused area.

5. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 1, characterized in that, The extraction of multiple key spatiotemporal nodes from the physical process model includes: Each key spatiotemporal node corresponds to a significant topological change in the contrast agent concentration field; Monitor the contrast agent concentration homogenization process described in the physical process model and record the moment when the concentration change rate changes abruptly; At the moment when the abrupt change occurs, a snapshot of the contrast agent concentration distribution on the cross-section of the blood vessel is taken; Analyzing the distribution snapshots identifies significant changes in the spatial distribution pattern of contrast agent concentration due to vascular bifurcation, stenosis, or dilation. The specific time and spatial location corresponding to each significant change event are defined as a key spatiotemporal node, and the concentration distribution pattern type of the key spatiotemporal node is recorded.

6. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 1, characterized in that, The step of feeding information from key spatiotemporal nodes into a recurrent neural network based on a temporal memory mechanism includes: For each key spatiotemporal node, its corresponding spatial location, occurrence time, and concentration distribution pattern type are encoded to form a node feature vector; Key spatiotemporal nodes are arranged in chronological order to form a sequence of node feature vectors, which serve as the input to the recurrent neural network. An attention mechanism is introduced into the hidden layer of the recurrent neural network, enabling the network to dynamically adjust the memory weights of historical node information based on the geometric morphological features of the blood vessel bifurcation structure. The recurrent neural network processes the node feature vector sequence and outputs the predicted trajectory of the contrast agent diffusion path at future time steps.

7. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 1, characterized in that, The process of integrating the predicted trajectory with the spatiotemporal concentration field to generate a contrast agent perfusion probability map containing multi-level vascular branches includes: The predicted trajectory output by the recurrent neural network is transformed into an initial field of probability density distribution in three-dimensional space; The initial field of the probability density distribution is superimposed on the spatiotemporal concentration field in the spatiotemporal correlation model; Based on the results of the superposition operation, different probability diffusion prior weights are assigned to the trunk vessels, main branches and terminal branches according to the pre-divided vascular branch hierarchy. By applying the graph propagation algorithm, the probability values ​​with prior weights are smoothly propagated along the vascular network topology, ultimately generating a contrast agent perfusion probability map that covers the entire region of interest and can distinguish between different levels of blood vessels.

8. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 1, characterized in that, The process of performing spatial probability density calibration, adjusting the probability value based on the local curvature of the blood vessel and the blood flow velocity distribution, includes: The vessel centerline is segmented from the angiography image, and the local curvature of each point on the vessel centerline is calculated. Acquire Doppler blood flow imaging data, or estimate the average blood flow velocity distribution in different segments of the blood vessel by contrast agent transit time; Establish a calibration function, the input of which is the local curvature of the blood vessel and the blood flow velocity, and the output is an adjustment coefficient for the baseline probability value; The calibration function is applied to each spatial location of the contrast agent perfusion probability map, and the probability values ​​in the map are calibrated point by point using adjustment coefficients, so that the probability values ​​are reduced in regions with large curvature and slow flow rate, and increased in regions with small curvature and fast flow rate.

9. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 1, characterized in that, The step of performing pixel-level probabilistic fusion of the calibrated contrast agent perfusion probability map with the original time-series image data includes: For each frame of the original time-series image data, its pixel intensity value is extracted and normalized to the range of zero to one, which is used as the observation probability. The calibrated contrast agent perfusion probability map is interpolated in the time dimension to generate a dynamic probability map sequence that corresponds one-to-one with the time points of the original image frames. For each pixel at each time point, the observed probability and the corresponding predicted probability in the dynamic probability map are weighted and averaged according to the preset fusion weights. The fusion probability values ​​of all pixels obtained after weighted averaging are subjected to spatiotemporal smoothing filtering, and finally reconstructed into a set of dynamic simulated images of intravascular contrast agent diffusion that are continuous in time and coherent in space.

10. The method for simulating intravascular contrast agent diffusion based on image analysis according to claim 7, characterized in that, The applied graph propagation algorithm smoothly propagates probability values ​​with prior weights along the vascular network topology, including: The vascular network topology is abstracted as a graph structure, where the intersection of vascular branches is defined as a graph node, and vascular branch segments are defined as graph edges. Based on the hierarchical relationship of blood vessel branches, a propagation weight coefficient is assigned to each graph edge, and the propagation weight coefficient is inversely proportional to the hierarchy of the blood vessel branches; The probability values ​​after assigning prior weights are used as the initial probability values ​​of the graph nodes; A graph propagation algorithm based on random walk is adopted to iteratively calculate the probability value of each graph node. In each iteration, the probability value of each graph node is updated to the average of its own probability value and the probability values ​​of its neighboring graph nodes multiplied by the corresponding propagation weight coefficient. When the change in the probability values ​​of all graph nodes in two consecutive iterations is less than a preset threshold, the iteration stops, and the probability values ​​of the graph nodes obtained in the final iteration are mapped back to the original spatial location of the vascular network, thus completing the smooth propagation of the probability values.

Images