Microvessel high-fidelity reconstruction system and method based on topological prior and neural implicit surface
By using topological priors and a neural implicit surface reconstruction system, the limitations of sensors, optical environment, and real-time performance in 3D reconstruction within microvessels were addressed, enabling the generation of high-fidelity 3D vascular models and improving lesion recognition capabilities.
Patent Information
- Application Number
- CN202610869091.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-16
- Publication Date
- 2026-08-25
AI Technical Summary
Existing technologies face challenges in 3D reconstruction within microvessels due to sensor physical limitations, spatial constraints, extreme optical environments, and real-time and interpretation requirements, making it difficult to achieve high-fidelity reconstruction.
A high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces is adopted, including hardware acquisition, image enhancement pipeline, topological constraint pose inference engine and generative neural implicit reconstruction module. Image enhancement and pose calculation are performed through point light source irradiance model, dense optical flow and neural symbol distance field to generate high-fidelity three-dimensional vascular model.
Under conditions of no IMU and single-LED follow-up illumination, high-fidelity three-dimensional reconstruction of microvessels was achieved, outputting a three-dimensional vascular model with geometric continuity, clear texture, and reliable blind-area completion anatomy, thus improving the ability to identify lesions.
Smart Images
Figure CN122636859A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of medical image processing and neurointerventional radiology, specifically to a high-fidelity microvascular reconstruction system and method based on topological priors and neural implicit surfaces. Background Technology
[0002] With the development of interventional radiology, doctors have an increasing need for direct assessment of lesions inside microvessels (with diameters between 1mm and 3mm). However, due to the narrow space of microvessels, the probe size must be limited to below 0.5mm, which leads to extremely stringent engineering limitations:
[0003] 1. Sensor physical limitations: Traditional digital circuits struggle to achieve high-definition sampling on photosensitive areas at the millimeter level, and currently, low-resolution analog video signals are mostly used. This extremely low resolution makes traditional feature-point-based matching algorithms (such as SIFT and ORB) almost completely ineffective when faced with smooth, repetitive textures on blood vessel walls.
[0004] 2. Spatial environment limitations: The probe (including a miniature camera and LED light source) cannot integrate pose sensing devices such as an IMU (Inertial Measurement Unit), which means that the camera's motion pose is completely unknown. Traditional monocular SLAM systems are prone to scale collapse and ambiguity in forward and backward motion directions in tubular environments.
[0005] 3. Extreme optical environment: Relying solely on follow-up LEDs for illumination resulted in the "flashlight effect," where light intensity decreased inversely with distance, and shadows and highlights changed drastically as the camera moved, violating the commonly used assumption of photometric consistency in 3D reconstruction.
[0006] 4. Real-time performance and interpretation requirements: Reconstruction not only requires speed, but also requires the results to have "interpretive value", that is, it cannot be just a blurry point cloud, but must be a high-fidelity surface with the texture of biological tissue.
[0007] Therefore, there is an urgent need for a solution that can achieve high-fidelity three-dimensional reconstruction of microvessels under conditions of low light, high noise, and no pose sensing, in order to overcome the shortcomings of the existing technologies. Summary of the Invention
[0008] One of the objectives of this invention is to provide a high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces, thereby solving the aforementioned problems.
[0009] To achieve the above objectives, a high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces is provided, including:
[0010] The hardware acquisition unit is used to acquire low-resolution analog video signals within microvessels.
[0011] The image enhancement pipeline is used to receive the low-resolution analog video signal, establish a point light source irradiance model that moves with the probe for dynamic brightness correction, and enhance the sub-pixel features of the image through a generative spatiotemporal super-resolution network to output the enhanced image sequence.
[0012] The topology-constrained pose inference engine is used to receive the enhanced image sequence, constrain the probe motion trajectory to the deformable cavity topology centerline under inertial guidance-free conditions, and perform inter-frame relative motion trend tracking and pose calculation based on dense optical flow within the local channel constrained by the centerline, and output inter-frame pose data.
[0013] The generative neural implicit reconstruction module is used to receive the enhanced image sequence and the inter-frame pose data, and to perform anatomically consistent geometric completion of the observation blind area using the neural symbol distance function to generate a high-fidelity three-dimensional blood vessel model.
[0014] Furthermore, the image enhancement pipeline includes:
[0015] The follow-up motion light and shadow inverse modeling unit is used to establish a point light source irradiance model, and perform inverse brightness compensation on each frame of the enhanced image sequence based on the point light source irradiance model to eliminate dynamic highlights and edge shadows caused by probe movement.
[0016] A generative spatiotemporal super-resolution network uses a recurrent neural structure to fuse multiple consecutive frames from the low-resolution analog video signal, extracting the subsurface scattering texture features of the blood vessel wall that are lost due to low resolution.
[0017] Furthermore, the point light source irradiance model is expressed as:
[0018]
[0019]
[0020] Where ρ is the reflectivity of the blood vessel wall, I0 is the initial intensity of the light source, r is the distance from point P to the optical center, and θ is the angle of incidence of the light ray. This represents the gain coefficient during the analog signal transmission process.
[0021] Furthermore, the topology-constrained pose inference engine includes:
[0022] Differentiable tubular centerline modeling unit is used to construct a differentiable tubular centerline equation based on spline curves, constraining the probe's search space within a local channel centered on the centerline;
[0023] The motion trend analysis unit is used to determine the relative motion trend between frames of the probe by analyzing the divergence and curl of the dense optical flow field between adjacent frames in the enhanced image sequence within the local channel. The motion trend includes forward movement, backward movement, or bending.
[0024] The pose calculation and optimization unit is used to calculate the inter-frame relative pose of the probe by minimizing the sum of the optical flow reprojection residual, the tube wall collision penalty term, and the orientation consistency constraint term. The tube wall collision penalty term is used to force the probe to return to the center of the tube cavity, and the orientation consistency constraint term is used to force the probe's line of sight to be consistent with the tangent vector of the equation of the differentiable tubular centerline.
[0025] Furthermore, the generative neural implicit reconstruction module includes:
[0026] The neural symbolic distance field expression unit is used to determine the coordinates of three-dimensional spatial sampling points by ray sampling based on the enhanced image sequence and the inter-frame pose data, and to map each three-dimensional spatial coordinate into a symbolic distance value using a multilayer perceptron, and output geometric information expressed in the form of a neural symbolic distance field.
[0027] The generative completion unit is used to receive the neural symbol distance field output by the neural symbol distance field expression unit, and perform three-dimensional geometric completion of the observation blind zone based on the diffusion model, and output the completed neural symbol distance field.
[0028] A curvature-sensitive loss unit is used to introduce a curvature-sensitive generative adversarial loss term during the joint optimization process of the neural symbol distance field expression unit and the generative completion unit. This term is configured to preferentially preserve local curvature abrupt points when completing missing geometric surfaces.
[0029] The generative neural implicit reconstruction module extracts the zero isosurface based on the completed neural symbol distance field to generate the high-fidelity three-dimensional blood vessel model.
[0030] One objective of this invention is to provide a high-fidelity reconstruction method for microvessels based on topological priors and neural implicit surfaces, comprising the following steps:
[0031] Step S1: Perform dynamic brightness correction and generative spatiotemporal super-resolution enhancement on the acquired low-resolution analog video signal using a point light source irradiance model that moves with the probe, to obtain an enhanced image sequence.
[0032] Step S2: Based on the obtained enhanced image sequence, constrain the probe motion trajectory to the topological center line of the deformable lumen under the condition of no inertial guidance, use dense optical flow to infer the pose trend, solve the probe's running trajectory in the microvascular, and output inter-frame pose data.
[0033] Step S3: Input the enhanced image sequence and the inter-frame pose data into a generative neural implicit network to perform anatomically consistent geometric completion of the observation blind area and update the neural symbol distance field parameters online;
[0034] Step S4: Based on the updated neural symbol distance field parameters, extract the zero isosurface through voxel space and output a high-fidelity three-dimensional blood vessel model.
[0035] Furthermore, step S1 specifically includes:
[0036] Inverse brightness compensation is performed based on a point light source irradiance model to eliminate dynamic highlights and edge shadows;
[0037] A generative spatiotemporal super-resolution network with a recurrent neural structure is used to fuse sub-pixel offset information from multiple consecutive frames to achieve super-resolution mapping. Furthermore, a tissue structure consistency term is introduced to ensure that the enhanced image can identify the microvascular endothelial texture.
[0038] Furthermore, step S2 specifically includes:
[0039] The blood vessels are preset to be continuous spline curves, and the probe pose search space is constrained within a local channel based on the spline curves.
[0040] Extract dense optical flow between adjacent frames and calculate the optical flow reprojection residual caused by changes in probe pose.
[0041] A joint energy functional is constructed that includes optical flow reprojection residuals, tube wall collision penalty terms, and orientation consistency constraints. The inter-frame relative pose of the probe is solved by iterative optimization.
[0042] Furthermore, step S3 specifically includes:
[0043] The neural symbol distance field representation steps are as follows: Based on the enhanced image sequence and the inter-frame pose data, the coordinates of the three-dimensional spatial sampling points are determined by ray sampling, and each three-dimensional spatial coordinate is mapped to a symbol distance value using a multilayer perceptron, thus outputting the neural symbol distance field.
[0044] Generative completion step: Based on the diffusion model, according to the neural symbol distance field output by the neural symbol distance field expression step, the observation blind area is geometrically completed in three dimensions, the completed neural symbol distance field is output, and the neural symbol distance field parameters are updated online;
[0045] Curvature-sensitive loss constraint step: A curvature-sensitive generative adversarial loss is introduced during the joint optimization process of the neural symbol distance field expression step and the generative completion step to preferentially preserve local curvature abrupt change points when completing missing geometric surfaces.
[0046] Furthermore, step S3 further includes:
[0047] Differential rendering step: Based on the neural symbol distance field output by the neural symbol distance field expression step, differential rendering is performed using the NeuS model. The image plane color is calculated through the volume rendering equation, where the rendering weight is derived from the symbol distance value, and the predicted color value is output.
[0048] Morphological constraint step: An anatomical morphological constraint term is added to the energy function of joint optimization. The anatomical morphological constraint term is used to prompt the model to automatically close the surface at the missing data based on the average curvature of the blood vessel wall in the reconstructed area, and output the morphological constraint loss value.
[0049] Principles and advantages:
[0050] 1. This invention utilizes the vascular anatomical topological centerline as a hard geometric constraint space, replacing the dependence on inertial navigation (IMU) or image feature points. By constructing a differentiable tubular centerline equation based on spline curves, the probe's six-degree-of-freedom search space is projected onto a local channel centered on the centerline, and the divergence and curl of the dense optical flow field are used to determine motion trends such as forward, backward, or bending. Based on this, a joint optimization equation is constructed, including optical flow reprojection residuals, tube wall collision penalties, and orientation consistency constraints, forcing the probe pose to conform to the bending characteristics of the anatomically flexible lumen. Even in smooth vascular wall regions with extremely sparse texture, this method provides robust, non-intrusive pose estimation, effectively solving the scale drift and pose collapse problems under IMU-less conditions, and providing a reliable spatial reference for subsequent multi-frame fusion.
[0051] 2. This invention employs Neural Symbolic Distance Field (Neural SDF) as the core of geometric representation, naturally possessing the ability to continuously and smoothly represent complex, tortuous vascular surfaces. When the probe moves too quickly or obstacle avoidance results in minimal information acquisition in certain areas, the system does not perform simple physical interpolation. Instead, it uses a generative completion module based on a diffusion model to supplement the blind areas with a continuous vascular wall surface that possesses medical tissue texture and reasonable anatomical structure, according to the vascular orientation and radius variation trend of the reconstructed area. A curvature-sensitive generative adversarial loss term ensures that local curvature abrupt changes are preferentially preserved during the completion process, preventing the generative algorithm from erasing clinically diagnostic microplaque features. The final reconstruction result is not only geometrically closed and anatomically reliable, but also visually presents a high-fidelity subsurface scattering texture, facilitating interactive interpretation by physicians.
[0052] 3. This invention addresses the "flashlight effect" and drastic changes in illumination caused by single-LED follow-up lighting by constructing a point source irradiance model that moves with the probe. Based on a coarse depth map of each frame, inverse brightness compensation is achieved to eliminate dynamic highlights and edge shadows, ensuring that subsequent neural modeling pipelines fuse a reflectance distribution with consistent texture (Albedo), rather than dynamic image intensity mixed with viewpoint bias. Building upon this, a generative spatiotemporal super-resolution network with a recurrent neural structure is employed to extract and "phantomize" high-resolution medical details (such as vascular endothelial cell arrangement and micro-collateral openings) from low-resolution analog video signals, thereby converting analog signals into high-definition images with interpretable value under low-light and high-noise conditions.
[0053] 4. This invention achieves end-to-end processing from raw low-quality analog signals to high-fidelity 3D models through the synergistic effect of image enhancement, topological constraint pose inference, and generative neural implicit reconstruction. Under conditions of 1-3mm microvascular depth, without IMU, and with single-LED servo illumination, this method can stably output geometrically continuous, textured, and anatomically reliable blind-zone completion 3D vascular models. The model surface clearly displays pathological features with clinical interpretation value, such as fine plaques and collateral openings. Compared with existing methods (such as traditional optical flow SLAM, 3D Gaussian splashing, and NeRF), this invention significantly improves trajectory reconstruction accuracy, surface continuity, and the rationality of blind-zone completion, providing intuitive and reliable 3D image data for surgical planning and intraoperative navigation.
[0054] 5. In the extremely limited scenario of microvascular interventional imaging, this invention is the first to systematically integrate physical illumination compensation, topological prior constraints and generative neural implicit reconstruction, achieving interpretation-level high-fidelity three-dimensional model output. This significantly improves the ability of interventional neurologists to intuitively identify intravascular lesions (such as plaques, thrombi, and dissections), and has important clinical value and commercial prospects. Attached Figure Description
[0055] Figure 1 This is a logic block diagram of a high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces, according to an embodiment of the present invention.
[0056] Figure 2 Figure 1 This is a flowchart of the high-fidelity reconstruction method for microvessels based on topological priors and neural implicit surfaces, according to an embodiment of the present invention. Detailed Implementation
[0057] The following detailed description illustrates the specific implementation method:
[0058] Example
[0059] A high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces, basically as follows: Figure 1 As shown, the system includes:
[0060] The hardware acquisition end is used to acquire low-resolution analog video signals within microvascular vessels via a CCD sensor. The CCD sensor includes a miniature endoscope lens and an intravascular imaging device. The data is then cleaned by a signal cleaning module (such as low-sampling-rate electromagnetic noise bandpass filtering and analog noise separation). Finally, the image preprocessing module performs preprocessing, including distortion correction, brightness normalization, and noise suppression.
[0061] An image enhancement pipeline is used to receive the low-resolution analog video signal, establish a point light source irradiance model that moves with the probe for dynamic brightness correction, and enhance the sub-pixel features of the image through a generative spatiotemporal super-resolution network, outputting an enhanced image sequence; the image enhancement pipeline includes:
[0062] The follow-up motion light and shadow inverse modeling unit is used to establish a point light source irradiance model, and perform inverse brightness compensation on each frame of the enhanced image sequence based on the point light source irradiance model to eliminate dynamic highlights and edge shadows caused by probe movement.
[0063] For establishing a point light source irradiance model:
[0064] Suppose the LED point light source is located at the optical center of the camera. For any point P on the blood vessel wall, what is the observed brightness I at the image plane (x,y)? obs It can be modeled as:
[0065]
[0066] Where ρ is the reflectivity of the blood vessel wall, I0 is the initial intensity of the light source, r is the distance from point P to the optical center, and θ is the angle of incidence of the light ray. This represents the gain coefficient during the analog signal transmission process.
[0067] Implementation steps:
[0068] (1) Use the coarse depth map generated from the preceding frame of the low-resolution analog video signal to estimate r and cos(θ);
[0069] (2) Through the formula Perform reverse brightness compensation to eliminate dynamic highlights and edge shadows caused by probe movement.
[0070] A generative spatiotemporal super-resolution network employs a recurrent neural structure to fuse multiple consecutive frames from a low-resolution analog video signal, extracting sub-pixel offsets in the temporal dimension, specifically extracting the subsurface scattering texture features of the blood vessel wall lost due to low resolution. This upscales the low-pixel count to a clarity level suitable for medical diagnosis. Implementation steps:
[0071] (1) Using the minute jitter in the continuous video of the low-resolution analog video signal as implicit sampling, and then using the residual generation network Achieve nearly double the super-resolution mapping.
[0072] (2) Introduce organizational structure consistency item This ensures that the enhanced image can clearly identify the fine vascular endothelial texture at the mm scale, thus obtaining the enhanced image sequence.
[0073] The image enhancement pipeline eliminates the uneven dynamic light intensity distribution caused by movement by establishing a physical photometric compensation model for the source light source of the motion-driven machine, and uses a generative super-resolution network to enhance extremely low logical pixels into high-definition spatiotemporal feature data; thus improving the low pixel level to a clarity level that meets the needs of medical diagnosis.
[0074] A topology-constrained pose inference engine is used to receive the enhanced image sequence, constrain the probe's motion trajectory to the centerline of the deformable lumen topology under inertial guidance-free conditions, and perform inter-frame relative motion trend tracking and pose calculation based on dense optical flow within the local channel constrained by this centerline, outputting inter-frame pose data. The topology-constrained pose inference engine, combined with tubular prior pose calculation, can output the probe's trajectory within the microvascular system. The topology-constrained pose inference engine includes:
[0075] A differentiable tubular centerline modeling unit is used to construct a differentiable tubular centerline equation based on a spline curve, constraining the probe's search space within a local channel centered on the centerline; since blood vessels are continuous spline curves, the probe pose search space is constrained within a local channel based on the spline curve.
[0076] For the equation of the centerline of a differentiable tubular vessel: assuming the vessel is a continuous spline curve C(s), its Frenet-Serret frame is defined as:
[0077]
[0078] Where T is the tangent vector, representing the forward trend of the probe; normal vector N, subnormal vector B.
[0079] The motion trend analysis unit is used to determine the relative motion trend between frames of the probe by analyzing the divergence and curl of the dense optical flow field between adjacent frames in the enhanced image sequence within the local channel. The motion trend includes forward movement, backward movement, or bending.
[0080] The pose calculation and optimization unit is used to calculate the inter-frame relative pose of the probe by minimizing the sum of the optical flow reprojection residual, the tube wall collision penalty term, and the orientation consistency constraint term. The tube wall collision penalty term is used to force the probe to return to the center of the tube cavity, and the orientation consistency constraint term is used to force the probe's line of sight to be consistent with the tangent vector of the equation of the differentiable tubular centerline.
[0081] The formula for calculating the optical flow reprojection residual caused by probe pose changes is as follows:
[0082]
[0083] in, These are two-dimensional pixel coordinates (i.e., the observed feature point positions) that are directly observed on the image through optical flow tracing or feature matching. For camera pose, , It is the homogeneous coordinate of a point in three-dimensional space in the world coordinate system, representing the three-dimensional position of the point.
[0084] The sum of minimizing the optical flow reprojection residual, the tube wall collision penalty term, and the orientation consistency constraint term is constructed as a joint optimization equation:
[0085]
[0086] in, This indicates the penalty for pipe wall collision. To ensure directional consistency, the camera's line of sight is forced. Must be tangent to the lumen Maintain a high cosine similarity.
[0087] In this scheme, the topology-constrained pose inference engine innovatively uses the tubular anatomical topology centerline as a hard geometric constraint space to replace physical sensors. Within the constraint path, it uses dense optical flow to achieve robust inter-frame relative motion trend tracking and pose calculation, thus solving the scale collapse caused by extreme drift.
[0088] A generative neural implicit reconstruction module is used to receive the enhanced image sequence and the inter-frame pose data, and to perform anatomically consistent geometric completion of the observation blind spots using a neural symbolic distance function to generate a high-fidelity three-dimensional vascular model. The generative neural implicit reconstruction module includes:
[0089] The Neural Symbolic Distance Field (NSDF) representation unit is used to determine the coordinates of three-dimensional spatial sampling points through ray sampling based on the enhanced image sequence and the inter-frame pose data. It then uses a multilayer perceptron (MLP) to map each three-dimensional spatial coordinate into a symbolic distance value, outputting geometric information expressed as a neural symbolic distance field. This unit employs a multi-resolution hash grid to encode the three-dimensional coordinates, inputting the encoded feature vector into the MLP and outputting the symbolic distance value. During online reconstruction, the hash grid parameters are updated synchronously with the MLP parameters. For the Neural Symbolic Distance Field (NSDF):
[0090] Using a multilayer perceptron (MLP) θ Map the 3D coordinate x to the signed distance value d:
[0091]
[0092] This constraint (Eikonal Loss) ensures that the geometric surface of the generated blood vessels is continuous and smooth, making it very suitable for describing biological blood vessel walls;
[0093] The generative completion unit is used to receive the neural symbol distance field output by the neural symbol distance field expression unit, and perform three-dimensional geometric completion of the observation blind zone based on the diffusion model, and output the completed neural symbol distance field.
[0094] A curvature-sensitive loss unit is used to introduce a curvature-sensitive generative adversarial loss term during the joint optimization process of the neural symbol distance field expression unit and the generative completion unit. This term is configured to preferentially preserve local curvature abrupt points when completing missing geometric surfaces.
[0095] Differential Rendering Unit: Based on the neural symbol distance field output from the neural symbol distance field expression step, differential rendering is performed using the NeuS model. Image plane color is calculated using the volumetric rendering equation, where the rendering weights are derived from the symbol distance values, and the predicted color value is output. The rendering equation for image plane color C is as follows:
[0096]
[0097] The weight w(h) is derived from the SDF value. h: usually represents the depth or distance of the sampling point along the ray (from the near plane h_near to the far plane h_far). v: usually represents the viewing direction (view direction) vector. c(h, v): the color value (or radiance) at a given depth h and viewing direction v, representing the color contribution of that sampling point along direction v.
[0098] Morphological constraint unit: An anatomical morphological constraint term is added to the energy function of the joint optimization. This term, based on the average curvature of the vessel wall in the reconstructed region, prompts the model to automatically close surfaces at missing data points, outputting a morphological constraint loss value. The anatomical morphological constraint term is as follows:
[0099]
[0100] This prompted the model to calculate the mean curvature of the surrounding blood vessel walls at locations with missing data. The surface is automatically closed to generate a complete three-dimensional blood vessel model.
[0101] The generative neural implicit reconstruction module, based on the completed neural symbol distance field, extracts zero isosurfaces in voxel space using the Marching Cubes algorithm to generate the high-fidelity 3D vascular model. The final output is a high-fidelity 3D vascular model that can be viewed from multiple angles by physicians.
[0102] The generative neural implicit reconstruction module establishes an online generative neural symbolic distance field (Neural SDF) based on hash acceleration. This network not only learns from measured image information, but also performs smooth surface completion of blind areas due to rapid movement or occlusion online based on vascular geometric priors, in accordance with medical logic. The final product is a high-fidelity 3D model with continuous tissue texture, which is conducive to clinical interactive interpretation.
[0103] A high-fidelity reconstruction method for microvessels based on topological priors and neural implicit surfaces, such as Figure 2 As shown, it includes the following steps:
[0104] Step S1: Perform dynamic brightness correction and generative spatiotemporal super-resolution enhancement on the acquired low-resolution analog video signal using a point light source irradiance model that moves with the probe, to obtain an enhanced image sequence; Step S1 specifically includes:
[0105] Step S101: Perform inverse brightness compensation based on the point light source irradiance model to eliminate dynamic highlights and edge shadows;
[0106] Step S102: A generative spatiotemporal super-resolution network with a recurrent neural structure is used to fuse sub-pixel offset information from multiple consecutive frames to achieve super-resolution mapping. A tissue structure consistency term is introduced to ensure the recognition of fine vascular endothelial texture in the enhanced image. This elevates the low-pixel level to a clarity level that meets the needs of medical diagnosis.
[0107] Step S2: Based on the obtained enhanced image sequence, the probe's motion trajectory is constrained to the center line of the deformable lumen topology under inertial guidance-free conditions. Pose trend inference is performed using dense optical flow to calculate the probe's trajectory within the microvascular system, and inter-frame pose data is output. Step S2 specifically includes:
[0108] Step S201: Preset the blood vessel as a continuous spline curve, and constrain the probe pose search space within a local channel based on the spline curve;
[0109] Step S202: Extract dense optical flow between adjacent frames and calculate the optical flow reprojection residual caused by probe pose change;
[0110] Step S203: Construct a joint energy functional that includes optical flow reprojection residuals, tube wall collision penalty terms, and orientation consistency constraints, and solve the inter-frame relative pose of the probe through iterative optimization.
[0111] Step S3: Input the enhanced image sequence and the inter-frame pose data into a generative neural implicit network to update the neural symbol distance field parameters online; Step S3 specifically includes:
[0112] Step S301: Neural symbol distance field expression step: Based on the enhanced image sequence and the inter-frame pose data, the coordinates of the three-dimensional spatial sampling points are determined by ray sampling, and each three-dimensional spatial coordinate is mapped to a symbol distance value using a multilayer perceptron, outputting the neural symbol distance field;
[0113] Step S302: Generative completion step: Based on the diffusion model, according to the neural symbol distance field output by the neural symbol distance field expression step, perform three-dimensional geometric completion on the observation blind area, output the completed neural symbol distance field, and update the neural symbol distance field parameters online;
[0114] Step S303: Curvature-sensitive loss constraint step: In the joint optimization process of the neural symbol distance field expression step and the generative completion step, a curvature-sensitive generative adversarial loss is introduced to preferentially preserve local curvature abrupt change points when completing missing geometric surfaces.
[0115] Step S304: Differential rendering step: Based on the neural symbol distance field output by the neural symbol distance field expression step, differential rendering is performed using the NeuS model. The image plane color is calculated through the volume rendering equation, where the rendering weight is derived from the symbol distance value, and the predicted color value is output.
[0116] Step S305: Morphological constraint step: Add an anatomical morphological constraint term to the energy function of joint optimization. The anatomical morphological constraint term is used to cause the model to automatically close the surface at the missing data based on the average curvature of the blood vessel wall in the reconstructed area, and output the morphological constraint loss value.
[0117] Step S4: Based on the updated neural symbol distance field parameters, extract the zero isosurface through voxel space and output a high-fidelity three-dimensional blood vessel model.
[0118] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, based on the guidance provided in this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces, characterized in that, The system includes: The hardware acquisition unit is used to acquire low-resolution analog video signals within microvessels. The image enhancement pipeline is used to receive the low-resolution analog video signal, establish a point light source irradiance model that moves with the probe for dynamic brightness correction, and enhance the sub-pixel features of the image through a generative spatiotemporal super-resolution network to output the enhanced image sequence. The topology-constrained pose inference engine is used to receive the enhanced image sequence, constrain the probe motion trajectory to the deformable cavity topology centerline under inertial guidance-free conditions, and perform inter-frame relative motion trend tracking and pose calculation based on dense optical flow within the local channel constrained by the centerline, and output inter-frame pose data. The generative neural implicit reconstruction module is used to receive the enhanced image sequence and the inter-frame pose data, and to perform anatomically consistent geometric completion of the observation blind area using the neural symbol distance function to generate a high-fidelity three-dimensional blood vessel model.
2. The high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces according to claim 1, characterized in that, The image enhancement pipeline includes: The follow-up motion light and shadow inverse modeling unit is used to establish a point light source irradiance model, and perform inverse brightness compensation on each frame of the enhanced image sequence based on the point light source irradiance model to eliminate dynamic highlights and edge shadows caused by probe movement. A generative spatiotemporal super-resolution network uses a recurrent neural structure to fuse multiple consecutive frames from the low-resolution analog video signal, extracting the subsurface scattering texture features of the blood vessel wall that are lost due to low resolution.
3. The high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces according to claim 2, characterized in that, The point light source irradiance model is expressed as follows: Where ρ is the reflectivity of the blood vessel wall, I0 is the initial intensity of the light source, r is the distance from point P to the optical center, and θ is the angle of incidence of the light ray. This is the gain coefficient in the analog signal transmission process.
4. The high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces according to claim 1, characterized in that, The topology-constrained pose inference engine includes: Differentiable tubular centerline modeling unit is used to construct a differentiable tubular centerline equation based on spline curves, constraining the probe's search space within a local channel centered on the centerline; The motion trend analysis unit is used to determine the relative motion trend between frames of the probe by analyzing the divergence and curl of the dense optical flow field between adjacent frames in the enhanced image sequence within the local channel. The motion trend includes forward movement, backward movement, or bending. The pose calculation and optimization unit is used to calculate the inter-frame relative pose of the probe by minimizing the sum of the optical flow reprojection residual, the tube wall collision penalty term, and the orientation consistency constraint term. The tube wall collision penalty term is used to force the probe to return to the center of the tube cavity, and the orientation consistency constraint term is used to force the probe's line of sight to be consistent with the tangent vector of the equation of the differentiable tubular centerline.
5. The high-fidelity microvascular reconstruction system based on topological priors and neural implicit surfaces according to claim 1, characterized in that, The generative neural implicit reconstruction module includes: The neural symbolic distance field expression unit is used to determine the coordinates of three-dimensional spatial sampling points by ray sampling based on the enhanced image sequence and the inter-frame pose data, and to map each three-dimensional spatial coordinate into a symbolic distance value using a multilayer perceptron, and output geometric information expressed in the form of a neural symbolic distance field. The generative completion unit is used to receive the neural symbol distance field output by the neural symbol distance field expression unit, and perform three-dimensional geometric completion of the observation blind zone based on the diffusion model, and output the completed neural symbol distance field. A curvature-sensitive loss unit is used to introduce a curvature-sensitive generative adversarial loss term during the joint optimization process of the neural symbol distance field expression unit and the generative completion unit. This term is configured to preferentially preserve local curvature abrupt points when completing missing geometric surfaces. The generative neural implicit reconstruction module extracts the zero isosurface based on the completed neural symbol distance field to generate the high-fidelity three-dimensional blood vessel model.
6. A method for high-fidelity reconstruction of microvessels based on topological priors and neural implicit surfaces, characterized in that, Includes the following steps: Step S1: Perform dynamic brightness correction and generative spatiotemporal super-resolution enhancement on the acquired low-resolution analog video signal using a point light source irradiance model that moves with the probe, to obtain an enhanced image sequence. Step S2: Based on the obtained enhanced image sequence, constrain the probe motion trajectory to the topological center line of the deformable lumen under the condition of no inertial guidance, use dense optical flow to infer the pose trend, solve the probe's running trajectory in the microvascular, and output inter-frame pose data. Step S3: Input the enhanced image sequence and the inter-frame pose data into a generative neural implicit network to perform anatomically consistent geometric completion of the observation blind area and update the neural symbol distance field parameters online; Step S4: Based on the updated neural symbol distance field parameters, extract the zero isosurface through voxel space and output a high-fidelity three-dimensional blood vessel model.
7. The method for high-fidelity reconstruction of microvessels based on topological priors and neural implicit surfaces according to claim 1, characterized in that, Step S1 specifically includes: Inverse brightness compensation is performed based on a point light source irradiance model to eliminate dynamic highlights and edge shadows; A generative spatiotemporal super-resolution network with a recurrent neural structure is used to fuse sub-pixel offset information from multiple consecutive frames to achieve super-resolution mapping. Furthermore, a tissue structure consistency term is introduced to ensure that the enhanced image can identify the microvascular endothelial texture.
8. The method for high-fidelity reconstruction of microvessels based on topological priors and neural implicit surfaces according to claim 1, characterized in that, Step S2 specifically includes: The blood vessels are preset to be continuous spline curves, and the probe pose search space is constrained within a local channel based on the spline curves. Extract dense optical flow between adjacent frames and calculate the optical flow reprojection residual caused by changes in probe pose. A joint energy functional is constructed that includes optical flow reprojection residuals, tube wall collision penalty terms, and orientation consistency constraints. The inter-frame relative pose of the probe is solved by iterative optimization.
9. The method for high-fidelity reconstruction of microvessels based on topological priors and neural implicit surfaces according to claim 1, characterized in that, Step S3 specifically includes: The neural symbol distance field representation steps are as follows: Based on the enhanced image sequence and the inter-frame pose data, the coordinates of the three-dimensional spatial sampling points are determined by ray sampling, and each three-dimensional spatial coordinate is mapped to a symbol distance value using a multilayer perceptron, thus outputting the neural symbol distance field. Generative completion step: Based on the diffusion model, according to the neural symbol distance field output by the neural symbol distance field expression step, the observation blind area is geometrically completed in three dimensions, the completed neural symbol distance field is output, and the neural symbol distance field parameters are updated online; Curvature-sensitive loss constraint step: A curvature-sensitive generative adversarial loss is introduced during the joint optimization process of the neural symbol distance field expression step and the generative completion step to preferentially preserve local curvature abrupt change points when completing missing geometric surfaces.
10. The method for high-fidelity reconstruction of microvessels based on topological priors and neural implicit surfaces according to claim 1, characterized in that, Step S3 further includes: Differential rendering step: Based on the neural symbol distance field output by the neural symbol distance field expression step, differential rendering is performed using the NeuS model. The image plane color is calculated through the volume rendering equation, where the rendering weight is derived from the symbol distance value, and the predicted color value is output. Morphological constraint step: An anatomical morphological constraint term is added to the energy function of joint optimization. The anatomical morphological constraint term is used to prompt the model to automatically close the surface at the missing data based on the average curvature of the blood vessel wall in the reconstructed area, and output the morphological constraint loss value.