Ischemic cerebrovascular disease angiogram image segmentation analysis method

Abstract

This invention relates to a method for segmenting and analyzing angiographic images of ischemic cerebrovascular diseases. It detects abnormal grayscale transitions, structural breaks, and artifact delay signals in angiographic images, extracts negative prior points indicating suspected ischemic areas, aggregates them to form prior abnormal regions, introduces a symmetric perturbation testing mechanism to determine potential occluded or abnormal vascular segments, and dynamically adjusts the segmentation threshold of the region. A local interference window is established within the determined region, and the frequency and rhythmic features of density stripes are extracted. Through the scaling strategy of the convolution kernel, discontinuous vascular segments caused by non-steady-state pulse changes are compensated for segmentation. Texture difference and frequency domain response offset are calculated. If the offset is within a preset physiological tolerance range, an interpolation completion mechanism is triggered to generate a reliable completion layer for subsequent calibration reference. By analyzing the multipath divergence degree of the vascular branch endpoints and the path offset changes in the image sequence, the previously segmented paths are dynamically adjusted and their continuity is calibrated.

Description

Technical Field

[0001] This invention relates to a method for segmenting and analyzing angiographic images, specifically a method for segmenting and analyzing angiographic images of ischemic cerebrovascular diseases. Background Technology

[0002] Based on Chinese patent CN111126403A, a method and system for cerebral vessel segmentation based on magnetic resonance angiography images, current methods for segmenting and analyzing angiography images in ischemic cerebrovascular diseases still suffer from several technical shortcomings and practical limitations, mainly in terms of segmentation accuracy, structural coherence, anomaly recognition ability, temporal adaptability, and result reliability. Firstly, existing methods primarily rely on magnetic resonance angiography (MRA) images for cerebral vessel segmentation. Although they introduce a double Gaussian model to model the grayscale distribution of cerebral vessels and utilize a three-dimensional weighted Markov random field to enhance the local continuity of the image, thus improving the overall morphological fidelity of cerebral vessel segmentation to some extent, their segmentation strategy still relies mainly on static image grayscale statistics, lacking an understanding of the dynamic perfusion process and the evolutionary behavior of time-series images. This static modeling + local smoothing approach is insufficient to address the complex pathological features commonly found in ischemic cerebrovascular diseases, such as perfusion instability, blood flow interruption, and artifact interference. For example, in clinical practice, due to the slow flow rate of contrast agent or the delayed filling of occluded segments of blood vessels, discontinuous gray bands or structural backfilling artifacts often appear in images. If only the dual Gaussian model is used to model the medium and high gray areas, it is very easy to misjudge the slowly perfused but structurally coherent microvessels as background, or mark the delayed artifacts as real pathways, resulting in both false negatives and false positives.

[0003] Secondly, although this invention introduces a three-dimensional Markov random field to attempt to improve the continuity of vascular structures, its core reliance is on the spatial similarity between image pixels and the co-occurrence relationship of local neighborhood gray levels. When faced with fractures, bifurcation, or abnormal extension paths, it lacks the ability to model structural directionality. Ischemic cerebrovascular diseases, especially in the branching regions of small and medium-sized vessels, often manifest as weak capillary perfusion and unstable course. If the segmentation system cannot determine based on the main direction of the vessel or the topological connection trend, it is highly prone to false connections or broken continuations. This invention does not establish a logical feedback mechanism based on the evolution of vascular path structures, and cannot perform historical judgment or score backtracking for abnormal paths, resulting in a lack of effective means to screen out structurally misleading paths. Thirdly, existing methods are basically based on segmentation decisions on single-frame images, lacking an inter-frame path continuity analysis mechanism. For digital subtraction angiography (DSA) or dynamic MRA images, cerebral vessels have obvious temporal evolution trajectories in the image sequence, such as perfusion sequence, flow velocity differences, and terminal arterial perfusion delays. This temporal information can be used to assist in judging the connectivity authenticity of vascular segments. However, the invention does not involve joint modeling between image sequences, nor does it set up any scoring module based on path temporal consistency or inter-frame orientation field consistency. This results in the model being unable to effectively identify non-structural changes such as orientation abrupt changes, transient occlusion, or break recovery caused by short-term artifacts, thereby affecting the stability and reliability of the final segmentation map.

[0004] Fourth, the method in this invention mainly relies on grayscale modeling and Markov neighborhood smoothing to adjust the segmentation boundaries, lacking multimodal feature fusion capabilities. For example, in actual angiography images, in addition to grayscale, there are multiple dimensions of information such as texture details, pulse stripe rhythm, and boundary response intensity that can jointly assist in segmentation judgment. However, this method does not introduce a rhythmic non-steady-state detection mechanism, frequency domain response analysis, or texture direction completion module, and cannot accurately reproduce key clinical manifestations such as microvascular rupture and unstable perfusion, severely limiting its application in pathological brain regions. Furthermore, in terms of confidence expression, this invention does not provide a confidence assessment mechanism for the segmentation results; the output of all paths or vascular segments is a fully trusted structure, which is not conducive to doctors quickly identifying suspicious areas or performing semi-automatic corrections in clinical practice.

[0005] In summary, while existing traditional cerebrovascular segmentation methods based on grayscale modeling and local random field optimization have basic practicality under specific conditions, they exhibit several technical shortcomings when facing disease scenarios such as ischemic cerebrovascular disease, which is characterized by changes in hemodynamics, structural fragmentation and ambiguity, perfusion rhythm abnormalities, and complex temporal evolution. These shortcomings include insufficient segmentation accuracy, incomplete structural analysis, weak anomaly identification ability, and ambiguous result reliability. Consequently, they cannot meet the high-quality requirements for accurate analysis of complex lesions and clinical decision support. Summary of the Invention

[0006] The purpose of this invention is to provide a method for segmenting and analyzing angiographic images of ischemic cerebrovascular diseases, thereby addressing some of the shortcomings and deficiencies pointed out in the background art.

[0007] The present invention addresses the aforementioned technical problems by employing the following technical solution: a method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease, comprising: extracting multiple negative segmentation prior points for indicating suspected ischemic areas by detecting gray-scale transition anomalies, structural breaks, and artifact delay signals in angiographic images; aggregating the prior points to form a prior abnormal region, and introducing a symmetric perturbation test mechanism in the prior abnormal region to analyze the response stability to the segmentation boundary, thereby determining potential occlusion or abnormal vascular segments, and dynamically adjusting the segmentation threshold of the region;

[0008] Within the determination area, a local interference window is established along the blood vessel direction to extract the frequency and rhythm features of density stripes. The discontinuous blood vessel segments caused by non-steady-state pulse changes are compensated and segmented by the scaling strategy of the convolution kernel. For the two ends of the identified broken blood vessel segments, the texture difference and frequency domain response offset are calculated. If the offset is within the preset physiological tolerance range, the interpolation completion mechanism is triggered to generate a reliable completion layer and retain it for subsequent calibration reference.

[0009] By analyzing the degree of multipath divergence of the vascular branch endpoints in the segmentation map and the changes in path offset in the image sequence, the previous segmentation path is dynamically adjusted and its continuity is calibrated.

[0010] Furthermore, the extraction process of the negative segmentation prior points includes establishing a gray-scale and texture contradiction tensor. Regions in the image that simultaneously have gradient breakage features but lack texture continuity are identified as structural pseudo-vascular abnormalities. The determination of the artifact delay signal adopts the inter-frame abnormal residual relationship of the image and performs reverse marking processing on the intensity recovery of static regions in the later stage of the perfusion time image.

[0011] Furthermore, the symmetric perturbation test introduces bidirectional mirror difference calculation, wherein when the symmetry error of the image response before and after the perturbation on both sides of the central axis of the blood vessel exceeds a set threshold, it is identified as an unstable segmentation region; the dynamically adjusted segmentation threshold is updated according to the response confidence field of the perturbation region, and the unstable response region will trigger the threshold to be lowered and the processing range of upstream and downstream blood vessels to be expanded.

[0012] Furthermore, the extraction of density stripe frequencies combines short-time Fourier transform and local frequency clustering to distinguish between real vascular pulse stripes and angiographic artifact stripes; the rhythmic features include spatial and temporal periodicity differences to detect rhythmic non-steady-state deformation of capillary segments caused by perfusion fluctuations, used for microvascular rupture compensation judgment; the scaling strategy of the convolution kernel adopts a learnable deformation kernel, the size of which is dynamically controlled by the rate of change of local frequency energy, to maintain the sensitivity of abnormal region segmentation.

[0013] Furthermore, the texture difference includes directional gradient field inversion calculation, and directional interpolation is performed to complete regions where the main directions of the two vascular segments are inconsistent but the texture is continuous; the judgment of the frequency domain response offset adopts joint matching of phase difference residual map and position offset; the reliable layer output by the completion mechanism also includes a completion confidence level map, in which different levels of completed segments selectively participate in subsequent path correction or manual review.

[0014] Furthermore, the multipath divergence judgment of the vascular branch endpoint adopts inter-frame orientation field difference clustering, and establishes an anomaly priority correction list for unstable endpoint paths; the feedback adjustment path process adopts a recursive path scoring system, which scores and updates the path credibility based on historical path continuity and structural rationality after each segmentation.

[0015] Furthermore, the inter-frame orientation field difference clustering adopts a joint clustering strategy of spatial orientation vector and inter-frame relative rotation angle; the calculation of the orientation field difference adopts a cosine similarity weighting scheme based on principal axis projection to suppress mis-clustering caused by abrupt changes in the orientation field in the vascular intersection region.

[0016] Furthermore, the anomaly priority correction list is sorted in multiple levels according to endpoint direction stability, path accessibility, and neighborhood vessel consistency score; the correction operation for abnormal paths includes minimum interpolation connection of structural differences to the direction of credible paths; in the process of constructing the anomaly correction list, the endpoint evolution trajectory of historical image frames is called to improve the tolerance for misidentification under short-term orientation field distortion conditions.

[0017] Furthermore, the recursive path scoring system introduces a path topology breakpoint penalty function to assign cumulative negative weights to segments in the historical path that exhibit repeated breakage behavior; the path continuity scoring introduces a smoothness index based on the continuity of path curvature to distinguish between physiological curvature and abnormal deviations caused by path misdirection.

[0018] A recursive path scoring system is used to analyze the segmented vascular paths segment by segment. A path topology breakpoint penalty function and a path curvature continuity smoothness index are introduced to jointly form a comprehensive path reliability scoring function, which identifies and optimizes discontinuous or structurally abnormal segmented paths. The core of this system lies in constructing the following comprehensive path scoring function:

[0019]

[0020] in:

[0021] For a path Each path segment After assessing its local curvature continuity and topological fracture history, the overall credibility score is accumulated. A higher score indicates a more stable and reasonable path; Representing a path Comprehensive credibility score, value range ; This represents the total length of the path, used for normalization. This represents the total number of segments into which the path is discretized; Indicates the path number The rate of change of curvature of a segment relative to the previous segment is defined as... , indicating the degree of local curvature jump; Indicates the first The offset of the local tangential angle of a segment from the global main direction reflects whether the segment deviates from the course of the main blood vessel (used to distinguish between physiological curvature and misdirection). The normalized frequency representing the number of times a path segment breaks in historical frames is used as a topological breakpoint penalty factor, with a range of... ; The instability coefficient of the segment boundary near the path segment is estimated from the response variance obtained from the perturbation test. This represents the boundary penalty control coefficient, which is an empirically set adjustment parameter (e.g., 0.5~2).

[0022] The scoring function is not a linear weighted combination, but rather a weighted penalty for unstable regions through exponential decay, exhibiting a greater negative impact in regions of high response volatility; through The proportional form suppresses unstructured bends, preventing misjudgment of physiological vascular tortuosity; [the following is introduced] It represents historical break behavior, not a single static judgment; this function can be used as the optimizer objective or path priority ranking basis to form a recursive update feedback loop.

[0023] The beneficial effects of this invention are as follows: By introducing mechanisms such as gray-scale and texture contradiction tensors, negative segmentation prior points, and artifact delay recognition, the interference of non-structural noise such as false vessels and perfusion residuals on the segmentation model is effectively suppressed, improving the recognition accuracy of occluded areas, broken segments, and small vessels, and significantly reducing the false positive rate and structural misclassification rate. Through modules such as density stripe frequency analysis, rhythmicity detection, and convolution kernel scaling strategies, the system can automatically adapt to image changes under physiological abnormal conditions such as perfusion fluctuations and low flow velocity regions, achieving stable segmentation and compensation of discontinuous vessels and microperfusion pathways, and improving the robustness of the model to actual ischemic states. The directional field difference clustering, abnormal path priority correction mechanism, and recursive path scoring system proposed in this invention can detect and repair topological breakpoints, misleading paths, and directional jump behaviors in segmentation in real time, maintaining the anatomical rationality and coherence of segmented vascular structures, and improving the credibility of atlas construction.

[0024] By invoking historical trajectories from multiple image frames and analyzing endpoint evolution trends, the system possesses the ability to handle short-term angiographic disturbances, unstable boundaries, and misidentification of fractures, forming a time-robust segmentation logic that effectively enhances the processing capability of dynamic angiographic image sequences. During interpolation and completion, the system automatically outputs completion confidence layers and distinguishes and marks completion segments of different levels, providing clinicians with visual references for high, medium, and low confidence paths. This facilitates a collaborative mechanism of "automatic segmentation + manual correction," improving the medical interpretability and clinical controllability of the segmentation results. The vascular structures generated by this method are not only highly accurate and continuous but can also serve as foundational data for subsequent high-level analyses such as brain perfusion mapping, occlusion length measurement, and collateral circulation assessment, providing quantitative support for clinical practice and significantly improving diagnostic efficiency and the scientific rigor of treatment decisions. Attached Figure Description

[0025] Figure 1 This is a simplified flowchart of ischemic cerebral angiography segmentation analysis according to the present invention.

[0026] Figure 2 This is a simplified functional relationship diagram of the intelligent segmentation of ischemic cerebrovascular system according to the present invention.

[0027] Figure 3 This is a diagram illustrating the adaptive correction mechanism for vessel endpoint segmentation and path scoring in this invention.

[0028] Figure 4 This is a flowchart of the DSA cerebral blood vessel segmentation example processing and confidence level stratification in Embodiment 1 of the present invention.

[0029] Figure 5 This is a flowchart of the branch endpoint anomaly identification and recursive correction-trusted path screening in Embodiment 2 of the present invention. Detailed Implementation

[0030] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0031] Combined with appendix Figure 1 This invention provides a method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease, which improves the accuracy of vascular segmentation and the ability to identify suspected ischemic areas by addressing the image structure abnormalities caused by occlusion, stenosis, and blood flow delay in cerebral angiography images. The raw cerebral angiography images were preprocessed, and a differential detection method based on regional grayscale distribution was used to identify pixel regions with abnormal grayscale transitions in the entire image. These abnormalities mainly manifested as abrupt brightness transitions within vascular structures, local discontinuous bright-dark boundaries, or signal inversions not supported by anatomical structures. Simultaneously, a structural consistency detection operator was used to scan the vascular branch paths of the entire image, marking pixel segments with structural breaks or abrupt terminations in the topological paths. Combined with high-brightness artifact delay signal regions that did not conform to physiological perfusion patterns in the image sequence, a candidate pixel set constituting negative segmentation prior points was comprehensively extracted. This pixel set represents a set of points in the image with a high probability of being non-real blood vessels or suspected lesions. Subsequently, the extracted prior points were combined through spatial aggregation and semantic normalization into a set of prior anomalous regions with morphological continuity and grayscale heterogeneity, forming the core region for subsequent intervention analysis. Within this prior anomalous region, a negative segmentation prior was introduced. This perturbation testing mechanism involves applying image perturbation factors based on mirroring, rotation, or scale perturbation in the lateral or radial direction. By comparing the response output of the segmentation model for the same boundary region before and after perturbation, the sensitivity of the boundary segmentation result to perturbation is evaluated, thereby calculating a boundary response stability index. This index quantifies the confidence fluctuation of the vascular structure in the model. Based on the stability analysis results, it is further determined whether there are occluded vascular segments, areas with extremely low flow velocities, or high-uncertainty structures that the model cannot identify within the prior abnormal region. If a potential abnormal vascular structure is identified, the segmentation threshold strategy for the corresponding region is dynamically adjusted. By appropriately reducing the model's response intensity requirement to boundary features within the region, the region has stronger fault tolerance and structural recovery possibility in subsequent segmentation processes. Ultimately, this achieves early identification and improved segmentation quality of ischemia-related regions in cerebral angiography images, providing a high-precision image analysis foundation for the subsequent diagnosis of ischemic cerebrovascular diseases.

[0032] For suspected abnormal regions identified by the anomaly detection mechanism, a local interference window is further established within these regions along the local orientation of the blood vessel. This window is constructed using an approximate axis obtained from a blood vessel centerline extraction algorithm for directional guidance, and its angle and length are adjusted in real-time based on the tangent direction at points along this axis. This ensures the window conforms spatially to the local contours of the blood vessel structure, avoiding analysis errors caused by directional deviations. Within the interference window, the density stripe distribution features of the blood vessel pixel region in the main direction are extracted. The local gray-level frequency changes in this direction are calculated using a sliding window method, and a short-time Fourier transform is applied. The Specialized Transform-Thermal Flip-Flattening (STFT) and Local Peak Detection (LTD) algorithms identify the dominant frequency and rhythmic fluctuation characteristics of stripes. Frequency represents the spatial oscillations caused by the discontinuity of vascular structures, while rhythmicity reflects the non-uniformity of contrast agent perfusion under blood flow conditions. Based on the joint frequency-rhythm feature results, an adaptive convolution kernel scaling strategy is triggered. A dynamic kernel scale control mechanism is introduced into the convolution segmentation module, adjusting the width of the convolution receptive field according to the frequency of the extracted stripes. If high-frequency discontinuous stripes are detected, the kernel scale is expanded to enhance cross-regional perception; if low-frequency stable stripes are detected, the kernel scale is contracted to preserve edge accuracy. This strategy achieves response to non-stationary... Spatial compensation segmentation of discontinuous vascular segments caused by impulsive changes in microvascular perfusion is used to effectively alleviate the problem of missegmentation of structural interruptions caused by insufficient microvascular perfusion. After initial compensation, the texture difference between the two ends of the identified broken vascular segment is calculated. The texture difference includes multiple dimensions such as local texture direction gradient, mean gray-level distribution, and texture correlation matrix features. At the same time, the frequency domain response offset of the two ends is compared, and phase correlation and spectral residual analysis methods are used to determine whether the two have a continuous trend in the frequency domain. If both the texture and frequency offset are within the preset physiological tolerance range, the tolerance value is determined based on the actual clinical situation of cerebral blood vessels. Statistical parameters for the bed structure are set, including indicators such as vessel diameter range, branch angle tolerance, and perfusion cycle. This triggers the interpolation completion mechanism, which uses Bezier direction interpolation, gradient-driven refinement, and local reconstruction strategies to complete the vascular structure in the intermediate fracture area. At the same time, a reliable completion layer is generated. This layer is preserved as an independent structural mask and marked with the completion confidence level. It does not directly participate in the structural output of the original master segmentation result, but serves as reference information for subsequent path continuity correction, vessel integrity scoring, and visualization analysis. This achieves functional repair and enhanced clinical interpretability of abnormal areas while ensuring segmentation accuracy.

[0033] To further improve the accuracy of vascular structure integrity assessment and the physiological consistency of segmentation paths, after initial vascular atlas segmentation, multi-path divergence analysis and image sequence path drift detection are performed on the vascular branch endpoint regions in the segmentation map. Vascular branch endpoints refer to all non-connected terminal pixel nodes in the segmented vascular structure; their spatial distribution often corresponds to distal capillaries, occlusion edges, or fake branch segments in the brain. To identify whether these endpoints are genuine vascular terminal structures or broken endpoints caused by segmentation errors, a path divergence map is constructed around each endpoint. The divergence map is back-analyzed to analyze all path trajectories from the endpoint to its superior vascular trunk, and the number of paths, directional differences, and branch angle changes are evaluated. If multiple sets of non-parallel paths with significant angular spans appear, and the endpoints of these sets cluster in the same region, the region is considered to have high divergence characteristics, indicating structural discontinuity or broken pseudo-branches. Simultaneously, the position of the endpoint in the image time series is changed... Using spatiotemporal analysis as an auxiliary criterion for judgment, multi-frame dynamic image data from cerebral angiography is used to track the spatial offset path of the endpoint in consecutive frames. The inter-frame drift trajectory is obtained through point tracking algorithm and optical flow consistency calculation, and the deviation is compared with the segmentation path of the current frame. If the endpoint position is found to have a significant directional drift over time and is inconsistent with the trend of known physiological and anatomical structures, it is further determined that there is a risk of segmentation path drift or discontinuity in the region. Based on the above spatiotemporal joint judgment, a dynamic feedback adjustment mechanism is constructed to re-score the confidence of the path generated by the early segmentation model and trigger a continuity calibration process. The calibration method includes introducing candidate paths into a probabilistic inversion search framework, selecting the path with stable structural orientation, high gray-level consistency and the greatest conformity with the adjacent vascular structure from the multiple path possibilities to replace the original segmentation path, thereby realizing path completion or retrospective repair. At the same time, all adjustment information is retained as a structural integrity evolution layer for model fine-tuning and subsequent analysis.

[0034] Combined with appendix Figure 2To improve the specificity and robustness of abnormal structure detection, a gray-level and texture contradiction tensor construction mechanism is introduced during the extraction of negative segmentation prior points. The tensor is used to measure the non-cooperative features between brightness variation and structural continuity in a region of an image. Local gradient field calculations and texture pattern modeling are performed on the original cerebral angiography image. The gray-level gradient amplitude along the main direction in the pixel neighborhood and the texture consistency index of adjacent pixels are obtained respectively. A contradiction mapping matrix is ​​established through tensor combination. When a pixel region in the image simultaneously exhibits significant gradient breakage (i.e., a sudden increase or abrupt change in edge gradient), and the texture direction consistency within this region is poor and there is no obvious periodic texture continuation trend, it is marked as a structural pseudo-vascular abnormality point. These points usually correspond to vascular occlusion ends, contrast agent diffusion areas, or venous residual artifact areas, and cannot constitute real vascular pathways. They are easily misidentified as small vascular branches in the segmentation model. Therefore, they are prioritized as exclusion points when constructing negative segmentation priors. The exclusion region is marked; at the same time, in order to enhance the ability to identify perfusion delay signals, an inter-frame residual abnormality judgment mechanism is further introduced between image frame sequences. That is, using the temporal data of multiple frames of images during cerebral angiography, a residual signal spectrum between consecutive frames is constructed. The residual calculation is based on pixel-level subtraction combined with local motion compensation to filter out grayscale changes caused by natural blood flow changes, and only retain the intensity recovery phenomenon of abnormal static areas in the time dimension. In particular, attention is paid to static tissues with strong reflective areas in frames in the later stages of perfusion time (such as more than 5 seconds after angiography). By forming an inverse marker mask with the difference of the earlier frames, the area covered by the mask is identified as perfusion delay artifact signal, which will be regarded as a low confidence area in subsequent segmentation processing, thereby further strengthening the model's ability to exclude non-vascular abnormal areas, and improving the overall accuracy of the segmentation model in identifying structural abnormalities and artifact interference in angiography images of ischemic cerebrovascular disease and the clarity of segmentation discrimination boundaries.

[0035] To further enhance the robustness of segmentation algorithms in identifying structurally uncertain regions, a symmetric perturbation testing strategy based on image perturbation and symmetry error feedback is proposed. This strategy performs perturbation testing on prior anomalous regions identified in the initial segmentation results, applying bidirectional mirror perturbation to both sides of the vascular structure's central axis while simultaneously recording changes in edge extraction, texture distribution, and segmentation contours before and after perturbation. To quantify the impact of this perturbation on the model response, a bidirectional mirror difference index is constructed. This index compares the structural response differences between the left and right sides of the vascular central axis under perturbation conditions, calculates the symmetry error coefficient, and compares it with a set threshold. When the symmetry error of a region exceeds the preset threshold (based on statistical results of normal vascular symmetry fluctuation range), the region is identified as an unstable segmentation boundary response region. Such regions are commonly found in areas with uneven contrast agent perfusion, abrupt narrowing of vascular branches, or proximity to occluded segments. Traditional models exhibit high errors in boundary identification for these regions. The system dynamically updates the segmentation threshold based on the response confidence field of the perturbed region after identifying unstable regions. The confidence field is a spatial distribution map composed of the confidence values ​​output by the model for the segmentation response of each pixel under perturbed conditions. By analyzing the variance, distribution density, and edge steepness of the built-in confidence in unstable regions, the system comprehensively judges the reliability of the model's judgment in these regions. When the confidence drops significantly or extreme fluctuations occur, the threshold adjustment module is triggered to lower the segmentation threshold for these regions, thereby reducing the rigidity of the model's boundary consistency requirements and allowing for more tolerant boundary fitting results. At the same time, to ensure consistent processing logic for upstream and downstream vascular segments in these regions, the system further expands the processing range along the vascular flow direction, including adjacent vascular segments in the adjustment range. By strengthening the rationality of local reasoning through continuous contextual information, the system prevents chain error propagation caused by local threshold adjustments, which leads to errors in the identification of the entire structure. Ultimately, this achieves dynamic robust correction and enhanced coherence of the segmentation path in complex lesion regions.

[0036] To address the issues of artifact interference, frequency misjudgment, and segmentation instability caused by rhythmic changes under perfusion fluctuations in angiography of microvascular strips, a vascular strip feature recognition mechanism integrating frequency analysis and rhythm modeling is proposed. The extraction of density stripe frequencies employs a joint analysis strategy of Short-Time Fourier Transform (STFT) and local frequency clustering. First, local frequency energy spectra are extracted from the angiography image along the vessel orientation using a sliding window approach. The frequency distribution and dominant frequency amplitude within each window are obtained using the STFT method. Based on this, a clustering algorithm is introduced to locally group all frequency response regions, aggregating regions with similar frequency characteristics and grayscale stripe morphology into one class. This effectively distinguishes between real vascular pulse stripes and angiography artifact stripes. Real vascular pulse stripes typically exhibit rhythmic enhancement, low-frequency fluctuations, and high spatial coherence, while angiography artifacts often have high-frequency isolated features or are distributed along non-vascular directions. Frequency clustering boundary determination effectively eliminates interfering stripe regions. Further rhythmic features are extracted based on these frequencies, encompassing periodic differences in both spatial and temporal dimensions. The method involves frequency and rhythm analysis, specifically calculating the consistency of local strip spacing changes in static images to determine spatial periodicity, and evaluating the periodic trend of vascular pixel grayscale changes over time in dynamic angiography frames. Temporal spectrum analysis combined with moving mean difference is used to assess the non-steady-state rhythmic deformation of capillary segments caused by perfusion fluctuations, thereby enabling the identification and compensation of microvascular ruptures or near-occlusion areas. Guided by these frequency and rhythmic features, a convolutional kernel scaling strategy suitable for non-steady-state structures is proposed. A learnable deformable convolutional kernel replaces the traditional fixed-shape kernel structure. The kernel size and deformation direction are controlled by the rate of change of local frequency energy. When frequency energy changes rapidly and texture perturbations are severe, the convolutional kernel increases its receptive field to enhance the overall perception of irregular areas; while in areas with stable texture and clear structure, the kernel range is compressed to improve edge recognition resolution. This mechanism maintains the model's response sensitivity in abnormal areas through frequency domain sensitivity adjustment, effectively improving the detection capability and segmentation accuracy of micro-ruptured vascular segments in angiography images, providing more stable, continuous, and reliable vascular structure support for subsequent ischemic lesion identification.

[0037] To achieve accurate structural completion of broken or blurred vascular segments, a completion mechanism integrating directional texture difference analysis and frequency domain offset matching is proposed. The texture difference analysis process relies not only on grayscale or intensity contrast but also introduces a directional gradient field inversion calculation method. Specifically, local gradient vector fields are established for the regions at both ends of the vascular break. The main direction vector is obtained by calculating the consistency between edge response and texture orientation, and the main directions at both ends are inverted and mapped. When there is a deviation in the directions at both ends but the regional texture features remain highly similar and there is a potential connectivity trend between the structures, the system performs structural completion using a directional interpolation method. During directional interpolation, a Bezier curve-based or minimum curvature path interpolation strategy is used to smoothly transition the direction vector, ensuring that the interpolation path has anatomical structural rationality and texture continuity in the image. To further verify whether the two structures belong to the same connected vessel, a frequency domain response offset judgment mechanism is introduced before completion. This is achieved by constructing a spectral response map and calculating the phase difference residual map at both ends of the break, combined with the position offset assessment. The spatial difference between the frequency energy centers of the two regions is used to capture the relative shift of the frequency response in direction and intensity using a phase difference residual map. If the residual is within a preset physiological threshold and the positional shift does not exceed the defined tolerance for branch connections (set according to the statistical model of vessel radius and branch angle), the fracture segment is considered to have a reliable basis for completion. After the completion is completed, the system will generate an output layer of the reliable completion process. The layer not only contains the spatial structure information of the completed segment, but also outputs a completion confidence rating map to classify and label the connection segments generated in different completion processes. The confidence level is based on a comprehensive score of multiple indicators such as directional consistency, texture similarity, frequency domain residual intensity, and spatial jump amplitude, and is divided into three levels: high confidence, medium confidence, and low confidence. High confidence segments can directly participate in the subsequent path tracing and correction process, medium confidence segments will serve as auxiliary structures to provide alternative path schemes, and low confidence segments will enter the manual review module by default, where clinical operators will confirm, correct, or remove them based on anatomical knowledge and visual judgment.

[0038] To effectively improve the identification and correction capabilities for misjudgments, ambiguous occlusions, or abnormal extension paths in cerebral vascular terminal regions, a dynamic path feedback mechanism based on multi-path divergence analysis is proposed. The multi-path divergence judgment of vascular branch endpoints first employs inter-frame orientation field difference clustering technology. Specifically, in the angiographic image sequence, the local orientation field of vascular pixels in each frame is extracted. The orientation field acquisition is based on structural tensor analysis and gradient principal direction estimation. The orientation vectors of the endpoint region in several frames are temporally superimposed to form a distribution map of the endpoint orientation trajectory over time. Based on this, density peak clustering or K-means clustering methods are used to cluster the orientation change patterns, identifying whether multi-path orientation splitting or instantaneous deflection behavior exists in the endpoint region. If a endpoint exhibits high difference, frequent orientation jumps, or lack of dominant orientation stability in the inter-frame orientation field, the path is marked as an unstable path. The system establishes an anomaly priority correction list accordingly, prioritizing such endpoints and their connecting paths for subsequent path scoring and structural repair. Subsequently, the feedback adjusts the path... During the process, a recursive path scoring system is used to dynamically evaluate the continuity and rationality of each segmentation path. After each vascular map segmentation, the historical continuity parameters of each path are recorded based on the established vascular topology, including multiple feature dimensions such as branch connectivity, path curvature angle, local grayscale consistency, and upstream and downstream structural integrity. A basic confidence score is assigned to each path, and this score is weighted and fused with the path scores from previous rounds to form a temporal scoring vector. The confidence level of each path in the structural map is recursively updated. When a path becomes discontinuous, deviates from the physiological and anatomical trend, or connects to an abnormal endpoint due to the current segmentation result, its score will decrease. If it falls below the set confidence threshold, the path backtracking module is triggered. It combines the unstable segments in the anomaly priority correction list to search for and reconstruct alternative paths, and re-inserts them into the topology map. Finally, a dynamically updated segmentation path map is formed, thereby improving the overall segmentation's fault tolerance, structural recovery, and topological rationality in the lesion region. This achieves the enhancement of continuity of terminal branch structures and the repair of pseudo-paths in angiographic images of ischemic cerebrovascular diseases.

[0039] Combined with appendix Figure 3To accurately identify the path divergence behavior of vascular endpoint regions in multi-frame angiography images and reduce orientation determination errors caused by vascular intersections or branch overlaps, an improved inter-frame orientation field differential clustering strategy is proposed. This strategy comprehensively considers both spatial orientation vectors and inter-frame relative rotation angles to construct a joint clustering feature space, enhancing the ability to describe the dynamic behavior of the terminal path. In each frame of angiography image, spatial orientation vectors of vascular region pixels are extracted using gradient principal direction estimation and local structure tensor calculation methods to form a single-frame orientation field map. Subsequently, the orientation vectors at corresponding positions in consecutive frames are compared, and the angle changes between inter-frame orientation vectors are calculated to obtain an inter-frame relative rotation angle sequence, which measures the orientation stability of the region in the time dimension. The spatial orientation vectors and inter-frame rotation angles are used as two-dimensional feature inputs, and clustering operations are performed using density clustering or K-means algorithms. This method identifies clusters of endpoint paths that exhibit consistent or divergent directional behavior across multiple image frames. To mitigate the interference of directional jumps caused by vascular junctions during clustering, a cosine similarity weighting scheme based on principal axis projection is introduced into the calculation of directional field differences. Specifically, the orientation vector of each pixel is projected onto the local principal axis of its respective vascular branch to obtain the local principal orientation component. The cosine similarity value between this component and the standard orientation is then calculated as the confidence weight of the pixel's orientation vector in the clustering. This approach enhances the clustering contribution along the consistent direction of the main trunk, while assigning lower weights to directional reversals or abrupt changes in structural regions such as lateral branches and junctions. This effectively alleviates the problem of mis-clustering of local abnormal orientations caused by drastic changes in vascular orientation, ensuring stable evaluation and reliable labeling of endpoint path divergence, and providing high-quality basic data support for subsequent path scoring and abnormal branch backtracking.

[0040] To further improve the stability and accuracy of abnormal endpoint path identification and correction, a mechanism for constructing an anomaly priority correction list and repairing paths based on multi-factor comprehensive scoring is proposed. This mechanism includes two stages: scoring and ranking, and structural connectivity repair, for vascular endpoint paths marked as directionally unstable or exhibiting divergent mutations in the segmentation results. The anomaly priority correction list is constructed using a multi-dimensional quantitative feature fusion approach. First, the endpoint directional stability score is calculated based on the inter-frame directional vector variability, evaluating the temporal consistency and principal direction shift of the endpoint's directional field across multiple consecutive angiographic frames; smaller directional fluctuations result in a higher score. Second, a path accessibility index is added, assessing the connectivity of the path from the main artery or proximal supply vessel to the endpoint; a high accessibility rate indicates the path has physiological blood supply rationality. Third, the consistency of neighboring vessels is evaluated, analyzing the similarity of other vessel segments in the vicinity of the endpoint in terms of texture direction, grayscale distribution, and topology to determine whether the endpoint is located within a real vascular branching system. Finally, a comprehensive score is established based on these three indicators, and all abnormal paths are ranked in multiple levels to form a correction priority. In the actual repair process, the system performs a minimum structural difference interpolation connection method for the top-ranked abnormal paths. Specifically, it searches for candidate connection segments with the minimum structural difference (including orientation consistency, texture similarity, and topological feasibility) in the adjacent reliable path region from the fracture endpoint. Then, it uses Bezier curve interpolation or the local curvature optimal path method to complete the structural completion connection, ensuring the naturalness and physiological rationality of the path repair. In addition, to prevent misidentification due to short-term image blurring, abnormal blood perfusion, or local signal-to-noise ratio decline, a historical image frame endpoint evolution trajectory analysis mechanism is introduced during the construction of the abnormal correction list. By calling the spatial movement trajectory and orientation change curve of the endpoint in the image sequence before and after multiple frames, combined with the trend stability estimation model under the time sliding window, the abnormal endpoints with large score fluctuations are given fault tolerance retention. This avoids mismarking the real vascular endpoint as an abnormal path under short-term orientation field distortion conditions, thereby improving the robustness of the system for boundary region path identification and the stability of correction processing. Finally, it achieves high-reliability segmentation and adaptive error correction of complex vascular endpoint paths in ischemic cerebrovascular disease angiography images.

[0041] To achieve refined judgment of the structural continuity, topological stability, and physiological rationality of segmented vascular paths, a recursive path scoring system is proposed. Its core lies in introducing a path topological breakpoint penalty function and a path curvature continuity smoothness index to perform segment-by-segment historical analysis and multi-dimensional evaluation of each segmented path, establishing a unified path comprehensive reliability scoring function. This allows for dynamic adjustment of path priorities in the case of multiple path candidates, or the elimination and correction of unreasonable paths. After each vascular map update, the recursive path scoring system updates the scores of all current path segments. The system records whether each path exhibits topological breakage behavior in historical image frames and records the breakage frequency through path segment number concatenation, forming a topological breakpoint penalty term. Simultaneously, the system calculates the rate of curvature change and principal direction offset between each path segment, constructing a smoothness index reflecting path curvature stability to distinguish between physiological curvature (such as normal vascular bends) and non-structural offsets caused by segmentation misdirection or artifact interference. Based on the above structural characteristics, the following path comprehensive reliability scoring function is proposed:

[0042]

[0043] Among them: for path Each fragment on The degree of local curvature jump was evaluated respectively. and the angular offset from the global main direction. It is used to measure whether a path has a sudden deviance or misguided trend; This represents the normalized frequency of path segment breaks in historical frames, used as a topological breakpoint penalty factor, with a value range of [value range missing]. Path segments that break frequently will be significantly downweighted; The boundary response instability coefficient exhibited by this path segment in the disturbance test is estimated using the disturbance-guided response variance. The boundary penalty control coefficient, set empirically, is used to control the influence range of the unstable region, and its value is typically taken as... between; The total path length is used for score normalization. The number of path segments is used; the design feature of this scoring function is the use of an exponential penalty strategy, that is, for unstable boundary regions... use The form is used to implement sharp decay control, thereby preventing unstable regions from dominating the score, while utilizing... The proportion of non-physiological mutations is suppressed to ensure that real vascular tortuosity is not misjudged as abnormal behavior; in addition, a fracture behavior record is constructed from historical frames. This method breaks through the limitations of traditional segmentation based on static structure judgment in a single frame, and realizes the cumulative scoring of path credibility in the time dimension. This function can be used as the basis for path candidate ranking or as an optimization objective function to participate in the feedback loop of the segmentation model, realize the dynamic updating and optimization of the continuity and topological reliability of blood vessel paths, and improve the stability and medical interpretability of segmentation paths in complex lesion areas.

[0044] Example 1:

[0045] Combined with appendix Figure 4In this embodiment, image data was acquired from a 65-year-old male patient who presented with sudden onset of left-sided limb weakness accompanied by speech impairment. A preliminary diagnosis suggested a possible occlusion of the M2 segment of the right middle cerebral artery. Digital subtraction angiography (DSA) was used to continuously acquire 14 frames at a frame rate of 2 frames per second, with an image resolution of 512×512 pixels. The image time coverage spanned from the start of contrast agent injection to 10 seconds later. Frames 5 through 10 represent the primary perfusion change phase. Firstly, a grayscale and texture inconsistency tensor was constructed for the entire image frame. During tensor calculation, a 3×3 pixel neighborhood was selected for gradient calculation, and the first and second-order gradient values ​​were obtained. A texture similarity matrix was then constructed based on the image after grayscale histogram equalization. The gradient breakage threshold was set at locations where the gradient change rate was greater than 40 grayscale units / pixel, and the texture consistency threshold was set to a local structural similarity index (SSIM) < 0.3. Through tensor cross-matching, the right insular region in the image... A region of approximately 28 consecutive pixels was detected that simultaneously met both of the above conditions, exhibiting abrupt changes in the vascular structure at the grayscale edge but a loss of texture direction connectivity. This region was identified as a suspected pseudo-vascular anomaly. Furthermore, its spatial location precisely corresponds to the course of segment M2 but lacks downstream perfusion filling, further suggesting that this segment is not a real open vessel but a false boundary caused by the enhancement of adjacent structures. This negative segmentation prior point was assigned a low-confidence weight label to prevent missegmentation as a real branch. Simultaneously, in terms of inter-frame delay signal detection, frames 10 and 6 were compared... Pixel-level differential analysis was used to construct inter-frame residual images. A difference threshold of 15 gray units was set within the dynamic window. In the lateral venous sinus region of the cerebral hemisphere, a region was detected where the brightness significantly increased in frame 10, while it showed low signal in previous frames. The gray level of this region jumped from 89 in frame 6 to 121 in frame 10, a difference of 32, far exceeding the threshold. Furthermore, this region did not appear as a main vascular axis connected to the arterial structure in the image, and was therefore identified as a perfusion delay artifact. Inverse labeling was then applied to subsequent segmentation images, assigning the label "low-confidence edge region." Additionally... The comparison results showed that if this area is not processed, it will be misidentified as a newly developed capillary dilatation area in the traditional U-Net model, resulting in false extension of the segmentation map. However, after applying the method of this invention, the segmentation model adjusts the boundary position under the constraints of negative prior points and delay artifact masks. The final segmented M2 segmentation occlusion path shows clear upstream vessels, smooth transition of the interruption point edge, and no misjudgment of downstream filling. The model confidence is improved by 12.6%, and the false positive rate is reduced by about 19.3%, providing a more accurate image basis and structural support for clinical judgment of the perfusion range of ischemic areas.

[0046] For the previously identified suspected M2 segment occlusion region, to further confirm its boundary stability and reduce the risk of misclassification of discontinuous vessels by the model, the system established a central axis around the vessel segment. Bidirectional mirror image segments were generated on both sides of the central axis and fed into the perturbation sensitivity calculation module to detect the morphological response differences of the segmentation boundaries on the left and right sides before and after perturbation. In the patient's 7th frame image, the perturbation edge on the left side of the central axis remained consistent, but the right side showed a morphological shift of approximately 8 pixels after perturbation. The calculated symmetry difference value exceeded the system's set error threshold of 7.5, thus this region was determined to be an unstable segmentation region. The system immediately triggered a dynamic threshold adjustment mechanism, lowering the model's boundary sensitivity threshold in this region. This allowed the segmentation model to accept more regions with blurred boundaries, avoiding vessel rupture due to a conservative high-threshold strategy. Simultaneously, the upstream and downstream processing range was expanded by 5 mm along the axis to improve the overall connectivity of the region. Subsequently, to address the issue of unclear strip identification caused by structural ambiguity in this region, the system invoked short-time Fourier transform to extract local frequencies of the density stripes of blood vessels in this segment of the image. In the original image, this segment exhibited obvious non-periodic high-frequency artifacts, and the region was excluded from the main structure because it did not match the frequency group of the main blood vessel in frequency clustering. However, the main frequency response of its upstream neighboring region was stable at 0.12Hz and consistent in direction. The rhythmicity detection module, through spatial periodicity difference and temporal gray-scale fluctuation analysis, discovered that the microcapillary region contained a low-frequency slowly varying structure that conformed to the physiological perfusion fluctuation rhythm. This structure was identified as a real but slowly perfused capillary pathway. Due to its structural fractures but strong rhythmicity, the system triggered a non-steady-state compensation mechanism, invoking a learnable deformable convolution kernel to stretch the originally too small kernel scale to a 5×9 receptive field, effectively encompassing the unstable perfusion blood vessel segment and significantly improving the integrity of the segmentation results at the boundary. Subsequently, texture interpolation calculations were performed on the obvious fracture structure in the region. Although the directional difference between the two ends reached 16 degrees, the similarity of the texture details was higher than 0.82. After directional gradient field inversion analysis, its interpolation direction curve was derived. The system completed the directional interpolation connection of the blood vessel segment and judged the connection of the segment to be reasonable by combining the frequency domain response offset and phase residual map. The position offset was about 6 pixels, and the residual map was smooth, verifying the potential connectivity of the structure. Therefore, this segment was included in the segmentation layer as a reliable completion structure. To ensure the controllability of the completion part in subsequent analysis, the system marked its confidence level as "medium confidence" and assigned it limited weight in the path optimization stage, only participating in path fitting when the confidence of the main path is insufficient.In this segmentation image, another vessel endpoint near the right basal ganglia showed an unstable extension trend. The system detected significant directional changes in this endpoint across multiple frame orientation fields and identified it as an unstable endpoint path after directional clustering, adding it to the anomaly priority correction list. Subsequently, the recursive path scoring system retrieved historical path records from the previous 6 frames and found that this path segment exhibited 3 breaks in past images, resulting in a low continuity score. Furthermore, its connection structure angle deviated from the main trunk direction by approximately 27 degrees, indicating a potentially misleading path. The system immediately triggered a correction mechanism, rematching adjacent structures with high accessibility and consistent with the main trunk direction, and replacing this path segment. In the final segmentation structure, the occlusion location of the patient's right M2 segment vessel was clearly marked, and surrounding microvascular segments were appropriately completed. The overall confidence map output showed that the proportion of high-confidence branches increased to 87%, and the low-confidence area was controlled to less than 6%. The system determined the degree of ischemic blood supply interruption to be moderate to severe, providing accurate data support for subsequent perfusion analysis and thrombolysis strategies.

[0047] Example 2:

[0048] Combined with appendix Figure 5In the aforementioned analysis of digital subtraction angiography of a 65-year-old male patient with ischemic cerebrovascular disease, the system has completed the structural segmentation, texture completion, and orientation interpolation of the suspected occlusion area in the M2 segment. In order to further improve the stability identification and abnormal path correction capabilities of the segmentation endpoint region, the method of this invention introduces an inter-frame orientation field differential clustering strategy in the structural identification of the vascular branch endpoint. By jointly analyzing the spatial orientation vector and the inter-frame relative rotation angle, the path divergence region is accurately located and an abnormal correction sequence is established to prevent pseudo-branches from misleading the segmentation results. In the patient's image sequence from frames 6 to 11, the system selected a small terminal blood vessel in the right basal ganglia region terminating in frame 9 as the analysis target. First, in each frame, the system used the structural tensor method to extract the principal direction vector of each pixel near the endpoint of this path, forming five sets of spatial direction fields. The principal direction vector remained stable between 48 and 52 degrees from frames 6 to 8, but abruptly increased to 74 degrees in frame 9, and then returned to 50 degrees in frame 10. Calculations showed that the endpoint direction exhibited a rotation angle fluctuation of up to 26 degrees between adjacent frames. The system jointly constructed a two-dimensional feature space of direction and rotation using the five frames of direction vector data and performed cluster identification using density clustering. The results showed that the endpoint direction belonged to two independent clusters, with only frame 9 falling into the anomalous cluster. This single-frame discontinuity behavior... The path was labeled as a "transient divergent path." To prevent mis-clustering due to overlapping directions in the vascular junction area, the system performed a principal axis projection operation on the direction vectors. The upstream vascular principal axis of the fitted endpoint connection path was used as a reference. All direction vectors were projected onto this principal axis, and the cosine similarity between the direction vectors was recalculated. Specifically, the angle between the direction and the principal axis in frame 9 was 23 degrees, with a similarity of approximately 0.53, while the similarity in the previous frames was 0.91 to 0.95. Combined with the cosine-weighted direction difference score, this abnormal direction fluctuation was further amplified in the clustering results, thus confirming that the direction abrupt change in frame 9 was an isolated pseudo-path extension. The system added negative weights to its breakage behavior in the recursive path scoring and replaced the segment with a candidate path with higher structural rationality in the subsequent path correction module. The final corrected path exhibits higher consistency and directional stability in the time series. The directional vector remains within ±7 degrees of the principal axis direction for six consecutive frames. No unnecessary spurious extensions appear in the segmented images. The confidence score of the entire path increases from 0.71 to 0.93. This indicates that the method can effectively eliminate transient abnormal paths in areas with complex branch endpoints and many intersection interferences through the principal axis projection-cosine weighted directional field clustering strategy, achieving a more stable, continuous, and reliable vascular segmentation structure. This provides a higher quality image basis for subsequent quantitative analysis of occlusion and blood flow reconstruction.

[0049] The system has successfully detected a pseudo-branch path with directional mutation behavior in the 9th frame through directional field difference clustering. In order to further repair such abnormal endpoint paths and enhance the continuity and medical rationality of the overall segmentation results, the abnormal priority correction list mechanism has been implemented. Its construction is based on endpoint direction stability, path accessibility and neighborhood blood vessel consistency score, and the evolution trajectory of historical image frames is introduced for temporal fault tolerance verification. First, the system scores all path endpoints marked as "abnormal directional fluctuations". In the pseudo-branch path in frame 9, its directional stability score is only 0.42, which is lower than the set stability threshold of 0.6. This score is calculated based on the variance of the directional vector between frames, and it was found that the standard deviation of directional change in frames 7 to 10 is 18.7 degrees. The path accessibility refers to the integrity and matching degree of the connection between the path's starting point and the main artery (M1 segment). Because this path connects to a non-main artery region, its accessibility is only 0.58. At the same time, the similarity of the vascular texture in its surrounding neighborhood and the structural topological alignment result in a consistency score of 0.61, which shows certain spatial offset and topological discontinuity problems. Based on the three indicators, the system assigns this path the second position in the abnormal priority ranking and enters the structural correction candidate sequence. During the correction process, the system searches for the connection target with the smallest structural difference among its neighboring reliable path candidate segments. The search uses the direction vector angle <15 degrees and texture grayscale variance <12 as matching conditions. In the 10th frame, a small artery path extending from the proximal end of M2 to the outer side of the insula is found. The path is similar in direction and texture to the pseudo-branch, with a structural difference score of 0.21 (0 is the maximum score, which means complete consistency). Based on this, the system performs interpolation connection and uses the minimum curvature interpolation algorithm to generate connection curves, so that the path forms a natural transition in terms of vision and structure, and updates the image segmentation layer in real time. To verify whether this completion connection was misjudged due to short-term image perturbation, the historical evolution trajectory of the endpoint between frames 6 and 12 was further retrieved. Through pixel-level coordinate tracking and direction vector backtracking, it was found that the endpoint fluctuated by more than 20 degrees from the main trunk direction in frames 8 and 10, but recovered to the direction consistent with the main vessel in frames 6 and 12. The trajectory showed a trend of "non-continuous deflection-return to the main axis". The system comprehensively judged it as a false endpoint path caused by short-term direction field distortion, confirming that the correction strategy was effective and had no risk of damaging the real vascular structure. Finally, after processing by this abnormal path correction mechanism, the original false extension in the region was replaced by a new path with high confidence and structural coherence. The upstream and downstream connectivity of the entire occlusion segment was enhanced, the vascular topology integrity score increased from the original 0.78 to 0.95, and the overall segmentation confidence of the system improved by 11.3%, providing a reliable basis for clarifying the endpoint of the occlusion segment and the blood supply recovery boundary.

[0050] In the angiographic image segmentation task of the aforementioned patient with right-sided M2 segment occlusion due to ischemic cerebrovascular disease, as the identification of branch structures, orientation field clustering, and abnormal path repair mechanisms were gradually completed, the system entered the global reliability assessment stage of the segmented paths. To reasonably sort all vascular paths, remove erroneous segmented segments, and optimize the complete connectivity of the occluded region, a recursive path scoring system was introduced. A complete reliability scoring function was constructed through path topological breakpoint penalty and path curvature continuity smoothness index to solve the common problem of "continuous path segmentation but unreliable structure". In this case, the system identified a total of 9 main vascular paths from frames 7 to 12, of which 3 branches were adjacent to the occlusion area and had unstable morphology. First, through historical frame path tracking records, it was found that one path named Path-6 had 3 breaks and 2 disappearances and reconstructions in 5 consecutive frames. Based on this, the system recorded its breakpoint penalty frequency as . Secondly, the curvature of each segment in the path is discretized and calculated. For example, the curvature of the 4th segment is... The previous paragraph is The rate of change of curvature is:

[0051]

[0052] The fourth segment has a local directional offset. The angle with the main direction is 12 degrees, which is substituted into the smoothness term. ,Right now:

[0053]

[0054] Meanwhile, the response variance of this segment boundary obtained from the perturbation test is: Set the penalty coefficient The exponential penalty term is:

[0055]

[0056] Substitute the above values ​​into the path score product of segment 4:

[0057]

[0058] This score indicates that the path segment contributes weakly to the stability and reliability assessment. A similar calculation accumulates the scores for the entire path and normalizes them by path length. The final comprehensive path reliability score function is shown below:

[0059]

[0060] in part, The calculated overall score for Path-6 is: The score is significantly lower than that of the main path Path-2 in this case. When the score difference exceeds 50%, the system classifies Path-6 as a low-confidence path, excluding it from structural trunk completion of the occluded segment, and moves it to an auxiliary layer for manual review. It is worth noting that although this path is topologically continuous, it suffers from drastic curvature jumps, frequent breaks, and unstable perturbation responses. The system effectively avoids erroneous high-confidence judgments caused by surface connectivity through continuous inter-frame dynamic analysis and a function recursive feedback mechanism. This achieves a three-dimensional path optimization closed loop that integrates structural rationality, segmentation continuity, and historical stability in vascular atlases. Ultimately, this scoring mechanism not only improves the overall structural stability of the segmentation map but also significantly reduces the workload of manual intervention at suspicious breaks, demonstrating the high practicality of this function scoring system in clinical scenarios for optimizing paths and eliminating false paths.

[0061] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease, characterized in that, Includes the following steps: By detecting grayscale transition anomalies, structural breaks, and artifact delay signals in angiography images, multiple negative segmentation prior points are extracted to indicate suspected ischemic areas. These prior points are aggregated to form a prior abnormal region, and a symmetric perturbation test mechanism is introduced into the prior abnormal region to analyze the response stability to the segmentation boundary, so as to determine potential occlusion or abnormal vascular segments and dynamically adjust the segmentation threshold of the region. Within the determination area, a local interference window is established along the blood vessel direction to extract the frequency and rhythm features of density stripes. The discontinuous blood vessel segments caused by non-steady-state pulse changes are compensated and segmented by the scaling strategy of the convolution kernel. For the two ends of the identified broken blood vessel segments, the texture difference and frequency domain response offset are calculated. If the offset is within the preset physiological tolerance range, the interpolation completion mechanism is triggered to generate a reliable completion layer and retain it for subsequent calibration reference. By analyzing the degree of multipath divergence of the vascular branch endpoints in the aforementioned segmented vascular image and their path offset changes in the image sequence, dynamic feedback adjustment and continuity calibration are performed on the previously segmented paths.

2. The method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease according to claim 1, characterized in that... The extraction process of the negative segmentation prior points includes establishing a gray-scale and texture contradiction tensor. Regions in the image that simultaneously have gradient breakage features but lack texture continuity are identified as structural pseudo-vascular abnormalities. The determination of the artifact delay signal adopts the inter-frame abnormal residual relationship of the image and performs reverse marking processing on the intensity recovery of static regions in the later stage of the perfusion time image.

3. The method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease according to claim 2, characterized in that... The symmetric perturbation test introduces bidirectional mirror difference calculation, in which the symmetry error of the image response before and after the perturbation on both sides of the blood vessel central axis exceeds a set threshold, and is identified as an unstable segmentation region; the segmentation threshold of the dynamically adjusted region is updated according to the response confidence field of the perturbation region, and the unstable response region will trigger the threshold to be lowered and the processing range of upstream and downstream blood vessels to be expanded.

4. The method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease according to claim 3, characterized in that... The extraction of the frequency of the density stripes combines short-time Fourier transform and local frequency clustering to distinguish between real vascular pulse stripes and angiographic artifacts; the rhythmic features include spatial and temporal periodicity differences to detect rhythmic non-steady-state deformation of capillary segments caused by perfusion fluctuations, which is used for microvascular rupture compensation judgment; the scaling strategy of the convolution kernel adopts a learnable deformation kernel, the size of which is dynamically controlled by the rate of change of local frequency energy to maintain the sensitivity of abnormal region segmentation.

5. The method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease according to claim 4, characterized in that... The texture difference includes directional gradient field inversion calculation, and directional interpolation is performed to complete regions where the main directions of the two vascular segments are inconsistent but the texture is continuous; the judgment of the frequency domain response offset adopts joint matching of phase difference residual map and position offset; the reliable layer output by the completion mechanism also includes a completion confidence level map, in which different levels of completed segments selectively participate in subsequent path correction or manual review.

6. The method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease according to claim 5, characterized in that... The multipath divergence judgment of the vascular branch endpoint is based on inter-frame orientation field difference clustering, and an anomaly priority correction list is established for unstable endpoint paths. The process of dynamically adjusting the previously segmented paths adopts a recursive path scoring system, which scores the path credibility based on historical path continuity and structural rationality after each segmentation.

7. The method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease according to claim 6, characterized in that... The inter-frame orientation field difference clustering adopts a joint clustering strategy of spatial orientation vector and inter-frame relative rotation angle; the calculation of the orientation field difference adopts a cosine similarity weighting scheme based on principal axis projection to suppress mis-clustering caused by abrupt changes in the orientation field in the vascular intersection region.

8. The method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease according to claim 7, characterized in that... The anomaly priority correction list is sorted in multiple levels according to endpoint direction stability, path accessibility, and neighborhood vessel consistency score; the correction operation for abnormal paths includes minimum interpolation connection of structural differences to the direction of credible paths; in the process of establishing the anomaly priority correction list, the endpoint evolution trajectory of historical image frames is called to improve the tolerance of misidentification under short-term orientation field distortion conditions.

9. The method for segmenting and analyzing angiographic images of ischemic cerebrovascular disease according to claim 8, characterized in that... The recursive path scoring system introduces a path topology breakpoint penalty function to assign cumulative negative weights to segments in the historical path that exhibit repetitive breakage behavior; the path continuity scoring introduces a smoothness index based on the continuity of path curvature to distinguish between physiological curvature and abnormal deviations caused by path misdirection.

Images