Cerebrovascular abnormality analysis method based on transcranial Doppler ultrasound image

Abstract

The invention discloses a cerebral vessel anomaly analysis method based on a transcranial Doppler ultrasound image, which comprises the steps of TCD image data acquisition, TCD image space-time filtering preprocessing, automatic blood vessel positioning and three-dimensional modeling, the method comprises the steps of generating a standardized blood vessel segmentation model and a static blood vessel network graph, extracting blood flow quantitative parameters and spectrum qualitative features, analyzing multi-blood vessel space correlation features through a graph attention network, performing primary classification on blood flow states, performing secondary classification on lesion types, making risk grading judgment, generating intervention suggestions and outputting a structured report. By setting an automatic parameter measurement method, the manual operation time of doctors is shortened, and the speed and accuracy of lesion judgment are improved through automatic standardization judgment of frequency spectrum forms and association of multi-dimensional parameters. Meanwhile, the experience requirement of doctors for judging the lesion is reduced, and reference is conveniently provided for batch screening of primary hospitals.

Description

Technical Field

[0001] This invention belongs to the technical field of medical image analysis, and relates to a method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging. Background Technology

[0002] Cerebral infarction and vascular stenosis are cerebrovascular diseases that affect blood vessels and blood circulation in the brain, often leading to brain tissue damage. Currently, the primary method for analyzing these conditions is through a combination of TCD (transcranial Doppler ultrasound) examination and the physician's experience.

[0003] Traditional TCD examinations rely on sonographers manually analyzing blood flow spectra, requiring segmental measurements of parameters such as blood flow velocity and resistance index. This process is time-consuming and inefficient. Furthermore, the analysis of the condition depends on the physician's subjective judgment based on experience. Differences in the interpretation of spectral morphology (such as turbulence and eddy current signals) among physicians can lead to missed or misdiagnosed cases (e.g., underestimating mild to moderate vascular stenosis). Additionally, current technology analyzes blood flow parameters in isolation from a single vessel, resulting in limited correlation of hemodynamics across multiple vessels (e.g., differences in flow velocity between bilateral arteries of the same name, and assessment of collateral circulation in the Circle of Willis), which is detrimental to a holistic analysis of the condition. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging, which effectively solves the problems in the prior art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging includes:

[0007] Data acquisition: The patient underwent a transcranial Doppler (TCD) examination, and the spectrum of the M1 segment of the middle cerebral artery was acquired.

[0008] Preprocessing: The ESTARFM algorithm was used to perform spatiotemporal filtering preprocessing on the TCD images;

[0009] Automatic blood vessel localization: A binary blood vessel tree structure is obtained through multi-scale blood vessel enhancement. The centerline is extracted and skeletonized from the binary blood vessel map. Corresponding blood vessel segments are selected from the skeleton fragments through spatial prior and region clustering. The skeleton nodes of the corresponding blood vessel segments are finely annotated by graph neural network.

[0010] 3D modeling: Based on automatic blood vessel localization, standardized blood vessel segmentation models are generated to form a static blood vessel network diagram;

[0011] Multi-feature parameter extraction: Extracting quantitative blood flow parameters based on variable window long Fourier transform and time domain analysis, constructing a spatial-frequency dual-flow network to extract spectral qualitative features, and analyzing multi-vessel spatial correlation features through graph attention network;

[0012] Abnormal classification: Primary classification is obtained through TCD spectral image processing, including normal, turbulent, and eddy current. The probability of lesion type is obtained by combining the primary classification with the results of multi-feature parameter extraction as secondary classification.

[0013] Risk grading assessment: Analyze the risk level by combining the results of secondary classification and the results of multi-feature parameter extraction, generate intervention recommendations, and output a structured report.

[0014] Optionally, the multi-scale vascular enhancement includes:

[0015] The response of blood vessels of different diameters in the original TCD image was enhanced using the Frangi vascular enhancement filter;

[0016] Adaptive threshold segmentation is applied to the preprocessed TCD images to initially separate candidate vascular regions.

[0017] By combining morphological closing and opening operations to remove noise and fill holes, a binarized vascular tree structure is obtained.

[0018] Optionally, the centerline extraction and skeletonization include:

[0019] Refine the binarized vascular image and extract the vascular skeleton;

[0020] By using distance transformation to mark the trunk and branch points on the skeleton, the continuous skeleton is segmented to obtain candidate vascular segments.

[0021] Extract the center lines of each main trunk on the weighted skeleton graph using minimum path search.

[0022] Optionally, the spatial prior and region clustering, and the fine annotation of the graph neural network include:

[0023] The ROI coarse localization box is provided using the B-mode model of the patient's skull or a pre-registered standard cerebral vascular topology template;

[0024] In the skeletal segment, the center line with the highest spatial overlap with each ROI is selected as the initial candidate for the corresponding vascular segment;

[0025] The remaining skeleton nodes are used to construct graph data;

[0026] Each skeleton node is classified and labeled using a pre-trained graph attention network;

[0027] Using the label sequence output by GAT, a continuous vascular centerline is formed and mislabeled short branches are eliminated.

[0028] Optionally, the extraction of quantitative blood flow parameters based on variable window long Fourier transform includes:

[0029] A variable-window-length fast Fourier transform is used to dynamically adjust the length of the spectrum analysis window. During the systolic phase, a short window is used to obtain the peak systolic velocity (PSV) through spectrum peak detection. During the diastolic phase, a long window is used to extract the end-diastolic velocity (EDV).

[0030] Calculate the mean velocity (MFV), resistance index (RI), pulsatility index (PI), and systolic / diastolic velocity ratio (S / D) based on peak systolic velocity and end-diastolic velocity values:

[0031]

[0032] Where T represents the length of a complete cardiac cycle; v(t) is the instantaneous blood flow velocity; and v(t) represents the cumulative value of blood flow velocity over time during the cardiac cycle, which is divided by the cycle time T to obtain the time-averaged flow velocity.

[0033] Optionally, the construction of the spatial-frequency dual-stream network to extract spectral qualitative features includes:

[0034] Spatial feature extraction: The ResNet-18 branch takes the original TCD spectrum image as input, enhances the high-frequency harmonic region feature response through the CBAM attention module, and outputs a 128-dimensional spatial feature vector.

[0035] Frequency domain feature extraction: The 1D convolution branch inputs the time domain signal and uses 5 layers of convolution to extract multi-scale time-frequency features. The LSTM module models the temporal dependencies and outputs a 64-dimensional frequency domain feature vector.

[0036] Feature fusion: Spatial and frequency domain features are concatenated by a fully connected layer and then input into an SVM classifier. The spectral morphology is determined by combining the SHAP value analysis.

[0037] Optionally, the analysis of multi-vessel spatial association features via graph attention networks includes:

[0038] By calculating the difference in quantitative blood flow parameters between the two bilateral arteries of the same name and comparing it with a preset threshold, it can be determined whether the blood flow symmetry is abnormal.

[0039] Based on the matching degree of blood flow quantitative parameters of each vessel in the Circle of Willis, graph attention network modeling is used to evaluate collateral circulation function by analyzing the hemodynamic correlation between multiple vessels and generate quantitative scores.

[0040] Optionally, the specific steps of the primary classification include:

[0041] The preprocessed TCD spectral images and time-domain blood flow signals are processed using a two-branch model;

[0042] The spatial branch uses a ResNet-18 network to output a 128-dimensional spatial feature vector.

[0043] The time-domain branch converts the framed signal into a time-spectrum graph via STFT, inputs a 4-layer Transformer encoder to model the long-term temporal dependence of diastolic velocity decay, and outputs a 64-dimensional time-domain feature vector.

[0044] After the dual-modal features are concatenated into 192-dimensional joint features, the probabilities of normal flow, turbulent flow, and eddy current are output through nonlinear mapping of a fully connected layer and a Softmax classifier.

[0045] The specific steps of the secondary classification include:

[0046] Input secondary classification results, vascular network topology data, temporal blood flow parameters, spatial correlation features, and microemboli detection results. Dynamic spatiotemporal graph attention network modeling is performed using a static vascular network graph. Vascular node features are fused with temporal data to construct a spatiotemporal adjacency matrix. Gated graph convolution is used to update node states, and a cross-attention mechanism is combined to strengthen the correlation between abnormal signals and the vascular network. Finally, the probability of lesion type is output.

[0047] Optionally, the risk classification determination step includes:

[0048] Inputs include secondary classification results, quantitative blood flow parameters, collateral circulation score, microemboli density, and modified ABCD2 score parameters:

[0049] Set the peak flow velocity during contraction, the collateral circulation score, and the microemboli density weights and scores, and then sum the three weighted values ​​as the total score.

[0050] Establish a correlation between risk levels and scores, with the total score corresponding to the corresponding risk level.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] By implementing automated parameter measurement methods, the manual operation time for physicians is reduced. By integrating parameters such as spectral morphology, blood flow parameters, and vascular network correlation characteristics, abnormalities are identified from multiple dimensions, and high-risk lesions are quickly marked. The automated standardized evaluation of spectral morphology and the correlation of multi-dimensional parameters improve the speed and accuracy of lesion diagnosis. Simultaneously, it reduces the experience required for physicians to determine lesions, facilitating mass screening in primary hospitals. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] This invention discloses a method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging, comprising:

[0055] Data acquisition: The patient underwent a transcranial Doppler (TCD) examination, and the spectrum of the M1 segment of the middle cerebral artery was acquired.

[0056] Preprocessing: The ESTARFM algorithm was used to perform spatiotemporal filtering preprocessing on the TCD images;

[0057] Automatic blood vessel localization: A binary blood vessel tree structure is obtained through multi-scale blood vessel enhancement. The centerline is extracted and skeletonized from the binary blood vessel map. Corresponding blood vessel segments are selected from the skeleton fragments through spatial prior and region clustering. The skeleton nodes of the corresponding blood vessel segments are finely annotated by graph neural network.

[0058] 3D modeling: Based on automatic blood vessel localization, standardized blood vessel segmentation models are generated to form a static blood vessel network diagram;

[0059] Multi-feature parameter extraction: Extracting quantitative blood flow parameters based on variable window long Fourier transform and time domain analysis, constructing a spatial-frequency dual-flow network to extract spectral qualitative features, and analyzing multi-vessel spatial correlation features through graph attention network;

[0060] Abnormal classification: Primary classification is obtained through TCD spectral image processing, including normal, turbulent, and eddy current. The probability of lesion type is obtained by combining the primary classification with the results of multi-feature parameter extraction as secondary classification.

[0061] Risk grading assessment: Analyze the risk level by combining the results of secondary classification and the results of multi-feature parameter extraction, generate intervention recommendations, and output a structured report.

[0062] Specifically, in the preprocessing stage, the ESTARFM algorithm is used to perform spatiotemporal filtering on TCD images to remove interference factors such as electromyography interference and vascular wall pulsation artifacts, thereby enhancing the ultrasound signal. Then, key cerebral vascular anatomical structures, such as the M1 segment of the middle cerebral artery and the V4 segment of the vertebral artery, are automatically identified and labeled through multi-scale vascular enhancement, centerline extraction, and graph neural network annotation, generating standardized vascular segmentation models and providing a unified spatial reference framework for subsequent feature extraction.

[0063] By using variable-window long Fourier transform and time-domain analysis, quantitative blood flow parameters such as peak systolic velocity (PSV), end-diastolic velocity (EDV), mean ventricular velocity (MFV), resistance index (RI), pulsatility index (PI), and systolic / diastolic velocity ratio (S / D) can be extracted. A spatial-frequency dual-flow network is constructed to extract qualitative spectral features and determine whether the spectral morphology is turbulent, eddy, or normal. Graph attention network analysis is used to analyze the spatial correlation features of multiple vessels, assess blood flow symmetry and bronchial circulation function, identify TCD blood flow patterns in the primary classification, and identify specific lesion types in the secondary classification, such as stenosis, spasm, and microemboli, based on static vascular network diagrams.

[0064] Based on the above analysis, a risk classification is generated, intervention recommendations are produced, and a structured report is output.

[0065] In this way, by setting up automated parameter measurement methods, the manual operation time of physicians is reduced. By integrating parameters such as spectral morphology, blood flow parameters, and vascular network correlation characteristics, abnormalities are identified from multiple dimensions, and high-risk lesions are quickly marked. Through the automatic standardized evaluation of spectral morphology and the correlation of multi-dimensional parameters, the speed and accuracy of lesion diagnosis are improved.

[0066] In some feasible approaches, the output structured report includes dynamically generating structured parameter tables using Python's Pandas and Plotly, labeling anomaly feature spectrum maps in real time on the original TCD spectrum using TensorFlow.js and Canvas, rendering a 3D vascular tree heatmap using Three.js, and constructing a dynamic network diagram of Willis rings based on Cytoscape.js.

[0067] The system triggers early warnings for high-risk lesions and pushes emergency intervention suggestions to the electronic medical record system. The early warning mechanism relies on the Redis stream processing engine to monitor thresholds in real time, triggering audible and visual alarms and pop-up notifications. Reports are pushed to the electronic medical record system through the HL7FHIR interface, and intervention suggestions are provided in conjunction with the Neo4j knowledge graph.

[0068] In spatiotemporal filtering, the TCD spectrogram can be viewed as a two-dimensional image sequence of "time-frequency". When applying the ESTARFM algorithm, low-resolution data corresponds to the original noisy temporal spectrum, while high-resolution data can be taken from the clear spectrum pre-generated by a deep denoising network at adjacent time points. First, for two reference times before and after the target frame, a high-resolution clear spectrum is generated by a deep denoising network, such as U-Net, and paired with the low-resolution original noisy frame. Then, on the spectrogram image, the spectrum is calculated and normalized based on three weights: spectral similarity, temporal similarity, and spatial distance, centered on the pixel.

[0069] Among them, spectral similarity S spec (p):

[0070]

[0071] These are the pixel values ​​of the original frame (including noisy pixels). For the corresponding clear frame (denoised) pixel value, σ spec The scaling parameter controls the sensitivity to amplitude differences and is set to 5-20.

[0072] Time similarity S temp (p):

[0073]

[0074] The pixel value of the current frame pixel. To correspond to the pixel value of that pixel in the previous frame, σ t The standard deviation of the time difference (adjustment scale) is set to 3-10.

[0075] Spatial distance D spat (p,q):

[0076]

[0077] p is the target pixel coordinate, q is the neighboring pixel coordinate, ||pq|| 2 For Euclidean distance, σ c σ is typically set to half the size of the neighboring window; for example, in a 3x3 window. c =1.5, the neighborhood window refers to taking a fixed-size neighborhood region (local small patch) centered on a certain target pixel for processing.

[0078] Calculate and normalize ω:

[0079]

[0080] Where r is any pixel in the neighborhood N(p) of the target pixel p.

[0081] Subsequently, the most representative neighboring pixels are dynamically selected, and then the low-resolution frames are fused with the high-resolution residuals. Then, the high-resolution values ​​of each reference frame are superimposed according to the weights, thereby significantly suppressing high-frequency electromyography interference and low-frequency vascular wall pulsation artifacts while preserving the real blood flow dynamics.

[0082] Taking pixel p at target time t as an example, assuming there are two reference sharp frames at times t1 and t2, the corresponding high-resolution images are... The original low-resolution image is First, construct a clear residual plot:

[0083]

[0084] For pixel p at target time t, these residual maps are fused by weighted fusion in the neighborhood:

[0085]

[0086] Where α1 and α2 are interpolation coefficients.

[0087] Finally, the sharpness estimate of the high-resolution frame is calculated as follows:

[0088]

[0089] in The original pixel value at pixel position p in the low-resolution image at target time t.

[0090] As a specific implementation of the cerebrovascular abnormality analysis method based on transcranial Doppler ultrasound imaging provided in the application, multi-scale vascular enhancement includes:

[0091] The response of blood vessels of different diameters in the original TCD image was enhanced using the Frangi vascular enhancement filter;

[0092] The preprocessed TCD images were segmented using adaptive thresholding methods such as Otsu's method to initially separate candidate vascular regions.

[0093] By combining morphological closing and opening operations to remove noise and fill holes, a binarized vascular tree structure is obtained.

[0094] Centerline extraction and skeletonization include:

[0095] Refine the binarized vascular image and extract the vascular skeleton;

[0096] By using distance transformation to mark the trunk and branch points on the skeleton, the continuous skeleton is segmented to obtain candidate vascular segments.

[0097] The center lines of each main trunk are extracted from the weighted skeleton graph by means of minimum path search. Dijkstra's algorithm can be used for minimum path search.

[0098] Spatial priors and region clustering, and fine-grained annotation of graph neural networks include:

[0099] The ROI coarse localization box is provided using the B-mode model of the patient's skull or a pre-registered standard cerebral vascular topology template;

[0100] In the skeletal segment, the center line with the highest spatial overlap with each ROI is selected as the initial candidate for the corresponding vascular segment;

[0101] The remaining skeleton nodes are used to construct graph data;

[0102] Each skeleton node is classified and labeled using a pre-trained graph attention network;

[0103] Using the label sequence output by GAT, a continuous vascular centerline is formed and mislabeled short branches are eliminated.

[0104] The skeleton node includes node features: local curvature, skeleton width, and average filter response intensity, and edge features: distance between adjacent nodes and direction vector. The nodes and edges of the skeleton node are classified and labeled to distinguish the locations of the anterior cerebral artery, the extension of the internal carotid artery, the vertebral artery, etc.

[0105] The specific steps of 3D modeling are as follows: After the centerline is located, rigid / affine registration based on mutual information is used to align the ultrasound coordinate system with the patient's preoperative CT / MRI model. Original ultrasound slices are sampled along each centerline, and elliptical models of blood vessel cross sections are fitted to restore blood vessel diameter information. The labeled centerlines and cross-sectional ellipses are reconstructed into tubular mesh models in 3D space to generate standardized blood vessel segmentation models.

[0106] As another specific implementation of the method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging provided in the application, the extraction of quantitative blood flow parameters based on variable window long Fourier transform includes:

[0107] The length of the spectrum analysis window is dynamically adjusted using a variable window length Fast Fourier Transform. During the systolic phase, a short window is used to obtain the peak systolic velocity (PSV) through peak spectral detection. During the diastolic phase, a long window is used to extract the end-diastolic velocity (EDV).

[0108] Calculate the mean velocity (MFV), resistance index (RI), pulsatility index (PI), and systolic / diastolic velocity ratio (s / D) based on the peak systolic velocity and end-diastolic velocity values:

[0109]

[0110] Where T represents the duration of a complete cardiac cycle; v(t) is the instantaneous blood flow velocity; This represents the cumulative value of blood flow velocity over time during a cardiac cycle. Dividing this value by the cycle time T yields the time-averaged flow velocity.

[0111] In some feasible approaches, a short window of 10ms can be used during systole to capture rapid changes in blood flow velocity with high temporal resolution, while a long window of 50ms can be used during diastole to improve frequency resolution and accurately extract low-frequency parameters.

[0112] Constructing a spatial-frequency dual-stream network to extract qualitative spectral features includes:

[0113] Spatial feature extraction: The ResNet-18 branch takes the original TCD spectrum image as input, enhances the high-frequency harmonic region feature response through the CBAM attention module, and outputs a 128-dimensional spatial feature vector.

[0114] Frequency domain feature extraction: The 1D convolution branch inputs the time domain signal and uses 5 layers of convolution to extract multi-scale time-frequency features. The LSTM module models the temporal dependencies and outputs a 64-dimensional frequency domain feature vector.

[0115] Feature fusion: Spatial and frequency domain features are concatenated by a fully connected layer and then input into an SVM classifier. The spectral morphology is determined by combining the SHAP value analysis.

[0116] In some feasible methods, the TCD spectral image is 256×256 pixels, the high-frequency harmonic region is selected as the 3rd harmonic and above, the time domain signal sampling rate is 10kHz, the 5-layer convolution kernel length is 3 / 5 / 7, the high-frequency harmonic energy contribution rate in SHAP value analysis is 35%, the spectral morphology includes turbulence, eddies and normal, and the preliminary qualitative output is obtained through SVM classification and SHAP analysis results, which quickly screens the morphology category and provides a basis for subsequent classification.

[0117] Graph attention network analysis revealed the following spatial association features of multiple blood vessels:

[0118] By calculating the difference in quantitative blood flow parameters between the two bilateral arteries of the same name and comparing it with a preset threshold, it can be determined whether the blood flow symmetry is abnormal.

[0119] Based on the matching degree of blood flow quantitative parameters of each vessel in the Circle of Willis, graph attention network modeling is used to evaluate collateral circulation function by analyzing the hemodynamic correlation between multiple vessels and generate quantitative scores.

[0120] In some feasible methods, the difference in flow velocity between the two anterior cerebral arteries is calculated by extracting the peak systolic velocity of the left and right anterior cerebral arteries:

[0121]

[0122] By setting a preset difference threshold, such as PSV 差 >20% is considered asymmetric, PSV 差 A stenosis rate of ≥35% is considered to indicate a moderate possibility of stenosis.

[0123] The model is built using a graph attention network, with each vessel in the Circle of Willis, such as the anterior communicating artery ACoA and the posterior communicating artery PCoA, as nodes. The synergy between vessels is dynamically calculated based on the blood flow pressure gradient as the edge weight, and the collateral score is finally output.

[0124] By acquiring the peak systolic velocity (PSV), such as V1 and V2 for two vessels, and calculating the spatial distance D between the two vessels through automatic vessel localization and 3D modeling (e.g., the distance between the left middle cerebral artery M1 and the anterior communicating artery ACoA), the velocity difference ΔV between adjacent vessels in the Circle of Willis, the pressure difference ΔP, and the pressure gradient G are calculated. Finally, the gradient is converted into the edge weight ω of the graph network, such as the distance from vessel i to vessel j.

[0125] ΔV = |V1 - V2|

[0126] ΔP≈4·(ΔV) 2

[0127]

[0128] Determine the vascular anatomy weights ω i (e.g., the left middle cerebral artery M1 is 40%), calculate the synergistic contribution value A of a single pair of vessels for all adjacent vessels within the Circle of Willis. i，j :

[0129] A i，j =ω i ·ω j ·ω i，j

[0130] The total contribution of all blood vessels is summed up as ∑A. i，j Calculate the lateral branch score S:

[0131]

[0132] Where B is the theoretical maximum total contribution, i.e., ω i，j Contribution when =1

[0133] As another specific implementation of the cerebrovascular abnormality analysis method based on transcranial Doppler ultrasound imaging provided in the application, the specific steps of primary classification include:

[0134] The preprocessed TCD spectral images and time-domain blood flow signals are processed using a two-branch model;

[0135] The spatial branch uses a ResNet-18 network to output a 128-dimensional spatial feature vector.

[0136] The time-domain branch converts the framed signal into a time-spectrum graph via STFT, inputs a 4-layer Transformer encoder to model the long-term temporal dependence of diastolic velocity decay, and outputs a 64-dimensional time-domain feature vector.

[0137] After the dual-modal features are concatenated into 192-dimensional joint features, the probabilities of normal flow, turbulent flow, and eddy current are output through nonlinear mapping of a fully connected layer and a Softmax classifier.

[0138] The specific steps for secondary classification include:

[0139] Input secondary classification results, vascular network topology data, temporal blood flow parameters, spatial correlation features, and microemboli detection results. Dynamic spatiotemporal graph attention network modeling is performed using a static vascular network graph. Vascular node features are fused with temporal data to construct a spatiotemporal adjacency matrix. Gated graph convolution is used to update node states, and a cross-attention mechanism is combined to strengthen the correlation between abnormal signals and the vascular network. Finally, the probability of lesion type is output. By comparing the probability with a threshold set for each lesion type, the possibility of the lesion can be considered when the probability exceeds the threshold.

[0140] In some feasible approaches, the ResNet-18 network embeds a CBAM attention module, with channel attention enhancing high-frequency harmonic features above 2kHz and spatial attention focusing on peak distortion regions during spectral contraction. The framed signal has a frame length of 50ms and an overlap rate of 50%.

[0141] The vascular network topology data consisted of a 3D model of Willis rings and node relationships. The temporal blood flow parameters were PSV sequences of 10 consecutive cardiac cycles. The vascular node features included PSV, RI, and vessel diameter. The edge weights of the spatiotemporal adjacency matrix were constructed based on the synergy between pressure gradient and flow velocity.

[0142] In some feasible approaches, TCD images are spatiotemporally filtered using the ESTARFM algorithm to suppress noise interference and retain high-frequency transient components (microemboli characteristic frequency bands). Short-term high-frequency abrupt changes in the temporal signal are captured by a frequency domain branch 1D convolution kernel (length 3 / 5 / 7), and the temporal dependency is modeled by an LSTM module to identify the burstiness and periodicity of microemboli signals. Then, a spatial domain branch ResNet-18+CBAM attention module is used to locate high-frequency bright spots >2kHz. Finally, signals with a main frequency >2kHz are extracted by STFT in the primary classification, candidate microemboli are labeled, and the number of effective signals within a unit time (default 10 minutes) is counted to output the "microemboli density" detection result.

[0143] The risk classification determination steps include:

[0144] Inputs include secondary classification results, quantitative blood flow parameters, collateral circulation score, microemboli density, and modified ABCD2 score parameters:

[0145] Set the peak flow velocity during contraction, the collateral circulation score, and the microemboli density weights and scores, and then sum the three weighted values ​​as the total score.

[0146] Establish a correlation between risk levels and scores, with the total score corresponding to the corresponding risk level.

[0147] In some feasible methods, the following weightings are set: PSV grading weight 40% (e.g., 200 cm / s = 3 points), collateral rating weight 25% (e.g., <0.6 = 2 points), microemboli weight 20% (e.g., ≥1 per 10 minutes = 2 points), and basic ABCD2 weight 15% (e.g., assumed to be 3 points). The total score calculation formula is as follows:

[0148] Total score = 0.4×3 + 0.25×2 + 0.2×2 + 0.15×3 = 2.55

[0149] When the total score threshold is set as low risk <2, intermediate risk 2-3, and high risk >3, the total score is mapped to the intermediate risk level. At the same time, the PSV value approaches the threshold, triggering a PSV value alarm. This is combined with the generation of intervention suggestions, such as anticoagulation therapy.

[0150] Simultaneously, a structured report will be output, integrating the primary classification results, secondary classification lesion types, risk levels, quantitative parameters, spatial correlation characteristics, and microemboli density into a structured table using Python's Pandas library. Abnormalities will be highlighted in red and associated with JSON-formatted metadata to ensure data standardization and traceability.

[0151] Multimodal charts are generated using visualization tools: Plotly and TensorFlow.js are combined with Canvas to dynamically draw the original TCD spectrum, overlaying red shading to mark turbulent areas and arrows to mark PSV peaks; Three.js renders a 3D heatmap based on a segmented vascular model, using blue-red gradient bands to map blood flow velocity; Cytoscape.js constructs a Willis loop dynamic network diagram, reflecting risk levels through node colors; a Redis stream processing engine monitors PSV and microemboli density in real time; Node.js links hardware LEDs and buzzers to respond to graded alarms, and displays PSV time-series curves and microemboli distribution trends in the interface pop-up using ECharts.

[0152] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging, characterized in that, include: Data acquisition: The patient underwent a transcranial Doppler (TCD) examination, and the spectrum of the M1 segment of the middle cerebral artery was acquired. Preprocessing: The ESTARFM algorithm was used to perform spatiotemporal filtering preprocessing on the TCD images; Automatic blood vessel localization: A binary blood vessel tree structure is obtained through multi-scale blood vessel enhancement. The centerline is extracted and skeletonized from the binary blood vessel map. Corresponding blood vessel segments are selected from the skeleton fragments through spatial prior and region clustering. The skeleton nodes of the corresponding blood vessel segments are finely annotated by graph neural network. 3D modeling: Based on automatic blood vessel localization, standardized blood vessel segmentation models are generated to form a static blood vessel network diagram; Multi-feature parameter extraction: Extracting quantitative blood flow parameters based on variable window long Fourier transform and time domain analysis, constructing a spatial-frequency dual-flow network to extract spectral qualitative features, and analyzing multi-vessel spatial correlation features through graph attention network; Abnormal classification: Primary classification is obtained through TCD spectral image processing, including normal, turbulent, and eddy current. The probability of lesion type is obtained by combining the primary classification with the results of multi-feature parameter extraction as secondary classification. Risk grading assessment: Analyze the risk level by combining the results of secondary classification and the results of multi-feature parameter extraction, generate intervention recommendations, and output a structured report.

2. The method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging according to claim 1, characterized in that: The multiscale vascular enhancement includes: The response of blood vessels of different diameters in the original TCD image was enhanced using the Frangi vascular enhancement filter; Adaptive threshold segmentation is applied to the preprocessed TCD images to initially separate candidate vascular regions. By combining morphological closing and opening operations to remove noise and fill holes, a binarized vascular tree structure is obtained.

3. The method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging according to claim 2, characterized in that: The centerline extraction and skeletonization include: Refine the binarized vascular image and extract the vascular skeleton; By using distance transformation to mark the trunk and branch points on the skeleton, the continuous skeleton is segmented to obtain candidate vascular segments. Extract the center lines of each main trunk on the weighted skeleton graph using minimum path search.

4. The method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging according to claim 3, characterized in that: The spatial priors and region clustering, and the fine-grained annotation of graph neural networks include: The ROI coarse localization box is provided using the B-mode model of the patient's skull or a pre-registered standard cerebral vascular topology template; In the skeletal segment, the center line with the highest spatial overlap with each ROI is selected as the initial candidate for the corresponding vascular segment; The remaining skeleton nodes are used to construct graph data; Each skeleton node is classified and labeled using a pre-trained graph attention network; Using the label sequence output by GAT, a continuous vascular centerline is formed and mislabeled short branches are eliminated.

5. The method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging according to claim 4, characterized in that: The extraction of quantitative blood flow parameters based on variable window Fourier transform includes: A variable-window-length fast Fourier transform is used to dynamically adjust the length of the spectrum analysis window. During the systolic phase, a short window is used to obtain the peak systolic velocity (PSV) through spectrum peak detection. During the diastolic phase, a long window is used to extract the end-diastolic velocity (EDV). Calculate the mean velocity (MFV), resistance index (RI), pulsatility index (PI), and systolic / diastolic velocity ratio (S / D) based on peak systolic velocity and end-diastolic velocity values: Where T represents the duration of a complete cardiac cycle; v(t) is the instantaneous blood flow velocity; This represents the cumulative value of blood flow velocity over time during a cardiac cycle. Dividing this value by the cycle time T yields the time-averaged flow velocity.

6. The method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging according to claim 5, characterized in that: The extraction of spectral qualitative features by constructing a spatial-frequency dual-stream network includes: Spatial feature extraction: The ResNet-18 branch takes the original TCD spectrum image as input, enhances the high-frequency harmonic region feature response through the CBAM attention module, and outputs a 128-dimensional spatial feature vector. Frequency domain feature extraction: The 1D convolution branch inputs the time domain signal and uses 5 layers of convolution to extract multi-scale time-frequency features. The LSTM module models the temporal dependencies and outputs a 64-dimensional frequency domain feature vector. Feature fusion: Spatial and frequency domain features are concatenated by a fully connected layer and then input into an SVM classifier. The spectral morphology is determined by combining the SHAP value analysis.

7. The method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging according to claim 6, characterized in that: The analysis of multi-vessel spatial association features using graph attention networks includes: By calculating the difference in quantitative blood flow parameters between the two bilateral arteries of the same name and comparing it with a preset threshold, it can be determined whether the blood flow symmetry is abnormal. Based on the matching degree of blood flow quantitative parameters of each vessel in the Circle of Willis, graph attention network modeling is used to evaluate collateral circulation function by analyzing the hemodynamic correlation between multiple vessels and generate quantitative scores.

8. The method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging according to claim 7, characterized in that: The specific steps of the primary classification include: The preprocessed TCD spectral images and time-domain blood flow signals are processed using a two-branch model; The spatial branch uses a ResNet-18 network to output a 128-dimensional spatial feature vector. The time-domain branch converts the framed signal into a time-spectrum graph via STFT, inputs a 4-layer Transformer encoder to model the long-term temporal dependency of diastolic velocity decay, and outputs a 64-dimensional time-domain feature vector. After the dual-modal features are concatenated into 192-dimensional joint features, the probabilities of normal flow, turbulent flow, and eddy current are output through nonlinear mapping of a fully connected layer and a Softmax classifier. The specific steps of the secondary classification include: Input secondary classification results, vascular network topology data, temporal blood flow parameters, spatial correlation features, and microemboli detection results. Dynamic spatiotemporal graph attention network modeling is performed using a static vascular network graph. Vascular node features are fused with temporal data to construct a spatiotemporal adjacency matrix. Gated graph convolution is used to update node states, and a cross-attention mechanism is combined to strengthen the correlation between abnormal signals and the vascular network. Finally, the probability of lesion type is output.

9. The method for analyzing cerebrovascular abnormalities based on transcranial Doppler ultrasound imaging according to claim 8, characterized in that: The risk classification and determination steps include: Inputs include secondary classification results, quantitative blood flow parameters, collateral circulation score, microemboli density, and modified ABCD2 score parameters: Set the peak flow velocity during contraction, the collateral circulation score, and the microemboli density weights and scores, and then sum the three weighted values ​​as the total score. Establish a correlation between risk levels and scores, with the total score corresponding to the corresponding risk level.

Images