A method and system for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering

Abstract

This invention discloses a method and system for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering. The method includes: simultaneously acquiring B-mode images and Doppler data of the heart and aligning them temporally; using a deep learning model to segment anatomical structures and generate a mask; constructing a clutter shield based on the mask to filter out strong blood flow signals from the cavities; extracting the epicardial boundary to locate candidate coronary artery regions; constructing a spatiotemporal matrix within the candidate regions; using an adaptive truncated singular value decomposition algorithm to separate tissue clutter and obtain coronary blood flow signals; and finally, outputting quantitative indicators after vascular morphology confirmation and angle correction based on the blood flow centerline. This invention effectively solves the problem of extracting weak coronary artery signals under strong interference through physical-level shielding and adaptive filtering, improving the accuracy and robustness of detection.

Description

Technical Field

[0001] This invention relates to the field of ultrasound medical image processing technology, and in particular to a method and system for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering. Background Technology

[0002] Functional screening for coronary artery disease is of paramount importance in clinical diagnosis, and ultrasound examination, with its non-invasive, real-time, and low-cost advantages, has become a highly promising screening tool. However, in practical applications, coronary artery ultrasound blood flow imaging faces severe physical and technical challenges. The coronary arteries are small in diameter and located deep within the heart, producing very weak blood flow signals that are closely aligned with the rapidly moving heart wall. Within the imaging field, the heart chambers, especially the left ventricle, are filled with high-energy blood flow signals. These strong signals easily overshadow the weak signals from adjacent coronary arteries, making it difficult for conventional detection methods to effectively capture target information.

[0003] In the complex dynamic acoustic environment of the heart, the continuous periodic beating of the heart causes real-time displacement of the region of interest (ROI), and the resulting tissue motion artifacts often overlap with low-velocity blood flow signals in terms of spectral characteristics. Conventional frequency domain wall filtering techniques struggle to simultaneously filter out clutter and preserve signal when processing such non-stationary signals, easily leading to the loss of weak blood flow information or the retention of strong interference. Furthermore, the determination and angle correction of the ROI in the coronary arteries are usually highly dependent on the physician's experience. Manual adjustments are not only time-consuming but also easily affected by subjective factors, resulting in insufficient repeatability and stability of the final quantitative indicators such as flow velocity and flow rate, making it difficult to meet the needs of precise diagnosis and treatment. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering, so as to solve the problems pointed out in the background art.

[0005] In a first aspect, the present invention provides a method for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering, comprising the following steps: Simultaneously acquire B-mode image sequences and Doppler data sequences of the heart under the same probe posture, and perform time alignment between the B-mode image sequences and the Doppler data sequences; The B-mode image sequence is segmented into anatomical structures using a deep learning model to generate an anatomical structure mask containing the heart cavity region and the myocardial region. Based on the anatomical structure mask, a clutter shield is generated, and the clutter shield is used to filter out the cavity blood flow signal in the Doppler data sequence; Extract the epicardial boundary of the myocardial region in the anatomical structure mask, and determine the candidate region of the coronary artery based on the epicardial boundary; A spatiotemporal matrix is ​​constructed within the candidate coronary artery region, and tissue clutter and blood flow signals are separated using an adaptive truncated singular value decomposition algorithm to obtain the coronary artery blood flow signal. The vascular morphology of the coronary artery blood flow signal is confirmed, and the angle is corrected based on the extracted blood flow centerline to output quantified coronary artery blood flow indicators.

[0006] Optionally, the step of using a deep learning model to perform anatomical structure segmentation on the B-mode image sequence includes: Acquire single-frame or short sequence B-mode images as input data and perform normalization preprocessing; The input data is fed into a pre-trained deep learning network based on the U-Net or EchoNet architecture; The anatomical structure mask is output, and the region segmented by the anatomical structure mask includes at least the left ventricular cavity, the left ventricular myocardium, the left atrial cavity, and the aortic root valve area.

[0007] Optionally, the generation of clutter shielding based on the anatomical structure mask includes: Select the left ventricular cavity region and the left atrial cavity region within the anatomical structure mask; The boundaries of the left ventricular cavity region and the left atrial cavity region are expanded outward with a pixel interval of 2mm to 5mm to form a safe shielding boundary; The region within the safety shield boundary and the aortic root valve region are merged into the clutter shield, and the signal weights of the Doppler data sequence located within the clutter shield are reset to zero or assigned low weights.

[0008] Optionally, determining the candidate coronary artery region based on the epicardial boundary includes: The outer boundary of the left ventricular myocardium is extracted as the epicardial boundary; The epicardial boundary is expanded outward by 2 mm to 6 mm to form a band-shaped region surrounding the epicardium as the candidate region for the coronary artery; Establish a cardiac coordinate system, and set priority angle sectors for the left anterior descending artery, left circumflex artery, or right coronary artery within the candidate coronary artery region according to different cross-sectional perspectives.

[0009] Optionally, the method further includes using myocardial motion information to assist filtering: The displacement field of each frame is obtained by performing speckle tracing or optical flow calculation on the left ventricular myocardial region in the anatomical structure mask. The myocardial strain rate is calculated based on the displacement field to estimate the timing of cardiac contraction. In the singular value decomposition algorithm, the information at the moment of cardiac contraction is used to identify and remove sheet-like tissue artifacts that vibrate synchronously with the heart wall.

[0010] Optionally, the separation of tissue clutter and blood flow signals using the adaptive truncation singular value decomposition algorithm includes: Singular value decomposition is performed on the spatiotemporal matrix, the logarithms of the singular values ​​are calculated and normalized, and the preliminary cutoff value is determined based on the inflection points of the curve. ; At the initial cutoff value Search within the preset neighborhood range, for each candidate Values ​​were reconstructed for blood flow signals and quality control indicators were calculated. ; Select the quality control indicators The largest The value is used as the final truncation value, and the signal is reconstructed using this final truncation value.

[0011] Optionally, the quality control indicators The calculation formula is: ; in, The similarity of the reconstructed signal to the vascular structure. This indicates the correlation between the reconstructed signal and the tissue background. This indicates the residual cavity energy. , , These are preset weighting coefficients.

[0012] Optionally, the step of confirming the vascular morphology of the coronary artery blood flow signal includes: The energy map of the separated coronary artery blood flow signal is binarized to obtain a vascular region mask; The central axis of the vascular region mask is extracted using a skeletonization algorithm. Short branches and isolated flashing points are removed by pruning, and the longest continuous skeleton is retained as the blood flow centerline.

[0013] Optionally, the output quantified coronary artery blood flow parameters include: Using the tangent direction of the blood flow centerline as the blood flow direction, calculate the angle between the blood flow direction and the direction of the ultrasound beam. ; According to the formula Calculate the actual blood flow velocity; where, The axial blood flow velocity component directly measured by ultrasound equipment; Regarding the included angle Perform a confidence assessment when the included angle When the preset threshold is exceeded, a low confidence warning message is output.

[0014] Secondly, the present invention provides a coronary artery blood flow signal extraction system based on anatomical structure segmentation and adaptive filtering, comprising: The data acquisition module is used to simultaneously acquire B-mode image sequences and Doppler data sequences under the same probe orientation; The anatomical segmentation module is used to segment the B-mode image sequence into anatomical structures using a deep learning model and generate an anatomical structure mask. A shielding processing module is used to generate a clutter shielding cover based on the anatomical structure mask to filter out cavity blood flow signals in the Doppler data sequence; The region localization module is used to determine candidate regions for coronary arteries based on the epicardial boundary; An adaptive filtering module is used to separate tissue clutter from blood flow signals within the candidate coronary artery region using an adaptively truncated singular value decomposition algorithm; The signal extraction module is used for vascular morphology confirmation, angle correction, and quantitative index output.

[0015] The present invention has achieved the following beneficial effects: This invention introduces deep learning-based anatomical segmentation technology to intelligently identify and construct physical-level shielding regions for strong blood flow signals within cardiac chambers. This mechanism can block the interference of high-energy blood flow within the chambers on weak coronary artery signals at the source. By safely expanding the segmented anatomical boundaries, it effectively solves the signal coverage problem caused by paravalve leakage in traditional methods, significantly improving the signal-to-noise ratio of the region of interest and creating favorable conditions for subsequent extraction of weak signals.

[0016] This invention employs an adaptive singular value decomposition (SVD) filtering algorithm based on multidimensional feature evaluation, overcoming the limitations of traditional fixed threshold filtering in adapting to non-stationary tissue motion. The system automatically optimizes the cutoff value based on singular value curve characteristics and indicators such as vascular structure similarity and tissue background correlation of the reconstructed signal. While effectively filtering out myocardial tissue motion clutter, it maximizes the preservation of low-velocity coronary blood flow components, achieving high-fidelity reconstruction of weak signals in complex flow field environments.

[0017] Furthermore, this invention automatically locates candidate coronary artery regions and performs angle correction using prior anatomical knowledge, achieving automation and intelligence in the detection process. Through epicardial boundary tracking and blood flow centerline skeletalization, it reduces reliance on human experience and subjective errors during operation. The quantitative indicators output by the confidence assessment mechanism have higher accuracy and clinical repeatability, providing reliable technical support for early screening and hemodynamic assessment of coronary heart disease.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of a coronary artery blood flow signal extraction system based on anatomical structure segmentation and adaptive filtering in an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] To address the technical challenges of existing coronary artery ultrasound blood flow imaging technologies, such as weak signals, strong tissue clutter interference, reliance on manual experience for region of interest localization, and poor repeatability of quantization results, this application provides a method and system for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering. The core of this technical solution lies in utilizing artificial intelligence deep learning technology to accurately identify and shield strong blood flow interference within the cardiac chambers, combining prior anatomical knowledge to pinpoint candidate regions of the coronary arteries, and introducing an adaptive singular value decomposition (SVD) filtering algorithm based on multidimensional feature evaluation, thereby achieving high-fidelity extraction and quantization of weak coronary artery blood flow signals.

[0023] Please see the appendix Figure 1 This embodiment provides a method for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering. This method is applied to ultrasound diagnostic equipment with high frame rate data acquisition capabilities or a medical image processing workstation connected to it. The method specifically includes the following steps: Step S100: Simultaneously acquire B-mode image sequences and Doppler data sequences of the heart under the same probe orientation, and time-align the B-mode image sequences and the Doppler data sequences.

[0024] In this embodiment, data acquisition is the foundation of the entire processing flow. To ensure spatial consistency between subsequent anatomical segmentation and blood flow signal extraction, the operator is required to maintain the stability of the handheld probe or use a robotic arm to fix the probe, performing data acquisition under the same probe posture. Commonly selected acoustic window locations include, but are not limited to, the apical four-chamber view, the apical two-chamber view, or the parasternal short-axis view. The specific view selection depends on the anatomical location of the coronary artery branch to be observed (such as the left anterior descending artery (LAD), the left circumflex artery (LCX), or the right coronary artery (RCA).

[0025] To capture the rapidly changing hemodynamic features within the coronary arteries and the rapid beating of the heart itself, this embodiment imposes stringent requirements on the imaging frame rate. The base frame rate for the two-dimensional B-mode image is set to no less than 30Hz, preferably between 50Hz and 80Hz. A higher frame rate provides fine temporal resolution, which is crucial for capturing the instantaneous displacement of the myocardium during the cardiac cycle and the filling process of coronary blood flow during diastole. Simultaneously, for Doppler data sequences (typically IQ data or radio frequency data), the pulse repetition frequency (PRF) setting must cover the expected maximum coronary blood flow velocity (typically in the range of 10cm / s to 100cm / s) to effectively prevent aliasing. For acquisition modes, plane-wave imaging or multi-line acquisition techniques can be used to maintain sufficient spatial resolution while ensuring the frame rate.

[0026] Because ultrasound systems typically employ time-division multiplexing or parallel processing to acquire B-mode and Doppler data, asynchrony may exist between the two on the time axis. For example, a B-mode image may appear at one time point... The data was collected, and the corresponding Doppler data may be available at a specific time point. Acquisition. This minute time difference can cause spatial misalignment between anatomical structures and blood flow signals during high-speed cardiac activity. Therefore, this step must perform strict time alignment. Specifically, the system reads the hardware timestamp of each frame of data and, using the frame time of the B-mode image as a reference, performs interpolation synchronization on the Doppler data sequence (such as linear interpolation or spline interpolation), or employs nearest neighbor matching to ensure that each frame of the anatomical image has a set of Doppler data with strictly corresponding times. This alignment operation enables cardiac cycle synchronization based on image frames without relying on external electrocardiogram (ECG) signals, laying the foundation for subsequent use of anatomical masks to shield clutter.

[0027] Step S200: Use a deep learning model to segment the B-mode image sequence into anatomical structures to generate an anatomical structure mask containing the heart cavity region and the myocardial region.

[0028] After acquiring the synchronized raw data, this step aims to utilize artificial intelligence technology to intelligently recognize the anatomical structure of the heart. Traditional signal processing methods often filter across the entire image, causing high-energy cardiac blood flow signals to mask weak coronary artery signals. This embodiment achieves anatomical structure perception through a deep learning model.

[0029] First, the acquired B-mode image sequences were preprocessed. The preprocessing included intensity normalization, mapping the pixel grayscale values ​​of the images to the standard range of [0, 1] or [0, 255] to eliminate brightness differences caused by different device gain settings; removal of extreme bright spots, such as artifacts caused by valve calcification or strong pericardial reflections, as these bright noises may interfere with the segmentation network's judgment; and lightweight denoising, employing anisotropic diffusion filtering or nonlocal mean filtering to smooth ultrasound speckle noise while preserving the edge sharpness of the endocardium and epicardium to the greatest extent possible.

[0030] Next, the preprocessed data is input into the deep learning model. To utilize temporal contextual information and improve segmentation robustness, this embodiment preferably uses a short sequence (e.g., 3 to 7 consecutive frames) of B-mode image stacks as the model input, rather than a single frame image. The deep learning model architecture can use the classic U-Net network or the EchoNet network optimized for ultrasound video segmentation. This network includes an encoder path and a decoder path. The encoder extracts multi-scale, high-dimensional semantic features of the image through consecutive convolutional and pooling layers, while the decoder restores the features to the original resolution through upsampling and skip connections, thereby achieving pixel-level classification prediction.

[0031] The output of the deep learning model is an anatomical structure mask. This mask semantically labels each pixel in the image, and the segmented regions include at least: the left ventricular cavity (LV cavity), left ventricular myocardium (LV myocardium), left atrial cavity (LA cavity), and the aortic root and valve region (AO / valve region). During the model training phase, a large amount of standard cardiac cross-sectional data annotated by experienced ultrasound physicians was used, and supervised training was performed using the Dice Loss loss function or cross-entropy loss function to address the problem of class imbalance between background and target regions in cardiac images.

[0032] To further improve the geometric quality of the mask, this step also includes post-processing operations. Morphological closing operations are used to fill in tiny holes inside the mask region, opening operations are used to remove isolated noise islands, and classification boundaries are smoothed to ensure that the generated anatomical structure mask is topologically continuous and closed, conforming to the physiological characteristics of cardiac anatomy.

[0033] Step S300: Generate a clutter shield based on the anatomical structure mask, and use the clutter shield to filter out the cavity blood flow signal in the Doppler data sequence.

[0034] This is one of the key innovations that distinguishes this application from existing technologies. The heart chambers (especially the left ventricle and left atrium) are filled with high-speed, high-energy blood flow signals, whose Doppler energy intensity is often tens or even hundreds of times greater than that of coronary blood flow. Without shielding, these strong signals can cause severe sidelobe leakage during frequency domain filtering or time domain processing, making coronary signals undetectable. This step constructs a physical-level "clutter rejection mask" to block interference at its source.

[0035] Specifically, the left ventricular cavity region and the left atrial cavity region are selected from the anatomical structure mask. Considering the potential pixel-level errors in deep learning segmentation models and the rapid displacement of the endocardial surface at the center of the cardiac cycle, simply shielding the segmented cavity regions is insufficient; a safety buffer zone must be established. Therefore, this embodiment performs morphological dilation on the boundaries of the left ventricular cavity region and the left atrial cavity region. The dilation amplitude is set to a pixel interval of 2mm to 5mm. This 2mm to 5mm range is an empirical value validated by a large amount of clinical data, which can effectively cover the complex flow field and segmentation errors that may exist under the endocardium without excessively eroding the myocardial region and thus inadvertently damaging the coronary artery signals attached to the epicardium.

[0036] Meanwhile, the aortic root and valve region are also key areas for shielding. This region typically exhibits strong scintillation artifacts caused by high-speed jets and valve opening and closing, and these broadband noises easily contaminate the coronary artery spectrum. Therefore, the segmented aortic root / valve region is also directly included in the shielding scope.

[0037] Finally, the expanded cavity region is merged with the aortic / valve region to generate a binary "clutter shield." When processing the Doppler data sequence, all Doppler signals located within this clutter shield (whether it's the amplitude of the IQ data or the power spectrum energy) are either set to zero or assigned a very low weighting coefficient (e.g., 0.01). Through this strong spatial domain suppression, the cardiac blood flow, which originally dominated the energy map, is completely removed, retaining only the signals from the myocardium and peripheral extracardiac regions, creating a high signal-to-noise ratio background environment for the extraction of weak coronary artery signals.

[0038] Step S400: Extract the epicardial boundary of the myocardial region in the anatomical structure mask, and determine the coronary artery candidate region based on the epicardial boundary.

[0039] Searching for coronary artery signals across the entire image is not only computationally intensive but also prone to introducing lung air interference or rib artifacts. Based on cardiac anatomy, the major coronary artery branches run within the fatty connective tissue on the epicardial surface. Therefore, this step aims to construct an intelligent region of interest (ROI) to focus signal processing on high-probability areas.

[0040] Specifically, the outer contour of the left ventricular myocardium (LV myocardium), i.e., the epicardial boundary, is extracted from the anatomical mask generated in step S200. Then, using this epicardial boundary as a reference, a region is expanded outwards from the heart. Considering the diameter of the coronary arteries (typically 2-4 mm) and positional fluctuations caused by heartbeats, the expansion distance is set to 2 mm to 6 mm. This band-shaped region surrounding the epicardium is defined as the coronary artery candidate region.

[0041] Furthermore, to improve the specificity of the detection, this embodiment combines anatomical orientation information for refined localization. A cardiac coordinate system is established with the long axis of the left ventricle as the reference. Based on the current scanning section (selected by the operator or automatically identified by the system), priority angular sectors for specific vessel branches are set within the candidate coronary artery region. For example, in the apical four-chamber view, the direction closer to the anterior interventricular groove is the potential region for the left anterior descending artery (LAD); the direction closer to the atrioventricular groove on the lateral wall is the potential region for the left circumflex artery (LCX); and the direction towards the right ventricular lateral wall is the potential region for the right coronary artery (RCA). Through this anatomical location prior, the algorithm can further narrow down the processing range and reduce false positive signals.

[0042] At this stage, myocardial motion information can be used as an aid. Speckle tracking or optical flow calculation is performed on the segmented left ventricular myocardial region to obtain the displacement field of the myocardium for each frame. The myocardial strain rate is calculated using the displacement field, thereby accurately estimating the contraction and relaxation times of the heart. This temporal information is then passed to subsequent filtering algorithms, serving as a crucial basis for distinguishing between persistent tissue artifacts and periodically changing blood flow signals.

[0043] Step S500: Construct a spatiotemporal matrix within the candidate coronary artery region, and use an adaptive truncated singular value decomposition algorithm to separate tissue clutter and blood flow signals to obtain coronary artery blood flow signals.

[0044] This is the core of the signal processing in this embodiment of the invention. Within the defined candidate coronary artery region, although blood flow within the cavity has been shielded, the signal still contains extremely strong myocardial tissue motion clutter. Traditional wall filters, typically based on a fixed cutoff frequency, cannot adapt to non-stationary tissue motion and easily filter out low-velocity blood flow or residual strong clutter. This embodiment employs a spatiotemporal filtering technique based on singular value decomposition (SVD) and innovatively proposes an adaptive truncation strategy.

[0045] First, the Doppler data within the coronary artery candidate region is reconstructed into a spatiotemporal matrix (Casorati Matrix). Assuming there are [missing information] within the candidate region... Each spatial pixel was collected. Frame data, then constructs a size of matrix For the matrix Perform singular value decomposition, i.e. In the formula middle: for The spatial feature vector matrix, whose column vectors represent different spatial patterns; for The transpose of the time feature vector matrix, whose row vectors represent the modulation curves of the corresponding spatial patterns as a function of time; for A diagonal matrix containing a series of singular values ​​arranged in descending order, with diagonal elements... That is, it is a singular value, and satisfies In the SVD domain, high-energy, highly spatiotemporally correlated tissue clutter typically corresponds to larger singular values ​​(lower-order components), while blood flow signals correspond to intermediate singular values ​​(higher-order components), and noise corresponds to the smallest singular value.

[0046] The key to SVD filtering lies in choosing an appropriate cutoff threshold. To remove the previous Each component represents clutter. To address the issues in traditional methods... To address the problem of difficulty in determining the value, this embodiment proposes an adaptive optimization method based on multidimensional feature evaluation.

[0047] This method first determines the initial cutoff value by analyzing the shape of the singular value distribution curve. Specifically, the logarithm of the singular values ​​is calculated and normalized to find the inflection point (Knee Point) or the point of maximum curvature of the curve, which usually corresponds to the energy boundary between tissue clutter and blood flow signal.

[0048] Subsequently, Preset neighborhood range (For example or Perform a fine-grained search within this range. For each candidate within this range... Value, using the value before removal The blood flow signal is reconstructed from the data after each component, and a comprehensive quality control index is calculated. This indicator The calculation formula is: ; in: This indicates the similarity of the reconstructed signal to the vascular structure. Frangi filtering or Hessian matrix eigenvalue analysis is performed on the reconstructed energy map to assess whether the signal exhibits the characteristic elongated, continuous tubular structure of the coronary arteries. When the value is selected appropriately, the blood vessel morphology is clearest, and the value of this item is the largest.

[0049] This indicates the correlation between the reconstructed signal and the tissue background. The correlation coefficient between the reconstructed signal and the myocardial grayscale changes in the original B-mode image is calculated. The true blood flow signal should not be completely synchronized with the myocardial motion pattern, resulting in a lower correlation. If... If the value is too small, there will be more residual noise and higher correlation; if the value of this item is large, it will lead to a decrease in the total score.

[0050] This indicates the residual cavity energy. Using the clutter shield generated in step S300, monitor whether there are still abnormally bright signals remaining at the edge of the shielded area. This item is used to penalize the cutoff value that causes noise leakage.

[0051] These are preset weighting coefficients used to balance the influence of each indicator.

[0052] The system traverses the search range and selects... The largest value This final cutoff value is used for signal reconstruction, thereby achieving adaptive filtering that effectively removes strong tissue clutter while preserving weak coronary blood flow signals to the maximum extent.

[0053] Step S600: Confirm the vascular morphology of the coronary artery blood flow signal, perform angle correction based on the extracted blood flow centerline, and output quantified coronary artery blood flow indicators.

[0054] While the signal-to-noise ratio is significantly improved after SVD filtering, a small amount of speckle noise may still remain. To obtain clinically usable diagnostic indicators, morphological confirmation and quantitative calculations are necessary.

[0055] First, the Doppler energy map reconstructed by SVD is binarized, and all independent connected regions are identified using a connected component analysis algorithm. Based on the anatomical characteristics of the coronary arteries, flicker signals that are too small (isolated noise), have too small aspect ratio (clump noise), or only appear in a single frame are removed, and candidate signals that are relatively continuous in time and spatially elongated and connected are retained as confirmed coronary arteries.

[0056] Next, the confirmed coronary blood flow region undergoes skeletonization. A thinning algorithm is used to progressively peel away the boundary pixels of the target region, shrinking the vessel region to a single-pixel-wide central axis while maintaining the topological structure. To eliminate false branches caused by edge spurs, pruning is also performed to remove short ends, ultimately retaining the longest and smoothest main skeleton as the blood flow centerline.

[0057] Based on this blood flow centerline, the system automatically performs Doppler angle correction. The tangent direction at each point on the blood flow centerline is calculated as the true blood flow direction vector. Combined with the known incident direction of the probe's sound beam, the angle between the sound beam and the blood flow is calculated point by point. According to the Doppler formula The measured frequency shift velocity is corrected to the true flow velocity. The axial blood flow velocity component is obtained directly from ultrasound equipment.

[0058] To prevent errors caused by overcorrection, this embodiment introduces a confidence assessment mechanism. When the calculated angle... When the angle exceeds a preset threshold (e.g., 60 degrees), due to the nonlinear amplification effect of the cosine function, even a small angle measurement error can lead to a large deviation in the velocity calculation. At this time, the system will output a low confidence warning message and mark the blood vessel segment on the display interface (e.g., displayed as a gray or yellow dashed line) to remind the doctor that the measurement value at this location may be unreliable.

[0059] Finally, the system outputs a series of quantitative indicators, including coronary flow reserve (CFVR), peak systolic velocity, peak diastolic velocity, time-velocity integral (VTI), and diastolic / systolic velocity ratio (DSVR), and generates a two-dimensional coronary color flow map with background clutter removed, which intuitively displays the perfusion status of the coronary arteries.

[0060] This embodiment provides a coronary artery blood flow signal extraction system based on anatomical structure segmentation and adaptive filtering. This system serves as the hardware and software carrier for the aforementioned method, and its functional modules logically correspond one-to-one with the steps of Embodiment 1. For example... Figure 2 As shown, the system mainly includes: The data acquisition module controls the ultrasound front-end hardware to simultaneously acquire B-mode image sequences and Doppler data sequences of the heart under the same probe orientation. This module integrates a high-precision clock synchronization unit to ensure strict alignment of each frame of anatomical image and Doppler data on the time axis.

[0061] The anatomical segmentation module, serving as the system's AI inference engine, is used to segment anatomical structures from B-mode image sequences using deep learning models (such as U-Net), generating anatomical masks that include the left ventricle, left atrium, myocardium, and aorta. This module typically runs on high-performance computing units such as GPUs or NPUs.

[0062] The shielding processing module is used to generate clutter shields based on anatomical structure masks. This module performs morphological dilation operations (2-5mm) and zeros or weights the signals within the shield in the Doppler data domain to physically filter out cavity blood flow interference.

[0063] The region localization module is used to extract the epicardial boundary based on a myocardial mask and expand it outward (2-6mm) to generate candidate regions for coronary arteries. This module also combines cross-sectional orientation information to set priority search sectors, achieving precise localization of the region of interest.

[0064] An adaptive filtering module is used to construct a spatiotemporal matrix within the candidate coronary artery region and employs an adaptively truncated singular value decomposition algorithm to separate tissue clutter from blood flow signals. This module incorporates built-in quality control metrics. The computational unit can automatically find the optimal cutoff value. This enables high-fidelity signal reconstruction.

[0065] The signal extraction module performs skeletonization on the reconstructed signal, extracts the blood flow centerline, and automatically calculates the Doppler angle and confidence level. This module ultimately outputs quantified hemodynamic parameters and enhanced blood flow images, which are then displayed to the user.

[0066] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for extracting coronary artery blood flow signals based on anatomical structure segmentation and adaptive filtering, characterized in that, Includes the following steps: Simultaneously acquire B-mode image sequences and Doppler data sequences of the heart under the same probe posture, and perform time alignment between the B-mode image sequences and the Doppler data sequences; The B-mode image sequence is segmented into anatomical structures using a deep learning model to generate an anatomical structure mask containing the heart cavity region and the myocardial region. Based on the anatomical structure mask, a clutter shield is generated, and the clutter shield is used to filter out the cavity blood flow signal in the Doppler data sequence; Extract the epicardial boundary of the myocardial region in the anatomical structure mask, and determine the candidate region of the coronary artery based on the epicardial boundary; A spatiotemporal matrix is ​​constructed within the candidate coronary artery region, and tissue clutter and blood flow signals are separated using an adaptive truncated singular value decomposition algorithm to obtain the coronary artery blood flow signal. The vascular morphology of the coronary artery blood flow signal is confirmed, and the angle is corrected based on the extracted blood flow centerline to output quantified coronary artery blood flow indicators.

2. The method according to claim 1, characterized in that, The step of using a deep learning model to segment the anatomical structure of the B-mode image sequence includes: Acquire single-frame or short sequence B-mode images as input data and perform normalization preprocessing; The input data is fed into a pre-trained deep learning network based on the U-Net or EchoNet architecture; The anatomical structure mask is output, and the region segmented by the anatomical structure mask includes at least the left ventricular cavity, the left ventricular myocardium, the left atrial cavity, and the aortic root valve area.

3. The method according to claim 2, characterized in that, The clutter shield generated based on the anatomical structure mask includes: Select the left ventricular cavity region and the left atrial cavity region within the anatomical structure mask; The boundaries of the left ventricular cavity region and the left atrial cavity region are expanded outward with a pixel interval of 2mm to 5mm to form a safe shielding boundary; The region within the safety shield boundary and the aortic root valve region are merged into the clutter shield, and the signal weights of the Doppler data sequence located within the clutter shield are reset to zero or assigned low weights.

4. The method according to claim 1, characterized in that, The determination of candidate coronary artery regions based on the epicardial boundary includes: The outer boundary of the left ventricular myocardium is extracted as the epicardial boundary; The epicardial boundary is expanded outward by 2 mm to 6 mm to form a band-shaped region surrounding the epicardium as the candidate region for the coronary artery; Establish a cardiac coordinate system, and set priority angle sectors for the left anterior descending artery, left circumflex artery, or right coronary artery within the candidate coronary artery region according to different cross-sectional perspectives.

5. The method according to claim 1, characterized in that, The method also includes using myocardial motion information to assist filtering: The displacement field of each frame is obtained by performing speckle tracing or optical flow calculation on the left ventricular myocardial region in the anatomical structure mask. The myocardial strain rate is calculated based on the displacement field to estimate the timing of cardiac contraction. In the singular value decomposition algorithm, the information at the moment of cardiac contraction is used to identify and remove sheet-like tissue artifacts that vibrate synchronously with the heart wall.

6. The method according to claim 1, characterized in that, The method of separating tissue clutter and blood flow signals using an adaptive truncation singular value decomposition algorithm includes: Singular value decomposition is performed on the spatiotemporal matrix, the logarithms of the singular values ​​are calculated and normalized, and the preliminary cutoff value is determined based on the inflection points of the curve. ; At the initial cutoff value Search within the preset neighborhood range, for each candidate Values ​​were reconstructed for blood flow signals and quality control indicators were calculated. ; Select the quality control indicators The largest The value is used as the final truncation value, and the signal is reconstructed using this final truncation value.

7. The method according to claim 6, characterized in that, The quality control indicators The calculation formula is: ； in, The similarity of the reconstructed signal to the vascular structure. This indicates the correlation between the reconstructed signal and the tissue background. This indicates the residual cavity energy. , , These are preset weighting coefficients.

8. The method according to claim 1, characterized in that, The process of confirming the vascular morphology of the coronary artery blood flow signal includes: The energy map of the separated coronary artery blood flow signal is binarized to obtain a vascular region mask; The central axis of the vascular region mask is extracted using a skeletonization algorithm. Short branches and isolated flashing points are removed by pruning, and the longest continuous skeleton is retained as the blood flow centerline.

9. The method according to claim 8, characterized in that, The output quantified coronary artery blood flow parameters include: Using the tangent direction of the blood flow centerline as the blood flow direction, calculate the angle between the blood flow direction and the direction of the ultrasound beam. ; According to the formula Calculate the actual blood flow velocity; where, The axial blood flow velocity component directly measured by ultrasound equipment; Regarding the included angle Perform a confidence assessment when the included angle When the preset threshold is exceeded, a low confidence warning message is output.

10. A coronary artery blood flow signal extraction system based on anatomical structure segmentation and adaptive filtering, characterized in that, include: The data acquisition module is used to simultaneously acquire B-mode image sequences and Doppler data sequences under the same probe orientation; The anatomical segmentation module is used to segment the B-mode image sequence into anatomical structures using a deep learning model and generate an anatomical structure mask. A shielding processing module is used to generate a clutter shielding cover based on the anatomical structure mask to filter out cavity blood flow signals in the Doppler data sequence; The region localization module is used to determine candidate regions for coronary arteries based on the epicardial boundary; An adaptive filtering module is used to separate tissue clutter from blood flow signals within the candidate coronary artery region using an adaptively truncated singular value decomposition algorithm; The signal extraction module is used for vascular morphology confirmation, angle correction, and quantitative index output.

Images