A method and system for identifying vascular pathway lesions based on target detection

Abstract

The present invention discloses a method and system for identifying vascular pathway lesions based on target detection, relating to the field of medical image processing technology. The method comprises: performing cardiac cycle phase division based on electrocardiogram signals, calculating the displacement vector field between adjacent frames of a vascular DSA image sequence, constructing a motion compensation model, and generating a stable DSA sequence; segmenting the stable DSA sequence into binary vascular mask images using a U-Net segmentation model, and using a particle image velocimetry algorithm to track the motion displacement of contrast agent particles in consecutive frames, generating a blood flow velocity vector field, and calculating an eddy current intensity feature map; channel-concatenating the binary vascular mask image and the eddy current intensity feature map to generate a vascular feature tensor, and generating lesion detection results by improving the YOLOv7 model. The present invention improves the accuracy, robustness, and clinical applicability of vascular lesion identification, achieving a closed-loop technology from multimodal data processing to automated diagnostic report generation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for identifying vascular pathway lesions based on target detection. Background Art

[0002] In the field of medical image analysis and cardiovascular disease diagnosis, with the rapid development of computer vision and artificial intelligence technologies, automated lesion identification methods based on digital subtraction angiography (DSA) have made significant progress in recent years. DSA can provide high-resolution images of vascular structures and is widely used in the clinical assessment of coronary, cerebral, and peripheral vascular diseases. Traditional vascular lesion identification relies primarily on the subjective judgment of radiologists on DSA images, which is time-consuming and susceptible to human factors. In recent years, breakthroughs in deep learning technology, particularly convolutional neural networks (CNNs), in image segmentation and object detection have promoted the development of automated DSA image analysis. For example, U-Net and its variants are widely used for vascular segmentation, and the YOLO series of models have been tried for rapid detection of lesion areas, improving diagnostic efficiency and consistency.

[0003] Although existing research has achieved certain results in vascular image analysis, several key technical bottlenecks remain. First, existing methods are mostly based on static images or do not consider the dynamic characteristics of blood flow, resulting in limited recognition of lesions related to hemodynamic abnormalities (such as stenosis and vortex anomalies). Traditional pixel-based segmentation methods have difficulty accurately depicting the spatiotemporal evolution of blood flow, thus affecting the accuracy of lesion localization. Second, most existing models ignore the vascular motion artifacts caused by cardiac pulsation in DSA image sequences, resulting in inconsistencies in the temporal dimension of the image sequence, affecting the stability of subsequent analysis. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for identifying vascular access lesions based on target detection to solve the problems of insufficient recognition ability of hemodynamic abnormalities and poor image sequence stability in the prior art.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for identifying vascular pathway lesions based on target detection, which includes synchronously acquiring an original vascular DSA image sequence and an original electrocardiogram signal, and performing preprocessing to generate a vascular DSA image sequence and an electrocardiogram signal; performing cardiac cycle phase division according to the electrocardiogram signal, and calculating the displacement vector field between adjacent frames of the vascular DSA image sequence, constructing a motion compensation model, and generating a stable DSA sequence; segmenting the stable DSA sequence into a binary vascular mask image through a U-Net segmentation model, and using a particle image velocimetry algorithm to track the motion displacement of contrast agent particles in consecutive frames, generating a blood flow velocity vector field, and calculating an eddy current intensity feature map; channel-concatenating the binary vascular mask image and the eddy current intensity feature map to generate a vascular feature tensor, and generating a lesion detection result through an improved YOLOv7 model; based on the lesion detection result, a pseudo depth map is obtained from the stable DSA sequence in combination with a single-view depth estimation network, and mapped to a three-dimensional vascular model to generate a structured report.

[0008] As a preferred solution of the target detection-based vascular access lesion identification method of the present invention, the preprocessing is performed to generate a vascular DSA image sequence and an electrocardiogram signal. The specific steps are as follows:

[0009] Time axis alignment of the original vascular DSA image sequence and the original electrocardiogram signal;

[0010] The aligned original vascular DSA image sequence is converted into a grayscale image, and size normalization, vascular area contrast enhancement, image noise removal, and time domain normalization are performed to generate a vascular DSA image sequence;

[0011] The aligned raw ECG signals are bandpass filtered to generate ECG signals.

[0012] As a preferred solution of the target detection-based vascular access lesion identification method of the present invention, wherein: the cardiac cycle phase is divided according to the electrocardiogram signal, and the displacement vector field between adjacent frames of the vascular DSA image sequence is calculated. The specific steps are as follows:

[0013] The R wave detection algorithm is used to identify the R wave peak point in the electrocardiogram signal, obtain the cardiac cycle, and divide the cardiac cycle into systole and diastole by phase division;

[0014] Two adjacent frames of images in the vascular DSA image sequence are input into the dense optical flow algorithm to calculate the displacement vector field.

[0015] As a preferred solution of the target detection-based vascular access lesion identification method of the present invention, the steps of constructing a motion compensation model and generating a stable DSA sequence are as follows:

[0016] The vascular DSA image sequence is classified into time phases according to the systolic and diastolic phases, and a motion compensation model is constructed based on the displacement vector field to generate motion compensation parameters.

[0017] Based on motion compensation parameters, image resampling is used to align the frames of vascular DSA image sequences, and a stable DSA sequence is generated after time domain filtering.

[0018] As a preferred embodiment of the target detection-based vascular access lesion identification method of the present invention, the stable DSA sequence is segmented into a binary vascular mask image using a U-Net segmentation model, and the motion displacement of contrast agent particles in consecutive frames is tracked using a particle image velocimetry algorithm. The specific steps are as follows:

[0019] A U-Net segmentation model with an encoder-decoder architecture is used to segment the stable DSA sequence into a binary vascular mask image.

[0020] Two consecutive frames of images in the stable DSA sequence and the binary vascular mask image are input into the particle image velocimetry algorithm to identify contrast agent particles and calculate their motion displacement.

[0021] As a preferred solution of the target detection-based vascular access lesion identification method of the present invention, the steps of generating the blood flow velocity vector field and calculating the eddy current intensity characteristic map are as follows:

[0022] Optimize the motion displacement of contrast agent particles with sub-pixel accuracy to generate blood flow velocity vector field;

[0023] The local rotation intensity of the blood flow velocity vector field is calculated by curl operation to generate the eddy current intensity value;

[0024] After the eddy current intensity characteristic value is normalized, it is spatially aligned with the binary blood vessel mask image to generate the eddy current intensity characteristic map.

[0025] As a preferred solution of the target detection-based vascular access lesion identification method of the present invention, the binary vascular mask image and the eddy current intensity feature map are channel-concatenated to generate a vascular feature tensor. The specific steps are as follows:

[0026] The binary blood vessel mask image is subjected to morphological dilation processing, and the eddy current intensity feature map is subjected to Gaussian filtering and smoothing processing;

[0027] The processed binary blood vessel mask image and eddy current intensity feature map are stitched into a dual-channel feature map through a deep stitching operation;

[0028] The dual-channel feature map is concatenated with the stable DSA sequence to generate a vascular feature tensor.

[0029] As a preferred solution of the target detection-based vascular access lesion identification method of the present invention, the lesion detection results are generated by improving the YOLOv7 model. The specific steps are as follows:

[0030] Add a multimodal feature fusion layer to the standard YOLOv7 model to generate an improved YOLOv7 model;

[0031] The improved YOLOv7 model is trained using a multi-task loss function, and the vascular feature tensor is input and forward propagation is performed to generate lesion detection results.

[0032] As a preferred solution of the target detection-based vascular access lesion identification method of the present invention, the pseudo depth map is obtained from the stable DSA sequence based on the lesion detection results in combination with the single-view depth estimation network, and mapped to the three-dimensional vascular model to generate a structured report. The specific steps are as follows:

[0033] The lesion detection results are formatted and normalized, and the stable DSA sequence is input into the single-view depth estimation network to output a pseudo depth map.

[0034] After bilateral filtering and contrast enhancement, the pseudo depth map is spatially aligned with the stable DSA sequence for verification.

[0035] Based on the verified pseudo-depth map and the stable DSA sequence, a 3D vascular model is constructed using a 3D reconstruction algorithm.

[0036] The processed lesion detection results are mapped to a 3D vascular model to generate a structured report.

[0037] In a second aspect, the present invention provides a vascular pathway lesion identification system based on target detection, including: a data acquisition module for synchronously acquiring an original vascular DSA image sequence and an original electrocardiogram signal, and performing preprocessing to generate a vascular DSA image sequence and an electrocardiogram signal; a phase compensation module for performing cardiac cycle phase division according to the electrocardiogram signal, and calculating the displacement vector field between adjacent frames of the vascular DSA image sequence, constructing a motion compensation model, and generating a stable DSA sequence; a particle velocimetry module for segmenting the stable DSA sequence into a binary vascular mask image through a U-Net segmentation model, and using a particle image velocimetry algorithm to track the motion displacement of contrast agent particles in consecutive frames, generate a blood flow velocity vector field, and calculate an eddy current intensity feature map; a lesion monitoring module for channel-concatenating the binary vascular mask image and the eddy current intensity feature map to generate a vascular feature tensor, and generate a lesion detection result through an improved YOLOv7 model; a report generation module for obtaining a pseudo depth map from the stable DSA sequence based on the lesion detection result in combination with a single-view depth estimation network, and mapping it to a three-dimensional vascular model to generate a structured report.

[0038] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the method for identifying vascular access lesions based on target detection as described in the first aspect of the present invention is implemented.

[0039] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the method for identifying vascular access lesions based on target detection as described in the first aspect of the present invention is implemented.

[0040] The present invention achieves the following benefits: It generates stable DSA image sequences through electrocardiogram-guided motion compensation technology, extracts hemodynamic features using a U-Net segmentation model and particle image velocimetry, and fuses these features with an improved YOLOv7 model to achieve high-precision lesion detection. Finally, it constructs a 3D vascular model using a single-view depth estimation network and generates a structured report. This improves the accuracy, robustness, and clinical applicability of vascular lesion identification, completing a closed-loop technology chain from multimodal data processing to automated diagnostic report generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 Flowchart of the vascular pathway lesion identification method based on target detection.

[0043] Figure 2 Schematic diagram of the vascular pathway lesion recognition system based on target detection.

[0044] Figure 3 Flowchart for preprocessing of vascular DSA image sequences and ECG signals.

[0045] Figure 4 Flowchart for the cardiac cycle phase partitioning and motion compensation model. DETAILED DESCRIPTION

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0049] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a method for identifying vascular access lesions based on target detection, comprising the following steps:

[0050] S1: synchronously acquire the original vascular DSA image sequence and the original electrocardiogram signal, and perform preprocessing to generate the vascular DSA image sequence and electrocardiogram signal;

[0051] S2.1: Synchronously acquire the original vascular DSA image sequence and original ECG signal;

[0052] Furthermore, the digital subtraction angiography device is set to high-speed acquisition mode, and the electrocardiograph adopts a standard twelve-lead configuration; the original vascular DSA image sequence acquisition and the original electrocardiogram signal recording are clock-aligned through a hardware synchronization signal, and the synchronization pulse signal simultaneously triggers the vascular DSA image sequence acquisition and electrocardiogram signal recording.

[0053] Digital subtraction angiography (DSA) equipment initiates rapid sequence acquisition after contrast agent injection, ensuring coverage of the complete vascular filling process. During acquisition, X-ray tube current and voltage are kept stable to prevent dose fluctuations that could cause image grayscale variations. Raw vascular DSA image sequences are stored in a lossless format, preserving full spatial resolution and dynamic range. Each raw vascular DSA image frame is accurately timestamped, ensuring the required time accuracy for subsequent time series analysis.

[0054] The electrocardiograph continuously records the entire raw vascular DSA image sequence, with the sampling frequency set to meet the Nyquist theorem's requirements for ECG characteristic frequencies. Lead impedance is maintained stable during raw ECG signal acquisition, and signal quality indicators are monitored in real time. Raw ECG signals are stored in floating-point format, preserving their complete voltage amplitude and waveform characteristics. The raw ECG signal record files are linked to the vascular DSA image sequence using a unified time base.

[0055] S2.2: preprocessing the acquired original vascular DSA image sequence and original electrocardiogram signal to generate a vascular DSA image sequence and an electrocardiogram signal;

[0056] Time axis alignment of the original vascular DSA image sequence and the original electrocardiogram signal;

[0057] Furthermore, each frame of the original vascular DSA image sequence is timestamped with millisecond-accurate timestamps, and the original ECG signal sampling points are also timestamped with precise timestamps. A linear interpolation algorithm is used to align the time axes of the original vascular DSA image sequence and the original ECG signal, ensuring that the phase of the original ECG signal corresponding to each frame of the original vascular DSA image precisely matches. This time alignment process preserves the complete frame sequence of the original vascular DSA image sequence and the continuous waveform characteristics of the original ECG signal.

[0058] The aligned original vascular DSA image sequence is converted into a grayscale image, and size normalization, vascular area contrast enhancement, image noise removal, and time domain normalization are performed to generate a vascular DSA image sequence;

[0059] Furthermore, the aligned original vascular DSA image sequence is first converted to grayscale images, and grayscale mapping uses window width and window position adjustment to optimize vascular region display. Size normalization uniformly scales the grayscale images and performs bicubic interpolation while maintaining the aspect ratio. Vascular region contrast enhancement uses an adaptive histogram equalization algorithm, focusing on enhancing the grayscale gradient at the vascular edge. Image noise reduction uses a non-local mean denoising method to suppress random noise while preserving vascular structure. Temporal normalization corrects grayscale consistency across the entire vascular DSA image sequence to eliminate inter-frame brightness fluctuations.

[0060] Band-pass filtering the aligned original electrocardiogram signal to generate an electrocardiogram signal;

[0061] Furthermore, the aligned original ECG signal is subjected to Butterworth bandpass filtering to eliminate baseline drift and high-frequency interference, and the generated ECG signal is stored in a floating-point data format.

[0062] S2: Divide the cardiac cycle phase according to the electrocardiogram signal, calculate the displacement vector field between adjacent frames of the vascular DSA image sequence, build a motion compensation model, and generate a stable DSA sequence;

[0063] S2.1: Use the R-wave detection algorithm to identify the R-wave peak point in the electrocardiogram signal, obtain the cardiac cycle, and divide the cardiac cycle into systole and diastole by phase division;

[0064] Furthermore, the ECG signal undergoes digital differentiation to enhance the R-wave characteristics and locate the R-wave peak. The interval between adjacent R-wave peaks is defined as a complete cardiac cycle, and each cycle is divided into systole and diastole based on cardiac electrophysiological characteristics. Systole begins at the R-wave peak and ends at the T-wave end; diastole extends from the T-wave end to the next R-wave start. The phase segmentation results are stored in a time-stamped file that precisely corresponds to the timestamps of the vascular DSA image series.

[0065] S2.2: Input two adjacent frames of the vascular DSA image sequence into the dense optical flow algorithm to calculate the displacement vector field, which is expressed as:

[0066] ;

[0067] in, is the pixel coordinate in the vascular DSA image sequence ( ) is in The displacement component on the axis, is the pixel coordinate in the vascular DSA image sequence ( ) is in The displacement component on the axis, For vascular DSA image sequences The spatial gradient on the axis, For vascular DSA image sequences The spatial gradient on the axis, For vascular DSA image sequence at time The time gradient on In pixel coordinates ( ) is a local neighborhood window centered on .

[0068] S2.3: Phase classification of the vascular DSA image sequence is performed according to systole and diastole, and a motion compensation model is constructed in combination with the displacement vector field to generate motion compensation parameters;

[0069] Furthermore, based on the ECG R-wave peak detection results, the vascular DSA image sequence was divided into systolic frames (corresponding to the cardiac ejection phase) and diastolic frames (corresponding to the cardiac filling phase). The displacement vector field sets between two adjacent frames in the systolic frame group and the displacement vector field sets between two adjacent frames in the diastolic frame group were extracted, respectively. The temporal noise in the systolic displacement vector field set was treated with a median filter, while the temporal noise in the diastolic displacement vector field set was treated with a mean filter. Both sets were spatially smoothed using a Gaussian convolution kernel. The systolic compensation weight coefficient was determined based on the proportion of the average systolic displacement amplitude in the total displacement amplitude, while the diastolic compensation weight coefficient was determined based on the proportion of the average diastolic displacement amplitude. The filtered systolic displacement vector field set was multiplied by the systolic compensation weight coefficient, and the filtered diastolic displacement vector field set was multiplied by the diastolic compensation weight coefficient. The resulting motion compensation parameter matrix containing the motion compensation parameters was generated by superposition. The motion compensation parameter matrix directly drives the inter-frame alignment and resampling operation of the vascular DSA image sequence, ensuring that the hemodynamic characteristics of the stable sequence are fully preserved.

[0070] S2.4: Based on the motion compensation parameters, image resampling is used to align the vascular DSA image sequences between frames, and a stable DSA sequence is generated after time-domain filtering.

[0071] Furthermore, motion compensation parameters are applied to inter-frame registration of the vascular DSA image sequence, and bilinear interpolation is used for image resampling. The registered vascular DSA image sequence undergoes temporal Gaussian filtering to eliminate residual jitter, generating a spatially consistent, stabilized DSA sequence. This stabilization process preserves the hemodynamic information of the original sequence, ensuring the accuracy of subsequent particle image velocimetry analysis. The resulting stabilized DSA sequence meets clinical accuracy requirements for vascular diameter measurement and stenosis assessment.

[0072] S3: The stable DSA sequence is segmented into a binary vascular mask image using the U-Net segmentation model. The particle image velocimetry algorithm is used to track the motion displacement of contrast agent particles in consecutive frames, generate the blood flow velocity vector field, and calculate the eddy current intensity characteristic map.

[0073] S3.1: A U-Net segmentation model with an encoder-decoder architecture is used to segment the stable DSA sequence into a binary vascular mask image.

[0074] The stabilized DSA sequence is then processed by a U-Net segmentation model, which employs a symmetric encoder-decoder architecture. The encoder consists of four downsampling stages, each employing two 3×3 convolutional layers with ReLU activation, followed by a 2×2 max pooling layer. The decoder consists of four upsampling stages, each employing transposed convolution to amplify the feature map and concatenate it with the feature map from the corresponding encoder layer. The output layer of the U-Net segmentation model uses 1×1 convolution with a sigmoid activation function to generate a probability map with the same resolution as the input image. This probability map is then binarized to produce a binary vascular mask image.

[0075] S3.2: Two consecutive frames of images from the stable DSA sequence and the binary vascular mask image are input into the particle image velocimetry algorithm to identify contrast agent particles;

[0076] Furthermore, two consecutive frames from the stable DSA sequence and a binary vascular mask image are fed into a particle image velocimetry algorithm. The algorithm detects contrast agent particles within the vascular region defined by the binary vascular mask image and uses Gaussian filtering preprocessing to enhance the signal-to-noise ratio of the contrast agent particles.

[0077] S3.3: Calculate the motion displacement of contrast agent particles and generate the blood flow velocity vector field after sub-pixel precision optimization;

[0078] Calculate the motion displacement of contrast agent particles, the expression is:

[0079] ;

[0080] in, Contrast agent particles in Displacement in the direction of motion, Contrast agent particles in Displacement in the direction of motion, and To stabilize two consecutive frames of images in the DSA sequence, is the contrast agent particle tracking window, To stabilize the image in the DSA sequence In the Contrast Agent Particle Tracking window The mean grayscale value of the image within To stabilize the image in the DSA sequence In the Contrast Agent Particle Tracking window The mean grayscale value of the image within ;

[0081] Furthermore, the motion displacement calculation uses cubic spline interpolation to achieve sub-pixel precision positioning, generating high-precision displacement field data. The displacement field data is then median filtered to remove outliers and output a smooth blood flow velocity vector field.

[0082] It should be noted that in the vascular scene, the contrast agent particle tracking window The setting of is directly related to the contour of the vascular region defined by the binary vascular mask image to cope with the constraints of vascular structure and the complexity of hemodynamics. Specifically, the tracking window The size of the tracking window is dynamically adjusted according to the local vessel diameter and curvature; increasing the window size in straight vessel segments improves the signal-to-noise ratio, and reducing the window size in curved vessel segments to avoid geometric distortion. The size of the window is optimized based on the blood flow velocity changes estimated by the initial displacement. In high-velocity blood flow areas, the window size is reduced to minimize motion blur, while in low-velocity blood flow areas, the window size is increased to enhance displacement capture accuracy. The particle image velocimetry algorithm is combined with Gaussian filtering preprocessing. These adjustments ensure accurate calculation of contrast agent particle motion displacement, ultimately generating a highly accurate blood flow velocity vector field.

[0083] For example, in the application scenario of the middle segment of the left anterior descending coronary artery, the blood vessel diameter is 2.8mm, including the straight segment length of 10mm and the curvature radius of the curved segment of 3mm; the contrast agent particle tracking window The initial size of the vascular region, defined based on the binary vascular mask image, was set to a rectangular area with a side length of 4.2 mm × 4.2 mm (1.5 times the vessel diameter). At the 3 mm radius of curvature in the curved section, the window size was reduced to 2.24 mm × 2.24 mm (0.8 times the vessel diameter) and adjusted to an elliptical shape, with the long axis aligned with the tangent of the vessel centerline. Based on the blood flow velocity estimated by the initial displacement, the window size of the high-speed 18 cm / s region was reduced to 2.8 mm × 2.8 mm (1.0 times the vessel diameter), and the window size of the low-speed 5 cm / s region was increased to 5.6 mm × 5.6 mm (2.0 times the vessel diameter). The particle image velocimetry algorithm was combined with Gaussian filter preprocessing to ensure that the displacement calculation error was reduced through these adjustments, ultimately generating a high-precision blood flow velocity vector field.

[0084] S3.4: Calculate the local rotation intensity of the blood flow velocity vector field through the curl operation to generate the eddy current intensity value, which is expressed as:

[0085] ;

[0086] in, is the eddy current intensity value, is the blood flow velocity vector field The direction component, is the blood flow velocity vector field The direction component, is the square root.

[0087] S3.5: After normalizing the eddy current intensity characteristic value, spatially align it with the binary blood vessel mask image to generate an eddy current intensity characteristic map;

[0088] Furthermore, the eddy current intensity eigenvalues ​​are linearly normalized to eliminate data scale differences. The normalization process calculates the scaling ratio based on the global maximum and minimum values ​​of the eddy current intensity eigenvalues ​​and converts the eddy current intensity eigenvalues ​​to a standard range. Then, the normalized eddy current intensity eigenvalues ​​are spatially aligned and initialized with the binary vascular mask image. The registration process uses the anatomical landmarks of the vascular branch points in the binary vascular mask image as a benchmark, specifically including the vascular bifurcation points as key registration reference points. The coordinates of the vascular bifurcation points are automatically identified and extracted through the image processing algorithm. The same anatomical landmarks are located in the eddy current intensity feature map to establish corresponding point pairs. The least squares method is used to optimize the rigid transformation parameters, calculate the translation matrix and rotation angle parameters, and minimize the coordinate position error of the anatomical landmark points. The optimized transformation parameters are applied to the eddy current intensity eigenvalues ​​to perform coordinate transformation. Finally, an eddy current intensity feature map that is completely consistent with the binary vascular mask image space is generated to ensure one-to-one correspondence between pixel coordinates.

[0089] Rigid registration is used to spatially align the eddy current intensity feature map and the binary vascular mask image. This registration process is based on anatomical landmarks of vascular branch points, and the least-squares method is used to optimize the transformation parameters. This alignment ensures that the eddy current intensity feature map and the binary vascular mask image have identical pixel coordinate correspondences, ensuring the accuracy of subsequent multimodal feature fusion.

[0090] S4: Perform channel concatenation on the binary vascular mask image and the eddy current intensity feature map to generate a vascular feature tensor, and generate lesion detection results by improving the YOLOv7 model;

[0091] S4.1: Perform morphological dilation on the binary blood vessel mask image and Gaussian filter smoothing on the eddy current intensity feature map;

[0092] Furthermore, the binary vascular mask image undergoes a morphological dilation operation using a circular structuring element to ensure complete coverage of the vascular edge region. The eddy current intensity feature map is smoothed using a Gaussian kernel, preserving macroscopic blood flow characteristics while suppressing local noise. The processed binary vascular mask image and eddy current intensity feature map maintain consistent spatial resolution and strictly correspond to each other in pixel coordinates.

[0093] S4.2: stitching the processed binary vascular mask image and eddy current intensity feature map into a dual-channel feature map through a deep stitching operation;

[0094] Furthermore, the processed binary vessel mask image is used as the first channel, and the eddy current intensity feature map as the second channel, to generate a dual-channel feature map through a deep stitching operation. The stitching process uses memory mapping technology to ensure processing efficiency even with large amounts of data, and outputs floating-point tensors.

[0095] S4.3: Concatenate the dual-channel feature map with the stable DSA sequence to generate a vascular feature tensor.

[0096] Furthermore, the two-channel feature map is concatenated with the RGB three channels of the stable DSA sequence to generate a five-channel vascular feature tensor.

[0097] S4.4: Add a multimodal feature fusion layer to the standard YOLOv7 model to generate an improved YOLOv7 model;

[0098] Furthermore, a multimodal feature fusion layer is added to the backbone of the standard YOLOv7 model, and a cross-modal attention mechanism is inserted after the ELAN module. This multimodal feature fusion layer calculates the correlation weights between the binarized vessel mask image channel and the eddy current intensity feature map channel, generating an inter-modal feature enhancement map. The head of the improved YOLOv7 model simultaneously outputs lesion bounding box coordinates and functional status classification results, supporting the joint detection of structural and functional stenosis.

[0099] S4.5: Use the multi-task loss function to train the improved YOLOv7 model and perform forward propagation after inputting the vascular feature tensor to generate lesion detection results;

[0100] Furthermore, the improved YOLOv7 model training utilizes a multi-task loss function, including a bounding box regression loss, a classification loss, and a blood flow feature consistency loss. The bounding box regression loss uses the CIoU metric to optimize localization accuracy, the classification loss uses a focal loss function to address class imbalance, and the blood flow feature consistency loss constrains the physical correlation between eddy current intensity features and lesion classification. After the vascular feature tensor is input into the improved YOLOv7 model, it undergoes three stages of feature extraction, fusion, and prediction to output lesion detection results.

[0101] It should be noted that functional status classification refers to the process of distinguishing the properties of vascular abnormalities based on the binary vascular mask image and eddy intensity feature map in the vascular feature tensor by the improved YOLOv7 model. Specific categories include normal (both morphology and blood flow are normal), structural stenosis (vascular morphological abnormalities such as sudden diameter reduction or irregular wall), and functional stenosis (hemodynamic abnormalities such as eddy intensity exceeding the physiological range). The label of each frame of the image is determined by the radiologist annotating the initial bounding box and category (based on the vascular diameter reduction rate threshold and eddy intensity threshold) on the stable DSA sequence, and then combining the eddy intensity feature map generated by the blood flow velocity vector field for spatial alignment verification. Finally, the multi-task loss function is optimized and confirmed in the training of the improved YOLOv7 model to ensure that the label is consistent with the actual nature of the lesion.

[0102] S5: Based on the lesion detection results, a single-view depth estimation network is combined to obtain a pseudo depth map from the stable DSA sequence, which is then mapped to a 3D vascular model to generate a structured report.

[0103] S5.1: Standardize the format of the lesion detection results and input the stable DSA sequence into the single-view depth estimation network to output a pseudo depth map;

[0104] Furthermore, the lesion detection results are first converted into a unified spatial coordinate system, and the bounding box coordinates are converted into a relative position description based on the centerline of the blood vessel to generate standardized lesion information. The detection box parameters output by the improved YOLOv7 model are geometrically corrected to eliminate the impact of projection deformation. The standardized lesion information contains attributes such as position, size, morphology, and functional score. The stable DSA sequence is input into a single-view depth estimation network based on an encoder-decoder architecture. The encoder uses ResNet50 to extract multi-level image features, and the decoder gradually restores the spatial resolution through transposed convolutional layers. The stable DSA sequence is input into the single-view depth estimation network, which outputs a pseudo depth map. Each pixel value represents the depth distance of the vascular structure relative to the imaging plane, and the numerical range is normalized to the range of 0-1.

[0105] It should be noted that the labels for training the single-view depth estimation network are determined through the following process: first, stable DSA sequences with multiple angles of the same blood vessel at intervals of more than 30° are collected, and binary blood vessel mask images of each angle are generated through the U-Net segmentation model; secondly, the motion recovery structure algorithm is used to calculate the camera pose parameters, and the initial point cloud is generated by combining multi-view stereo matching, and then a three-dimensional blood vessel mesh model is constructed through the Poisson surface reconstruction algorithm; then the three-dimensional blood vessel mesh model is orthogonally projected onto a two-dimensional plane along the target view direction, and the shortest distance from the pixel to the model surface is calculated to generate a depth truth map; finally, the depth truth map is bilaterally filtered and contrast enhanced to match its resolution with the stable DSA sequence, generating depth label data for supervised training of the single-view depth estimation network. This process is based on the three-dimensional reconstruction algorithm and is coordinated with the spatial alignment verification step.

[0106] S5.2: After bilateral filtering and contrast enhancement, the pseudo depth map is spatially aligned with the stable DSA sequence for verification.

[0107] Furthermore, the pseudo-depth map undergoes bilateral filtering to suppress noise while preserving vessel edge sharpness. Contrast enhancement employs an adaptive histogram stretching algorithm to emphasize depth differences between vascular branches. The processed pseudo-depth map is then reprojected and verified against a stabilized DSA sequence to ensure that spatial consistency of key anatomical landmarks remains within acceptable limits.

[0108] S5.3: Based on the verified pseudo-depth map and the stable DSA sequence, a 3D vascular model is constructed using a 3D reconstruction algorithm.

[0109] The verified pseudo-depth map and the stabilized DSA sequence are then fed into the 3D reconstruction process. An initial point cloud is generated from the pseudo-depth map using the marching cubes algorithm, and the point cloud is then colored using the grayscale information from the stabilized DSA sequence. A Poisson surface reconstruction algorithm converts the point cloud into a triangular mesh model, preserving the topological structure of the vascular branches. The reconstructed 3D vascular model undergoes Laplace smoothing to eliminate patch noise.

[0110] S5.4: Map the processed lesion detection results to the three-dimensional vascular model and generate a structured report;

[0111] Furthermore, the standardized lesion detection results are mapped onto the surface of the 3D vascular model via perspective projection. This mapping process takes into account the depth information provided by the pseudo-depth map and the characteristics of vascular curvature to ensure accurate location of the lesion. Each lesion is marked as a colored area on the 3D model, with varying degrees of stenosis using a red-yellow color coding. Lesion information within the 3D vascular model is automatically extracted to generate a structured report.

[0112] This embodiment also provides a target detection-based vascular access lesion identification system, comprising: a data acquisition module for synchronously acquiring and preprocessing raw vascular DSA image sequences and raw electrocardiogram signals to generate vascular DSA image sequences and electrocardiogram signals; a phase compensation module for performing cardiac cycle phase division based on the electrocardiogram signals, calculating the displacement vector field between adjacent frames of the vascular DSA image sequence, constructing a motion compensation model, and generating a stable DSA sequence; a particle velocimetry module for segmenting the stable DSA sequence into binary vascular mask images using a U-Net segmentation model, and using a particle image velocimetry algorithm to track the motion displacement of contrast agent particles in consecutive frames, generating a blood flow velocity vector field, and calculating an eddy current intensity feature map; a lesion monitoring module for performing channel cascade on the binary vascular mask image and the eddy current intensity feature map to generate a vascular feature tensor, and generating lesion detection results using an improved YOLOv7 model; and a report generation module for obtaining a pseudo depth map from the stable DSA sequence based on the lesion detection results in combination with a single-view depth estimation network, mapping the pseudo depth map to a three-dimensional vascular model, and generating a structured report.

[0113] This embodiment further provides a computer device applicable to the target detection-based vascular access lesion identification method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the target detection-based vascular access lesion identification method proposed in the above embodiment.

[0114] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0115] This embodiment further provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the program implements the target detection-based vascular access lesion identification method proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0116] In summary, this invention utilizes electrocardiogram-guided motion compensation to generate stable DSA image sequences, combines a U-Net segmentation model with a particle image velocimetry algorithm to extract hemodynamic features, fuses these features with an improved YOLOv7 model to achieve high-precision lesion detection, and finally utilizes a single-view depth estimation network to construct a 3D vascular model and generate a structured report. This improves the accuracy, robustness, and clinical applicability of vascular lesion identification, completing a closed-loop technology chain from multimodal data processing to automated diagnostic report generation.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying vascular access lesions based on target detection, characterized by: include, Synchronously acquiring original vascular DSA image sequences and original electrocardiogram signals, and performing preprocessing to generate vascular DSA image sequences and electrocardiogram signals; The cardiac cycle phase is divided according to the electrocardiogram signal, and the displacement vector field between adjacent frames of the vascular DSA image sequence is calculated to build a motion compensation model and generate a stable DSA sequence. The stable DSA sequence was segmented into a binary vascular mask image using the U-Net segmentation model. The particle image velocimetry algorithm was used to track the motion displacement of contrast agent particles in consecutive frames, generate the blood flow velocity vector field, and calculate the eddy current intensity characteristic map. The binary vascular mask image and the eddy current intensity feature map are channel-concatenated to generate a vascular feature tensor, and the lesion detection results are generated by improving the YOLOv7 model. Based on the lesion detection results, a single-view depth estimation network is combined to obtain a pseudo depth map from the stable DSA sequence, which is mapped to a three-dimensional vascular model to generate a structured report.

2. The method for identifying vascular access lesions based on target detection according to claim 1, characterized in that: The preprocessing is performed to generate a vascular DSA image sequence and an electrocardiogram signal. The specific steps are as follows: Time axis alignment of the original vascular DSA image sequence and the original electrocardiogram signal; The aligned original vascular DSA image sequence is converted into a grayscale image, and size normalization, vascular area contrast enhancement, image noise removal, and time domain normalization are performed to generate a vascular DSA image sequence; The aligned raw ECG signals are bandpass filtered to generate ECG signals.

3. The method for identifying vascular access lesions based on target detection according to claim 2, wherein: The cardiac cycle phase is divided according to the electrocardiogram signal, and the displacement vector field between adjacent frames of the vascular DSA image sequence is calculated. The specific steps are as follows: The R wave detection algorithm is used to identify the R wave peak point in the electrocardiogram signal, obtain the cardiac cycle, and divide the cardiac cycle into systole and diastole by phase division; Two adjacent frames of images in the vascular DSA image sequence are input into the dense optical flow algorithm to calculate the displacement vector field.

4. The method for identifying vascular access lesions based on target detection according to claim 3, wherein: The specific steps of constructing a motion compensation model and generating a stable DSA sequence are as follows: The vascular DSA image sequence is classified into time phases according to the systolic and diastolic phases, and a motion compensation model is constructed based on the displacement vector field to generate motion compensation parameters. Based on motion compensation parameters, vascular DSA image sequences are aligned between frames through image resampling, and stable DSA sequences are generated after time domain filtering.

5. The method for identifying vascular access lesions based on target detection according to claim 4, characterized in that: The stable DSA sequence is segmented into a binary vascular mask image using the U-Net segmentation model, and the particle image velocimetry algorithm is used to track the motion displacement of contrast agent particles in consecutive frames. The specific steps are as follows: A U-Net segmentation model with an encoder-decoder architecture is used to segment the stable DSA sequence into a binary vascular mask image. Two consecutive frames of images in the stable DSA sequence and the binary vascular mask image are input into the particle image velocimetry algorithm to identify contrast agent particles and calculate their motion displacement.

6. The method for identifying vascular access lesions based on target detection according to claim 5, characterized in that: The steps of generating the blood flow velocity vector field and calculating the eddy current intensity characteristic map are as follows: Optimize the motion displacement of contrast agent particles with sub-pixel accuracy to generate blood flow velocity vector field; The local rotation intensity of the blood flow velocity vector field is calculated by curl operation to generate the eddy current intensity value; After the eddy current intensity characteristic value is normalized, it is spatially aligned with the binary blood vessel mask image to generate the eddy current intensity characteristic map.

7. The method for identifying vascular access lesions based on target detection according to claim 6, wherein: The binary blood vessel mask image and the eddy current intensity feature map are channel-concatenated to generate a blood vessel feature tensor. The specific steps are as follows: The binary blood vessel mask image is subjected to morphological dilation processing, and the eddy current intensity feature map is subjected to Gaussian filtering and smoothing processing; The processed binary blood vessel mask image and eddy current intensity feature map are stitched into a dual-channel feature map through a deep stitching operation; The dual-channel feature map is concatenated with the stable DSA sequence to generate a vascular feature tensor.

8. The method for identifying vascular access lesions based on target detection according to claim 7, wherein: The lesion detection results are generated by improving the YOLOv7 model. The specific steps are as follows: Add a multimodal feature fusion layer to the standard YOLOv7 model to generate an improved YOLOv7 model; The improved YOLOv7 model is trained using a multi-task loss function, and the vascular feature tensor is input and forward propagation is performed to generate lesion detection results.

9. The method for identifying vascular access lesions based on target detection according to claim 8, characterized in that: Based on the lesion detection results, the single-view depth estimation network is combined to obtain a pseudo depth map from the stable DSA sequence, and mapped to the three-dimensional vascular model to generate a structured report. The specific steps are as follows: The lesion detection results are formatted and normalized, and the stable DSA sequence is input into the single-view depth estimation network to output a pseudo depth map. After bilateral filtering and contrast enhancement, the pseudo depth map is spatially aligned with the stable DSA sequence for verification. Based on the verified pseudo-depth map and the stable DSA sequence, a 3D vascular model is constructed using a 3D reconstruction algorithm. The processed lesion detection results are mapped to a 3D vascular model to generate a structured report.

10. A target detection-based vascular access lesion identification system, based on the target detection-based vascular access lesion identification method according to any one of claims 1 to 9, characterized in that: include, A data acquisition module is used to synchronously acquire the original vascular DSA image sequence and the original electrocardiogram signal, and perform preprocessing to generate the vascular DSA image sequence and the electrocardiogram signal; Phase compensation module, used to divide the cardiac cycle phase according to the electrocardiogram signal, calculate the displacement vector field between adjacent frames of the vascular DSA image sequence, build a motion compensation model, and generate a stable DSA sequence; The particle velocimetry module is used to segment the stable DSA sequence into a binary vascular mask image using the U-Net segmentation model, and use the particle image velocimetry algorithm to track the motion displacement of contrast agent particles in consecutive frames, generate the blood flow velocity vector field, and calculate the eddy current intensity characteristic map; The lesion monitoring module is used to perform channel cascade on the binary vascular mask image and the eddy current intensity feature map to generate a vascular feature tensor, and then generate lesion detection results by improving the YOLOv7 model; The report generation module is used to obtain a pseudo depth map from a stable DSA sequence based on the lesion detection results in combination with a single-view depth estimation network, and map it to a three-dimensional vascular model to generate a structured report.

Images