DSA coronary artery interlayer detection method and system based on R (2 +1) D-18 neural network

Abstract

The invention discloses a DSA coronary artery interlayer detection method and system based on an R (2 + 1) D-18 neural network. According to the method, the IVUS detection result is used as a gold standard, automatic and accurate detection and blood vessel path visualization of the true and false coronary dissection cavities are realized by decoupling the spatial and temporal characteristics of the DSA image, and the problems that the traditional technology depends on experience, the precision is limited, and dynamic characteristics cannot be captured are solved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical image analysis technology, and specifically relates to a method and system for detecting and visualizing the path of coronary artery dissection using digital subtraction angiography (DSA) based on an R(2+1)D-18 neural network. Background Technology

[0002] Cardiovascular disease has become a major global health burden, and PCI surgery is the mainstream treatment for coronary heart disease. During PCI surgery, the blood vessel wall is prone to tearing due to external forces such as the operation of interventional instruments. Blood will enter between the layers of the blood vessel wall along the tear, forcing the wall structure to separate and form a "false lumen," while the blood vessel lumen that normally carries blood is the "true lumen."

[0003] This "coexistence of true and false lumens" poses a significant challenge to PCI surgery: on the one hand, if instruments such as guidewires or stents are mistakenly inserted into the false lumen during the procedure, it will directly lead to the expansion of the false lumen and aggravate the dissection injury; on the other hand, the presence of the false lumen will interfere with the doctor's judgment of the true course of the blood vessel, significantly increasing the difficulty and risk of the surgery.

[0004] Currently, there are two main types of techniques used in clinical practice to observe coronary artery dissection:

[0005] OCT / IVUS technology: As the "gold standard" for aortic dissection detection, it can provide high-resolution images of the vascular wall structure, but it is an invasive procedure that may aggravate vascular damage. Moreover, the test results are highly dependent on the doctor's operating experience and image interpretation ability, making it difficult to achieve real-time, non-invasive dynamic monitoring during the operation.

[0006] DSA technology: As a commonly used safety imaging tool in PCI surgery, it can present the overall course of blood vessels in real time. However, traditional DSA image analysis relies entirely on the doctor's experience and judgment, and cannot accurately identify the true and false lumen boundaries of the dissection. Furthermore, single-frame static image analysis cannot capture dynamic features such as contrast agent flow and dissection area fluctuations, resulting in limited accuracy in dissection detection and difficulty in meeting the monitoring needs for dynamic evolution of the dissection during surgery.

[0007] In summary, existing technologies suffer from problems such as "high risk of invasive detection, low accuracy of non-invasive detection, and lack of dynamic feature capture." There is an urgent need for a technology that can rely on safe DSA imaging to achieve non-invasive, real-time, and accurate coronary artery dissection detection and path visualization, providing reliable intelligent assessment and decision support for PCI surgery. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a DSA coronary artery dissection detection and path visualization method based on R(2+1)D-18 neural network. Using IVUS detection results as the gold standard, this method achieves automatic and accurate detection of the true and false lumens of coronary artery dissection and visualization of the vascular path by decoupling the spatiotemporal features of DSA images, thus solving the problems of traditional techniques relying on experience, limited accuracy, and inability to capture dynamic features.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a DSA coronary artery dissection detection method based on an R(2+1)D-18 neural network, comprising the following steps:

[0010] (1) Collect DSA coronary angiography videos acquired synchronously during PCI surgery, and collect, label, preprocess and divide the DSA coronary angiography video data based on the IVUS gold standard;

[0011] (2) Construct an R(2+1)D-18 neural network model, and decouple the spatial and temporal features of DSA video through (2+1)D convolution to extract the features of real and fake cavities;

[0012] (3) The model is trained using a multi-task loss function and the Adam optimizer to obtain the best-performing mezzanine detection model;

[0013] (4) Input the DSA video collected in real time during the operation into the trained model to realize the automatic identification and localization of the true and false lumens of the dissection, and reconstruct the vascular path based on the true lumen boundary to output the visualization results.

[0014] Furthermore, the determination of the IVUS gold standard in step (1) adopts the following rules:

[0015] In IVUS imaging, a true lumen is defined as one in which the blood flow direction is consistent with that of the main coronary artery, the lumen is continuous and intact, and clear normal vessel wall structures are visible.

[0016] In IVUS imaging, the lumen is an abnormal blood flow channel, with branches or reversals from the main blood flow direction, interruptions in continuity, and visible torn intima sheets or hematomas surrounding it. When there is no intact normal vessel wall structure, it is called a false lumen.

[0017] Furthermore, the preprocessing in step (1) includes the following steps:

[0018] Gaussian filtering is used to remove background noise from DSA images;

[0019] Histogram equalization enhances the grayscale difference between blood vessels and the background.

[0020] Temporal enhancement processing, including random frame shifting and time-scale scaling, is applied to video sequences to simulate differences in contrast agent flow rates during different surgeries, thereby improving the model's generalization ability.

[0021] Furthermore, the R(2+1)D-18 neural network model in step (2) includes: an input layer, a (2+1)D convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0022] The (2+1)D convolutional layer contains 18 convolutional units. Each convolutional unit first extracts single-frame static features through a 3×3×1 2D spatial convolution, and then extracts inter-frame dynamic features through a 1×1×3 1D temporal convolution. The ReLU activation function is used after each convolution.

[0023] Furthermore, in step (3), the task loss function is a weighted sum of classification loss and localization loss. The classification loss adopts the cross-entropy loss function, and the localization loss adopts the smooth L1 loss function. The weight ratio of classification loss to localization loss is 1:1.5.

[0024] Furthermore, the automatic identification and positioning of the true and false cavities in the interlayer in step (4) specifically adopts the following steps:

[0025] Slice the real-time DSA video during the operation;

[0026] Input the video clip into the model and output the probability map of the true and false cavity categories and the boundary coordinates for each frame;

[0027] The category probability threshold is set to 0.7, and only the region with a probability ≥ 0.7 is retained as the effective true and false cavity region. The boundary coordinates of the true and false cavities in consecutive frames are smoothed. If the boundary offset of the same cavity between adjacent frames exceeds 5 pixels, linear interpolation is used for correction.

[0028] The grade of the interstitial space is assessed based on the ratio of the false cavity area to the true cavity area.

[0029] Furthermore, the vascular path reconstruction based on the true lumen boundary in step (4) includes the following steps:

[0030] The geometric center of the true lumen region in each frame is extracted as the path control point. The control points are connected in chronological order and the three-dimensional true lumen path is reconstructed by combining the DSA spatial coordinate system. The true lumen region, false lumen region, and dissection tear are marked in the navigation interface and the three-dimensional vascular path is overlaid and displayed, while the dissection level information is also displayed.

[0031] A DSA coronary artery dissection detection system based on an R(2+1)D-18 neural network, used to implement the method described in any of the above-mentioned embodiments, characterized in that it includes:

[0032] The data acquisition and preprocessing module is used to collect DSA coronary angiography videos acquired synchronously during PCI surgery. Based on the IVUS gold standard, it performs DSA coronary angiography video data acquisition, annotation, preprocessing and dataset partitioning.

[0033] The R(2+1)D-18 model module is used to construct an R(2+1)D-18 neural network model, which decouples the spatial and temporal features of DSA video through (2+1)D convolution to extract the features of real and fake cavities.

[0034] The model training module is used to train the model using a multi-task loss function and the Adam optimizer to obtain the best-performing mezzanine detection model.

[0035] The automatic detection and visualization output module is used to input the DSA video acquired in real time during the operation into the trained model to realize the automatic identification and localization of the true and false lumens of the dissection, and to reconstruct the vascular path based on the boundary of the true lumen and output the visualization results.

[0036] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0037] An electronic device includes a processor and a memory, the memory storing a computer program, wherein when the processor executes the computer program, it implements the method of any one of claims 1 to 7.

[0038] Beneficial effects

[0039] Compared with the prior art, the present invention has the following significant advantages:

[0040] High detection accuracy: By decoupling the spatiotemporal features of DSA images through the R(2+1)D-18 model, it captures both the static features of vascular structures and the dynamic features of contrast agent flow and dissection fluctuations, breaking through the limitations of traditional single-frame analysis; the model is trained and validated using IVUS as the gold standard. After validation on the test set, the accuracy of identifying true and false cavities is ≥92%, and the boundary positioning error is ≤3 pixels, which is significantly better than the accuracy of traditional DSA manual judgment.

[0041] High safety: It relies on non-invasive DSA imaging to detect aortic dissection, eliminating the need for additional invasive procedures such as OCT / IVUS, thus avoiding further damage to blood vessels and meeting the safety requirements of PCI surgery.

[0042] High real-time performance: The R(2+1)D-18 model reduces computational complexity through spatiotemporal decoupling convolution, and the model inference speed reaches 30fps, which can meet the needs of real-time detection and visualization during surgery without significant delay.

[0043] Highly clinically applicable: It automatically outputs the results of true and false lumen localization, dissection level, and visualized images of vascular pathways, reducing doctors' reliance on experience, assisting doctors in accurately judging the course of blood vessels, reducing the risk of instruments accidentally entering the false lumen, and improving the efficiency and safety of PCI surgery. Detailed Implementation

[0044] To enable those skilled in the art to better understand the present invention, a detailed description will be provided below in conjunction with embodiments of this application. It should be understood that the embodiments described herein are merely one implementation of this application, and not all. For those skilled in the art, any other technical solutions that can be obtained based on the embodiments disclosed in this application without inventive effort should fall within the protection scope of the claims of this application.

[0045] 1. Overview of Technical Solution

[0046] This invention uses IVUS detection results as the gold standard for defining true and false lumens. It constructs an R(2+1)D-18 neural network model to process DSA coronary angiography videos: first, it extracts vascular structural features (such as lumen morphology and dissection contour) from single-frame images through 2D spatial convolution, and then captures temporal dynamic features between consecutive frames (such as differences in contrast agent filling speed and fluctuations in the dissection area) through 1D temporal convolution. Finally, it achieves automatic identification and localization of true and false lumens in the dissection and outputs visualized vascular path results.

[0047] 2. Gold Standard Establishment and Data Preparation

[0048] 2.1 The Gold Standard for Defining True and False Cavities

[0049] Using intravascular ultrasound (IVUS) results as the core diagnostic criterion, combined with morphological and hemodynamic characteristics, the rules for defining true and false lumens are clarified:

[0050] True lumen determination: In IVUS imaging, the lumen is consistent with the blood flow direction of the main coronary artery, continuous and intact, and the normal vessel wall structure (intima, media and adventitia are clearly visible).

[0051] False lumen identification: In IVUS imaging, the lumen is an abnormal blood flow channel, with branches or reversals from the main blood flow direction, interruptions in continuity, and visible torn intimal flaps or hematomas surrounding it, without intact normal vessel wall structures.

[0052] 2.2 Data Preparation

[0053] Data acquisition: Collect DSA coronary angiography video and IVUS image data acquired simultaneously during PCI surgery. The video frame rate is set to 25-30fps, and the resolution of a single frame image is adjusted to 512×512 pixels to ensure that the images cover different branches of the coronary artery and typical scenarios where dissection occurs.

[0054] Data annotation: Based on the above IVUS gold standard, two or more doctors with more than five years of clinical experience in cardiology manually annotated the true and false cavity regions in the DSA video frames. The annotation content included the coordinates of the true and false cavity boundaries and the location of the aortic tear.

[0055] Data augmentation and preprocessing: Preprocessing the labeled DSA video data, including:

[0056] Image denoising: Gaussian filtering is used to remove background noise from DSA images;

[0057] Contrast enhancement: Improves the grayscale difference between blood vessels and the background through histogram equalization;

[0058] Temporal enhancement: Random frame shifts (±2 frames) and temporal scaling (0.8-1.2 times) are applied to video sequences to simulate differences in contrast agent flow rates during different surgeries, thereby improving the model's generalization ability;

[0059] Dataset partitioning: The preprocessed dataset is divided into training, validation, and test sets in a 7:2:1 ratio for model training, parameter optimization, and performance validation.

[0060] 3. Construction of R(2+1)D-18 Neural Network Model

[0061] 3.1 Model Structure Design

[0062] To address the issues of traditional 2D convolution ignoring dynamic features and pure 3D convolution having high spatiotemporal coupling complexity, this invention employs the R(2+1)D-18 model. It models the spatiotemporal features of DSA videos through a "space-time" decoupled convolution method. The model structure includes:

[0063] Input layer: Receives pre-processed DSA video segments. The input format is (batch size, time step, height, width, number of channels), where the time step is set to 8 (i.e., 8 consecutive frames of video are input each time) and the number of channels is 1 (grayscale video).

[0064] (2+1)D convolutional layer: This is the core feature extraction module of the model, containing 18 convolutional units. Each convolutional unit consists of two steps:

[0065] 2D spatial convolution: A 3×3×1 convolution kernel (3×3 in spatial dimension and 1 in temporal dimension) is used to extract spatial features from a single frame image, capturing static features such as vascular lumen morphology and dissection contour. After convolution, nonlinearity is introduced through the ReLU activation function.

[0066] 1D Temporal Convolution: Based on spatial convolution, a 1×1×3 convolution kernel (1×1 in spatial dimension and 3 in temporal dimension) is used to perform temporal convolution on feature maps between consecutive frames to capture dynamic features such as changes in contrast agent filling speed and fluctuations in the interstitial region with blood flow. It is also activated by the ReLU activation function.

[0067] Pooling layer: A 2×2×2 max pooling layer (2×2 spatial dimension, 2 temporal dimension) is set after each (2+1)D convolutional unit to reduce the dimension of the feature map, reduce the amount of computation, and at the same time retain key spatiotemporal features;

[0068] Fully connected layers: Three fully connected layers are set in the back end of the model to map the feature vectors output by the pooling layer to class probabilities (real cavity, fake cavity, background) and position coordinates (boundary between real and fake cavities).

[0069] Output layer: The softmax activation function is used to output the class probabilities of the true and false cavity regions, and the regression loss function is used to output the coordinate values ​​of the boundary between the true and false cavities, thus realizing the joint task of "classification + localization".

[0070] 3.2 Model Training and Optimization

[0071] Loss function design: A multi-task loss function is adopted, including:

[0072] Classification loss: The cross-entropy loss function is used to optimize the model's classification accuracy for real cavities, fake cavities, and background;

[0073] Localization loss: A smooth L1 loss function is used to optimize the model's prediction accuracy of the boundary coordinates of the real and fake cavities; The total loss function is a weighted sum of the classification loss and the localization loss (with a weight ratio of 1:1.5) to ensure that the model takes into account both classification accuracy and localization precision.

[0074] Optimizer selection: The Adam optimizer is used, with an initial learning rate of 1e-4. Every 5 epochs (training rounds), the learning rate is reduced to 0.5 times the original value to avoid overfitting or gradient vanishing during model training.

[0075] Training process: Input the training set data into the model in batches of 16, set the number of training rounds to 50, evaluate the model performance using the validation set after each round of training (using the accuracy of true and false cavity recognition and boundary positioning error as indicators), and save the model parameters with the best performance on the validation set.

[0076] 4. Implementation of Coronary Artery Dissection Detection and Path Visualization

[0077] 4.1 Automatic Detection Process for Mezzanine Layers

[0078] Real-time DSA video input: During PCI surgery, the DSA coronary angiography video acquired in real time during the operation is sliced ​​into continuous video segments according to a time step of 8 frames.

[0079] Model inference: Input the video clip into the trained R(2+1)D-18 model, and the model outputs the class probability map and boundary coordinates of the real cavity and the fake cavity in each frame of the image;

[0080] Post-processing of results: Post-processing of the model output results, including:

[0081] Probability threshold screening: Set the category probability threshold to 0.7, and only retain the region with a probability ≥ 0.7 as the effective true and false cavity region, and exclude low confidence prediction results;

[0082] Temporal consistency verification: The boundary coordinates of the true and false cavities in consecutive frames are smoothed. If the boundary offset of the same cavity between adjacent frames exceeds 5 pixels, linear interpolation is used to correct it to ensure the temporal stability of the detection results.

[0083] Dissection severity assessment: Based on the ratio of the detected false cavity area to the true cavity area, dissections are classified into mild (false cavity area / true cavity area < 0.3), moderate (0.3 ≤ false cavity area / true cavity area < 0.6), and severe (false cavity area / true cavity area ≥ 0.6), providing doctors with a reference for the severity of the dissection.

[0084] 4.2 Visualization of Vascular Pathways

[0085] Based on the coordinates of the true lumen boundary detected by the model, a coronary artery pathway visualization module is constructed:

[0086] True cavity center trajectory extraction: Calculate the pixel coordinates of the true cavity region in each frame of the image, and take the geometric center of the true cavity region as the blood vessel path control point of that frame;

[0087] 3D path reconstruction: Connect the vascular path control points of consecutive frames in chronological order, and combine them with the spatial coordinate system of DSA images (with the coronary artery ostium as the origin) to reconstruct the three-dimensional true lumen path of the coronary artery.

[0088] Visual output: The true lumen area is marked in red, the false lumen area in blue, and the dissection tear in yellow. The three-dimensional vascular path is overlaid on the surgical navigation interface and the dissection level information is updated in real time, providing doctors with intuitive intraoperative navigation reference.

[0089] Example

[0090] Data preparation and gold standard application

[0091] Data Acquisition: Simultaneous DSA video and IVUS imaging data of 200 patients who underwent PCI surgery in our Department of Cardiology from January 2023 to January 2024 were collected. This included 120 cases diagnosed with coronary artery dissection (40 mild, 50 moderate, and 30 severe) and 80 normal cases without dissection. The DSA video frame rate was 25fps with a single frame resolution of 512×512 pixels, and the IVUS image resolution was 400×400 pixels.

[0092] Gold standard annotation: Two chief physicians with 10 years of clinical experience in cardiology annotated the true and false cavities of DSA video frames based on IVUS images. LabelMe was used as the annotation tool. For frames with inconsistent annotation results, a third chief physician arbitrated the results, and finally 150,000 annotated DSA images were obtained.

[0093] Data preprocessing: Gaussian filtering (standard deviation σ=1.0) was used to remove image noise, and histogram equalization was used to improve contrast; random frame shift (±2 frames) and time scale scaling (0.8-1.2 times) were applied to the video sequence to generate 200,000 frames of enhanced data; the data was divided into training set (140,000 frames), validation set (40,000 frames), and test set (20,000 frames) in a 7:2:1 ratio.

[0094] Model training and performance validation

[0095] Model training environment: NVIDIA Tesla V100 GPU, 32GB memory, software environment: Python 3.8, PyTorch 1.10, training batch size set to 16, initial learning rate 1e-4, training epochs 50.

[0096] Performance verification metrics: Accuracy, Sensitivity (true cavity recognition rate), Specificity (false cavity recognition rate), and Mean Absolute Error (MAE) are used as verification metrics.

[0097] Validation results: After training, the performance metrics on the test set are as follows:

[0098] Accuracy: 83.5%;

[0099] Sensitivity: 84.2%;

[0100] Specificity: 82.8%;

[0101] Boundary localization error: 2.3 pixels; Model inference speed is 32fps, which meets the real-time requirements during surgery.

[0102] Intraoperative application process

[0103] Image input: During a patient's PCI procedure, coronary angiography video (25fps) was acquired in real time using a DSA device, and the video signal was transmitted to the intelligent analysis terminal of this invention via the hospital's imaging system.

[0104] Automatic detection: The terminal slices the video into 8 frames / segments, inputs the R(2+1)D-18 model, and the model outputs the true and false cavity detection results for each frame within 0.03 seconds: The patient has a moderate dissection in the left anterior descending artery (false cavity area / true cavity area = 0.45), the true cavity boundary coordinates are (120,150)-(380,320), the false cavity boundary coordinates are (135,165)-(360,300), and the tear is located at (250,240).

[0105] Path visualization: The terminal automatically extracts the center trajectory of the true lumen, reconstructs the three-dimensional vascular path, and overlays it on the surgical navigation interface: the true lumen is marked in red, the false lumen in blue, and the tear in yellow, while displaying the "moderate dissection" prompt information.

[0106] Clinical application: Based on the visualization results, doctors can accurately determine the course of blood vessels and insert the guidewire into the lesion site along the true lumen path, successfully implanting the stent. No instruments were accidentally inserted into the false lumen during the operation, and the operation time was shortened by 15 minutes compared with the traditional method.

[0107] The embodiments described above are merely one approach to the present invention and do not limit the scope of the invention. Within the scope of the claims of the present invention, any equivalent substitutions, technical modifications, or detailed improvements based on the present invention, as long as they do not depart from the core design concept of the present invention, shall fall within the protection scope of the present invention.

Claims

1. A DSA coronary artery dissection detection method based on R(2+1)D-18 neural network, characterized in that... Includes the following steps: (1) Collect DSA coronary angiography videos acquired synchronously during PCI surgery, and collect, label, preprocess and divide the DSA coronary angiography video data based on the IVUS gold standard; (2) Construct an R(2+1)D-18 neural network model, and decouple the spatial and temporal features of DSA video through (2+1)D convolution to extract the features of real and fake cavities; (3) The model is trained using a multi-task loss function and the Adam optimizer to obtain the best-performing mezzanine detection model; (4) Input the DSA video collected in real time during the operation into the trained model to realize the automatic identification and localization of the true and false lumens of the dissection, and reconstruct the vascular path based on the true lumen boundary to output the visualization results.

2. The DSA coronary artery dissection detection method based on R(2+1)D-18 neural network according to claim 1, characterized in that: The determination of the IVUS gold standard in step (1) adopts the following rules: In IVUS imaging, a true lumen is defined as one in which the blood flow direction is consistent with that of the main coronary artery, the lumen is continuous and intact, and clear normal vessel wall structures are visible. In IVUS imaging, the lumen is an abnormal blood flow channel, with branches or reversals from the main blood flow direction, interruptions in continuity, and visible torn intima sheets or hematomas surrounding it. When there is no intact normal vessel wall structure, it is called a false lumen.

3. The DSA coronary artery dissection detection method based on R(2+1)D-18 neural network according to claim 1, characterized in that: The preprocessing in step (1) includes the following steps: Gaussian filtering is used to remove background noise from DSA images; Histogram equalization enhances the grayscale difference between blood vessels and the background. Temporal enhancement processing, including random frame shifting and time-scale scaling, is applied to video sequences to simulate differences in contrast agent flow rates during different surgeries, thereby improving the model's generalization ability.

4. The DSA coronary artery dissection detection method based on R(2+1)D-18 neural network according to claim 1, characterized in that: The R(2+1)D-18 neural network model in step (2) includes: an input layer, a (2+1)D convolutional layer, a pooling layer, a fully connected layer, and an output layer; The (2+1)D convolutional layer contains 18 convolutional units. Each convolutional unit first extracts single-frame static features through a 3×3×1 2D spatial convolution, and then extracts inter-frame dynamic features through a 1×1×3 1D temporal convolution. The ReLU activation function is used after each convolution.

5. The DSA coronary artery dissection detection method based on R(2+1)D-18 neural network according to claim 1, characterized in that: The task loss function in step (3) is a weighted sum of classification loss and localization loss. The classification loss adopts the cross-entropy loss function, and the localization loss adopts the smooth L1 loss function. The weight ratio of classification loss to localization loss is 1:1.

5.

6. The DSA coronary artery dissection detection method based on R(2+1)D-18 neural network according to claim 1, characterized in that: The automatic identification and positioning of the true and false cavities in the interlayer in step (4) specifically adopts the following steps: Slice the real-time DSA video during the operation; Input the video clip into the model and output the probability map of the true and false cavity categories and the boundary coordinates for each frame; The category probability threshold is set to 0.7, and only the region with a probability ≥ 0.7 is retained as the effective true and false cavity region. The boundary coordinates of the true and false cavities in consecutive frames are smoothed. If the boundary offset of the same cavity between adjacent frames exceeds 5 pixels, linear interpolation is used for correction. The grade of the interstitial space is assessed based on the ratio of the false cavity area to the true cavity area.

7. The DSA coronary artery dissection detection method based on R(2+1)D-18 neural network according to claim 1, characterized in that: The vascular path reconstruction based on the true lumen boundary in step (4) includes the following steps: The geometric center of the true lumen region in each frame is extracted as the path control point. The control points are connected in chronological order and the three-dimensional true lumen path is reconstructed by combining the DSA spatial coordinate system. The true lumen region, false lumen region, and dissection tear are marked in the navigation interface and the three-dimensional vascular path is overlaid and displayed, while the dissection level information is also displayed.

8. A DSA coronary artery dissection detection system based on an R(2+1)D-18 neural network, used to implement the method according to any one of claims 1 to 7, characterized in that... include: The data acquisition and preprocessing module is used to collect DSA coronary angiography videos acquired synchronously during PCI surgery. Based on the IVUS gold standard, it performs DSA coronary angiography video data acquisition, annotation, preprocessing and dataset partitioning. The R(2+1)D-18 model module is used to construct an R(2+1)D-18 neural network model, which decouples the spatial and temporal features of DSA video through (2+1)D convolution to extract the features of real and fake cavities. The model training module is used to train the model using a multi-task loss function and the Adam optimizer to obtain the best-performing mezzanine detection model. The automatic detection and visualization output module is used to input the DSA video acquired in real time during the operation into the trained model to realize the automatic identification and localization of the true and false lumens of the dissection, and to reconstruct the vascular path based on the boundary of the true lumen and output the visualization results.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method of any one of claims 1 to 7.