A high-frequency electrocardiogram coronary culprit vessel positioning method and system

Abstract

The application relates to the technical field of intelligent medical treatment, and discloses a high-frequency electrocardiogram coronary culprit vessel positioning method and system. The construction process of the positioning model used by the method comprises the following steps: extracting multi-lead images from an electrocardiogram image; performing self-supervised learning pre-training on the positioning model; randomly selecting part of the lead images to perform lead-level mask, performing time sequence mask along a time axis in the unmasked lead image to generate a masked image; implementing a space-time dependence modeling process on the masked image; obtaining a reconstructed image through a reconstruction head based on the context representation corresponding to the masked image, and minimizing the reconstruction loss; performing supervised learning on the positioning model to output positive probabilities of different coronaries; and calculating a patient comprehensive blockage judgment based on the positive probabilities. The application supports a scenario with only image carriers, can complete lead-level structured reconstruction and time alignment without original electrocardiogram waveforms, and significantly improves the usability of historical data and grassroots equipment data.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, specifically to a method and system for locating culprit vessels in coronary arteries using high-frequency electrocardiography. Background Technology

[0002] In clinical practice, a large amount of high-frequency electrocardiogram (ECG) data is generated from resting or stress testing procedures. A significant portion of this data is stored only as images, and the original time-series signals are not always available. To accurately locate the culprit vessels (such as the left anterior descending artery, left circumflex artery, and right coronary artery) in patients with coronary artery disease, it is necessary to collaboratively model translead relationships and time dependencies. Current technologies have not yet established a standardized image-to-time reconstruction and alignment process, and lack sufficient joint characterization of high-frequency components and translead dependencies. Furthermore, existing lead information combinations are mostly simple concatenation or outer products, or involve separate discrimination by lead followed by voting, which fails to meet the need for a shared understanding of "time-lead" and integrated localization output (both coronary and patient-level). Summary of the Invention

[0003] To address the aforementioned technical issues, this invention provides a method and system for locating culprit vessels in coronary arteries using high-frequency electrocardiography (ECG). It proposes an integrated framework of "high-frequency ECG image-to-temporal reconstruction—translead alignment—shared representation—multi-lead discrimination": This framework enables structured reconstruction in real-world scenarios where only image data is available. It effectively aligns multi-lead information within a unified representation space and outputs the location results and patient-level comprehensive judgments for the left anterior descending artery (LAD), left circumflex artery (LCX), and right coronary artery (RCA) based on shared representations. This improves the accuracy and clinical usability of culprit vessel location even in the absence of original temporal signals.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a method for locating culprit vessels in coronary arteries using high-frequency electrocardiography, wherein the process of constructing the localization model includes:

[0006] Extracting multi-lead images from electrocardiogram (ECG) images;

[0007] The localization model undergoes self-supervised pre-training: randomly selected lead images are masked at the lead level, and temporal masking is performed along the time axis in the unmasked lead images to generate mask images; a spatiotemporal dependency modeling process is implemented on the mask images: each mask image is segmented along the time axis to form a cross-lead aligned segment sequence, each segment is encoded into a vector, and lead position embedding and temporal position embedding are added before being input into the encoder. Spatiotemporal dependency modeling is performed through spatial attention and temporal attention to obtain a contextual representation; based on the contextual representation corresponding to the mask images, a reconstructed image is obtained through a reconstruction head, and the reconstruction loss is minimized;

[0008] Supervised learning is performed on the localization model. Multi-scale morphological features and edge features are extracted from each lead image and adaptively fused with the original lead image to obtain a feature-enhanced image. The spatiotemporal dependency modeling process is implemented on the feature-enhanced image, and the context representation corresponding to the obtained feature-enhanced image is passed through a linear classifier to output the positive probability of different coronary arteries.

[0009] The patient's overall blockage is determined based on the positive probability.

[0010] In one embodiment, the extraction of multi-lead images from electrocardiogram images specifically includes:

[0011] Multi-lead images are extracted from the patient's electrocardiogram images and de-identified. The multi-lead images contain multiple lead images. The lead images are organized into a lead image sequence according to a standard order, and the orientation, resolution and pixel coordinate system of the lead images are unified so that the horizontal direction corresponds to the time axis and the vertical direction corresponds to the amplitude.

[0012] In one embodiment, the random selection of a portion of the lead image for lead-level masking specifically includes:

[0013] For each patient's multi-lead images, a set of lead image indices is randomly selected. To perform lead-level masking, the image blocks corresponding to the masked leads are zeroed out or replaced with mask marker patches. Simultaneously, the mask marker matrix is ​​recorded, generating a lead-level mask. :

[0014] ;

[0015] in, For the first Index of lead images, express exist Value at position, This represents the pixel index in the height direction of the lead image. Indicates the pixel index along the width direction of the lead image;

[0016] The step of performing temporal masking along the time axis in the unmasked lead image to generate a masked image specifically includes:

[0017] In each lead image that is not masked at the lead level, samples are taken along the time axis. consecutive segments To perform occlusion, a temporal mask is formed. :

[0018] ;

[0019] in, For the first A continuous time-occlusion interval, The first The start and end coordinates of a continuous segment. express exist The value at the position;

[0020] mask image ;in, It represents the Hadamah accumulation. Indicates the first One lead image.

[0021] In one embodiment, segmenting each mask image along the time axis to form a transconductor aligned sequence of segments specifically includes:

[0022] For each mask image By dividing the sequence along the time axis at equal steps, the following fragment sequences are obtained:

[0023] ;

[0024] Indicates the first The sequence of segments obtained by segmenting the mask image corresponding to each lead image. express The first in A segment, This represents the index at time t. Indicates the total length of the segment sequence. Indicates the first Index of each lead image.

[0025] In one embodiment, encoding each segment into a vector and adding lead position embedding and temporal position embedding before inputting it into the encoder, and performing spatiotemporal dependency modeling through spatial attention and temporal attention, specifically includes:

[0026] Each segment is passed through an embedding function Encode as a vector In addition to the first Embedded lead positions of individual lead images Embedded with the time position at time t , obtained the The image of the first lead at the first lead Input representation at each time step ; Indicates the first The mask image corresponding to the nth lead image is obtained by segmenting the nth lead image. A segment; will Feed into Transformer encoder Obtain contextual representation The spatial attention in the Transformer encoder uses segments at the same time on multiple leads as lexical units to model inter-lead correlation, while the temporal attention uses segments of the same lead in adjacent time slices as lexical units to model temporal dependence.

[0027] In one embodiment, obtaining the reconstructed image through the reconstruction head and minimizing the reconstruction loss specifically includes:

[0028] The context representation Through multilayer perceptron Output reconstructed image and the original image In contrast, the reconstruction loss based on mean square error was minimized. To learn:

[0029] ;

[0030] in, For the first Index of lead images, For the first Lead-level masking for each lead image. For the first Temporal masking of individual lead images, It is the L2 norm. It represents the Hadamardi (or Hadama) stack.

[0031] In one embodiment, the step of extracting multi-scale morphological features and edge features from each lead image and performing adaptive feature fusion with the original lead image to obtain a feature-enhanced image specifically includes:

[0032] Using convolution kernels with several different receptive fields Convolution is performed on the lead images to obtain feature maps. , , The kernel size is determined by convolution, and the data is concatenated along the channel dimension to obtain multi-scale morphological features. ,in, This represents the convolution operation. Indicates a splicing operation;

[0033] Apply the Laplacian operator to the lead image to extract edge features. :

[0034] ;

[0035] Multi-scale morphological features and edge features are stitched together with the lead image in the channel dimension, and then processed through a multilayer perceptron using the Softmax function. Obtain normalized feature weights :

[0036] ;

[0037] These normalized weights are used to linearly weight and aggregate information between features to obtain enhanced features. :

[0038] ;

[0039] Enhanced features With lead images By fusing through gating, a feature-enhanced image is obtained. :

[0040] ;

[0041] in, For weight fusion.

[0042] In one embodiment, the step of passing the context representation corresponding to the obtained feature-enhanced image through a linear classifier to output the positive probability of different coronary arteries specifically includes:

[0043] Contextual representation The positive probability of the left anterior descending branch is output using a trihead linear discriminant. The positive probability of levorotatory branch The probability of a positive result for the right coronary artery. :

[0044] ;

[0045] in, For the discriminator weight parameters, For bias, Indicating transposition, the positive probabilities of the left anterior descending artery, the left circumflex artery, and the right coronary artery are collectively denoted as: , LAD is the left anterior descending artery, LCX is the left circumflex artery, and RCA is the right coronary artery;

[0046] During supervised learning, the binary cross-entropy loss is minimized. Come and learn:

[0047] ;

[0048] in, These are the positive label weights and negative label weights, used to mitigate the differences in sample proportions across different coronary arteries; For the first The label value of each sample.

[0049] In one embodiment, the calculation of the patient's overall blockage based on the positive probability specifically includes:

[0050] First, determine the positive probability for each coronary artery. Thresholding is performed to obtain the predicted coronary artery-level label. :

[0051] ;

[0052] in, For indicator functions, As a threshold, when hour, =1, otherwise =0; Including coronary labeling of the left anterior descending artery The coronary artery grade label of the left circumflex branch Coronary label of the right coronary artery ;

[0053] Based on the coronary artery-level labels of the patient's three coronary arteries, output a comprehensive patient-level label. :

[0054] .

[0055] In a second aspect, the present invention provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any embodiment of the first aspect.

[0056] Compared with the prior art, the beneficial technical effects of the present invention are:

[0057] This invention supports scenarios with only image carriers, enabling lead-level structured reconstruction and time alignment without requiring the original ECG waveform, significantly improving the usability of historical data and data from primary equipment.

[0058] This invention introduces self-supervised reconstruction pre-training, which learns basic ECG representations without relying on additional labels, reducing annotation costs, and providing a more robust initialization for subsequent supervised fine-tuning.

[0059] This invention employs multi-scale convolution and Laplacian edge detection in parallel to extract morphology and edge cues. Combined with adaptive feature fusion and gating decision, it can automatically adjust weights for different image qualities and layouts, enhancing sensitivity and robustness to subtle waveform changes.

[0060] This invention is based on Transformer-based translead-temporal joint modeling, which combines lead and temporal position encoding to complete the characterization of dependencies in a unified representation space, thereby improving the ability to locate and discriminate criminal blood vessels. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0062] Figure 2 This is a flowchart illustrating the modeling process in an embodiment of the present invention. Detailed Implementation

[0063] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0064] like Figure 1 As shown, the high-frequency electrocardiogram coronary culprit vessel localization method of the present invention includes the following steps in the construction process of the localization model:

[0065] S1, extract multi-lead images from electrocardiogram images;

[0066] S2, perform self-supervised pre-training of the localization model: randomly select some lead images for lead-level masking, perform temporal masking along the time axis in the unmasked lead images to generate mask images; perform spatiotemporal dependency modeling on the mask images: segment each mask image along the time axis to form a cross-lead aligned segment sequence, encode each segment into a vector and add lead position embedding and temporal position embedding before inputting it into the encoder, perform spatiotemporal dependency modeling through spatial attention and temporal attention to obtain contextual representation; based on the contextual representation corresponding to the mask images, obtain the reconstructed image through the reconstruction head and minimize the reconstruction loss;

[0067] S3. Supervised learning is performed on the localization model. Multi-scale morphological features and edge features are extracted from each lead image and adaptively fused with the original lead image to obtain a feature-enhanced image. The spatiotemporal dependency modeling process is implemented on the feature-enhanced image. The context representation corresponding to the obtained feature-enhanced image is passed through a linear classifier to output the positive probability of different coronary arteries.

[0068] S4, Calculate the patient's overall blockage judgment based on the positive probability.

[0069] The objective of this invention is to automatically provide a binary classification of the three coronary arteries (LAD / LCX / RCA) in real-world scenarios where only high-frequency electrocardiogram images are available and the original waveforms may not be accessible, and to calculate a patient-level comprehensive occlusion assessment for comparison with clinical gold standards and statistical analysis. The overall processing and modeling workflow is as follows: Figure 2As shown, the dashed lines of the branches represent the self-supervised learning pre-training process, while the solid lines represent the supervised learning process.

[0070] 1. Lead image extraction and desensitization:

[0071] In real-world clinical settings, many high-frequency electrocardiograms are archived only as static images. Therefore, the primary task is to extract the patient's corresponding 12-lead images completely and accurately, and then perform de-identification processing (masking names, hospital numbers, examination numbers, dates, and all other identifiable information) to meet privacy and compliance requirements. The extracted lead images need to be organized into a lead image sequence according to a standard order. The 12 leads include lead I, lead II, lead III, aVR, lead aVL, lead aVF, lead V1, lead V2, lead V3, lead V4, lead V5, and lead V6. The orientation, resolution, and pixel coordinate system of the images are unified, so that the horizontal direction strictly corresponds to the time axis and the vertical direction corresponds to the amplitude, thus laying the baseline for subsequent image-to-time sequence segmentation, alignment, and modeling.

[0072] 2. Self-supervised learning pre-training – lead-level masking:

[0073] The 12 leads show strong correlation in ECG vector projection; the same pathological change (such as abnormal high-frequency components caused by ischemia / scarring in the blood supply area) will present similar or complementary morphological cues within a specific lead cluster. Therefore, a set of lead images is randomly selected during the self-supervised phase. By performing full-lead occlusion, the model reconstructs the occluded lead image from the remaining visible lead images. This is equivalent to maximizing the predictable mutual information between leads, forcing the encoder to learn the cross-lead coupling structure instead of memorizing specific pixels. In implementation, the image patch corresponding to the occluded lead image can be zeroed out or replaced with a mask marker patch, while the mask marker matrix is ​​recorded for use in the reconstruction head's supervision signal. Specifically, this invention generates the following lead-level mask. :

[0074] ;

[0075] in, This is the set of indexes for lead images that require lead-level masking.

[0076] 3. Self-supervised learning pre-training – temporal mask:

[0077] Simply performing lead-level masking allows the model to primarily learn the correspondences between different leads; to simultaneously enhance the understanding of temporal dependencies within a single lead, several consecutive segments are sampled along the time axis from the lead images that are not masked by the entire lead. To perform occlusion, a timing mask is formed. :

[0078] ;

[0079] in, For the first A continuous time occlusion interval (column coordinates). For the first The starting coordinates of a continuous segment For the first The termination coordinates of each continuous segment. Unlike pixel-by-pixel random loss, segmental occlusion more closely resembles the loss patterns in real-world recordings, such as electrode detachment, motion artifacts, and page overlay, and better tests the model's temporal context recovery capabilities. The final input of the self-supervised method is the mask image resulting from the combined action of the lead-level mask and the temporal mask. :

[0080] ;

[0081] in, It represents the Hadamardi (or Hadama) stack.

[0082] 4. Divide along the time axis:

[0083] For each mask image By dividing the sequence along the time axis at equal steps, the following fragment sequences are obtained:

[0084] , ;

[0085] The window width is 16 pixels and the stride is 8 pixels. Simultaneously with the segmentation, a unified time index is used. Cross-lead alignment is then completed to form a two-dimensional sequence raster of "lead × time". The calculation method for the feature-enhanced image is given later.

[0086] 5. Spatiotemporal dependency modeling based on Transformer:

[0087] Each segment By embedding functions Encode as a vector:

[0088] ;

[0089] In addition to the lead position embedding Embedded with time and location The input representation is obtained as follows:

[0090] ;

[0091] Input representation Feed into the Transformer encoder Obtaining contextual representation:

[0092] ;

[0093] The Transformer encoder contains two types of attention:

[0094] Spatial attention (along the lead dimension) uses 12 lead blocks in the same time slice as lexical units to model the correlation between precordial leads and limb leads, as well as mirror lead pairs, and learns the spatial coupling of blood supply patterns.

[0095] Temporal attention (along the time dimension) uses blocks of the same lead in adjacent time slices as terms to capture the evolution of periodic and short-term anomalies (high-frequency spikes, QRS detail changes).

[0096] Dual attention employs alternating stacking, resulting in contextual representations. To ensure consistency in representation between self-supervision and supervision tasks for subsequent reconstruction and classification sharing.

[0097] In this invention, time-axis segmentation and Transformer-based spatiotemporal dependency modeling are two processes: self-supervised learning pre-training and supervised learning sharing. In self-supervised learning pre-training, for the masked image... The process of performing time-axis block segmentation and Transformer-based spatiotemporal dependency modeling; the process of performing time-axis block segmentation and Transformer-based spatiotemporal dependency modeling for feature-enhanced images during self-supervised learning pre-training.

[0098] 6. Self-supervised learning pre-training – Head reconstruction:

[0099] During the pre-training phase, the context representation output by the Transformer encoder is... Through multilayer perceptron Output reconstructed image and compared with the original lead image. In contrast, the reconstruction loss based on mean square error was minimized. To learn:

[0100] ;

[0101] in, It uses the L2 norm. The occluded portion is completed using images from unoccluded leads and the context of time slices, thereby driving the encoder to learn a reducible structure across leads and across time. The weights after pre-training are used as initialization for supervised learning, which significantly accelerates convergence and improves the robustness of downstream discrimination with a small number of annotations (reducing sensitivity to the quality of individual leads).

[0102] 7. Supervised Learning – Multi-scale Feature Extraction and Edge Detection:

[0103] During the labeled supervised learning phase, two classes of interpretable low-level features are extracted in parallel for each lead image:

[0104] Multi-scale convolutional morphological features: using convolutional kernels with several different receptive fields. Convolve the image to obtain and splicing in the channel dimension ,in, This represents the convolution operation. This indicates a stitching operation. This process can explicitly capture patterns of high-frequency QRS texture at different time scales. Smaller kernels focus on sharp details (fragmentation, jaggedness), while larger kernels cover contour changes of the entire complex wave, which helps to identify morphological distortions related to the blood supply area.

[0105] Laplacian edge features: Applying the Laplacian operator to the image to enhance abrupt transitions (equivalent to second-order difference) and extracting edge features. :

[0106] ;

[0107] This processing can highlight the rising and falling edges of the QRS wave and fine streaks, which are particularly sensitive to high-frequency anomalies.

[0108] 8. Supervised Learning – Feature Fusion:

[0109] Multi-scale morphological features and edge features are stitched together with the original lead image in the channel dimension, and then processed through a multilayer perceptron and a function is used. Obtain the normalized feature weights:

[0110] ;

[0111] These normalized weights are used to linearly weight and aggregate information between features to obtain enhanced features. :

[0112] ;

[0113] Based on image quality, noise, and layout differences, more reliable feature branches can be adaptively emphasized (e.g., edge branches with higher weights are given to those with more reliable edges in certain cases). Subsequently... Further comparison with the original lead images Image fusion through gating, feature enhancement :

[0114] ;

[0115] in, To integrate weights, This ensures that the weight value is between 0 and 1. This operation preserves the interpretable shape of the original image while injecting robust reinforcing cues, avoiding overfitting caused by single-path distortion.

[0116] 9. Supervised Learning – Output of a Tri-head Linear Classifier:

[0117] Finally, the supervised learning will be based on feature-enhanced images. The context representation obtained through time-axis segmentation and Transformer-based spatiotemporal dependency modeling The positive probability of the left anterior descending branch is output through a three-headed linear discriminant. The positive probability of levorotatory branch The probability of a positive result for the right coronary artery. :

[0118] ;

[0119] in, For the discriminator weight parameters, For bias, Indicates transpose. LAD is the left anterior descending artery, LCX is the left circumflex artery, and RCA is the right coronary artery.

[0120] During supervised learning, the binary cross-entropy loss is minimized. Come and learn:

[0121] ;

[0122] Label weights are used to mitigate differences in sample proportions across different coronary arteries (e.g., LAD positivity is relatively more common in some data). Tri-head decoupling facilitates learning specific criteria for each blood supply zone, avoids mutual interference, and maintains a consistent understanding across leads and time by sharing the backbone.

[0123] 10. Calculate the patient's overall obstruction assessment:

[0124] The inference phase first calculates the positive probability for each coronary artery. Thresholding is performed to obtain the predicted coronary artery-level labels. :

[0125] ;

[0126] in, This is an indicator function, by default. .

[0127] Finally, based on the determination of the three coronary arteries, a comprehensive patient-level label is output. :

[0128] ;

[0129] This corresponds to the clinical interpretation that "if any criminal's blood vessel is positive, the patient's level is positive."

[0130] 11. Comparison and statistical analysis with clinical gold standards:

[0131] Using the degree of blockage and stenosis recorded in clinical records as the gold standard, "greater than 70% or clear occlusion" is mapped as a positive label 1, and the rest as a negative label 0.

[0132] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0133] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0134] In one embodiment, a computer system is provided, which may be a server. The computer system includes a processor, memory, and a network interface 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, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data used in the methods described above. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the methods described above.

[0135] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0136] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0137] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for locating culprit vessels in the coronary arteries using high-frequency electrocardiography, characterized in that, The process of constructing the positioning model used includes: Extracting multi-lead images from electrocardiogram (ECG) images; The localization model undergoes self-supervised pre-training: randomly selected lead images are masked at the lead level, and temporal masking is performed along the time axis in the unmasked lead images to generate mask images; a spatiotemporal dependency modeling process is implemented on the mask images: each mask image is segmented along the time axis to form a cross-lead aligned segment sequence, each segment is encoded into a vector, and lead position embedding and temporal position embedding are added before being input into the encoder. Spatiotemporal dependency modeling is performed through spatial attention and temporal attention to obtain a contextual representation; based on the contextual representation corresponding to the mask images, a reconstructed image is obtained through a reconstruction head, and the reconstruction loss is minimized; Supervised learning is performed on the localization model. Multi-scale morphological features and edge features are extracted from each lead image and adaptively fused with the original lead image to obtain a feature-enhanced image. The spatiotemporal dependency modeling process is implemented on the feature-enhanced image, and the context representation corresponding to the obtained feature-enhanced image is passed through a linear classifier to output the positive probability of different coronary arteries. The patient's overall blockage is determined based on the positive probability.

2. The method for locating the culprit vessel in coronary artery using high-frequency electrocardiography according to claim 1, characterized in that, The extraction of multi-lead images from electrocardiogram images specifically includes: Multi-lead images are extracted from the patient's electrocardiogram images and de-identified. The multi-lead images contain multiple lead images. The lead images are organized into a lead image sequence according to a standard order, and the orientation, resolution and pixel coordinate system of the lead images are unified so that the horizontal direction corresponds to the time axis and the vertical direction corresponds to the amplitude.

3. The method for locating the culprit vessel in coronary artery using high-frequency electrocardiography according to claim 1, characterized in that, The random selection of a portion of the lead images for lead-level masking specifically includes: For each patient's multi-lead images, a set of lead image indices is randomly selected. To perform lead-level masking, the image blocks corresponding to the masked leads are zeroed out or replaced with mask marker patches. Simultaneously, the mask marker matrix is ​​recorded, generating a lead-level mask. : ； in, For the first Index of lead images, express exist Value at position, This represents the pixel index in the height direction of the lead image. Indicates the pixel index along the width direction of the lead image; The step of performing temporal masking along the time axis in the unmasked lead image to generate a masked image specifically includes: In each lead image that is not masked at the lead level, samples are taken along the time axis. consecutive segments To perform occlusion, a temporal mask is formed. : ； in, For the first A continuous time-occlusion interval, The first The start and end coordinates of a continuous segment. express exist The value at the position; mask image ;in, It represents the Hadamah accumulation. Indicates the first One lead image.

4. The method for locating the culprit vessel in coronary artery using high-frequency electrocardiography according to claim 1, characterized in that, The segmentation of each mask image along the time axis to form a cross-lead aligned segment sequence specifically includes: For each mask image By dividing the sequence along the time axis at equal steps, the following fragment sequences are obtained: ； Indicates the first The sequence of segments obtained by segmenting the mask image corresponding to each lead image. express The first in A segment, This represents the index at time t. Indicates the total length of the segment sequence. Indicates the first Index of each lead image.

5. The method for locating the culprit vessel in coronary artery using high-frequency electrocardiography according to claim 1, characterized in that, The process of encoding each segment into a vector, adding lead position embedding and temporal position embedding, and then inputting it into the encoder, followed by spatiotemporal dependency modeling through spatial attention and temporal attention, specifically includes: Each segment is passed through an embedding function Encode as a vector In addition to the first Embedded lead positions of individual lead images Embedded with the time position at time t , obtained the The image of the first lead at the first lead Input representation at each time step ; Indicates the first The mask image corresponding to the nth lead image is obtained by segmenting the nth lead image. A segment; will Feed into Transformer encoder Obtain contextual representation The spatial attention in the Transformer encoder uses segments at the same time on multiple leads as lexical units to model inter-lead correlation, while the temporal attention uses segments of the same lead in adjacent time slices as lexical units to model temporal dependence.

6. The method for locating the culprit vessel in coronary artery using high-frequency electrocardiography according to claim 1, characterized in that, The process of obtaining a reconstructed image through a reconstruction head and minimizing reconstruction loss specifically includes: The context representation Through multilayer perceptron Output reconstructed image and the original image In contrast, the reconstruction loss based on mean square error was minimized. To learn: ； in, For the first Index of lead images, For the first Lead-level masking for each lead image. For the first Temporal masking of individual lead images, It is the L2 norm. It represents the Hadamardi (or Hadama) stack.

7. The method for locating the culprit vessel in coronary artery using high-frequency electrocardiography according to claim 1, characterized in that, The process of extracting multi-scale morphological and edge features from each lead image and then adaptively fusing them with the original lead image to obtain a feature-enhanced image specifically includes: Using convolution kernels with several different receptive fields Convolution is performed on the lead images to obtain feature maps. , , The kernel size is determined by convolution, and the data is concatenated along the channel dimension to obtain multi-scale morphological features. ,in, This represents the convolution operation. Indicates a splicing operation; Apply the Laplacian operator to the lead image to extract edge features. : ； Multi-scale morphological features and edge features are stitched together with the lead image in the channel dimension, and then processed through a multilayer perceptron using the Softmax function. Obtain normalized feature weights : ； These normalized weights are used to linearly weight and aggregate information between features to obtain enhanced features. : ； Enhanced features With lead images By fusing through gating, a feature-enhanced image is obtained. : ； in, For weight fusion.

8. The method for locating the culprit vessel in coronary artery using high-frequency electrocardiography according to claim 1, characterized in that, The context representation corresponding to the obtained feature-enhanced image is passed through a linear classifier to output the positive probability of different coronary arteries, specifically including: Contextual representation The positive probability of the left anterior descending branch is output using a trihead linear discriminant. The positive probability of levorotatory branch The probability of a positive result for the right coronary artery. : ； in, For the discriminator weight parameters, For bias, Indicating transposition, the positive probabilities of the left anterior descending artery, the left circumflex artery, and the right coronary artery are collectively denoted as: , LAD is the left anterior descending artery, LCX is the left circumflex artery, and RCA is the right coronary artery; During supervised learning, the binary cross-entropy loss is minimized. Come and learn: ； in, These are the positive label weights and negative label weights, used to mitigate the differences in sample proportions across different coronary arteries; For the first The label value of each sample.

9. A method for locating culprit coronary arteries using high-frequency electrocardiography according to claim 8, characterized in that, The calculation of the patient's comprehensive blockage judgment based on the positive probability specifically includes: First, determine the positive probability for each coronary artery. Thresholding is performed to obtain the predicted coronary artery-level label. : ； in, For indicator functions, As a threshold, when hour, =1, otherwise =0; Including coronary labeling of the left anterior descending artery The coronary artery grade label of the left circumflex branch Coronary label of the right coronary artery ; Based on the coronary artery-level labels of the patient's three coronary arteries, output a comprehensive patient-level label. : 。 10. A computer system comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

Images