A method and system for detecting abnormalities in cardiac ultrasound data

Abstract

This invention discloses a method and system for anomaly detection in cardiac ultrasound data, relating to the field of ultrasound data monitoring. The method includes acquiring inter-frame dynamic feature values ​​and image embedding feature values ​​for each frame; identifying each systolic-diastolic fluctuation cycle in the original cardiac ultrasound image sequence; determining whether to enter the rhythm anomaly judgment process based on the structural cycle deviation judgment result; and implementing corresponding operations based on the rhythm anomaly judgment result. This invention effectively improves rhythm stability and anomaly recognition accuracy by integrating dual-channel feature extraction and intelligent cycle recognition mechanisms, employing a two-level strategy of structural deviation and rhythmic fluctuation for anomaly judgment.

Description

Technical Field

[0001] This invention relates to the field of ultrasound data monitoring, specifically to a method and system for detecting abnormalities in cardiac ultrasound data. Background Technology

[0002] Echocardiography, as a non-invasive, real-time imaging medical diagnostic tool, has been widely used for structural assessment and functional diagnosis of the cardiovascular system. While observing the dynamic changes of various cardiac structures during systole and diastole, current clinical interpretation of echocardiogram images still relies heavily on human experience. Existing technologies often focus on anatomical structural changes, while neglecting the analysis of time-series features such as motion rhythm and periodic stability, making it difficult to identify subtle rhythmic fluctuations. When comparing different systolic-diastolic cycles, existing systems struggle to achieve standardized alignment and difference measurement of inter-frame image embedding features, leading to inaccurate periodic consistency judgments. In cases of mild structural deviations or local rhythm disturbances, traditional systems struggle to effectively identify, label, or automatically correct them. In summary, current echocardiogram data analysis methods still have significant shortcomings, necessitating an automated anomaly detection method and system that integrates image structure embedding, motion dynamic feature analysis, and periodic rhythm modeling to improve the intelligent interpretation and diagnostic assistance efficiency of echocardiogram data.

[0003] Existing technologies, such as the invention patent CN114469176B (public account), are a method and related device for detecting fetal cardiac ultrasound images. The method includes acquiring multimodal echocardiograms; determining the fetal cardiac morphology, fetal heart rhythm indicators, and fetal atrioventricular motion patterns based on the acquired echocardiograms; and determining the type of fetal arrhythmia based on the fetal cardiac morphology, fetal heart rhythm indicators, and fetal atrioventricular motion patterns.

[0004] Existing technology, such as the invention patent CN107847173B (public account), is an ultrasound sequencing system and method, including a catheter configured for delivery into a body cavity defined by surrounding tissue; multiple ultrasound transducers connected to the distal end of the catheter; and an electronic module configured to selectively turn each ultrasound transducer on / off according to a predetermined activation sequence, and to process signals received from each ultrasound transducer to at least generate a 2D display of the surrounding tissue. Users can selectively calculate and display various aspects of cardiac activity. Users can display dipole density (DDM), charge density (CDM), or voltage (VV). The shape and location of the chamber (surface) and the potential recorded on the electrodes can be displayed.

[0005] Current technologies in ultrasound data monitoring and processing focus on using variational autoencoders and semi-supervised learning to extract standard sections, but they only consider the static features of ultrasound images and lack exploration of inter-frame temporal relationships, failing to meet the needs for detecting abnormalities caused by the periodic changes in the heart. Furthermore, current technologies only analyze ultrasound images at a macroscopic level and cannot identify subtle rhythmic abnormalities. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for detecting abnormalities in cardiac ultrasound data. To achieve the above objectives, this invention utilizes the following technical solution: A method for detecting abnormalities in cardiac ultrasound data, comprising:

[0007] Dual-channel feature extraction and detection were performed on the original cardiac ultrasound image sequence to obtain the inter-frame dynamic feature values ​​and image embedding feature values ​​of each frame in the original cardiac ultrasound image sequence.

[0008] Based on the inter-frame dynamic features and image embedding feature values ​​of each frame, the systolic-diastolic oscillation cycles in the original cardiac ultrasound image sequence are identified.

[0009] The mean structural alignment error of each systolic-diastolic oscillation cycle is obtained, the structural periodic deviation judgment result of each systolic-diastolic oscillation cycle is obtained, and the decision on whether to enter the rhythm abnormality judgment process is based on the structural periodic deviation judgment result.

[0010] The image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed. After introducing the influence of the structural periodic deviation judgment result, the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained, and corresponding operations are performed based on the rhythm abnormality judgment result.

[0011] As a preferred technical solution, the inter-frame dynamic feature values ​​and image embedding feature values ​​of each frame in the original cardiac ultrasound image sequence are obtained. The specific process is as follows:

[0012] In the inter-frame dynamic channel, motion information parameters of each frame in the original cardiac ultrasound image sequence are extracted, including the mean local optical flow, boundary displacement intensity, frame difference energy density, and displacement direction abrupt change rate. The unit weight factors corresponding to the motion information parameters are extracted from the database, including the unit weight factor of the mean local optical flow, the unit weight factor of the boundary displacement intensity, the unit weight factor of the frame difference energy density, and the unit weight factor of the displacement direction abrupt change rate. The motion information parameters of each frame in the original cardiac ultrasound image sequence are corrected and coupled with the corresponding unit weight factors to obtain the inter-frame dynamic feature values ​​of each frame in the original cardiac ultrasound image sequence.

[0013] In the image embedding feature channel, embedding information parameters of each frame in the original B-ultrasound sequence are extracted, including embedding norm, feature entropy, and principal component projection amplitude. The unit weight factors corresponding to the embedding information parameters are extracted from the database, including the embedding norm unit weight factor, feature entropy unit weight factor, and principal component projection amplitude unit weight factor. The embedding information parameters of each frame in the original B-ultrasound sequence are modified and coupled with the corresponding unit weight factors to obtain the image embedding feature values ​​of each frame in the original cardiac B-ultrasound image sequence.

[0014] As a preferred technical solution, the systolic-diastolic fluctuation cycle in the original cardiac ultrasound image sequence is identified based on the inter-frame dynamic features and image embedding feature values ​​of each frame. The specific process is as follows:

[0015] The inter-frame dynamic features and image embedding features of each frame are synchronized and aligned on the time axis to construct a unified periodic perception feature curve.

[0016] By performing first-order derivative analysis on the periodic perception characteristic curve, the changing trend of the characteristic trend is identified. Further detection of the positive and negative changes of its derivative is used to obtain the trend reversal point, and potential peaks and valleys are preliminarily identified.

[0017] Based on the trend reversal point, local extreme points are determined by combining the peak and valley detection algorithm. At the same time, the wavelet transform method is introduced to extract the dominant wave rhythm from the periodic sensing feature curve. By combining the peak and valley detection results with the wavelet dominant frequency information, preliminary candidate points of the boundary of each contraction-diplation wave cycle are obtained, which are denoted as the preliminary candidate points of each boundary.

[0018] Arrange the preliminary candidate points of each boundary in ascending order of frame number, calculate the inter-frame distance between each candidate point and its adjacent points, and if the inter-frame time is less than the set time redundancy threshold, then these points are considered to be repeated responses of the same boundary event, and the average frame position is selected as the merged preliminary candidate point of the boundary.

[0019] Calculate the period length between two preliminary candidate boundary points, and calculate the mean and standard deviation of all period lengths. Remove abnormal period segments and record the remaining period segments as candidate period segments. Align the embedded trajectory of each candidate period segment on the time axis and compare its trajectory similarity with that of the same-order candidate period segments. Filter based on trajectory similarity and remove candidate period segments with a trajectory similarity lower than a preset threshold. Use the period boundary points of each selected candidate period segment as the final period division result. Output the start and end frame numbers of each contraction-diastole cycle, and number each frame within each contraction-diastole cycle according to its standardized position. Output the standardized frame position number of each frame in each contraction-diastole cycle.

[0020] As a preferred technical solution, the mean structural alignment error of each contraction-diastole cycle is obtained, specifically including:

[0021] The image embedding amount of each frame in each systolic-diastolic fluctuation cycle is obtained. The image embedding amount of each frame with the same standardized frame position number in each systolic-diastolic fluctuation cycle is averaged and used as the reference value of the image embedding amount of that standardized frame position number. The difference between the image embedding amount of each frame in a certain systolic-diastolic fluctuation cycle and the corresponding image embedding amount reference value is obtained to obtain the image embedding amount difference of each frame in that systolic-diastolic fluctuation cycle. The average value of the structure alignment error of that cycle is obtained by averaging the image embedding amount differences of each frame.

[0022] As a preferred technical solution, the structural periodic deviation judgment results for each systolic-diastolic fluctuation cycle are obtained, specifically including:

[0023] The mean structural alignment error of each systolic-diastolic oscillation cycle is compared with the pre-stored mean structural alignment error threshold in the database. If the mean structural alignment error of a certain systolic-diastolic oscillation cycle is greater than or equal to the mean structural alignment error threshold, then the structural periodicity deviation of that systolic-diastolic oscillation cycle is determined to be structural periodicity deviation.

[0024] If the mean structural alignment error of a certain contraction-diastolic oscillation cycle is less than the mean structural alignment error threshold, then the structural periodicity deviation of that contraction-diastolic oscillation cycle is judged as not having any structural periodicity deviation.

[0025] As a preferred technical solution, the decision on whether to proceed to the rhythm abnormality judgment process is based on the structural periodic deviation judgment results, specifically including:

[0026] If a structural periodic deviation exists in a certain systolic-diastolic fluctuation cycle, the difference in image embedding of each frame in the systolic-diastolic fluctuation cycle is obtained. Frames with an image embedding difference greater than or equal to a threshold value are selected as deviation frames. The number of consecutive deviation frames in the systolic-diastolic fluctuation cycle is counted. If the number of consecutive deviation frames is greater than or equal to a preset consecutive number threshold, the structural periodic deviation of the systolic-diastolic fluctuation cycle is defined as a serious deviation, and an early warning is issued. The cycle is not entered into the rhythm abnormality judgment process.

[0027] If the number of consecutive deviation frames is less than a preset threshold, the structural periodic deviation of the systolic-diastolic fluctuation cycle is defined as a slight deviation. The number of consecutive deviation frames and the distribution skewness are obtained. The number of consecutive deviation frames is mapped and matched in a pre-stored mapping set of consecutive number-image enhancement parameters in the database to obtain the image enhancement parameters for the systolic-diastolic fluctuation cycle. The image enhancement parameters for the systolic-diastolic fluctuation cycle are used to enhance the image of the systolic-diastolic fluctuation cycle. The distribution skewness is mapped in a pre-stored mapping set of distribution skewness-temporal position adjustment parameters in the database to obtain the temporal position adjustment parameters for the systolic-diastolic fluctuation cycle. The temporal position of each deviation frame in the systolic-diastolic fluctuation cycle is adjusted using the temporal position adjustment parameters for the systolic-diastolic fluctuation cycle. After completing the image enhancement and temporal position adjustment, the process enters the rhythm abnormality judgment process.

[0028] As a preferred technical solution, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the influence of the structural periodic deviation judgment result is introduced to obtain the rhythmic abnormality judgment result of each systolic-diastolic fluctuation cycle, specifically including:

[0029] Embedded information parameters of each frame within each systolic-diastolic fluctuation cycle are extracted, including the embedding norm, feature entropy, and principal component projection amplitude of each frame. The embedded information parameters are arranged in the frame sequence within the cycle, and time series analysis is performed to calculate the index parameters of the characteristic curves within each systolic-diastolic fluctuation cycle, including the first derivative, local volatility, and principal component reconstruction error.

[0030] If the structural periodic deviation judgment result of a certain systolic-diastolic fluctuation cycle is that a structural periodic deviation exists, then the difference between the consecutive number of deviation frames of the systolic-diastolic fluctuation cycle and the consecutive number threshold is calculated to obtain the consecutive number difference value. This value is then entered into the pre-stored consecutive number difference-penalty factor mapping set in the database for mapping and matching to obtain the penalty factor of the systolic-diastolic fluctuation cycle.

[0031] The penalty factor based on the contraction-diastolic oscillation cycle is used to correct the index parameters of the internal characteristic curve.

[0032] If any indicator of a certain systolic-diastolic fluctuation cycle exceeds the corresponding preset threshold, the rhythm abnormality judgment result of the systolic-diastolic fluctuation cycle is determined to be rhythm abnormality.

[0033] If all indicators of a certain systolic-diastolic fluctuation cycle do not exceed the corresponding preset threshold, then the rhythm abnormality judgment result of the systolic-diastolic fluctuation cycle is determined to be normal.

[0034] As a preferred technical solution, corresponding operations are performed based on the rhythm abnormality judgment results, and the specific processing conditions are as follows:

[0035] If the rhythm abnormality judgment result of the systolic-diastolic fluctuation cycle is a rhythm abnormality, then all systolic-diastolic fluctuation cycles with rhythm abnormality judgment result are counted. If the number of consecutive cycles is greater than the preset abnormality threshold, then the rhythm abnormality is determined to be a global rhythm abnormality, and then the main frequency matching loss mechanism is implemented.

[0036] If the number of consecutive cycles is less than or equal to the preset abnormal threshold, the rhythm abnormality is determined to be a local cycle abnormality. All systolic-diastolic fluctuation cycles with rhythm abnormality judgment results are marked as abnormal cycles and an early warning is issued.

[0037] As a preferred technical solution, a frequency matching loss mechanism is implemented, specifically including:

[0038] By performing Fourier transform on the inter-frame dynamic parameter sequence of each systolic-diastolic fluctuation cycle, the dominant frequency of each frame is extracted, and the spectral multipeak ratio of each systolic-diastolic fluctuation cycle is obtained. If the spectral multipeak ratio of a certain systolic-diastolic fluctuation cycle is greater than the preset spectral multipeak ratio threshold, it is marked as an abnormal continuous systolic-diastolic fluctuation cycle, and an early warning is issued.

[0039] If the spectral multipeak ratio of all systolic-diastolic fluctuation cycles is less than or equal to the spectral multipeak ratio threshold, then the dominant frequency of each frame in each systolic-diastolic fluctuation cycle is compared with the dominant frequency with the highest proportion in the entire cycle. Based on the comparison results, a loss factor is constructed. The loss factor of each frame is compared with the loss factor threshold pre-stored in the database. If the loss factor of each frame does not exceed the threshold, it is used in the correct rhythm frequency generation algorithm to assist the algorithm in generating rhythm frequencies and correcting each frame.

[0040] If the loss factor of a frame in a certain systolic-diastolic fluctuation cycle exceeds the threshold, an early warning will be issued.

[0041] Additionally, a system for detecting abnormalities in cardiac ultrasound data includes:

[0042] The feature extraction module is used to perform dual-channel feature extraction and detection on the original cardiac ultrasound image sequence to obtain the inter-frame dynamic feature values ​​and image embedding feature values ​​of each frame in the original cardiac ultrasound image sequence.

[0043] The period recognition module is used to identify each systolic-diastolic fluctuation period in the original cardiac ultrasound image sequence based on the inter-frame dynamic features and image embedding feature values ​​of each frame.

[0044] The structural periodic deviation judgment module is used to obtain the average structural alignment error of each systolic-diastolic fluctuation cycle, obtain the structural periodic deviation judgment result of each systolic-diastolic fluctuation cycle, and determine whether to enter the rhythm abnormality judgment process based on the structural periodic deviation judgment result.

[0045] The rhythm anomaly detection module analyzes the image embedding feature parameters of each systolic-diastolic fluctuation cycle, introduces the influence of the structural periodic deviation detection results, obtains the rhythm anomaly detection results of each systolic-diastolic fluctuation cycle, and performs corresponding operations based on the rhythm anomaly detection results.

[0046] Compared with the prior art, the embodiments of the present invention have at least the following beneficial effects:

[0047] (1) This invention provides an anomaly detection method for cardiac ultrasound data. Through a "dual-channel feature extraction" architecture, it simultaneously acquires inter-frame dynamic parameters and image embedding information, and introduces a unit weight factor for coupling processing, effectively improving the specificity and noise resistance of feature expression. Furthermore, a unified periodic perception curve is constructed through feature alignment, and multiple methods such as derivative analysis, wavelet extraction, and extreme value screening are employed to ensure the accuracy and continuity of systolic-diastolic period identification. In addition, systematically filtering abnormal periodic segments, standardizing the periodic boundary frame order, and eliminating periods with low similarity enhances the robustness of period division, providing a stable foundation for subsequent analysis.

[0048] (2) This invention constructs a two-stage judgment logic for structural periodic deviation and rhythm anomaly. First, the mean value of the structural alignment error is used to identify periodic structural deviation, and the number of deviation frames is used to determine whether the deviation is severe or mild. In the case of severe deviation, a warning is issued directly and subsequent judgments are skipped; in the case of mild deviation, the process proceeds to the image enhancement and temporal adjustment stage before rhythm anomaly analysis. In the rhythm anomaly judgment, multiple indicators such as principal component reconstruction error, derivative change, and volatility are further introduced for evaluation, and correction is made in conjunction with the penalty factor for structural deviation. This progressive judgment path can effectively distinguish between local short-term anomalies and potential systemic fluctuation problems, achieving a more comprehensive and reasonable judgment strategy.

[0049] (3) This invention further classifies anomalies into local periodic anomalies or global rhythmic anomalies based on their continuity, and provides early warning or initiates a dominant frequency matching loss mechanism accordingly. The dominant frequency mechanism obtains spectral information through Fourier transform and constructs a loss factor by combining the multi-peak ratio of the spectrum and the dominant frequency comparison. If no abnormal conditions are triggered, it further assists in rhythmic frequency repair. This mechanism realizes closed-loop adjustment capability after anomalies, and can adaptively adjust the rhythmic trend of abnormal signals. In addition, this method performs frame-by-frame dominant frequency correction on periodic frames, providing stable and structured frequency information.

[0050] Of course, any product implementing this invention does not necessarily need to achieve all of the above advantages at the same time. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0052] Figure 2This is a schematic diagram of the system modules of the present invention.

[0053] Figure 3 This is a schematic diagram of the logic flow of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] In the description of this invention, it should be understood that the terms "opening", "upper", "lower", "thickness", "top", "middle", "length", "inner", "around", etc., which indicate orientation or positional relationship, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the components or elements referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting this invention.

[0056] Please see Figure 1 and Figure 3 As shown, this embodiment of the invention provides a method for detecting abnormalities in cardiac ultrasound data, including:

[0057] Dual-channel feature extraction and detection were performed on the original cardiac ultrasound image sequence to obtain the inter-frame dynamic feature values ​​and image embedding feature values ​​of each frame in the original cardiac ultrasound image sequence.

[0058] The process of obtaining inter-frame dynamic feature values ​​and image embedding feature values ​​for each frame in the original cardiac ultrasound image sequence is as follows:

[0059] In the inter-frame dynamic channel, motion information parameters of each frame in the original cardiac ultrasound image sequence are extracted, including the mean local optical flow, boundary displacement intensity, frame difference energy density, and displacement direction abrupt change rate. The mean local optical flow is calculated. Optical flow estimation is performed between adjacent frames (e.g., using Farneback optical flow, RAFT, or PWC-Net) to obtain the motion vector of each pixel. Each frame image is divided into multiple sub-regions, and the optical flow vector within each region is averaged to obtain the mean local optical flow distribution for that region. By summing the mean optical flow values ​​from multiple regions within a frame, a low-dimensional motion representation reflecting the overall displacement trend of that frame can be formed.

[0060] Assess the boundary displacement intensity. Anatomical structure boundaries are extracted in each frame using edge detection algorithms (such as Canny or Sobel). The movement distances of these boundaries between adjacent frames are then matched and measured. In this embodiment, optical flow tracing is used to obtain the movement distance of each edge point between adjacent frames. The average movement distance of each edge point between adjacent frames is then calculated to obtain the average movement distance of the structural boundaries, which is the boundary displacement intensity. This intensity represents the activity level of cardiac tissue deformation and movement, and is particularly suitable for identifying critical moments of systole and diastole.

[0061] Calculate the frame difference energy density. For each pair of adjacent frames, calculate the pixel-level difference image, square the difference images, and then average them to obtain the energy density value of the inter-frame image difference. This value is the average squared value of the pixel intensity change, quantifying the overall intensity of the image change between two adjacent frames. The larger this value, the more drastic the image change between two adjacent frames, indicating a strong jump or rhythm shift in heartbeat.

[0062] The displacement direction abrupt change rate is extracted. The motion direction of pixels is extracted from the optical flow vector field, and a direction histogram of the entire frame is constructed. By tracking these direction changes in the time dimension, it is possible to identify whether there are abrupt direction changes in local regions across multiple consecutive frames. In this embodiment of the invention, the specific determination criteria include: in the direction histogram of the entire frame, if the change in direction between time t+1 and time t is greater than or equal to a preset direction change threshold, then abrupt direction changes are determined. The abrupt change rate is quantified by statistically analyzing the pixels in the entire frame whose direction changes are greater than the preset threshold and calculating their proportion among all pixels in the entire frame. It is an important indicator for judging rhythmic perturbations or abnormal motion patterns.

[0063] The unit weighting factors corresponding to motion information parameters are extracted from the database, including the unit weighting factor for the local optical flow mean, the unit weighting factor for the boundary displacement intensity, the unit weighting factor for the frame difference energy density, and the unit weighting factor for the displacement direction abrupt change rate. The motion information parameters of each frame in the original cardiac ultrasound image sequence are then corrected and coupled with their corresponding unit weighting factors to obtain the inter-frame dynamic feature values ​​of each frame in the original cardiac ultrasound image sequence. Specifically, these include:

[0064] ;

[0065] in, Let be the inter-frame dynamic feature value of the t-th frame. Let be the mean local optical flow of frame t. Let be the boundary displacement intensity of frame t. Let be the frame difference energy density of frame t. Let be the abrupt change rate of the displacement direction in frame t. The local optical flow mean is a unit weighting factor. The boundary displacement strength unit weight factor, The unit weighting factor for frame difference energy density. The displacement direction abrupt change rate is a unit weighting factor. The frames in the original cardiac ultrasound image sequence are numbered. , This represents the total number of frames in the original cardiac ultrasound image sequence.

[0066] It should be noted that the local optical flow mean unit weight factor is used to normalize the intensity of local displacement between frames (in pixels / frame), giving it dimensionless properties in subsequent feature fusion. Simultaneously, this factor is also used to adjust the proportion of local overall displacement in dynamic feature representation. When the overall cardiac motion trend is strong but structural deformation is weak, increasing the weight of this factor helps enhance its sensitivity to rhythm determination.

[0067] The boundary displacement intensity unit weighting factor measures the average distance the boundaries of cardiac anatomy structures move between adjacent frames, also in pixels per frame. This factor, after eliminating units, assigns a higher or lower influence to structural boundary deformation, making it suitable for enhancing the dynamic representation of systolic-diastolic rhythm changes on key structures such as the heart wall and valves.

[0068] The frame difference energy density unit weighting factor is a coefficient that normalizes pixel-level image intensity changes (in units of the square of gray values). It is mainly used to adjust the weighting of dynamic features by the degree of image change. This factor has important regulatory significance for identifying jumps, abnormal interruptions, or abrupt changes in cardiac rhythm during cardiac motion.

[0069] The displacement direction abrupt change rate unit weight factor is used to regulate the influence of the degree of pixel orientation abrupt change (itself a dimensionless ratio). In the original image sequence, this parameter reflects the degree of drastic change in the motion direction of a local region and has a significant indicative role in abnormal motion patterns, rhythmic perturbations, premature beats, etc. By adjusting this factor, the contribution ratio of orientation anomalies to the overall inter-frame dynamic features can be controlled.

[0070] It should also be noted that the motion information parameters of each frame, including the mean local optical flow, boundary displacement intensity, frame difference energy density, and displacement direction abrupt change rate, are correlated. The mean local optical flow reflects the pixel-level motion trend of the overall image, while the boundary displacement intensity focuses on the degree of deformation of the anatomical edges. Both usually increase simultaneously during strong contraction or relaxation, reflecting the consistency between the overall cardiac motion and local structural changes. Local optical flow is greatly affected by global motion (such as slight probe movement), while boundary displacement intensity is more sensitive to changes in cardiac chamber volume and valve motion, and therefore more representative of morphological and structural changes. When the image motion is intense (such as rapid contraction), it is usually accompanied by increased pixel motion amplitude and enhanced image grayscale changes, manifested as an increase in optical flow amplitude and frame difference. Therefore, under conditions without other interference (such as probe jitter or noise abrupt changes), the mean local optical flow and frame difference energy density may show a certain degree of positive correlation under specific motion modes. Optical flow describes displacement velocity (vector field), while frame difference energy emphasizes pixel intensity changes (brightness difference). Therefore, frame difference is sensitive to image noise, while optical flow is more stable. During the rhythmic contraction-dipsia process, the directional abrupt change rate is usually low. Abnormal motion (such as premature beats and tremors) can cause discontinuities in the optical flow direction, increasing the abrupt change rate. This may be accompanied by abnormal changes in optical flow amplitude and enhanced differences in image brightness, thus causing fluctuations in optical flow distribution and an increase in frame difference energy. The directional abrupt change rate is a morphological "abrupt change" indicator. Unlike the "intensity" attribute of other parameters, it is particularly sensitive to abnormal, non-periodic motion and is an important supplement to the identification of abnormalities such as arrhythmias.

[0071] In the image embedding feature channel, embedding information parameters for each frame of the original ultrasound sequence are extracted, including embedding norm, feature entropy, and principal component projection magnitude. In this channel, a deep image encoder (such as a convolutional neural network (CNN), visual transformer, or hybrid encoder) is used to perform high-dimensional feature extraction on each frame of the original ultrasound sequence, obtaining embedding vectors reflecting cardiac structure information. Further, inter-frame embedding information parameters with diagnostic or identification value are extracted from these embedding vectors to characterize the dynamic changes and temporal consistency of inter-frame structures, specifically including the following aspects:

[0072] For each frame, the L2 norm of the extracted embedding vector is calculated to measure the overall structural feature strength of that frame. Variations in the embedding norm reflect the periodic fluctuations in the structural complexity of the heart at different stages (such as systole and diastole), which helps identify frames with weak function or ambiguous structure.

[0073] Feature entropy measures the structural complexity and stability of the current frame in the embedding space by normalizing the embedding vector of each frame and calculating its information entropy value. High feature entropy indicates rich structural information or the potential presence of uncertainties such as motion blur or anomalous deformation. By comparing the fluctuations in entropy values ​​across consecutive frames, structural jumps and inconsistencies can be identified.

[0074] Principal component projection amplitudes are obtained by performing principal component analysis (PCA) on an embedded vector sequence to obtain the projection results of the first few principal components. The projection amplitudes reflect the main directions and amplitudes of structural changes in the embedded space and can be used to measure the regularity and integrity of the entire periodic structural activity. Abrupt changes or sharp decreases in amplitude in the principal component trajectories may indicate abnormal periodic motion or weakened local functions.

[0075] The unit weight factors corresponding to the embedded information parameters are extracted from the database, including the embedding norm unit weight factor, the feature entropy unit weight factor, and the principal component projection amplitude unit weight factor. The embedded information parameters of each frame in the original ultrasound sequence are then modified and coupled with their corresponding unit weight factors to obtain the image embedding feature values ​​of each frame in the original cardiac ultrasound image sequence. Specifically, these include:

[0076] ;

[0077] in, Embed feature values ​​for the image in frame t. Let be the embedding norm of the t-th frame. Let be the feature entropy of frame t. Let be the principal component projection magnitude of the t-th frame. For the embedding norm unit weight factor, The feature entropy is the unit weight factor. Principal component projection amplitude unit weight factor, The frames in the original cardiac ultrasound image sequence are numbered. , This represents the total number of frames in the original cardiac ultrasound image sequence.

[0078] It should be noted that the embedding norm unit weight factor is used to eliminate the unit of the overall magnitude (i.e., L2 norm, in units of vector magnitude) of the embedding vector for each frame, making this feature dimensionless under different encoder architectures and scale distributions. This factor can also adjust the proportion of the overall structural intensity of the image in the periodic dynamic features. When the overall cardiac structure contracts or relaxes significantly, but local feature changes are not obvious, appropriately increasing the weight of this factor helps to enhance the significance of that frame in the periodic structural fluctuation analysis.

[0079] The feature entropy unit weight factor is used to standardize the information entropy value (in bits / frame or nats / frame) of the image embedding vector to remove biases related to the number of embedding dimensions or normalization strategies. Feature entropy reflects the information complexity and uncertainty of embedded features, and setting this factor can adjust the weight of complexity variations in periodic detection and anomaly recognition. Especially when recognizing blurred frames, frames with abnormal motion, or structures with instantaneous jumps, increasing this factor helps improve the model's sensitivity to structurally chaotic regions.

[0080] The principal component projection amplitude unit weight factor is used to eliminate the dimensional influence of different amplitude components during principal component analysis, making the changes in the principal direction comparable under different cycles or lesion conditions. This factor also controls the dynamic expression weight of the principal deformation direction of the structure (such as the contraction axis). When there are phenomena such as decreased integrity of the systolic / diastolic trajectory of the cardiac structure or abrupt changes in projection amplitude, adjusting this factor can effectively enhance its expression in the detection of abnormal cycles and improve the recognition accuracy of functionally weakened frames.

[0081] It should also be noted that the embedding information parameters, including embedding norm, feature entropy, and principal component projection amplitude, are related. Embedding norm represents the overall "strength" or "amplitude" of the embedding vector, reflecting the energy intensity of the image frame activated in the high-dimensional feature space. Feature entropy describes the "complexity" or "uniformity" of the embedding vector distribution, used to measure the uncertainty of the image frame in the embedding feature dimension. Principal component projection amplitude reflects the concentration or dominant trend of the embedded features in the principal direction (e.g., the PCA principal axis), highlighting the main deformation patterns. When there is noise or invalid features in the image, entropy may increase while the norm does not increase significantly, reflecting redundant distribution or abnormal perturbation. When image features are concentrated in a specific direction (e.g., a clear deformation trend of a certain anatomical structure), the norm is high and the principal component projection amplitude is large; if the image has a lot of noise or the features are divergent, the norm is high but the principal component amplitude may be low, indicating that the energy distribution is not concentrated; when the feature distribution is concentrated on a few principal components (large principal projection amplitude), the entropy may be low (concentrated distribution); if the entropy is high but the principal component amplitude is not high, it indicates that the features are scattered, the structure is unclear, or there is interference.

[0082] Identify each systolic-diastolic oscillation cycle in the original cardiac ultrasound image sequence based on the inter-frame dynamic feature values ​​and image embedding feature values ​​of each frame.

[0083] The inter-frame dynamic feature values ​​and image embedding feature values ​​of each frame are synchronized and aligned on the time axis to construct a unified periodic perception feature curve.

[0084] It should be noted that the original cardiac ultrasound image sequence has a natural frame sequence number, and the frame sequence number can be used as a time index. Since the inter-frame dynamic features and image embedding features are extracted frame by frame based on this image sequence, the frame number can be directly used as the time axis reference to make a one-to-one correspondence between the features of the two channels in the time dimension.

[0085] Because the dimensions and numerical ranges of the two types of features differ significantly, normalization is required for each type of feature. In this embodiment of the invention, an activation function is used to map the inter-frame dynamic feature values ​​and image embedding feature values ​​of each frame to the [0,1] interval. The inter-frame dynamic feature values ​​are multiplied by the inter-frame dynamic feature weighting factor, and the image embedding feature values ​​are multiplied by the image embedding feature weighting factor, and then coupled. The periodicity-aware feature curve is obtained by plotting the curve according to the frame sequence number. The periodicity-aware feature curve integrates the joint information of motion trend and structural expression, and has a good periodic fluctuation structure, which is the basis for subsequent peak and valley detection, wavelet analysis, and periodic boundary determination.

[0086] By performing first-order derivative analysis on the periodic perception characteristic curve, the changing trend of the characteristic trend is identified. Further detection of the positive and negative changes of its derivative is used to obtain the trend reversal point, and potential peaks and valleys are preliminarily identified.

[0087] It should be noted that performing first-order derivative analysis on the periodic sensing characteristic curve aims to identify the trend of this feature changing over time, thereby locating key fluctuation nodes in the periodic rhythm. The specific operation process is as follows: Obtain the periodic sensing characteristic curve. This curve represents the feature values ​​of each frame of the image after fusion on the time axis. Perform first-order difference operations to approximate its first derivative. Specifically:

[0088] ;

[0089] This derivative reflects the growth and decay trends of the eigenvalues, when... When, it indicates that the feature is in an upward trend, when When this occurs, it indicates that the characteristic is in a downward trend.

[0090] Furthermore, the positive and negative changes of the derivative are identified to obtain trend reversal points, namely the turning point from positive to negative and the turning point from negative to positive. The turning point from positive to negative usually corresponds to the local peak of the characteristic curve, which in this embodiment corresponds to the peak of cardiac contraction. The turning point from negative to positive usually corresponds to the local valley of the characteristic curve, which in this embodiment corresponds to the low point of cardiac diastole.

[0091] Based on the trend reversal point, local extreme points are determined by combining the peak and valley detection algorithm. At the same time, the wavelet transform method is introduced to extract the dominant wave rhythm from the periodic sensing feature curve. By combining the peak and valley detection results with the wavelet dominant frequency information, preliminary candidate points of the boundary of each contraction-diplation wave cycle are obtained, which are denoted as the preliminary candidate points of each boundary.

[0092] It should be noted that at the identified trend reversal points, a peak-valley detection algorithm is used to further pinpoint local maxima and local minima. Local maxima reflect the cardiac systolic peak, and local minima reflect the cardiac diastolic trough, ultimately resulting in a set of time point sequences representing local systolic peaks and diastolic troughs, forming a preliminary extreme value time series. To enhance the ability to identify the dominant rhythm, wavelet transform is introduced to perform multi-scale frequency domain analysis on the periodic sensing feature curve. The specific process includes: selecting a mother wavelet (in this embodiment, the Morlet wavelet is selected), extracting the amplitude response and power spectral density at the dominant wavelet frequency, analyzing the dominant frequency components with concentrated energy in the time-frequency graph, and using the dominant frequency components to reflect the basic systolic-diastolic cycle length to obtain an approximate period range for period boundary constraints.

[0093] The peak and valley detection results are fused with the wavelet dominant frequency rhythm for analysis. Specifically, under the constraint of dominant frequency period, extreme value pairs with excessively small or large intervals between adjacent extreme points are screened out, and the start and end frames of each candidate period segment are used as preliminary candidate boundary points.

[0094] Arrange the preliminary candidate points of each boundary in ascending order of frame number, calculate the inter-frame distance between each candidate point and its adjacent points, and if the inter-frame time is less than the set time redundancy threshold, then these points are considered to be repeated responses of the same boundary event, and the average frame position is selected as the merged preliminary candidate point of the boundary.

[0095] Calculate the period length between two preliminary candidate boundary points, and calculate the mean and standard deviation of all period lengths. Remove abnormal period segments and record the remaining period segments as candidate period segments. Align the embedded trajectory of each candidate period segment on the time axis and compare its trajectory similarity with that of the same-order candidate period segments. Filter based on trajectory similarity and remove candidate period segments with a trajectory similarity lower than a preset threshold. Use the period boundary points of each selected candidate period segment as the final period division result. Output the start and end frame numbers of each contraction-diastole cycle, and number each frame within each contraction-diastole cycle according to its standardized position. Output the standardized frame position number of each frame in each contraction-diastole cycle.

[0096] The mean structural alignment error of each systolic-diastolic oscillation cycle is obtained, the structural periodic deviation judgment result of each systolic-diastolic oscillation cycle is obtained, and the decision on whether to enter the rhythm abnormality judgment process is based on the structural periodic deviation judgment result.

[0097] The mean structural alignment error for each systolic-diastolic oscillation cycle was obtained, specifically including:

[0098] The image embedding amount of each frame in each systolic-diastolic fluctuation cycle is obtained. The image embedding amount of each frame with the same standardized frame position number in each systolic-diastolic fluctuation cycle is averaged and used as the reference value of the image embedding amount of that standardized frame position number. The difference between the image embedding amount of each frame in a certain systolic-diastolic fluctuation cycle and the corresponding image embedding amount reference value is obtained to obtain the image embedding amount difference of each frame in that systolic-diastolic fluctuation cycle. The average value of the structure alignment error of that cycle is obtained by averaging the image embedding amount differences of each frame.

[0099] It should be noted that image embedding refers to the numerical representation of the high-dimensional feature vector obtained in the embedding space after each frame of echocardiogram image is input into a deep coding network (such as CNN, ViT, etc.). It is a compressed representation of the structural information in that frame of image, reflecting deep information such as texture features, boundary shapes, and anatomical structure layout. In this embodiment of the invention, the deep network, CLIP (Contrastive Language-Image Pretraining), inputs each frame of two-dimensional echocardiogram image into the model and extracts the output of the intermediate layer or the penultimate layer as the image embedding vector.

[0100] The decision to proceed to the rhythm abnormality assessment process is based on the results of the structural periodic deviation assessment, specifically including:

[0101] If a structural periodic deviation exists in a certain systolic-diastolic fluctuation cycle, the difference in image embedding of each frame in the systolic-diastolic fluctuation cycle is obtained. Frames with an image embedding difference greater than or equal to a threshold value are selected as deviation frames. The number of consecutive deviation frames in the systolic-diastolic fluctuation cycle is counted. If the number of consecutive deviation frames is greater than or equal to a preset consecutive number threshold, the structural periodic deviation of the systolic-diastolic fluctuation cycle is defined as a serious deviation, and an early warning is issued. The cycle is not entered into the rhythm abnormality judgment process.

[0102] If the number of consecutive deviation frames is less than a preset threshold, the structural periodic deviation of the systolic-diastolic fluctuation cycle is defined as a slight deviation. The number of consecutive deviation frames and the distribution skewness are obtained. The number of consecutive deviation frames is mapped and matched in a pre-stored mapping set of consecutive number-image enhancement parameters in the database to obtain the image enhancement parameters for the systolic-diastolic fluctuation cycle. The image enhancement parameters for the systolic-diastolic fluctuation cycle are used to enhance the image of the systolic-diastolic fluctuation cycle. The distribution skewness is mapped in a pre-stored mapping set of distribution skewness-temporal position adjustment parameters in the database to obtain the temporal position adjustment parameters for the systolic-diastolic fluctuation cycle. The temporal position of each deviation frame in the systolic-diastolic fluctuation cycle is adjusted using the temporal position adjustment parameters for the systolic-diastolic fluctuation cycle. After completing the image enhancement and temporal position adjustment, the process enters the rhythm abnormality judgment process.

[0103] It should be noted that distribution skewness is a statistical indicator used to measure the degree of asymmetry in data distribution. It describes the direction and extent of the skewness of the data distribution relative to the mean. In this embodiment of the invention, within the contraction-diastole cycle, frames with structural deviations are analyzed. The calculation of the distribution skewness of these deviated frames describes whether their distribution over time is uniform or biased towards a specific part of the cycle. If the distribution skewness of the deviated frames is close to zero, it indicates that the deviated frames are relatively evenly distributed throughout the cycle; if the skewness is positive, it indicates that the deviated frames are mainly concentrated in the latter half of the cycle; if the skewness is negative, it indicates that the deviated frames are mainly concentrated in the first half of the cycle.

[0104] The structural periodic deviation judgment results for each systolic-diastolic oscillation cycle are obtained, specifically including:

[0105] The mean structural alignment error of each systolic-diastolic oscillation cycle is compared with the pre-stored mean structural alignment error threshold in the database. If the mean structural alignment error of a certain systolic-diastolic oscillation cycle is greater than or equal to the mean structural alignment error threshold, then the structural periodicity deviation of that systolic-diastolic oscillation cycle is determined to be structural periodicity deviation.

[0106] If the mean structural alignment error of a certain contraction-diastolic oscillation cycle is less than the mean structural alignment error threshold, then the structural periodicity deviation of that contraction-diastolic oscillation cycle is judged as not having any structural periodicity deviation.

[0107] The image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed. After introducing the influence of the structural periodic deviation judgment result, the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained, and corresponding operations are performed based on the rhythm abnormality judgment result.

[0108] By analyzing the image embedding feature parameters of each systolic-diastolic fluctuation cycle and introducing the influence of structural periodic deviation judgment results, the rhythmic anomaly judgment results of each systolic-diastolic fluctuation cycle are obtained, specifically including:

[0109] Embedded information parameters of each frame within each systolic-diastolic fluctuation cycle are extracted, including the embedding norm, feature entropy, and principal component projection amplitude of each frame. The embedded information parameters are arranged in the frame sequence within the cycle, and time series analysis is performed to calculate the index parameters of the characteristic curves within each systolic-diastolic fluctuation cycle, including the first derivative, local volatility, and principal component reconstruction error.

[0110] It should be noted that the first derivative reflects the rate of change of the characteristic curve over time, i.e., the slope of the inter-frame embedding parameters. It reveals the dynamic changes in cardiac structure during the systolic-diastolic cycle, such as rapid rises or falls. The first derivative is calculated using numerical differencing methods for the embedding parameter sequence within the cycle.

[0111] Local volatility describes the amplitude and instability of the characteristic curve within the systolic-diastolic oscillation cycle, reflecting subtle changes in structural characteristics and possible anomalous jumps within the cycle. Higher volatility usually indicates drastic local structural changes, potentially corresponding to abnormal cardiac motion or noise effects. It is obtained by calculating the standard deviation within the systolic-diastolic oscillation cycle.

[0112] Principal component reconstruction error measures the magnitude of the error when reconstructing the original embedded feature sequence using a selected number of principal components. It reflects the explanatory power of Principal Component Analysis (PCA) for dynamic structural changes within a period; a large error may indicate that important anomalous patterns have not been captured by the principal components. It is obtained by performing PCA on the embedded information parameters, selecting principal components, and then calculating the reconstruction error after reconstructing the original data.

[0113] If the structural periodic deviation judgment result of a certain systolic-diastolic fluctuation cycle is that a structural periodic deviation exists, then the difference between the consecutive number of deviation frames of the systolic-diastolic fluctuation cycle and the consecutive number threshold is calculated to obtain the consecutive number difference value. This value is then entered into the pre-stored mapping set of consecutive number difference-penalty factor in the database for mapping and matching to obtain the penalty factor of the systolic-diastolic fluctuation cycle. The matching rule is that the smaller the consecutive number difference value, the larger the penalty factor.

[0114] The penalty factor based on the contraction-diplation oscillation cycle is used to correct the index parameters of the internal characteristic curve. The correction process includes multiplying the penalty factor with the index parameters.

[0115] If any indicator of a certain systolic-diastolic fluctuation cycle exceeds the corresponding preset threshold, the rhythm abnormality judgment result of the systolic-diastolic fluctuation cycle is determined to be rhythm abnormality.

[0116] If all indicators of a certain systolic-diastolic fluctuation cycle do not exceed the corresponding preset threshold, then the rhythm abnormality judgment result of the systolic-diastolic fluctuation cycle is determined to be normal.

[0117] Based on the rhythm abnormality assessment results, corresponding operations are performed, with the specific processing conditions as follows:

[0118] If the rhythm abnormality judgment result of the systolic-diastolic fluctuation cycle is a rhythm abnormality, then all systolic-diastolic fluctuation cycles with rhythm abnormality judgment result are counted. If the number of consecutive cycles is greater than the preset abnormality threshold, then the rhythm abnormality is determined to be a global rhythm abnormality, and then the main frequency matching loss mechanism is implemented.

[0119] If the number of consecutive cycles is less than or equal to the preset abnormal threshold, the rhythm abnormality is determined to be a local cycle abnormality. All systolic-diastolic fluctuation cycles with rhythm abnormality judgment results are marked as abnormal cycles and an early warning is issued.

[0120] The main frequency matching loss mechanism includes:

[0121] By performing Fourier transform on the inter-frame dynamic parameter sequence of each systolic-diastolic fluctuation cycle, the dominant frequency of each frame is extracted, and the spectral multipeak ratio of each systolic-diastolic fluctuation cycle is obtained. If the spectral multipeak ratio of a certain systolic-diastolic fluctuation cycle is greater than the preset spectral multipeak ratio threshold, it is marked as an abnormal continuous systolic-diastolic fluctuation cycle, and an early warning is issued.

[0122] It should be noted that performing a Fourier transform on the inter-frame dynamic parameter sequence of each systolic-diastolic fluctuation cycle can convert it from the time domain to the frequency domain, extracting the energy distribution of the sequence at different frequency components. The dominant frequency is the frequency component with the highest energy in the spectrum, corresponding to the most dominant heart rhythm in that cycle. If the dominant frequency is stable and clear, it indicates that the rhythm of the cycle is regular; if the dominant frequency is blurred or shifted, it may indicate abnormal structural activity. The multi-peak ratio refers to the proportion of the sum of the energy of other larger frequency components in the spectrum besides the dominant frequency to the total energy. It measures whether there are multiple "competing" rhythms or motion instability in the cycle. The Fourier transform can reveal rhythm changes that are difficult to detect on the surface and is very effective for highly periodic signals such as cardiac motion. As an anomaly indicator, the multi-peak ratio can identify those cycle segments that appear to have normal period division but have complex or abnormal internal rhythms.

[0123] If the spectral multipeak ratio of all systolic-diastolic fluctuation cycles is less than or equal to the spectral multipeak ratio threshold, then the dominant frequency of each frame in each systolic-diastolic fluctuation cycle is compared with the dominant frequency with the highest proportion in the entire cycle. Based on the comparison results, a loss factor is constructed. The loss factor of each frame is compared with the loss factor threshold pre-stored in the database. If the loss factor of each frame does not exceed the threshold, it is used in the correct rhythm frequency generation algorithm to assist the algorithm in generating rhythm frequencies and correcting each frame.

[0124] It should be noted that the most prevalent frequency throughout the entire cycle is extracted by: performing histogram analysis on the prevalent frequencies of all frames to identify the frequency with the highest percentage. Constructing the loss factor for each frame involves: comparing the prevalent frequency of each frame with the most prevalent frequency and simultaneously calculating its loss factor. ,in, Let $\frac{i}{i}$ be the loss factor between the main frequency of the i-th frame and the main frequency with the highest percentage. Let i be the main frequency of the i-th frame. The frequency with the highest proportion is used. The loss factor of each frame is fed into the correct rhythm frequency generation algorithm, which in this embodiment of the invention is frequency smoothing interpolation + multi-period regression algorithm, to perform rhythm curve fitting and generate a smooth periodic rhythm frequency curve, which is used to reposition some slightly offset frames based on the rhythm frequency curve.

[0125] If the loss factor of a frame in a certain systolic-diastolic fluctuation cycle exceeds the threshold, an early warning will be issued.

[0126] like Figure 2 As shown, in this embodiment, the present invention provides an abnormality detection system for cardiac ultrasound data, comprising:

[0127] The feature extraction module is used to perform dual-channel feature extraction and detection on the original cardiac ultrasound image sequence to obtain the inter-frame dynamic feature values ​​and image embedding feature values ​​of each frame in the original cardiac ultrasound image sequence.

[0128] The period recognition module is used to identify each systolic-diastolic fluctuation period in the original cardiac ultrasound image sequence based on the inter-frame dynamic features and image embedding feature values ​​of each frame.

[0129] The structural periodic deviation judgment module is used to obtain the average structural alignment error of each systolic-diastolic fluctuation cycle, obtain the structural periodic deviation judgment result of each systolic-diastolic fluctuation cycle, and determine whether to enter the rhythm abnormality judgment process based on the structural periodic deviation judgment result.

[0130] The rhythm anomaly detection module analyzes the image embedding feature parameters of each systolic-diastolic fluctuation cycle, introduces the influence of the structural periodic deviation detection results, obtains the rhythm anomaly detection results of each systolic-diastolic fluctuation cycle, and performs corresponding operations based on the rhythm anomaly detection results.

[0131] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0132] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementation methods. Clearly, many modifications and variations can be made based on the content of this specification. The selection and detailed description of these embodiments in this specification are intended to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. Any modifications or variations that do not deviate from the structure of the invention or exceed the scope defined by the invention should fall within the protection scope of the invention.

Claims

1. A method of anomaly detection for cardiac ultrasound data, characterized in that, The method comprises the following steps: dual-channel feature extraction detection is performed on the original heart B-ultrasound image sequence to obtain inter-frame dynamic feature values and image embedding feature values of each frame in the original heart B-ultrasound image sequence; each systolic-diastolic fluctuation cycle in the original heart B-ultrasound image sequence is identified based on the inter-frame dynamic features and the image embedding feature values of each frame; a structural alignment error mean of each systolic-diastolic fluctuation cycle is obtained, a structural periodicity deviation judgment result of each systolic-diastolic fluctuation cycle is obtained, and it is determined whether to enter a rhythm abnormality judgment process based on the structural periodicity deviation judgment result; image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, a rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural periodicity deviation judgment result is introduced, and corresponding operations are performed based on the rhythm abnormality judgment result; the inter-frame dynamic feature values and the image embedding feature values of each frame in the original heart B-ultrasound image sequence are obtained, and the specific process is as follows: in the inter-frame dynamic channel, motion information parameters of each frame in the original heart B-ultrasound image sequence are extracted, including local optical flow mean, boundary displacement intensity, frame difference energy density and displacement direction mutation rate, unit weight factors corresponding to the motion information parameters are extracted from a database, including local optical flow mean unit weight factor, boundary displacement intensity unit weight factor, frame difference energy density unit weight factor and displacement direction mutation rate unit weight factor, and the motion information parameters of each frame in the original heart B-ultrasound image sequence are respectively coupled with the corresponding unit weight factors after correction to obtain the inter-frame dynamic feature values of each frame in the original heart B-ultrasound image sequence; in the image embedding feature channel, embedding information parameters of each frame in the original B-ultrasound sequence are extracted, including embedding norm, feature entropy and principal component projection amplitude, unit weight factors corresponding to the embedding information parameters are extracted from a database, including embedding norm unit weight factor, feature entropy unit weight factor and principal component projection amplitude unit weight factor, and the embedding information parameters of each frame in the original B-ultrasound sequence are respectively coupled with the corresponding unit weight factors after correction to obtain the image embedding feature values of each frame in the original heart B-ultrasound image sequence.

2. The method of claim 1, wherein: each systolic-diastolic fluctuation cycle in the original heart B-ultrasound image sequence is identified based on the inter-frame dynamic features and the image embedding feature values of each frame, and the specific process is as follows: the inter-frame dynamic features and the image embedding features of each frame are synchronously aligned on a time axis to construct a unified cycle perception feature curve; by performing first-order derivative analysis on the cycle perception feature curve, the change trend of the feature trend is identified, the positive and negative changes of the derivative are further detected to obtain trend reversal points, and potential peak values and potential valley values are preliminarily identified; based on the trend reversal points, local extreme points are determined in combination with a peak-valley detection algorithm, and a dominant fluctuation rhythm is extracted from the cycle perception feature curve by introducing a wavelet transform method, the peak-valley detection result and wavelet main frequency information are comprehensively considered, boundary preliminary candidate points of each systolic-diastolic fluctuation cycle are obtained, and each boundary preliminary candidate point is recorded. Arranging the preliminary candidate points of each boundary in ascending order of frame number, calculating the inter-frame distance between each boundary candidate point and adjacent points, and if the inter-frame time is less than a set time redundancy threshold, considering that these points belong to the same boundary event repeated response, and selecting the average frame position as the merged preliminary candidate point of the boundary; Calculating the period length between two preliminary candidate points of the boundary, and calculating the mean and standard deviation of all period lengths, eliminating abnormal period segments, and recording the remaining period segments as candidate period segments. Aligning the embedded trajectory of each candidate period segment on the time axis, comparing the trajectory similarity between it and the same sequence candidate period, and screening based on the trajectory similarity, eliminating candidate period segments below the preset trajectory similarity threshold. The period boundary points of the screened candidate period segments are the final period division results, and the start and end frame numbers of each systolic-diastolic period are output. Each frame in each systolic-diastolic period is numbered according to the standardized position, and the standardized frame position number of each frame of each systolic-diastolic period is output.

3. The method of claim 1, wherein: The structure alignment error mean of each systolic-diastolic fluctuation period is obtained, specifically including: Obtaining the image embedding amount of each frame in each systolic-diastolic fluctuation period, and obtaining the image embedding amount of each frame in each systolic-diastolic fluctuation period. The same standardized frame position number is processed as an image embedding amount reference value of the standardized frame position number. The image embedding amount of each frame in a certain systolic-diastolic fluctuation period is subtracted from the corresponding image embedding amount reference value to obtain the image embedding amount difference value of each frame in the systolic-diastolic fluctuation period. The structure alignment error mean of each frame is obtained by averaging the image embedding amount difference value.

4. The method of claim 1, wherein: The structure alignment error mean of each systolic-diastolic fluctuation period is obtained, specifically including: The structure alignment error mean of each systolic-diastolic fluctuation period is compared with the pre-stored structure alignment error mean threshold in the database. If the structure alignment error mean of a certain systolic-diastolic fluctuation period is greater than or equal to the structure alignment error mean threshold, the structure period deviation judgment result of the systolic-diastolic fluctuation period is that there is a structural period deviation. If the structure alignment error mean of a certain systolic-diastolic fluctuation period is less than the structure alignment error mean threshold, the structure period deviation judgment result of the systolic-diastolic fluctuation period is that there is no structural period deviation.

5. The method of claim 1, wherein: The structure alignment error mean of each systolic-diastolic fluctuation period is obtained, specifically including: If a certain systolic-diastolic fluctuation period has a structural period deviation, the image embedding amount difference value of each frame in the systolic-diastolic fluctuation period is obtained, and the frames with an image embedding amount difference value greater than or equal to an image embedding amount difference value threshold are selected as deviation frames. The continuous number of deviation frames in the systolic-diastolic fluctuation period is counted. If the continuous number of deviation frames is greater than or equal to a preset continuous number threshold, the structural period deviation of the systolic-diastolic fluctuation period is defined as a serious deviation, and a warning is given. If the continuous number of the deviation frames is less than the preset continuous number threshold, the structural period deviation of the systolic-diastolic fluctuation cycle is defined as mild deviation, the continuous number and distribution skewness of the deviation frames are obtained, the continuous number of the deviation frames is mapped into the mapping set of the pre-stored continuous number-image enhancement parameter in the database for mapping matching, the image enhancement parameter of the systolic-diastolic fluctuation cycle is obtained, the systolic-diastolic fluctuation cycle is image enhanced by using the image enhancement parameter of the systolic-diastolic fluctuation cycle, the distribution skewness is input into the mapping set of the pre-stored distribution skewness-time sequence position adjustment parameter in the database, the time sequence position adjustment parameter of the systolic-diastolic fluctuation cycle is obtained, and the time sequence position of each deviation frame of the systolic-diastolic fluctuation cycle is adjusted by using the time sequence position adjustment parameter of the systolic-diastolic fluctuation cycle. After the image enhancement and the time sequence position adjustment are completed, the rhythm abnormality judgment process is entered.

6. The method of claim 1, wherein: The image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation judgment result is introduced. Specifically, the image embedding feature parameters of each systolic-diastolic fluctuation cycle are analyzed, and the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle is obtained after the influence of the structural period deviation ​ ​ ​ ​ ​ 7. The method of claim 1, wherein: ​ ​ ​ 8. The method of claim 7, wherein: ​ The Fourier transform is performed on the inter-frame dynamic parameter sequence of each systolic-diastolic fluctuation cycle to extract the main frequency of each frame, and the spectral multi-peak ratio of each systolic-diastolic fluctuation cycle is obtained; if the spectral multi-peak ratio of a certain systolic-diastolic fluctuation cycle is greater than a preset spectral multi-peak ratio threshold, the abnormal continuous systolic-diastolic fluctuation cycle is marked and a warning is given; If the spectral multi-peak ratios of all systolic-diastolic fluctuation cycles are less than or equal to the spectral multi-peak ratio threshold, the main frequencies of each frame in each systolic-diastolic fluctuation cycle are compared with the main frequency with the highest proportion in the whole cycle, and a loss factor is constructed based on the comparison result; the loss factors of each frame are compared with the pre-stored loss factor threshold in the database, and if the loss factors of each frame do not exceed the threshold, the correct rhythm frequency generation algorithm is used to generate the rhythm frequency and correct each frame; If the loss factor of a certain frame in a certain systolic-diastolic fluctuation cycle exceeds the threshold, a warning is given.

9. System for applying the anomaly detection method of cardiac ultrasound data according to any one of claims 1 to 8, characterized in that, The system specifically comprises: a feature extraction module, a cycle identification module, a structural cycle deviation judgment module, and a rhythm abnormality judgment module; The feature extraction module is configured to perform double-channel feature extraction detection on the original heart B-ultrasound image sequence to obtain the inter-frame dynamic feature value and the image embedding feature value of each frame in the original heart B-ultrasound image sequence. The cycle identification module is configured to identify each systolic-diastolic fluctuation cycle in the original heart B-ultrasound image sequence based on the inter-frame dynamic feature and the image embedding feature value of each frame. The structural cycle deviation judgment module is configured to obtain the structural alignment error mean of each systolic-diastolic fluctuation cycle, obtain the structural cycle deviation judgment result of each systolic-diastolic fluctuation cycle, and determine whether to enter the rhythm abnormality judgment process based on the structural cycle deviation judgment result. The rhythm abnormality judgment module analyzes the image embedding feature parameters of each systolic-diastolic fluctuation cycle, obtains the rhythm abnormality judgment result of each systolic-diastolic fluctuation cycle after introducing the influence of the structural cycle deviation judgment result, and performs corresponding operations based on the rhythm abnormality judgment result.

Images