An intelligent auxiliary diagnosis system for children DCM based on deep learning of cardiac ultrasound images

Abstract

This invention discloses an intelligent auxiliary diagnostic system for pediatric dilated cardiomyopathy (DCM) based on deep learning of cardiac ultrasound images, belonging to the field of intelligent auxiliary diagnostic technology. It includes a multi-source synchronous acquisition module, a passive displacement recognition module, a time-series baseline rearrangement module, an apical stabilization window module, and a risk assessment output module. The multi-source synchronous acquisition module acquires continuous cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals, writing these data sequentially into the same time series chain to establish a traceable time anchor for subsequent dynamic analysis. This invention ensures temporal consistency of cardiac ultrasound data through multi-source synchronous acquisition and time anchor alignment. Combined with passive displacement feature bands and an apical stabilization window, it enables the identification and removal of abnormal segments, effectively suppressing non-physiological deformation interference and improving the stability of cardiac dynamics analysis and the accuracy of dilated cardiomyopathy diagnosis in children.

Description

Technical Field

[0001] This invention relates to the field of intelligent assisted diagnostic technology, specifically to an intelligent assisted diagnostic system for pediatric DCM based on deep learning of cardiac ultrasound images. Background Technology

[0002] The Pediatric DCM Intelligent Auxiliary Diagnostic System Based on Deep Learning of Echocardiogram Images is an intelligent sensing system designed for screening and diagnosing dilated cardiomyopathy in children. Using echocardiogram images as the core input, the system combines highly sensitive ultrasound intelligent sensing units with deep learning algorithms. Through model structures such as deep convolutional networks, temporal feature extraction networks, or multimodal fusion networks, it automatically identifies, quantifies, and assesses the risk of key ultrasound indicators such as left ventricular dilation, abnormal wall motion, and changes in ejection fraction. The system comprises an ultrasound image acquisition and sensing module, an image quality enhancement and structured segmentation module, a deep learning feature extraction and assessment module, and a diagnostic output module. It can automatically identify changes in cardiac structure and functional abnormalities without requiring manual frame-by-frame measurements, and provides DCM risk warnings based on real-time pathological feature signals acquired by the intelligent sensing system, thus providing pediatric echocardiologists with auxiliary and interpretable diagnostic evidence.

[0003] The existing technology has the following shortcomings: In existing technologies, intelligent auxiliary diagnostic systems for pediatric diabetic cerebrovascular disease (DCM) based on echocardiography images typically rely on deep learning models to extract features and analyze dynamic changes in continuous cardiac cycle images to identify chamber expansion trends and myocardial motion abnormalities. However, during actual ultrasound acquisition, children are highly susceptible to transient chest wall compression due to respiratory fluctuations and probe contact pressure fluctuations. This phenomenon causes non-physiological deformation in a localized area at the apex of the heart, resulting in passive displacement trajectories in the image frame sequence. Existing models lack a mechanism to distinguish these transient non-physiological deformations, often misinterpreting them as a continuous evolution of left ventricular dilation, thus creating a false chamber expansion trend within the dynamic analysis interval. This misidentification not only distorts the system's temporal understanding of left ventricular function but may also lead to abnormally high DCM risk indices output by the model, severely impacting the reliability and safety of clinical diagnosis.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a pediatric DCM intelligent auxiliary diagnostic system based on deep learning of cardiac ultrasound images, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a pediatric DCM intelligent auxiliary diagnostic system based on deep learning of cardiac ultrasound images, comprising a multi-source synchronous acquisition module, a passive displacement recognition module, a temporal baseline rearrangement module, an apical stabilization window module, and a risk assessment output module: The multi-source synchronous acquisition module acquires continuous cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals. It writes the cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals into the same time series chain in chronological order, establishing a traceable time anchor for subsequent dynamic analysis to achieve time alignment of multi-source information. The passive displacement recognition module performs correlation analysis on cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals based on time anchors. It identifies the trajectory of non-physiological deformation of the apex of the heart caused by instantaneous chest wall compression and generates a passive displacement feature band representing the abnormal deformation range, thus marking the location of abnormal segments in the time series chain. The temporal rearrangement baseline module rearranges the temporal order of cardiac ultrasound image sequences based on the passive displacement feature band, delays the arrangement of the identified abnormal segments, forms a stable temporal baseline composed of normal cardiac motion segments, and maintains the consistency between the temporal anchor and the multi-source signals. The apical stabilization window module establishes an apical stabilization window on the stable time baseline, removes abnormal segments outside the stable time baseline from the image analysis path, and generates a continuous chain of left ventricular contours using only continuous image frames within the apical stabilization window, ensuring the temporal stability of left ventricular structural change analysis. The risk assessment output module uses the left ventricular contour continuum as input to drive the deep learning risk assessment module, suppresses residual feature interference caused by passive displacement, outputs the risk judgment result of dilated cardiomyopathy based on the apical stability window, and forms clinically interpretable auxiliary diagnostic information.

[0007] Preferably, the steps for establishing a multi-source information time series chain and constructing a time anchor are as follows: The cardiac ultrasound data of the subject is continuously acquired. Through the signal triggering channel connected to the timing control unit of the medical ultrasound equipment, the cardiac ultrasound image sequence is continuously recorded at a predetermined frame rate within the cardiac cycle. At the same time, probe contact pressure information and chest wall displacement signal are acquired synchronously with the image frames, so that the pressure information and displacement signal are synchronized with the cardiac ultrasound image sequence. After acquisition, the cardiac ultrasound image sequence, probe contact pressure information and chest wall displacement signal are time-stamped. By assigning the same timestamp to each frame of image, pressure information and displacement signal, the correspondence of multi-source data in the time dimension is realized. Based on time markers, cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals are sequentially integrated to form a continuous time series chain, ensuring constant time intervals and eliminating time drift. A traceable time anchor is established on the basis of the time series chain, and a time series reference frame with physical meaning is formed by time integration calibration, providing a time benchmark for subsequent dynamic analysis.

[0008] Preferably, the time anchor is established with the starting moment of the time series chain as the reference point. By integrating and calibrating the continuous time intervals, the cardiac ultrasound image information, probe contact pressure information and chest wall displacement signal at each moment are mapped to a unified time axis. This ensures that the time anchor remains consistent in the subsequent dynamic change identification and deformation trajectory analysis process, thereby guaranteeing the temporal logic consistency of multi-source information in the dynamic analysis stage.

[0009] Preferably, the passive displacement feature band generation steps are as follows: Based on the time anchor, the correspondence between each time node is established. The cardiac ultrasound image frame, probe contact pressure information and chest wall displacement signal are matched by the timestamp, so that each time unit in the time series chain has both spatial morphology and external force response information. After forming a complete time unit, the temporal variation trend of probe contact pressure information and chest wall displacement signal is analyzed to identify the instantaneous changes in the same direction of pressure rise and displacement increase in a short period of time, and the non-physiological deformation trend is determined by combining the changes in the image frame of the apex region. Time periods with non-physiological deformation trends are continuously marked and aggregated in the time series chain to form abnormal intervals and record the start and end times. A passive displacement feature band is generated based on the abnormal intervals. The abnormal intervals are then arranged continuously on the time axis to mark the location of abnormal segments and form a feature expression with temporal continuity.

[0010] Preferably, when generating the passive displacement feature band, the time anchor is used as the main index to merge the cardiac ultrasound image frames, probe contact pressure information and chest wall displacement signals in the abnormal interval according to the time sequence, so that each abnormal interval forms a continuous time band segment and maintains the time interval with the normal interval in the time series chain, thereby ensuring that the passive displacement feature band has continuity and traceability in the time dimension.

[0011] Preferably, the steps for generating a stable time baseline are as follows: After obtaining the passive displacement feature band, the time series chain is divided into time segments. Based on the start and end times of the abnormal interval, the time series chain is divided into continuous time periods, and the time correspondence between cardiac ultrasound image frames, probe contact pressure information and chest wall displacement signals is maintained. Based on the time anchor, the time sequence of each time period is reconstructed, normal segments are arranged continuously in the original order, and abnormal segments are arranged in a delayed order to form a continuous time structure. The continuous time structure composed of normal segments is extended to smooth the time transition between adjacent segments to form a stable time baseline and maintain time consistency with multi-source signals. The rearranged time anchors are kept consistent by remapping them to a new time structure, ensuring a strict time correspondence between cardiac ultrasound image frames, probe contact pressure information, and chest wall displacement signals.

[0012] Preferably, during the time-series chain rearrangement, the time intervals of all time periods are normalized by using a unified time reference, so that adjacent time segments are continuously connected on the time axis. After a stable time baseline is formed, the time anchor sequence of abnormal segments is shifted backward, thereby ensuring the continuity of the stable time baseline and the temporal consistency of cardiac ultrasound image frames, probe contact pressure information, and chest wall displacement signals.

[0013] Preferably, the steps for generating the left ventricular contour continuum are as follows: The temporal distribution range of the apical region is determined on a stable time baseline. By traversing continuous normal cardiac motion segments over time, time intervals containing complete cardiac cycles are identified, and time periods of stable apical motion are selected based on time anchors to define the initial time range of the apical stability window. The continuity of image frames within the apical stable window is screened, and adjacent image frames are sequentially associated by time anchors to identify the temporal coherence of the apical structure. Frame groups that conform to the continuous systolic and diastolic change pattern are identified as effective continuous segments. Abnormal segments outside the stable time baseline are removed from the image analysis path. The image frames, probe contact pressure information and chest wall displacement signals corresponding to the abnormal segments are removed synchronously by comparing the time anchor index. The continuous image frames within the apical stable window are temporally reassembled according to the time anchor sequence to generate a continuous chain of left ventricular contours, so that the structural changes of the left ventricle form a continuous mapping in the time dimension.

[0014] Preferably, when the time range of the apical stability window is determined based on the stability of apical movement, the continuous time period with the lowest pressure fluctuation amplitude and the smallest displacement change is selected by analyzing the synchronous change trend of the probe contact pressure information and chest wall displacement signal corresponding to the time anchor, so as to ensure that the movement of the apical region within the window only reflects the physiological contraction and relaxation process of the heart.

[0015] Preferably, the risk assessment steps using the left ventricular profile continuum as input are as follows: The time nodes in the continuous chain of left ventricular contours are organized temporally and the left ventricular structure of each frame is arranged in chronological order by time anchor index, so that the spatial contour changes are consistent with the time baseline. Secondly, a fusion correspondence is established between the continuous chain of left ventricular contours and the probe contact pressure information and chest wall displacement signal, so that the structural change information and the external mechanical signal form a synchronous mapping in the time dimension, thereby distinguishing between physiological deformation and external pressure. Based on the multi-source correspondence, a dynamic feature expression structure is formed with the left ventricular contour continuous chain as the core input, which suppresses the residual feature interference caused by passive displacement, so that the input data only reflects the real physiological motion. The risk assessment results are mapped to specific time points within the apical stability window using time anchors, generating time-related risk assessment curves and outputting clinically interpretable auxiliary diagnostic information.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention establishes a multi-source synchronous acquisition and time-anchor correspondence mechanism, enabling precise alignment of cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals in the time dimension. This fundamentally ensures the temporal consistency of data during dynamic analysis. In this way, time drift errors caused by respiratory fluctuations and probe contact fluctuations are effectively eliminated, allowing the identification results of cardiac structural changes to truly reflect physiological motion states, thereby improving the stability and reliability of cardiac motion feature extraction in children.

[0017] This invention achieves the identification, isolation, and delayed processing of abnormal segments by constructing passive displacement feature bands, performing temporal rearrangement, and establishing an apical stable window, enabling structural analysis of the apical region to be performed solely based on a stable time baseline. This design significantly enhances the resistance of dynamic analysis to non-physiological deformations, suppresses the interference of passive displacement on model learning results, and allows the risk assessment process for dilated cardiomyopathy to focus on real cardiac dynamic changes, thereby improving the accuracy and clinical interpretability of diagnostic results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a schematic diagram of a module of a pediatric DCM intelligent auxiliary diagnostic system based on deep learning of cardiac ultrasound images according to the present invention. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0021] This invention provides, for example Figure 1 The illustrated pediatric DCM intelligent auxiliary diagnostic system based on deep learning of cardiac ultrasound images includes a multi-source synchronous acquisition module, a passive displacement recognition module, a temporal baseline rearrangement module, an apical stabilization window module, and a risk assessment output module.

[0022] Continuous acquisition of cardiac ultrasound data from the subject is performed. During acquisition, a signal trigger channel connected to the timing control unit of the medical ultrasound equipment continuously records each frame of the cardiac ultrasound image within the cardiac cycle according to predetermined frame rate parameters, ensuring that the cardiac structure is completely captured throughout the entire systole and diastole process. Simultaneously, at the same time reference, probe contact pressure information synchronized with the image frames is acquired. This information originates from a pressure sensor deployed at the contact end of the ultrasound probe. The pressure sensor generates synchronized pressure data points at the instant of each image frame acquisition, ensuring that each image frame has a corresponding pressure signal value. At the same time, a highly sensitive displacement detection unit is deployed on the chest wall surface of the subject to record minute displacement changes on the chest wall surface during probe contact. Through the parallel operation of the pressure sensor and the displacement detection unit, the pressure information and displacement signal are synchronized with the ultrasound image sequence, thus forming a consistent initial sampling basis in the time dimension.

[0023] Based on continuously sampled data, time-stamping processing is applied to the cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals. Each frame of ultrasound image is assigned a unique timestamp at the moment of acquisition, and the signals generated by the pressure sensor and displacement detection unit are also recorded with the same timestamp number. Through this time stamping, a correspondence is established between the three types of data on the same time dimension, creating a one-to-one mapping between image frames, pressure points, and displacement signals. During this process, a constant sampling interval time scale ensures that the arrangement of each data stream in the time sequence has strict temporal continuity. The resulting time-stamped sequence not only records the physical state at each sampling moment but also provides a traceable temporal information structure for establishing a dynamic analysis benchmark. This time stamping processing allows sampling errors and sampling delays from different data sources to be uniformly described on the time axis, thus creating an accurate temporal reference basis for subsequent multi-source fusion analysis.

[0024] After the time stamps are established, the cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals are sequentially integrated according to their respective time stamps to form a continuous time series chain. This time series chain uses timestamps as the primary index, merging the three types of signal data at each moment in chronological order, ensuring that each time point contains synchronized image frames, pressure values, and displacement amounts. By constructing the time series chain, data from different sources can present a continuous and traceable logical structure in the temporal dimension. To avoid time drift caused by device sampling delays or sensor response differences, a unified time reference is used to normalize all timestamps during the integration process, ensuring that the interval between adjacent time points remains constant and that the entire time series chain macroscopically appears as a continuous and uninterrupted temporal record. The resulting time series chain not only accurately reflects the structural and mechanical changes of the heart during a complete cardiac cycle but also includes pressure fluctuations and displacement responses caused by external contact, providing unified data support for subsequent dynamic feature correlation analysis.

[0025] After the time series chain is integrated, a traceable time anchor is established as a reference benchmark for dynamic analysis based on the synchronization structure of the three types of data in the time series chain. This time anchor uses the starting moment of the time series chain as the reference point. By integrating the time intervals of consecutive frames, the image information, pressure information, and displacement information at each moment are mapped to the same time axis, forming a physically meaningful temporal reference framework. The establishment of the time anchor not only records the true temporal relationship of the data during the acquisition phase but also provides a positioning basis for subsequent analysis. When the system needs to analyze the dynamic changes of the heart structure during subsequent processing, it can trace back to the original acquisition state at any moment through the time anchor, ensuring that the extraction of each dynamic feature is based on an accurate temporal correspondence. Through the establishment of this time anchor, a stable reference system can be formed in the time dimension, comprehensively ensuring the temporal consistency of multi-source information. This time anchor runs through the entire data processing flow and can be continuously used as a time benchmark in subsequent dynamic change identification, deformation trajectory analysis, and temporal rearrangement processes, thereby ensuring the temporal logical consistency of all data during the dynamic analysis phase.

[0026] The passive displacement recognition module performs correlation analysis on cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals based on time anchors. It identifies the trajectory of non-physiological deformation of the apex of the heart caused by instantaneous chest wall compression and generates a passive displacement feature band representing the abnormal deformation range, thus marking the location of abnormal segments in the time series chain. After establishing the time anchor, to ensure a consistent analytical foundation across multiple sources of information in the temporal dimension, and to further analyze the dynamic correlation between echocardiogram image sequences, probe contact pressure information, and chest wall displacement signals, it is necessary to perform correlation analysis on the temporal correspondence and changing trends of various signals under the time anchor. This allows for the identification of non-physiological deformation trajectories at the apex caused by instantaneous chest wall compression, and the generation of passive displacement feature bands in the form of continuous time intervals. The locations of abnormal segments are clearly marked in the time series chain, providing a basis for subsequent temporal rearrangement and baseline construction. The specific steps are as follows:

[0027] Based on time anchors, a time-by-time correspondence is established between each frame of a cardiac ultrasound image sequence and its corresponding probe contact pressure information and chest wall displacement signal. Specifically, using the timestamps recorded by the time anchors, image frames, pressure data points, and displacement data points at the same time point are matched, ensuring that the three types of signals constitute a complete time unit at each moment. During this process, for the continuously acquired time series chain, the time intervals are kept consistent, ensuring that adjacent time units are continuously distributed along the time axis. This time-by-time correspondence method allows for the synchronous correlation of imaging information of cardiac structural changes with the mechanical changes of external forces, giving each data node in the time series chain dual information of spatial morphology and external force response. The construction of this time unit provides a fine-grained temporal foundation for subsequent dynamic correlation analysis, making physical signals from different sources comparable and traceable in the time dimension.

[0028] After forming a complete time unit, the temporal correspondence between the changes in image frames and the changes in external force signals within that time unit is continuously analyzed. By comparing the temporal synchronous change trends of probe contact pressure information and chest wall displacement signals in each time unit, it is observed whether there are instantaneous changes in the same direction of pressure increase and displacement increase within a continuous period. When the probe contact pressure rises rapidly in a short period of time, and the chest wall displacement signal shows an outward shift trend within the same time interval, it indicates that there is an instantaneous chest wall compression phenomenon caused by external force within that time interval. At this time, the ultrasound image frame change characteristics corresponding to that time interval are included in the key analysis scope, and the image frame changes in adjacent time intervals are compared. If the spatial position of the apical region of the heart shows a non-periodic displacement in the image frame that is different from normal myocardial contraction, it can be preliminarily judged that there is a non-physiological deformation trend within that time interval. Through this comparison of continuous temporal trends, the changes in external force and changes in image structure can be organically linked, thereby identifying time intervals that may contain passive deformation on the time axis.

[0029] After identifying time periods exhibiting non-physiological deformation trends, the distribution of these time periods within the time series chain is continuously calibrated, and these time periods are aggregated as anomalous intervals. Using a time anchor as a reference, adjacent time segments experiencing continuous external force changes accompanied by abnormal apical morphological changes are merged to form continuous anomalous intervals. Then, using these anomalous intervals as boundaries, the time series chain is divided into intervals containing normal dynamic changes and intervals containing passive deformation. To ensure clear identification of these anomalous intervals in subsequent analysis, they are marked with time indices within the time series chain, and the start and end times of each anomalous interval are recorded on the time axis. Through this continuous calibration and interval aggregation, the time segments of passive deformation have a clear position within the overall temporal structure and form a clear boundary with adjacent normal time periods. This operation not only achieves temporal localization of non-physiological deformations but also lays the foundation for generating physically meaningful feature representations.

[0030] After identifying the abnormal intervals, a passive displacement feature band representing the abnormal deformation intervals is generated based on the identified abnormal intervals in the time series chain. This feature band uses the time axis as the main line, arranging all abnormal intervals continuously in chronological order, so that each abnormal interval corresponds to a continuous time segment on the time axis. Each time segment corresponds to a set of linked changes in three types of data: echocardiogram image frames, probe contact pressure information, and chest wall displacement signals. Therefore, this feature band includes both the temporal span information of non-physiological deformation and the multi-source signal characteristics during the abnormality. The generation process of the passive displacement feature band transforms the temporal distribution of abnormal deformation into a continuous temporal feature expression, allowing it to exist as an independent information layer in the time series chain. When the time series chain is called by subsequent processing stages, the location and duration of abnormal segments can be directly determined through the passive displacement feature band, thereby achieving accurate tracking and temporal localization of non-physiological deformation. Through the construction of this passive displacement feature band, non-physiological deformation has a clear expression in the temporal dimension, providing a clear basis for subsequent temporal rearrangement and stable baseline construction.

[0031] The temporal rearrangement baseline module rearranges the temporal order of cardiac ultrasound image sequences based on the passive displacement feature band, delays the arrangement of the identified abnormal segments, forms a stable temporal baseline composed of normal cardiac motion segments, and maintains the consistency between the temporal anchor and the multi-source signals. After generating the passive displacement feature bands and temporally calibrating the abnormal segments, to ensure the temporal continuity and physiological consistency of subsequent cardiac structural change analysis, the cardiac ultrasound image sequence needs to be rearranged chronologically based on the passive displacement feature bands. This delays the processing of the calibrated abnormal segments in the temporal dimension, thereby forming a stable time baseline composed of normal cardiac motion segments. During this process, it is also necessary to maintain the consistency between the time anchor and the multi-source signals, ensuring that the rearranged time series chain has a strict temporal reference relationship with the original acquired data in terms of dynamic characteristics. The specific steps are as follows:

[0032] After obtaining the passive displacement feature band, the time series chain is divided into normal and abnormal segments. By reading the start and end times of the abnormal intervals marked in the passive displacement feature band, the time series chain is divided into several continuous time periods according to the boundaries between the abnormal and normal intervals. Each time period maintains the temporal correspondence between three types of data: cardiac ultrasound image frames, probe contact pressure information, and chest wall displacement signals, ensuring that the segmented time structure maintains the overall continuity of the time anchor. During this process, the time index of each segment is carefully preserved to ensure accurate recombination of segments during subsequent reordering. This time segmentation operation based on the passive displacement feature band establishes the temporal boundary between normal and abnormal segments, providing a basic time unit for subsequent reordering operations.

[0033] After time segmentation, the order of each time segment on the time axis is reconstructed. Using the time anchor as the core reference, all normal segments are kept in their original acquisition time order, ensuring their continuous arrangement on the time axis. Simultaneously, segments identified as abnormal are temporarily removed from the original order and placed at the end of the time series chain in a delayed manner. During reconstruction, the inter-frame interval within each time segment is kept constant to ensure that the inter-frame temporal characteristics of the cardiac ultrasound image sequence are not disrupted. In this way, the reordered time series chain is composed entirely of normal cardiac motion segments in the initial section, with continuous frame order and uniform time steps, thus forming a time series backbone free from artifact interference. During reconstruction, the time anchor remains as a global time reference, allowing the delayed abnormal segments to still maintain a correspondence with the original multi-source signals through the time anchor, ensuring the traceability and consistency of the entire time structure.

[0034] After time rearrangement, a stable time baseline is constructed for the resulting continuous structure of normal segments. Based on the rearranged time series chain, time intervals composed of normal cardiac motion segments are temporally extended to smooth the temporal transitions between adjacent segments. By uniformly connecting the time intervals of continuous time segments, the rearranged time structure exhibits uninterrupted and abrupt continuity in the time dimension. At this point, the stable time baseline not only reflects the actual contraction and relaxation rhythms of the heart on the time axis but also maintains a synchronous correspondence with probe contact pressure information and chest wall displacement signals, thereby achieving overall coordination among multi-source signals in the time dimension. This stable time baseline serves as the basis for subsequent apical stabilization window establishment and left ventricular contour continuous chain extraction. Its role is to provide a time reference unaffected by external non-physiological deformations, enabling the continuous expression of the entire cardiac motion process under physiological conditions.

[0035] After a stable time baseline is established, the rearranged time anchors are maintained for consistency to ensure that the temporal correspondence between the time anchors and multi-source signals remains unchanged. By recalibrating the time index of each time node in the rearranged time series chain, the time anchors are mapped from the original sequence to the new time structure, ensuring that the cardiac ultrasound image frames, probe contact pressure information, and chest wall displacement signals corresponding to each time anchor remain strictly consistent in time. For aberrant segments that are delayed, their time anchors maintain their relative order but are shifted sequentially on the global time axis to ensure that the temporal attributes of aberrant segments do not affect the continuity of the stable time baseline. Through this time anchor consistency maintenance operation, the entire time series chain still possesses a complete temporal reference system after rearrangement, ensuring that in subsequent analyses, whether normal or aberrant segments are accessed, the corresponding state at the original acquisition time can be traced in the time dimension. In this way, the rearranged time structure logically maintains the continuity and traceability of the original data, while functionally achieving the stabilization of cardiac motion sequences and the isolation of non-physiological interferences.

[0036] The apical stabilization window module establishes an apical stabilization window on the stable time baseline, removes abnormal segments outside the stable time baseline from the image analysis path, and generates a continuous chain of left ventricular contours using only continuous image frames within the apical stabilization window, ensuring the temporal stability of left ventricular structural change analysis. After establishing a stable time baseline, to further ensure the temporal continuity of cardiac structural changes and the physiological consistency of the analysis, an apical stable window needs to be established on the stable time baseline. This involves filtering the dynamic characteristics of each time segment in the echocardiogram image sequence, selecting image frames with continuous cardiac motion within the stable time baseline range, and removing abnormal segments outside the stable time baseline from the image analysis path. Only continuous image frames within the apical stable window are retained for subsequent generation of the left ventricular structural contour chain, thus ensuring the stability and continuity of the left ventricular structural change analysis in the temporal dimension. The specific steps are as follows:

[0037] The temporal distribution range of the apical region is determined within the overall structure of the stable time baseline. By traversing continuous segments of normal cardiac motion within the stable time baseline, time intervals containing complete cardiac cycles are identified, and the relative motion amplitude of the apex at different time points is located within these intervals. Using the correspondence of time anchors, the change in apical position at each time point is mapped to probe contact pressure information and chest wall displacement signals, thus selecting the time segment where the apical motion is most stable and least affected by external displacement throughout the entire time baseline. The initial time range of the apical stability window is defined using the start and end times of this time segment as time boundaries. Determining this time range allows subsequent analysis to focus on the period of stable cardiac motion in the temporal dimension, avoiding the influence of non-physiological interference on morphological change analysis.

[0038] After determining the initial range of the apical stable window, image frames within this time range are screened for continuity. By reading the ultrasound image frame sequence corresponding to this window from the stable time baseline, adjacent image frames are temporally correlated according to their time anchor order, ensuring that the inter-frame time interval remains consistent with the original acquisition. Based on this, the temporal coherence of each consecutive image frame group is evaluated to identify whether the structural changes between frames remain continuous and consistent in direction. When the apical structure between adjacent image frames exhibits regular contraction and relaxation changes in spatial location and is consistent with the temporal sequence direction on the time baseline, this frame group is included in the valid continuous segment of the apical stable window. Segments with inter-frame abrupt changes, morphological shifts, or asynchrony with the pressure signal are excluded from the apical stable window. This screening process ensures that the apical stable window contains only image frames that maintain a true physiological rhythm on the stable time baseline, providing a highly consistent image sequence basis for subsequent left ventricular structure extraction.

[0039] After determining the effective continuous segment of the apical stable window, abnormal segments outside the stable time baseline are completely removed from the image analysis path. Using the index relationship of time anchors, the time nodes corresponding to the abnormal segments are compared with the time nodes within the stable window, and the image frames, probe contact pressure information, and chest wall displacement signals corresponding to the abnormal segments are simultaneously removed from the data processing path on the time axis. During the removal operation, the temporal continuity of the stable time baseline is maintained, ensuring that the time series within the apical stable window remains continuously distributed according to the time anchor order. After removal, the entire time series structure consists of two parts: the former is the continuous frame sequence within the stable window, and the latter is the isolated abnormal segment. At this point, the apical stable window forms a one-to-one correspondence with the stable time baseline in terms of time structure, ensuring that the time anchors remain consistent throughout the analysis process. This operation allows the subsequent left ventricular contour extraction step to be performed entirely based on the data within the stable window on the time axis, avoiding interference from abnormal segments in the structural analysis.

[0040] After removing abnormal fragments, the continuous image frames within the apical stable window are temporally reassembled to generate a continuous chain of left ventricular contours. Using the starting frame within the apical stable window as a reference, adjacent frames are arranged sequentially according to time anchors, allowing the structural changes of the left ventricle at different time points to unfold continuously along a timeline. By temporally mapping the spatial distribution of the left ventricular morphology in each frame, the structural boundaries of the left ventricle in different systolic and diastolic phases are continuously mapped onto the time axis, thus constructing a continuous chain of left ventricular structural changes in the temporal dimension. This continuous chain not only records the morphological changes of the left ventricle within a complete stable cardiac cycle but also maintains synchronization with time anchors, probe contact pressure information, and chest wall displacement signals. Through the generation of this continuous chain, the temporal characteristics of cardiac structural changes are stably expressed, providing precise temporal and morphological evidence for subsequent risk analysis and diagnostic output.

[0041] The risk assessment output module uses the left ventricular contour continuum as input to drive the deep learning risk assessment module, suppresses the residual feature interference caused by passive displacement, outputs the risk judgment result of dilated cardiomyopathy based on the apical stability window, and forms clinically interpretable auxiliary diagnostic information. After obtaining the left ventricular contour continuum, to achieve intelligent assisted risk assessment of dilated cardiomyopathy in children, it is necessary to use the left ventricular contour continuum as input information to drive the deep learning risk assessment process, while ensuring temporal continuity and structural consistency. This suppresses residual feature interference caused by passive displacement within a temporally stable physiological reference frame, outputting a dilated cardiomyopathy risk assessment result based on an apical stable window, and forming clinically interpretable diagnostic information. The specific steps are as follows:

[0042] After generating the continuous chain of left ventricular contours within the apical stable window, the spatial contours of each time node in the continuous chain are temporally structured. Using the index relationship of time anchors, each left ventricular structural frame in the continuous chain is arranged sequentially in chronological order, ensuring that each frame corresponds to a specific time node and that the time interval is consistent with the stable time baseline. At this point, each contour in the continuous chain of left ventricular contours not only reflects the geometric state of the cardiac structure at a specific moment but also records the dynamic evolution of the apical region within the stable time baseline. This temporal structuring ensures that the left ventricular contours exhibit a continuous contraction and relaxation process in the temporal dimension, and that all morphological changes are consistent with the physiological rhythm within the stable window. During the structuring process, the correspondence between the left ventricular boundary and the probe contact pressure information and chest wall displacement signal at each time node is maintained, ensuring that spatial structural changes and mechanical responses remain consistent in the temporal dimension. This temporal structuring provides a complete and temporally coherent input basis for subsequent risk assessment.

[0043] Based on temporal structuring, a fusion correspondence is established between the continuous chain of left ventricular contours and multi-source signals. Through time anchoring, the probe contact pressure information and chest wall displacement signal corresponding to each frame of the left ventricular contour are realigned, ensuring that at the same time point, both cardiac structural changes and corresponding external mechanical effects are present. This approach allows for a coordinated mapping of structural motion and external forces in the temporal dimension, distinguishing between deformation caused by physiological contraction and passive displacement caused by external pressure. During this process, time points within the apical stability window are iterated sequentially, ensuring a complete record of the relationship between all contour changes and corresponding signal changes. Since abnormal segments have been removed in the previous step, the current time series structure is entirely composed of real physiological activity, guaranteeing that the correspondence between multi-source signals purely reflects the heart's own dynamic characteristics. This temporal synchronization and information fusion enables subsequent risk assessment based on the actual cardiac motion state and suppresses the residual influence of external artifact signals in the temporal dimension.

[0044] After establishing multi-source correspondences, a dynamic feature expression structure oriented towards risk output is constructed using the left ventricular contour continuum as the core input. Through contour connections between consecutive frames, the cardiac contraction and relaxation processes are unfolded temporally, allowing the volume change trend of the left ventricular cavity within a complete cardiac cycle to be presented. In this dynamic process, a time anchor ensures the accurate temporal location of each contour change, thus making the morphological changes in the time series traceable and temporally consistent. Based on this, the temporally continuous information reflecting cardiac dynamics characteristics from the contour continuum is input into the risk assessment process, enabling the risk assessment to capture the true motion pattern of the heart within a stable window. Since the abnormal segments corresponding to passive displacement feature bands have been delayed and removed from the analysis path, the current input data no longer includes transient deformations caused by external force interference, effectively suppressing the influence of residual interference on the results. Through this input structure based on a stable time baseline and an apical stable window, the risk assessment process can focus on analyzing the heart's own dynamic abnormalities, ensuring consistency between the output risk judgment and the actual pathological state.

[0045] After completing the risk analysis of dynamic characteristics, the risk assessment results are time-mapped and diagnostic information is output. Using a time anchor, the risk assessment results are correlated with the time range of the apical stability window, ensuring that each risk value corresponds to a specific time point within the stability window, thus forming a time-related risk assessment curve. This curve allows for a visual observation of the functional changes in the heart throughout the entire stable cardiac cycle and identifies areas of abnormal contraction or delayed dilation of the left ventricle at different stages. During this process, the risk assessment results are combined with a continuous chain of left ventricular contours, allowing each risk value to correspond to a specific cardiac structural morphology, resulting in a diagnostic result with both temporal continuity and spatial visualization. Through this output method, clinical users can not only obtain risk warning values ​​for dilated cardiomyopathy but also intuitively understand the structural basis for risk formation. This output result uses a stable time baseline as the core time frame, the apical stability window as the dynamic range, and the continuous chain of left ventricular contours as the structural basis, achieving a full-link correspondence from time to space to risk result, making the auxiliary diagnostic information interpretable and clinically readable.

[0046] This invention establishes a multi-source synchronous acquisition and time-anchor correspondence mechanism, enabling precise alignment of cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals in the time dimension. This fundamentally ensures the temporal consistency of data during dynamic analysis. In this way, time drift errors caused by respiratory fluctuations and probe contact fluctuations are effectively eliminated, allowing the identification results of cardiac structural changes to truly reflect physiological motion states, thereby improving the stability and reliability of cardiac motion feature extraction in children.

[0047] This invention achieves the identification, isolation, and delayed processing of abnormal segments by constructing passive displacement feature bands, performing temporal rearrangement, and establishing an apical stable window, enabling structural analysis of the apical region to be performed solely based on a stable time baseline. This design significantly enhances the resistance of dynamic analysis to non-physiological deformations, suppresses the interference of passive displacement on model learning results, and allows the risk assessment process for dilated cardiomyopathy to focus on real cardiac dynamic changes, thereby improving the accuracy and clinical interpretability of diagnostic results.

[0048] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A pediatric DCM intelligent auxiliary diagnostic system based on deep learning of cardiac ultrasound images, characterized in that, It includes a multi-source synchronous acquisition module, a passive displacement recognition module, a time-series baseline rearrangement module, an apical stabilization window module, and a risk assessment output module: The multi-source synchronous acquisition module acquires continuous cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals, and writes the cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals into the same time series chain in chronological order to establish a traceable time anchor. The passive displacement recognition module performs correlation analysis on cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals based on time anchors to identify the non-physiological deformation trajectory of the apex caused by instantaneous chest wall compression and generate passive displacement feature bands. The temporal rearrangement baseline module rearranges the temporal order of cardiac ultrasound image sequences based on the passive displacement feature band, delaying the arrangement of the identified abnormal segments to form a stable temporal baseline composed of normal cardiac motion segments. The apical stabilization window module establishes an apical stabilization window on the stable time baseline, removes abnormal segments outside the stable time baseline from the image analysis path, and generates a continuous chain of left ventricular contours using only continuous image frames within the apical stabilization window. The risk assessment output module uses the continuous chain of the left ventricular contour as input to drive the deep learning risk assessment module, suppresses the interference of residual features caused by passive displacement, and outputs the risk judgment result of dilated cardiomyopathy based on the apical stability window.

2. The intelligent auxiliary diagnostic system for pediatric DCM based on deep learning of cardiac ultrasound images according to claim 1, characterized in that, The steps to establish a multi-source information time series chain and construct a time anchor are as follows: The cardiac ultrasound data of the subject is continuously acquired. Through the signal triggering channel connected to the timing control unit of the medical ultrasound equipment, the cardiac ultrasound image sequence is continuously recorded at a predetermined frame rate within the cardiac cycle. At the same time, probe contact pressure information and chest wall displacement signal are acquired synchronously with the image frames, so that the pressure information and displacement signal are synchronized with the cardiac ultrasound image sequence. After acquisition, the cardiac ultrasound image sequence, probe contact pressure information and chest wall displacement signal are time-stamped by assigning the same timestamp to each frame of image, pressure information and displacement signal. Based on time markers, cardiac ultrasound image sequences, probe contact pressure information, and chest wall displacement signals are sequentially integrated to form a continuous time series chain, ensuring constant time intervals and eliminating time drift. A traceable time anchor is established on the basis of the time series chain, and a time series reference frame with physical meaning is formed by time integration calibration.

3. The intelligent auxiliary diagnostic system for pediatric DCM based on deep learning of cardiac ultrasound images according to claim 2, characterized in that, The time anchor is established with the starting moment of the time series chain as the reference point. By integrating and calibrating the continuous time intervals, the cardiac ultrasound image information, probe contact pressure information and chest wall displacement signal at each moment are mapped to a unified time axis, so that the time anchor remains consistent in the subsequent dynamic change identification and deformation trajectory analysis process.

4. The intelligent auxiliary diagnostic system for pediatric DCM based on deep learning of cardiac ultrasound images according to claim 2, characterized in that, The steps for generating the passive displacement feature band are as follows: Based on the time anchor, the correspondence between each time node is established. The cardiac ultrasound image frame, probe contact pressure information and chest wall displacement signal are matched by the timestamp, so that each time unit in the time series chain has both spatial morphology and external force response information. After forming a complete time unit, the temporal variation trend of probe contact pressure information and chest wall displacement signal is analyzed to identify the instantaneous changes in the same direction of pressure rise and displacement increase in a short period of time, and the non-physiological deformation trend is determined by combining the changes in the image frame of the apex region. Time periods with non-physiological deformation trends are continuously marked and aggregated in the time series chain to form abnormal intervals and record the start and end times. A passive displacement feature band is generated based on the abnormal intervals, and the abnormal intervals are arranged continuously on the time axis.

5. The intelligent auxiliary diagnostic system for pediatric DCM based on deep learning of cardiac ultrasound images according to claim 4, characterized in that, When generating passive displacement feature bands, the time anchor is used as the main index to merge cardiac ultrasound image frames, probe contact pressure information and chest wall displacement signals in abnormal intervals according to time order, so that each abnormal interval forms a continuous time band segment and maintains the time interval with the normal interval in the time series chain.

6. The intelligent auxiliary diagnostic system for pediatric DCM based on deep learning of cardiac ultrasound images according to claim 4, characterized in that, The steps for generating a stable time baseline are as follows: After obtaining the passive displacement feature band, the time series chain is divided into time segments. Based on the start and end times of the abnormal interval, the time series chain is divided into continuous time periods, and the time correspondence between cardiac ultrasound image frames, probe contact pressure information and chest wall displacement signals is maintained. Based on the time anchor, the time sequence of each time period is reconstructed, normal segments are arranged continuously in the original order, and abnormal segments are arranged in a delayed order to form a continuous time structure. The continuous time structure composed of normal segments is extended to smooth the time transition between adjacent segments to form a stable time baseline and maintain time consistency with multi-source signals. The rearranged time anchors are kept consistent by remapping them to a new time structure, ensuring a strict time correspondence between cardiac ultrasound image frames, probe contact pressure information, and chest wall displacement signals.

7. The intelligent auxiliary diagnostic system for pediatric DCM based on deep learning of cardiac ultrasound images according to claim 6, characterized in that, In the process of reordering the time series chain, the time intervals of all time periods are normalized by using a unified time base, so that adjacent time segments are connected continuously on the time axis, and after a stable time baseline is formed, the time anchor order of abnormal segments is shifted backward.

8. The intelligent auxiliary diagnostic system for pediatric DCM based on deep learning of cardiac ultrasound images according to claim 6, characterized in that, The steps for generating the left ventricular contour continuum are as follows: The temporal distribution range of the apical region is determined on a stable time baseline. By traversing continuous normal cardiac motion segments over time, time intervals containing complete cardiac cycles are identified, and time periods of stable apical motion are selected based on time anchors to define the initial time range of the apical stability window. The continuity of image frames within the apical stable window is screened, and adjacent image frames are sequentially associated by time anchors to identify the temporal coherence of the apical structure. Frame groups that conform to the continuous systolic and diastolic change pattern are identified as effective continuous segments. Abnormal segments outside the stable time baseline are removed from the image analysis path. The image frames, probe contact pressure information and chest wall displacement signals corresponding to the abnormal segments are removed synchronously by comparing the time anchor index. The continuous image frames within the apical stable window are temporally reassembled according to the time anchor sequence to generate a continuous chain of left ventricular contours.

9. A pediatric DCM intelligent auxiliary diagnostic system based on deep learning of cardiac ultrasound images according to claim 8, characterized in that, When the time range of the apical stability window is determined based on the stability of apical movement, the continuous time period with the lowest pressure fluctuation amplitude and the smallest displacement change is selected as the final range of the apical stability window by analyzing the synchronous change trend of the probe contact pressure information corresponding to the time anchor and the chest wall displacement signal.

10. A pediatric DCM intelligent auxiliary diagnostic system based on deep learning of cardiac ultrasound images according to claim 8, characterized in that, The steps for risk assessment using a continuous chain of left ventricular contours as input are as follows: The time nodes in the continuous chain of left ventricular contours are organized temporally and the left ventricular structure of each frame is arranged in chronological order by time anchor index. Secondly, a fusion correspondence is established between the continuous chain of left ventricular contours and the probe contact pressure information and chest wall displacement signal, so that the structural change information and the external mechanical signal are synchronously mapped in the time dimension. Based on the multi-source correspondence, a dynamic feature expression structure is formed with the left ventricular contour continuous chain as the core input, which suppresses the residual feature interference caused by passive displacement, so that the input data only reflects the real physiological motion. The risk assessment results are mapped to specific time points within the apical stability window using time anchors, generating time-related risk assessment curves and outputting clinically interpretable auxiliary diagnostic information.

Images