Artificial intelligence-based system and method for identifying neonatal umbilical recovery

Abstract

The present application relates to the technical field of data processing, and is a new-born umbilical recovery condition recognition system and method based on artificial intelligence, specifically comprising: acquiring umbilical image data of a new-born; pre-processing and quality control are performed on the umbilical image data to obtain effective umbilical image data; candidate anatomical anchor points are extracted based on the effective umbilical image data, and an umbilical local anatomical reference system is established according to the candidate anatomical anchor points; under the umbilical local anatomical reference system, the umbilical image data at the current acquisition time and the umbilical image data at the historical acquisition time are executed cross-time alignment; recovery condition features are extracted based on the aligned umbilical image data; recovery condition recognition is performed based on the recovery condition features to obtain a new-born umbilical recovery condition recognition result; the present application solves the problems of distortion in umbilical recovery feature extraction, large fluctuation in recognition result and unstable trend analysis in the prior art.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and is a newborn umbilical cord recovery recognition system and method based on artificial intelligence. Background Technology

[0002] Newborn umbilical cord care is crucial throughout the early postnatal recovery period. From the moment the umbilical cord stump is ligated, it typically undergoes several consecutive stages, including drying, shrinking, color change, local rupture, detachment, and continued tissue healing after detachment. Since much of the observation and care of newborn umbilical cords occurs in home settings, how to continuously and stably identify umbilical cord recovery in non-standardized environments has become a significant technical challenge in the intelligentization of newborn care.

[0003] In existing technologies, most methods for recognizing umbilical cord images of newborns employ single-frame image classification, umbilical region segmentation, or static feature judgment. These methods typically assume that umbilical cord images of the same newborn taken at different times can be directly compared, or establish a correspondence between different time points through methods such as center cropping, outer contour detection, and rigid registration. While these approaches are effective when the umbilical cord appearance is relatively stable and the acquisition conditions are standardized, they still have significant limitations in real-world nursing scenarios. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] It needs to be further explained that the technical problem to be solved by this invention specifically addresses the following issues: Firstly, in home photography environments, there are common factors such as lens angle deviation, changes in shooting distance, unstable focus, local shadows, uneven exposure, operator hand tremors, diaper edge obstruction, and slight undulations of the newborn's abdomen. These factors can cause changes in the position, size, texture, and shadow appearance of the same umbilical region in images at different time points, thus affecting the stability of the recognition results. Secondly, around the 5th to 14th day after birth, the umbilical cord stump usually enters a transitional stage before and after detachment. This stage is not simply a matter of overall shrinkage or color change, but is often accompanied by phenomena such as deflection, lifting, collapse, local scab rupture and curling, separation of one side of the root contact interface while the other side remains attached, and redistribution of periumbilical folds and abdominal wall textures due to changes in posture and tension. As a result, the visual center, outer contour centroid, and shadow structure of the same umbilical region may all drift in images at different time points, and the key structures that truly require continuous attention, namely the root contact area of ​​the stump and its adjacent stable tissues, are difficult to reliably lock using conventional visual center positioning methods.

[0006] Existing methods often input the entire umbilical region into the model uniformly, lacking modeling of the differences in local structural stability. Therefore, they are prone to misinterpreting changes in the posture of the free end of the stump as changes in the degree of recovery, or misjudging changes in periumbilical folds and projected shadows as local structural anomalies. This leads to abnormal jumps in continuous monitoring results between adjacent time points, reducing the reliability of the identification.

[0007] Furthermore, during the transitional phase before and after the umbilical cord stump falls off, the surface of the scab layer is prone to rupture, lamellar peeling, and edge rolling, resulting in enhanced local high-frequency texture energy. If the system cannot distinguish between this type of surface texture evolution and changes in the actual root connection structure, it is prone to generating spurious change responses, making single-time identification and trend analysis unstable.

[0008] Therefore, this invention proposes an artificial intelligence-based system and method for recognizing the recovery status of a newborn's umbilical cord.

[0009] The technical solution of the newborn umbilical cord recovery recognition method based on artificial intelligence of the present invention includes the following steps: Step S101: Obtain umbilical image data of the newborn, the umbilical image data including umbilical image data at the current acquisition time and umbilical image data at at least one historical acquisition time; Step S102: Preprocess and quality control the umbilical image data to obtain valid umbilical image data; Step S103: Extract candidate anatomical anchor points based on the effective umbilical image data, and establish a local anatomical reference system for the umbilical region based on the candidate anatomical anchor points; Step S104: Under the local anatomical reference frame of the umbilicus, perform cross-time alignment between the umbilicus image data at the current acquisition time and the umbilicus image data at the historical acquisition time; Step S105: Extract recovery features based on the aligned umbilical image data. The recovery features include at least root contact interface features, periumbilical fold reconstruction features, and residual surface texture evolution features. Step S106: Perform recovery status identification based on the recovery status features to obtain the neonatal umbilical cord recovery status identification result; Step S107: Based on at least one of the current identification results and historical identification results, output at least one of the following: recovery stability information, trend analysis results, and collected feedback information; The establishment of the umbilical local anatomical reference system includes determining the reference origin, the principal axis direction, and the secondary axis direction. The cross-temporal alignment includes performing differentiated alignment constraints based on the stability differences of different structural regions within the umbilical region.

[0010] Specifically, in step S101, acquiring newborn umbilical image data includes acquiring a short-time image sequence of the newborn umbilical region, wherein the acquisition duration of the short-time image sequence is [duration missing]. The number of image frames is Frames, capture frame rate During the acquisition process, the angle between the principal optical axis of the acquisition terminal lens and the local normal of the abdominal wall should not exceed [the specified value]. .

[0011] Specifically, in step S102, the preprocessing and quality control of the umbilical image data includes: calculating the sharpness index, exposure balance index, and average displacement index between adjacent frames of each frame image, and based on the sharpness index, exposure balance index, and average displacement index, filtering out blurry frames, overexposed frames, underexposed frames, and frames with large drift, and determining the target image frame or target image frame set for dissection anchoring based on the remaining image frames after filtering. The sharpness index is calculated based on the Laplacian response or gradient response of the image; the exposure balance index is calculated based on the mean gray level and standard deviation of the gray level in the target area; and the average displacement index is calculated based on the feature point tracking results between adjacent frames.

[0012] Specifically, step S102 further includes constructing a comparability prior scoring model to comprehensively evaluate the overall quality of the image data acquired in this study; The comparability prior scoring model is configured as follows: a state vector is constructed based on representative features of the available representative frames after filtering. And based on the mean vector obtained in advance from qualified samples under standard collection conditions. With covariance matrix Calculate Mahalanobis distance and generate comparability prior scores. ; when If the image quality falls below a preset threshold, the image acquisition is deemed substandard, a retake prompt is output, and cross-time-series trend analysis of the acquired image data is stopped.

[0013] Specifically, step S103, extracting candidate anatomical anchor points and establishing a local anatomical reference system for the umbilicus, includes: extracting one or more of the following from the umbilicus image: candidate point of the umbilical fossa center, convergence point of the periumbilical folds, endpoint of the contact arc segment of the stump root, candidate line of the local principal axis of symmetry of the abdominal wall, and abrupt change zone of local texture direction; scoring the candidate anatomical anchor points using an anatomical anchor point credibility model; and selecting a target anchor point combination to establish a local anatomical reference system for the umbilicus based on the scoring results. The umbilical local anatomical reference system takes the geometric center of the root contact arc segment or the umbilical fossa stability center as the reference origin, the main symmetry direction of the abdominal wall as the main axis direction, and the direction orthogonal to the main axis direction as the secondary axis direction. The configuration of the anatomical anchor point credibility model is as follows: for the first... A number of candidate anchor points are selected based on the eigenvalues ​​of the Harris matrix formed by the image gradients within their neighborhood windows. and Calculate the corner response of pure shape ; Based on the displacement variance of the candidate anchor points in the short-time image sequence Calculate the motion stability penalty term ; Calculate the confidence level of candidate anchor points ,in Parameters used to control the intensity of the penalty for motion stability.

[0014] Specifically, in step S104, cross-temporal alignment includes: dividing the umbilicus region into a root contact stability region, a periumbilical fold elastic change region, and a stump free end drift region, and applying a strong constraint to the root contact stability region, applying an elastic constraint to the periumbilical fold elastic change region, and not applying an additional geometric constraint to the stump free end drift region. Wherein, the positional consistency constraint strength applied to the root contact stability region is greater than the positional consistency constraint strength applied to the residual free end drift region; Among them, the root contact stability region With reference origin The elastic variation zone of the periumbilical folds is defined by taking the root contact arc segment as the center and the characteristic scale as the radius. For self The outer edge extends outward to a ring-shaped region containing the convergence point of the periumbilical folds, the free end drift region of the stump. Except for the umbilical region and The area where the remaining parts are located.

[0015] Specifically, in step S104, cross-time alignment further includes: A partitioned non-rigid alignment model is constructed, and the optimal displacement field is solved by functional energy optimization. The objective function for functional energy optimization is composed of a weighted sum of a data fidelity term, an elastic regularization term, and a rigid regularization term. The data fidelity term is constructed based on the local normalized cross-correlation of the entire map region. The elastic regularization term corresponds to the elastic variation region of the periumbilical fold and is constructed based on the strain tensor. The rigid regularization term corresponds to the root contact stability region and is constructed based on the displacement norm. After alignment is completed, cross-temporal comparability scoring parameters are calculated based on the alignment quality and anatomical pose deviation of the root contact stability zone. ; ; in The normalized cross-correlation coefficient is the result of alignment of the root contact stable region. As a reference to the origin offset, Main axis included angle deviation, and These are the acceptable positional and angular deviation characteristic scales, respectively; When the cross-time series comparability scoring parameter is lower than the preset threshold, the trend analysis results are stopped, and a retake prompt or only a single recovery identification result is output.

[0016] Specifically, in step S105, extracting the recovery status features includes: Extract the root contact interface features and construct structural parameters based on the root contact interface features; The root contact interface features include contact arc length. Average width of contact band Standard deviation of contact strip width and exposed zone area ; The structural parameters include the relative width parameter of the contact strip. Contact band non-uniformity parameters and relative area parameters of the exposed zone ; Extracting periumbilical fold reconstruction features, the periumbilical fold reconstruction features including the dispersion of the periumbilical fold direction. ; Extracting the surface texture evolution features of the remnant, the surface texture evolution features of the remnant include the proportion of high-frequency texture energy. ; Construct the restored structure state vector at the current moment .

[0017] Specifically, step S106 includes: Calculate the current restored structure state vector Minimum statistical deviation relative to a pre-built set of normal recovery states ; ;in This is a set of normal recovery states. The covariance matrix corresponding to the normally recovered sample; Based on the minimum statistical deviation distance and the cross-time series comparability scoring parameters Generate recovery stability score parameters , Output stability recovery information; The recovery status identification is performed by spatial feature extraction and temporal feature fusion. When performing temporal feature fusion, only the structural state vectors corresponding to historically acquired image data that meet the cross-temporal comparability requirements are included.

[0018] In addition, the newborn umbilical cord recovery recognition system based on artificial intelligence of the present invention includes the following modules: The module includes an image acquisition module, a preprocessing and quality control module, an anatomical anchoring module, a cross-temporal alignment module, a restoration feature recognition module, and a data analysis and feedback module. The image acquisition module is used to acquire umbilical image data of newborns; The preprocessing and quality control module is used to evaluate and filter the umbilical image data to obtain valid image data. The anatomical anchoring module is used to extract candidate anatomical anchor points of the umbilicus and establish a local anatomical reference system for the umbilicus. The cross-temporal alignment module is used to perform cross-temporal alignment between the umbilical image at the current acquisition time and the umbilical image at the historical acquisition time under the local anatomical reference system of the umbilicus, so as to establish a stable anatomical correspondence between different time points. The restoration feature recognition module is used to extract restoration features based on the aligned umbilicus image and perform restoration recognition. The restoration features include at least one or more of the following: root contact interface features, periumbilical fold reconstruction features, and stump surface texture evolution features. The data analysis and feedback module is used to output at least one of the following based on at least one of the current identification results and historical identification results: recovery stability information, trend analysis results, and collected feedback information.

[0019] A storage medium storing instructions that, when read by a computer, cause the computer to execute the artificial intelligence-based method for recognizing the recovery status of a newborn's umbilical cord.

[0020] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described artificial intelligence-based method for recognizing the recovery status of a newborn's umbilical cord.

[0021] Compared with the prior art, the technical effects of the present invention are as follows: This invention establishes a specialized structural analysis mechanism for the critical scenario in continuous monitoring where misjudgment is most likely to occur—the transition phase before and after umbilical cord stump detachment. This mechanism replaces the existing end-to-end judgment method based on full images, improving the relevance and interpretability of umbilical cord monitoring in real-world home environments.

[0022] This invention improves the structural comparability between umbilical images at different time points by replacing the visual center positioning method with a local anatomical reference frame, and reduces the cropping center drift problem caused by the deflection of the free end of the stump.

[0023] This invention uses a comparability prior scoring model based on Mahalanobis distance to pre-judge the acquisition quality, thus ensuring the comparability of the historical image database from the source.

[0024] This invention introduces a non-rigid alignment mechanism for the umbilicus region and employs similarity-driven functional energy optimization to ensure strong constraints in the high-stability region of the root, elastic manifold deformation in the folded region, and cut off interference from the drift region of the free end, so that the measurement of key tissues has a unified anatomical correspondence.

[0025] This invention is fully compatible with networked applications for both home and healthcare settings. It establishes a time-domain tracking and physical measurement curve tracking mechanism in cloud collaboration, and achieves attribution analysis of trend fluctuations through quality records at each stage, thus creating a barrier for smart home care. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1This is a flowchart illustrating the artificial intelligence-based method for recognizing the recovery status of a newborn's umbilical cord. Figure 2 This is a schematic diagram illustrating the variation of the comparability prior scoring parameters of the present invention with the sample number; Figure 3 This is a schematic diagram illustrating the establishment of the local anatomical reference system for the umbilicus and the selection of candidate anatomical anchor points according to the present invention. Figure 4 This is a schematic diagram of the partitioned non-rigid alignment model of the present invention; Figure 5 This is a schematic diagram illustrating the change of the cross-time comparability scoring parameters of the present invention over time; Figure 6 This is a schematic diagram comparing the comparability of cross-temporal structures before and after the non-rigid alignment of the partitions in this invention; Figure 7 This is a schematic diagram illustrating the trend of the recovery stability scoring parameters over time in this invention; Figure 8 This is a schematic diagram of the structure of the artificial intelligence-based neonatal umbilical cord recovery recognition system of the present invention. Detailed Implementation

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0029] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0030] Example 1: like Figure 1 As shown in the embodiment of the present invention, the method for recognizing the recovery status of a newborn's umbilical cord based on artificial intelligence is as follows: Figure 1 As shown, the specific steps include the following: Step S101: Obtain umbilical image data of the newborn, the umbilical image data including umbilical image data at the current acquisition time and umbilical image data at at least one historical acquisition time; In step S101, acquiring newborn umbilical image data includes acquiring a short-time image sequence of the newborn umbilical region, wherein the acquisition duration of the short-time image sequence is [duration missing]. The number of image frames is Frames, capture frame rate During the acquisition process, the angle between the principal optical axis of the acquisition terminal lens and the local normal of the abdominal wall should not exceed [the specified value]. .

[0031] It should be noted that during the acquisition process, the angle between the principal optical axis of the acquisition terminal lens and the local normal of the abdominal wall should not exceed [a certain value]. The aim is to reduce the impact of changes in shooting posture and close-range perspective distortion on the structural characterization of the root contact area.

[0032] Preferably, in step S101 of another embodiment, the newborn's umbilical cord image data is acquired via a smartphone, tablet terminal, nursing acquisition terminal, or medical acquisition device.

[0033] The preferred method is short-time image sequence acquisition, which involves continuous image capture in a single acquisition. Image sequences, preferably acquired Frame image, frame rate preferably It should be noted that this setting can preserve the timing information of the newborn's slight abdominal undulations, the operator's slight movements of holding the device, and the instantaneous swing of the free end of the stump; More specifically, during actual data acquisition, the system preferably provides a framing guide frame on the terminal interface to prompt the operator to position the navel in the central area and to ensure that the angle between the principal optical axis of the lens and the local normal of the abdominal wall does not exceed [the specified value]. When the terminal detects that the image is tilted too much or the shooting distance is too close, the system can prompt the user to adjust the angle or distance.

[0034] In the implementation that introduces a flexible calibration patch, a pixel length scaling factor can also be established in the image to enable absolute scaling. In the implementation that does not introduce an external calibration component, the system adopts a relative scaling method under the same terminal conditions to ensure the consistency of cross-time-series comparisons of the same object.

[0035] For example, the relative scale quantization method specifically involves: after establishing the anatomical reference frame for the current acquisition, selecting the pixel distance between the center of the umbilicus depression and any endpoint of the arc segment contacting the root in the image. As a reference scale; it is assumed that the absolute change in the size of the umbilical fossa structure of the same newborn during continuous monitoring is less than the pixel scale change caused by shooting distance and zoom. Therefore, for any two consecutive measurements of the newborn, its dimensional parameters (such as...) Multiply by a coefficient uniformly After being normalized to the same scale for comparison, and after relative scale quantization, all ratio parameters based on this scale (e.g.) It is no longer affected by the absolute scale.

[0036] Step S102: Preprocess and quality control the umbilical image data to obtain valid umbilical image data; In step S102, the preprocessing and quality control of the umbilical image data includes: calculating the sharpness index, exposure balance index, and average displacement index between adjacent frames of each frame; and based on the sharpness index, exposure balance index, and average displacement index, filtering out blurred frames, overexposed frames, underexposed frames, and frames with large drift; and based on the remaining image frames after filtering, determining the target image frames or target image frame set for anatomical anchoring. The sharpness index is calculated based on the Laplacian response or gradient response of the image; the exposure balance index is calculated based on the mean gray level and standard deviation of the gray level in the target area; and the average displacement index is calculated based on the feature point tracking results between adjacent frames.

[0037] In step S102 of another embodiment, the image quality of the image sequence acquired in step S101 is evaluated frame by frame to filter out image frames that cannot be used for subsequent recognition and to determine the overall usability of this acquisition, as follows: For the The Laplacian sharpness index for the frame image is: ; in, For the first Frame image in pixels grayscale value at that location The area to be evaluated. This represents the number of pixels in the region.

[0038] Meanwhile, the exposure uniformity index is calculated as the coefficient of variation. : ; in, The grayscale mean of the target area. The standard deviation of grayscale values ​​for the target region.

[0039] It should be noted that, Used to evaluate whether the local contrast of the target region is suitable for boundary analysis and texture extraction; For adjacent frame images and The system can obtain [data / information] through corner detection and sparse optical flow tracking. Displacement vectors of feature points And calculate the average displacement : ; In this embodiment, the system performs frame-by-frame removal based on a preset engineering threshold, as follows: when a frame of image... ,or Exceeding the allowed range ,or When this happens, the frame is marked as an unusable frame.

[0040] After elimination, the preferred ones are retained. Frame candidate frames are used for dissection anchoring and representative frame selection.

[0041] It should also be noted that after frame-level screening, the system selects one frame from the retained candidate frames as the target reference frame for dissection anchoring and restoration identification. The selection criteria are the highest sharpness index or the frame with the smallest inter-frame displacement among multiple high-resolution frames as the representative frame. Step S102 further includes constructing a comparability prior scoring model to comprehensively evaluate the overall quality of the image data collected this time. The comparability prior scoring model is configured as follows: a state vector is constructed based on representative features of the available representative frames after filtering. And based on the mean vector obtained in advance from qualified samples under standard collection conditions. With covariance matrix Calculate Mahalanobis distance and generate comparability prior scores. ; when If the image quality falls below a preset threshold, the image acquisition is deemed substandard, a retake prompt is output, and cross-time-series trend analysis of the acquired image data is stopped.

[0042] In another embodiment, based on frame-level filtering, the system performs an overall quality assessment of the current acquisition. Here, a comparability prior scoring model is constructed to measure the degree of deviation between the current acquisition and the ideal acquisition conditions, as detailed below: Representative features from a single data collection are grouped into a vector. Each component is taken as the mean or optimal value of the representative frames that are available after filtering.

[0043] Define the mean vector obtained by statistical analysis of a large number of qualified samples under standard collection conditions. Covariance Matrix .

[0044] Exemplarily, in another embodiment, the mean vector Covariance Matrix The images were obtained as follows: Before system deployment, at least 300 qualified umbilical cord images of different newborns, from different acquisition terminals, and under different lighting conditions were collected. Samples labeled as having acceptable image quality by medical professionals were then selected, and the images from each sample were extracted. , , A set of feature vectors is formed, and the mean vector and covariance matrix of this set are calculated as initial parameters; during system use, a sliding window approach can be used to continuously update the parameters. and It should also be noted that when the sample size is less than 30, it will revert to the preset default diagonal covariance matrix.

[0045] Constructing a score using Mahalanobis distance: ; It should be noted that when the deviation in any dimension is too large, the Mahalanobis distance increases sharply. It drops precipitously to near 0, achieving an effect of overall rejection due to excessively poor quality. When When the image quality falls below a preset threshold, the system directly determines that the current acquisition quality is unqualified, outputs a retake prompt, and stops performing cross-time series trend analysis on the acquired image data.

[0046] In another embodiment, the The values ​​are also stored in the cloud as the basic quality record for this collection, which is used to analyze the attribution of trend fluctuations.

[0047] like Figure 2 As shown, after the data collection is completed, this invention first uses comparability prior scoring parameters. A preliminary assessment of the basic acquisition quality of the image data was conducted. In the figure, the horizontal axis represents the sample number, and the vertical axis represents the corresponding... The score, indicated by the dashed line, represents the system's preset acceptable threshold. When a sample score is higher than the threshold, it means that the acquisition meets the basic requirements for subsequent structural analysis in terms of clarity, exposure balance, and inter-frame displacement stability, and can proceed to the anatomical anchoring and cross-temporal alignment process. When the score is lower than the threshold, the system outputs a retake prompt, preventing the acquisition from directly entering the historical trend analysis link. This is to avoid low-quality images contaminating the historical sample database. This not only helps reduce false fluctuations caused by acquisition problems but also gives the system a stronger self-correction capability in parent-user scenarios. That is, when the image itself is unreliable, it prioritizes prompting a retake rather than directly outputting prompts that may mislead care decisions.

[0048] Step S103: Extract candidate anatomical anchor points based on the effective umbilical image data, and establish a local anatomical reference system for the umbilical region based on the candidate anatomical anchor points; Step S103, extracting candidate anatomical anchor points and establishing a local anatomical reference system for the umbilicus, includes: extracting one or more of the following from the umbilicus image: candidate point of the umbilical fossa center, convergence point of the periumbilical folds, endpoint of the contact arc segment of the stump root, candidate line of the local principal axis of symmetry of the abdominal wall, and abrupt change zone of local texture direction; scoring the candidate anatomical anchor points using an anatomical anchor point credibility model; and selecting a target anchor point combination based on the scoring results to establish a local anatomical reference system for the umbilicus. The umbilical local anatomical reference system takes the geometric center of the root contact arc segment or the umbilical fossa stability center as the reference origin, the main symmetry direction of the abdominal wall as the main axis direction, and the direction orthogonal to the main axis direction as the secondary axis direction. It should be noted that the target anchor point combination includes at least an anchor point for determining the reference origin and an anchor point for determining the direction of the main axis.

[0049] like Figure 3 As shown, this invention does not directly use the visual geometric center as the basis for clipping and alignment in the umbilical region. Instead, it first extracts candidate anatomical anchor points such as the candidate point of the umbilical depression center, the convergence point of the periumbilical folds, the endpoint of the contact arc segment of the stump root, the candidate line of the local principal axis of symmetry of the abdominal wall, and the local texture direction change zone. Then, it uses the anatomical anchor point credibility model to select anchor point combinations with higher stability to establish a local anatomical reference system.

[0050] Among them, the geometric center of the root contact arc segment or the stable center of the umbilicus is used as the reference origin, the local principal symmetry direction of the abdominal wall is used as the principal axis direction, and the direction perpendicular to the principal axis is used as the secondary axis direction, thus forming a unified coordinate reference suitable for cross-time series comparison.

[0051] The lower part of the figure further compares the differences between the traditional visual center localization method and the anatomical reference system method of the present invention. The traditional method is easily affected by the deflection of the free end of the stump, local rotation and viewpoint disturbance, which causes the cropping area and reference direction to drift, resulting in a lack of stable anatomical correspondence between images at different time points. However, after the present invention establishes a local reference system by establishing stable anatomical anchor points, even if the posture of the free end of the stump changes, the key area of ​​the root can still maintain the consistency of position and orientation under a unified coordinate system.

[0052] The configuration of the anatomical anchor point credibility model is as follows: for the first... A number of candidate anchor points are selected based on the eigenvalues ​​of the Harris matrix formed by the image gradients within their neighborhood windows. and Calculate the corner response of pure shape ; Based on the displacement variance of the candidate anchor points in the short-time image sequence Calculate the motion stability penalty term ; Calculate the confidence level of candidate anchor points ,in Parameters used to control the intensity of the penalty for motion stability.

[0053] In one implementation, Take as the equivalent pixel diameter of the umbilicus region For example, if the equivalent diameter of the umbilical region is approximately 200 pixels, then The resolution can be set to 6 to 16 pixels, which allows anchor points with displacement variance below this level to obtain a high stability score, while drift points with displacement significantly exceeding this range are effectively eliminated, thereby selecting reliable anchor points that maintain consistent anatomical position in consecutive frames.

[0054] In step S103 of another embodiment, the system performs dissection anchoring processing on the retained representative frames.

[0055] This invention does not use the center of brightness or darkness or the center of the outer contour of the umbilicus image as the sole cropping center, but instead focuses on establishing a local reference system around the root contact area and its surrounding stable tissues.

[0056] In its specific implementation, the system extracts at least one of the following candidate anatomical anchor points or anchoring structures in the local area of ​​the umbilicus: candidate point at the center of the umbilical depression, convergence point of the periumbilical folds, endpoint of the contact arc segment at the root of the stump, candidate line of the local principal axis of symmetry of the abdominal wall, abrupt change zone of local texture direction, and minimum curvature change zone.

[0057] To screen out anchor points with stable anatomical significance, an anatomical anchor point credibility model was constructed. The evaluation metrics used in this model are designed to avoid interference from differences in brightness and contrast.

[0058] For the For each candidate anchor point, calculate the Harris matrix formed by the image gradients within its neighborhood window to obtain two eigenvalues. and The pure shape corner response is calculated as follows: ; It should be noted that this response value falls within The value within the image reflects only the angularity of the local structure and is unrelated to the absolute brightness and contrast of the image.

[0059] Calculate the displacement variance using the optical flow trajectory of the anchor point in a short-time image sequence. And construct a motion stability penalty term, and then calculate the confidence of the candidate anchor points as follows: ; in Parameters used to control the intensity of the penalty for motion stability.

[0060] The system selects from all candidate anchor points. Anchor point combinations that exceed the preset confidence threshold and satisfy the minimum spacing constraint are used as target anchor points to establish a local anatomical reference system, and the average confidence of the selected anchor point group is calculated. Stored in the cloud, it serves as an indicator of the quality of the evaluation system.

[0061] In one implementation, the confidence preset threshold is set to The minimum spacing constraint requires that the Euclidean distance between any two selected anchor points is not less than the equivalent diameter of the umbilicus region. This is to ensure that the anchor points are not concentrated in a single local area, thus providing sufficient geometric constraints for the establishment of the reference frame.

[0062] In this embodiment, the system uses the geometric center of the root contact arc segment or the umbilicus stability center as the reference origin. Taking the principal axis of local symmetry of the abdominal wall as the principal axis Take the direction orthogonal to it as the secondary axis. .

[0063] Furthermore, within this local reference frame, the current image undergoes rotation correction, translation normalization, and local resampling.

[0064] Step S104: Under the local anatomical reference system of the umbilicus, perform cross-time alignment between the umbilicus image data at the current acquisition time and the umbilicus image data at the historical acquisition time to establish a stable anatomical correspondence between umbilicus images at different time points. It should be noted that the cross-temporal alignment in step S104 aims to reduce the interference of the attitude change of the free end of the stub on the identification of the key root structure.

[0065] In step S104, cross-temporal alignment includes: dividing the umbilicus region into a root contact stability region, a periumbilical fold elastic change region, and a stump free end drift region; applying a strong constraint to the root contact stability region, applying an elastic constraint to the periumbilical fold elastic change region, and not applying an additional geometric constraint to the stump free end drift region. Wherein, the positional consistency constraint strength applied to the root contact stability region is greater than the positional consistency constraint strength applied to the residual free end drift region; Among them, the root contact stability region With reference origin The elastic variation zone of the periumbilical folds is defined by taking the root contact arc segment as the center and the characteristic scale as the radius. For self The outer edge extends outward to a ring-shaped region containing the convergence point of the periumbilical folds, the free end drift region of the stump. Except for the umbilical region and The area where the remaining parts are located.

[0066] In step S104, cross-time alignment further includes: A partitioned non-rigid alignment model is constructed, and the optimal displacement field is solved by functional energy optimization. The objective function for functional energy optimization is composed of a weighted sum of a data fidelity term, an elastic regularization term, and a rigid regularization term. The data fidelity term is constructed based on the local normalized cross-correlation of the entire map region. The elastic regularization term corresponds to the elastic variation region of the periumbilical fold and is constructed based on the strain tensor. The rigid regularization term corresponds to the root contact stability region and is constructed based on the displacement norm. like Figure 4 As shown, when performing cross-temporal registration between the current image and historical images, this invention divides the umbilicus region into root contact stability regions based on structural stability and deformation characteristics. Periumbilical folds and areas of elastic change and the drift region of the free end Among them, the root contact stability zone is the most critical and geometrically stable area in the restoration analysis. Therefore, strong constraints and low displacement restrictions are applied during the alignment process to maintain the accurate correspondence of this area as much as possible. The elastic variation zone of the periumbilical folds allows for continuous and smooth elastic manifold deformation to adapt to the natural reconstruction of local abdominal wall tissues at different time points. The drift zone of the free end of the stump is subject to only weak constraints to reduce the interference of its posture swing, tilting or deflection on the overall alignment result.

[0067] It should be noted that, by Figure 4 It can also be seen that the displacement field solved by this invention Instead of applying a completely uniform rigid transformation across the entire region, this model assigns differentiated geometric constraint strengths to different partitions. This ensures accurate correspondence of the root structure while allowing reasonable local deformation in non-critical areas. Compared to traditional global rigid alignment or indiscriminate registration methods, this partitioned non-rigid alignment model better reflects the actual tissue changes during the umbilical cord stump detachment transition period and effectively reduces mismatch problems caused by free end drift.

[0068] After alignment is completed, cross-temporal comparability scoring parameters are calculated based on the alignment quality and anatomical pose deviation of the root contact stability zone. ; ; in The normalized cross-correlation coefficient is the result of alignment of the root contact stable region. As a reference to the origin offset, Main axis included angle deviation, and These are the acceptable positional and angular deviation characteristic scales, respectively; In one implementation, the acceptable positional deviation characteristic scale Taken as the pixel length of the root contact arc segment For example, if the pixel length of the arc segment is times, pixels, then for Pixels. It should be noted that the residual offset of the reference origin after cross-temporal alignment mainly originates from the slight difference in viewing angle between the two acquisitions, the subtle undulations of the neonatal abdominal wall, and the sub-pixel residuals of the alignment algorithm. The magnitude of this type of measurement noise is typically on the scale of the arc segment feature. Within this ratio; if the offset exceeds this ratio, it is highly likely due to changes in the anatomical structure of the root contact interface itself or alignment failure, and should be judged as incomparable. The selection of this ratio achieves adaptive noise tolerance independent of shooting distance and anatomical scale.

[0069] In one implementation, the acceptable angular deviation characteristic scale Take as (about It should be noted that the maximum angle between the lens principal optical axis and the normal of the abdominal wall is limited by the acquisition specifications. The maximum rotation angle of the visual axis caused by the orthogonal directions of the two acquisitions is as follows: Even considering the sub-pixel positioning error of the superimposed anchor point and the influence of the newborn's micro-movements, the total angular deviation is within the normal range and does not exceed [a certain value]. ;and The following deviations are mainly caused by anchor point positioning accuracy and normal body movement, and are considered acceptable measurement noise. Therefore, this range can effectively distinguish between normal acquisition angle fluctuations and abnormal rotations caused by serious acquisition violations or incorrect reference frame construction.

[0070] When the cross-time series comparability scoring parameter is lower than the preset threshold, the trend analysis results are stopped, and a retake prompt or only a single recovery identification result is output.

[0071] like Figure 5 As shown, during continuous monitoring, this invention further utilizes cross-time series comparability scoring parameters. The structural consistency between the current image and historical valid images is quantitatively evaluated. As shown in the figure, after adopting the partitioning non-rigid alignment method of this invention, the corresponding time points are... The values ​​remained at a high level overall, with small fluctuations in the curve, indicating that the established local anatomical correspondences were relatively stable and could continuously support subsequent trend analysis. In contrast, traditional methods, lacking stable anchoring and zonal deformation control, ... The values ​​are generally low and fluctuate significantly, even approaching or falling below the comparability threshold at certain points in time, reflecting a lack of fundamental structural comparison between images.

[0072] The results shown in the figure demonstrate that the present invention can not only reduce errors in single registration but also maintain a high level of consistency in continuous multi-time-point monitoring. This means that when constructing historical recovery trajectories, the system can more reliably determine whether a structural change originates from actual recovery progress or from non-pathological factors such as posture changes, acquisition angle disturbances, or free end swinging of the stump, thereby effectively improving the credibility and interpretability of trend analysis.

[0073] For example, in another preferred embodiment, in step S104, the system performs cross-temporal alignment between the currently acquired anatomically standardized image and historically valid acquired images; In particular, to ensure the continuity of measurements throughout the recovery timeline, the alignment process preferably uses the most recent valid acquisition image as the reference for the current image. This means performing frame-by-frame serial alignment; or, within a time period where the morphology of the residual ends is relatively stable, using the first image acquired within that time period as a common baseline reference. .

[0074] First, divide the navel area into root contact stability zones. Periumbilical folds and areas of elastic change and the drift region of the free end .

[0075] For example, in one implementation, the division of the three regions is based on the local anatomical reference system established in step S103 and the extracted candidate anatomical anchor points: The root contact stability zone Defined with reference origin Centered on the origin, extending from any endpoint of the root contact arc segment to the reference origin. A circular neighborhood with a radius of a preset multiple (e.g., 1.2 to 1.5 times) of the distance, which mainly covers the connection interface between the umbilical cord root and the abdominal wall and its adjacent stable tissues; The elastic variation area of ​​the periumbilical folds Defined as an annular region extending outward from the outer edge of the root contact stability zone to include the convergence point of the periumbilical folds, the outer boundary of which is determined by the outer edge of the umbilicus or by the outer envelope of the line connecting the convergence points of the periumbilical folds; The free end drift region of the residual end Defined as the area of ​​the umbilical region outside the elastic change zone of the periumbilical folds, mainly covering the free end of the umbilical cord stump, the rolled-up part of the scab layer, and structures that may not be directly connected to the abdominal wall; Construct a partitioned non-rigid alignment model and solve for the optimal displacement field using functional energy optimization. .

[0076] Let the historical reference image be The current image is Displacement field The optimization objective is: ; It should be understood that the aforementioned energy functional is not a direct simulation of the physical system, but rather a variational image registration target functional. Its construction logic fully corresponds to the partitioning differentiation alignment concept of this technical solution. Specifically, this energy functional consists of a weighted sum of three cost functions with explicit optimization objectives. The physical meaning and design basis of each term are as follows: For the first item (Data fidelity item), based on the entire map area Average local normalization cross-correlation quantization of the current image After displacement field transformation and compared with historical reference image The degree of local similarity between them, where the NCC value range is Taking a negative sign results in lower energy for more similar regions during optimization, thus driving the displacement field to evolve in the direction of maximizing image content matching. This term is divided by the total image area. This makes it an average similarity metric independent of image size, ensuring that its contribution to optimization does not shift with changes in image resolution.

[0077] Among them, the local normalized cross-correlation For example, the standard similarity metric operator in the field of image registration, if for two images and In pixel coordinates Centered preset local window (For example or Within a pixel, it is defined as: ; in and Images and In the window The average gray level within the range. The range of this operator is... , This indicates that the gray-level distributions of the two images are completely positively correlated within this local window, meaning they have a perfect structural match. Indicates that it is irrelevant; This indicates that the grayscale distributions are completely negatively correlated.

[0078] It should also be noted that NCC eliminates the grayscale difference between two images caused by changes in illumination or overall brightness shift by subtracting the local mean and dividing by the local standard deviation. Therefore, it has an important robust advantage in umbilical image registration under non-standardized illumination conditions. For the second item (Elastic regularization term), corresponding to the elastic variation zone of the periumbilical folds. Its physical prototype is the average strain energy density in linear elasticity, where the strain tensor The squared Frobenius norm characterizes the elastic deformation energy. This term as a whole measures the average degree of deformation in the folded region. The periumbilical tissue in this region allows for a certain degree of non-rigid deformation due to changes in posture and abdominal wall tension. Therefore, only elastic constraints are applied to limit excessive stretching or twisting, without forcing it to remain stationary. This is achieved by adjusting the weights. The tolerance for deformation in this area can be controlled.

[0079] Regarding the third item (Rigid regularization term), corresponding to the root contact stability region. Its physical prototype is the quadratic penalty potential of constrained rigid body displacement, with the square of the displacement norm. Divide by characteristic length The square of the absolute displacement is converted into a relative displacement independent of the anatomical scale, and then normalized by the area of ​​the region to obtain the average displacement energy of that region. The umbilical cord root contact area should have high anatomical stability during recovery observation, and perceptible pixel displacement due to changes in the acquisition angle is not allowed. Therefore, a strong constraint is imposed on this region, directly penalizing the displacement itself (here, Typically, a larger value is chosen to reflect a strong rigid constraint, which aims to force the region to remain anchored in place during cross-temporal alignment.

[0080] For example, the feature length Used to convert absolute pixel displacement into relative displacement that is independent of anatomical scale; in one implementation, It is taken as the equivalent diameter of the umbilicus region in the current image or the arc length of the root contact arc segment; furthermore, in another preferred embodiment, Alternatively, a fixed empirical value can be used, such as the average pixel diameter of the umbilical region at a standard shooting distance (e.g., 200 pixels), to ensure that the constraint strength between different images is comparable; In summary, it is necessary to further explain that the first term, the data fidelity term, measures the structural similarity between the current image after displacement field transformation and the historical reference image. This term utilizes the characteristic that local normalized cross-correlation is insensitive to changes in illumination, ensuring that images taken under different exposure conditions in a home environment can still establish a reliable similarity benchmark. Taking a negative sign, the more similar the two images are, the lower the energy of this term, thereby driving the displacement field to evolve in the direction of image content matching. This term acts on the entire image area, providing the basic image matching dynamics for the entire alignment process. The second term, the elasticity regularization term, corresponds to the elastic variation area of ​​the periumbilical folds. The abdominal wall skin of newborns is soft, and under different postures, breathing states, and abdominal tension, the periumbilical skin folds will undergo natural elastic deformation. This deformation is a normal phenomenon rather than a pathological change. The role of this term is to allow smooth and continuous deformation in this area to absorb the above-mentioned physiological differences, while penalizing excessive stretching, twisting, or tearing, preventing the displacement field from introducing unnatural distortions in pursuit of image similarity. The third term is a rigid regularization term, corresponding to the root contact stability zone. The connection area between the umbilical cord root and the abdominal wall should have the highest anatomical stability during cross-temporal observations. Regardless of the acquisition angle or the swinging of the free end of the stump, the position of this area in the image should not undergo any perceptible shift. This term directly penalizes any pixel displacement deviating from its original position within this area. Its mathematical form is similar to the spring potential energy constraining the motion of a rigid body in physics; that is, the smaller the displacement, the lower the energy, and the larger the displacement, the more the energy increases sharply. This forcibly anchors the area to the anatomical reference position, ensuring consistent correspondence of key tissue structures in cross-temporal comparisons.

[0081] In one implementation, the minimum value of the energy functional is solved using a multi-resolution Gaussian-Newton iterative method. During the solution process, the image is downsampled using a three-level Gaussian pyramid, and optimization is performed layer by layer from coarse to fine. Each layer is iterated until the displacement field update amount is lower than the preset convergence threshold or the maximum number of iterations is reached.

[0082] The optimal displacement field is obtained by solving the energy functional optimization problem. Then, the displacement field is applied to the currently acquired image. Registered images are generated through spatial transformation and resampling. ,Right now .

[0083] It should be noted that the registered image is in the root contact stable region. Internal and historical reference images The corresponding anatomical structures have a one-to-one pixel-level mapping relationship, thus providing an anatomically comparable basis for subsequent structural feature extraction and reconstruction identification.

[0084] like Figure 6As shown, to verify the improvement effect of the partitioned non-rigid alignment strategy of the present invention on the comparability of cross-temporal structures, in another preferred embodiment, the average registration errors of the three methods—unaligned, global rigid alignment, and partitioned non-rigid alignment of the present invention—in the root stable region were statistically compared. The box plot in the figure shows that the overall error is large in the unaligned case, and the distribution range is wide, indicating a significant structural positional shift between images at different time points. After adopting global rigid alignment, the error decreased somewhat, but because it cannot adapt to the reconstruction of periumbilical folds and changes in free end pose, the median and mean of the error remained at a high level. However, after adopting the partitioned non-rigid alignment of the present invention, the error distribution converged significantly, and the median, mean, and upper and lower quartile ranges were all significantly reduced, indicating that the anatomical correspondence of the key root structure in cross-temporal images was significantly enhanced.

[0085] This figure further illustrates that the present invention does not simply pursue surface overlap in overall shape, but focuses on improving the registration accuracy of the root stable region, which is most relevant to the determination of the recovery state. Since the extraction of the subsequent five structural parameters all depend on the stability measures of the root contact interface, exposed area, and periumbilical tissue relationship, the result of this error reduction directly corresponds to the improvement of the consistency of subsequent feature extraction, which is in line with the design concept of strong constraints in the stable region and weak constraints in the drift region of the present invention.

[0086] After alignment, the system further calculates cross-time series comparability score parameters.

[0087] The score is determined jointly by the structural fidelity of the root contact stabilization zone and the anatomical pose deviation. The normalized cross-correlation coefficient after stabilization zone alignment is calculated. and the offset of the reference origin deviation from the spindle angle .definition: ; in and These are the acceptable positional and angular deviation characteristic scales, respectively.

[0088] when When the value is below the preset threshold, the system determines that the cross-time series comparability is not valid, stops outputting trend analysis results, and outputs a retake prompt and / or only outputs the single recovery status identification result.

[0089] It should be noted that, For the aforementioned localized cross-correlation in the root contact stability region The spatial average value on the surface, i.e.: ; It should be noted that it comprehensively measures the overall structural consistency between the aligned current image and the historical reference image within the root contact stability region; because this region has high anatomical stability, the aligned... Expected value close to (usually in) (within the range) If this value is significantly low, it indicates that even after non-rigid registration, the structures of the two images in this critical region cannot be reliably matched, which may be due to poor acquisition quality, substantial changes in root anatomy, or incorrect reference frame construction.

[0090] Step S105: Extract restoration features based on the aligned umbilical image data. These restoration features include at least root contact interface features, periumbilical fold reconstruction features, and stump surface texture evolution features, including: Extract the root contact interface features and construct structural parameters based on the root contact interface features; The root contact interface features include contact arc length. Average width of contact band Standard deviation of contact strip width and exposed zone area ; The structural parameters include the relative width parameter of the contact strip. Contact band non-uniformity parameters and relative area parameters of the exposed zone ; Extracting periumbilical fold reconstruction features, the periumbilical fold reconstruction features including the dispersion of the periumbilical fold direction. ; Extracting the surface texture evolution features of the remnant, the surface texture evolution features of the remnant include the proportion of high-frequency texture energy. ; Construct the restored structure state vector at the current moment .

[0091] For example, in another embodiment, in step S105, the system extracts structured features from the current image based on the local anatomical reference frame and the cross-temporal alignment results. The restored features include root contact interface features, periumbilical fold reconstruction features, and stump surface texture evolution features.

[0092] The root contact interface features include at least the following: contact arc length. Average width of contact band Standard deviation of contact strip width and the area of ​​the exposed zone .

[0093] The root contact interface features are converted into relative morphological parameters independent of shooting distance and device focal length. Specifically, the relative width parameter of the contact zone, the non-uniformity parameter of the contact zone, and the relative area parameter of the exposed zone are calculated as follows: ; in, Used to characterize the relative thickness of the contact band at the root of the stub over the overall contact arc length. Used to characterize the degree of non-uniformity of the contact band width along the arc direction. Used to characterize the area percentage of the exposed root region relative to the contact arc length scale.

[0094] These three correspond to the overall contraction state, local stability state, and exposure and detachment state of the root contact structure, respectively, and they do not substitute for each other.

[0095] The system further extracts the reconstruction parameters of the periumbilical folds and the evolution parameters of the surface texture of the remnant. It should be noted that the dispersion of the periumbilical fold direction... This is used to quantify the degree of disorder in the direction of periumbilical skin folds. During normal recovery, periumbilical folds are usually distributed in an orderly manner, approximately radially or concentrically, with the umbilicus as the center. The principal direction angle of each texture unit is... The texture units are relatively concentrated with low directional dispersion. However, when uneven local tissue contraction, scab formation, or inflammatory reactions occur before or after the stump falls off, the direction of the periumbilical skin folds may become locally distorted, abruptly change direction, or redistribute regionally, leading to a dispersion of the main directions of each texture unit and an increase in directional dispersion. Therefore, The increase in size can serve as an auxiliary indicator of changes in mechanical state or local abnormal reactions during the remodeling of periumbilical tissues. Calculate the dispersion of the periumbilical fold direction for: ; in The principal orientation angle of each texture unit. For the average direction, This represents the number of detectable texture units within the periumbilical region.

[0096] High-frequency texture energy ratio This method is used to quantify the proportion of high-frequency texture components on the surface of the umbilical cord stump caused by scab rupture, lamellar peeling, and edge curling. During the drying and healing process, the surface of the umbilical cord stump undergoes an evolution from a relatively smooth and moist state to a rough and dry state. Especially in the transition stage before and after detachment, the scab layer is prone to cracking, lifting, and fragmentation. These phenomena manifest as an enhancement of high-frequency energy in the image frequency domain. After Fourier transforming the image, high-frequency subbands representing fine textures and edge structures are delineated. The proportion of energy within this sub-band to the total energy of the entire image is the high-frequency texture energy proportion. An increase in this parameter indicates that the surface layer of the remnant is in an active structural evolution period, while an abnormal surge or fluctuation may indicate unstable crusting or local abnormalities.

[0097] Calculate the proportion of high-frequency texture energy for: ;in These are the two-dimensional Fourier transform coefficients of the umbilical region image. For the full frequency domain, This parameter represents the preset high-frequency sub-band range and is the power spectral density ratio.

[0098] Therefore, the system constructs the restored structure state vector at the current moment: ; in, The relative thickness of the contact band at the root stump over the overall contact arc length reflects the contraction state of the root connection. It characterizes the degree of non-uniformity in the width of the contact zone, reflecting whether the root detachment is locally consistent. The area percentage of the exposed root region reflects the spatial extent of the detachment process; It characterizes the degree of disorder in the direction of periumbilical skin folds, reflecting the changes in mechanical state accompanying tissue remodeling; The proportion of high-frequency texture components on the surface of the stump reflects the activity of crust evolution. The five parameters cover five dimensions: root connectivity, root uniformity, exposure degree, surrounding tissue reconstruction, and surface texture evolution. Together, they constitute a complete description of the current umbilical recovery state. This structural state vector is used to fully describe the current umbilical recovery state.

[0099] Step S106: Perform recovery status identification based on the recovery status features to obtain the neonatal umbilical cord recovery status identification result; Calculate the current restored structure state vector Minimum statistical deviation relative to a pre-built set of normal recovery states ; ;in This is a set of normal recovery states. The covariance matrix corresponding to the normally recovered sample; Based on the minimum statistical deviation distance and the cross-time series comparability scoring parameters Generate recovery stability score parameters , Output stability recovery information; The recovery status identification is performed by spatial feature extraction and temporal feature fusion. When performing temporal feature fusion, only the structural state vectors corresponding to historically acquired image data that meet the cross-temporal comparability requirements are included.

[0100] For example, in step S106 of another embodiment, the system uses a method of fusing spatial features and temporal features to perform preliminary recovery status classification and identification, and obtains the current recovery status identification result.

[0101] Meanwhile, in order to objectively quantify the stability of the recovery process, this embodiment introduces an evaluation mechanism based on statistical distance.

[0102] The system forms a set of normal recovery states based on records of normal recovery processes in historically valid collected samples or training samples. .

[0103] For example, in another embodiment, the set of normal recovery states The specific construction method is as follows: 500 sets of historical valid samples are obtained from the cloud management module, records marked as normal recovery by medical experts or nursing staff are selected, and then classified according to their corresponding age. and key events (such as signs of detachment of the stump) The data is then stratified. For each stratum, the structural state vectors of all samples are extracted. Calculate the covariance matrix of this layer. and mean vector And filter out those that are related to The Mahalanobis distance is less than a preset threshold (e.g., ) sample points The set of normal recovery states that constitute this stratification It should be noted that if a stratum has insufficient valid samples, it will fall back to a coarser-grained group. It should also be noted that the current vector to be evaluated... The stratification will be matched with the newborn's current age and shedding status. To calculate the deviation distance .

[0104] The set of normal recovery states can be grouped and established according to the number of days after birth, whether the stump has fallen off, the root contact status, or the stage of medical and nursing labeling.

[0105] For the current structural state vector Calculate its minimum statistical deviation distance relative to the set of normal recovery states: ; in, This is the reference state vector in the set of normal recovery states. This is the covariance matrix corresponding to the normally recovered sample. This covariance matrix is ​​used to reflect the natural fluctuation range and correlation of different structural parameters during the normal recovery process.

[0106] Continue calculating the recovery stability score parameters: ; in, For cross-time series comparability scoring parameters, This represents the statistical deviation distance of the current structural state vector relative to the set of normal recovery states.

[0107] If the currently acquired image is not highly comparable to historical images, then The decrease in stability score indicates a trend that may lead to a decrease or cessation of output. If the current structural state deviates significantly from the normal recovery trajectory, then An increase in the exponential term indicates a decrease in the stability of the current recovery state.

[0108] like Figure 7 As shown, after obtaining the current recovered structural state vector and completing the statistical deviation calculation from the normal recovered state set, this invention generates recovery stability score parameters. This is used to describe how close the current recovery status is to the normal recovery trajectory. The horizontal axis in the graph represents days postnatal, and the vertical axis represents the recovery stability score. As the recovery process progresses, most valid data collection points... The data shows a gradual upward trend or stabilization in a higher range, indicating that the umbilical structure is gradually approaching a normal recovery pattern. Data points excluded due to insufficient acquisition quality or inadequate cross-time series comparability are distinguished by special markings in the figure. Instead of providing trend conclusions, these points are prompted to be reacquired to avoid unreliable data interfering with the recovery trajectory.

[0109] also, Figure 7 It also provides two decision boundaries: a verification threshold and a stability threshold. When... When the value is below the stability threshold but still above the review threshold by 0.5, it indicates that although the current recovery status has not reached the ideal stability range, it can still be observed; when When the value approaches or falls below the review threshold, especially when accompanied by an increase in structural deviation or a sudden decline in trend, the system can further output information suggesting medical staff review to indicate a risk of abnormal recovery. Therefore, the present invention... It is not an isolated single-point score, but rather forms a continuous trend together with historical valid data collection records. This is used to distinguish between normal recovery fluctuations, abnormal data collection, and potential abnormal states requiring medical intervention, thereby enhancing the practical significance of the system's output results.

[0110] Step S107: Based on at least one of the current identification results and historical identification results, output at least one of the following: recovery stability information, trend analysis results, and collected feedback information; More specifically, in step S107 of another embodiment, the system outputs result information based on the current identification result, the recovery stability score, and historical records.

[0111] For the recent The recovery stability sequence from the first effective acquisition is used to define the trend volatility. for: ; It should be noted that when When the fluctuation exceeds the preset threshold, the system prioritizes backtracking and checking each historical data collection record. and average confidence level of anchor points If a certain collection or If the image quality is significantly lower than the normal range, the trend fluctuation is primarily determined to be caused by poor image quality or a poorly constructed reference system. A prompt will be displayed to the user indicating that the image quality or anchoring effect of that particular image is subpar and suggesting a retake. Only when the quality of each captured image is significantly lower than the normal range will the trend fluctuation be considered to be caused by poor image quality or a poorly constructed reference system. and All are normal Only when the number remains high will a warning be sent to healthcare professionals.

[0112] In one implementation, the preset fluctuation threshold is: The empirical value between them, if calculated twice consecutively If all exceed this threshold, a backtracking check process will be triggered.

[0113] Based on the above, in another embodiment, an example is provided of umbilical cord monitoring of a newborn on the 9th day after birth, conducted by parents using a mobile phone in a home environment based on the present invention. The parents use their mobile phone to collect short-term image sequences of the umbilical cord in a home environment, collecting a total of... Frames, duration approximately .

[0114] The system first calculates the sharpness and exposure balance of each frame in step S102, and then discards those frames. Slightly blurry frames caused by hand tremors are retained. Candidate frames.

[0115] The Mahalanobis distance is calculated using the retained frames to obtain... If the value is higher than the threshold, the quality is deemed acceptable, and the value is stored in the cloud.

[0116] In step S103, the system extracts candidate points for the center of the umbilical depression, the endpoints of the root contact arc segment, and the convergence points of the folds from the candidate frames. It then uses the anatomical anchor point confidence model to calculate the corner response and motion variance of each point, selects the anchor point group with the highest confidence level that meets the distribution requirements, establishes a local anatomical reference system, and sets the average confidence level... Store in the cloud.

[0117] In step S104, the current image is non-rigidly aligned with the valid umbilical images stored in the cloud from the previous two days, and then the alignment is calculated. The origin offset and angular deviation are both small, resulting in If the time series is deemed comparable, it is allowed to be included in trend analysis.

[0118] In step S105, the system extracts the root contact interface features of the current image to obtain: contact arc length. Average width of contact band Standard deviation of contact strip width Area of ​​exposed zone Number of local detachment points The calculation yields: ; Simultaneously, the dispersion of the fold direction was calculated. High-frequency texture energy ratio Forming a structural state vector .

[0119] In step S106, the system compares the state vector with the set of normal recovery states and calculates the minimum Mahalanobis distance. , combined The recovery stability score was obtained. .

[0120] This score is analyzed in conjunction with historical data to determine trend volatility. Located in the low fluctuation range, the system feedback indicates that the current recovery status is stable. It is recommended to continue observation at the existing frequency.

[0121] Example 2: like Figure 8 As shown in the figure, the artificial intelligence-based neonatal umbilical cord recovery recognition system of this invention includes the following modules: The module includes an image acquisition module, a preprocessing and quality control module, an anatomical anchoring module, a cross-temporal alignment module, a restoration feature recognition module, and a data analysis and feedback module. The image acquisition module is used to acquire umbilical image data of newborns; The preprocessing and quality control module is used to evaluate and filter the umbilical image data to obtain valid image data. The anatomical anchoring module is used to extract candidate anatomical anchor points of the umbilicus and establish a local anatomical reference system for the umbilicus. The cross-temporal alignment module is used to perform cross-temporal alignment between the umbilical image at the current acquisition time and the umbilical image at the historical acquisition time under the local anatomical reference system of the umbilicus, so as to establish a stable anatomical correspondence between different time points. The restoration feature recognition module is used to extract restoration features based on the aligned umbilicus image and perform restoration recognition. The restoration features include at least one or more of the following: root contact interface features, periumbilical fold reconstruction features, and stump surface texture evolution features. The data analysis and feedback module is used to output at least one of the following based on at least one of the current identification results and historical identification results: recovery stability information, trend analysis results, and collected feedback information.

[0122] like Figure 8 As shown, the system of this invention adopts an architecture in which a parent's mobile terminal, a cloud server, and a medical care terminal work collaboratively. The parent's mobile terminal is used to acquire and upload umbilical cord image sequences and receive system-returned prompts for re-image capture, nursing suggestions, and scoring results. The cloud server is used to handle image data storage, structural parameter management, matching of normal recovery status sets, trend analysis, and anomaly attribution processing. The medical care terminal is used to receive anomaly records, view historical trend curves, and manually review anomaly samples and provide nursing feedback. This three-terminal linkage architecture effectively combines daily continuous acquisition in a home environment with professional review in a medical setting, enabling the system to not only output single recognition results but also establish a continuous temporal recovery monitoring link around the same newborn.

[0123] Furthermore, Figure 8 The image library, parameter library, normal recovery state set, trend analysis, and anomaly attribution functional units shown correspond to the unified cloud management process of the output results of each module in this invention. In other words, the system does not only make immediate judgments on single images, but also associates and stores the image data generated from each acquisition, the prior comparability scoring parameters, the cross-time series comparability scoring parameters, the recovery stability scoring parameters, and the medical staff review results. This allows for retrospective analysis of historical states, identification of trend anomalies, and output of more explanatory prompts to parents in subsequent analyses. This system architecture facilitates a closed-loop application of home-based initial screening—cloud analysis—medical staff review—nursing feedback, improving the practicality and scalability of this invention in real-world nursing scenarios.

[0124] Example 3: This embodiment provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor; The processor executes the aforementioned AI-based method for recognizing the recovery status of a newborn's umbilical cord by calling a computer program stored in memory.

[0125] The electronic device can vary considerably depending on its configuration or performance. It may include one or more Central Processing Units (CPUs) and one or more memories, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the artificial intelligence-based neonatal umbilical cord recovery recognition method provided in the above-described embodiment. The electronic device may also include other components for implementing its functions; for example, it may have wired or wireless network interfaces and input / output interfaces for data input and output. Further details are omitted in this embodiment.

[0126] Example 4: This embodiment proposes a computer-readable storage medium on which an erasable and rewritable computer program is stored. When the computer program runs on the computer device, it causes the computer device to execute the aforementioned artificial intelligence-based method for recognizing the recovery status of a newborn's umbilical cord.

[0127] For example, computer-readable storage media can be read-only memory (ROM), random access memory (RAM), compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage devices.

[0128] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired network and / or wireless network. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives (SSDs).

[0129] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0130] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

Claims

1. A method for recognizing the recovery status of a newborn's umbilical cord based on artificial intelligence, characterized in that, The method includes: Step S101: Obtain umbilical image data of the newborn, the umbilical image data including umbilical image data at the current acquisition time and umbilical image data at at least one historical acquisition time; Step S102: Preprocess and quality control the umbilical image data to obtain valid umbilical image data; Step S103: Extract candidate anatomical anchor points based on the effective umbilical image data, and establish a local anatomical reference system for the umbilicus based on the candidate anatomical anchor points; Step S104: Under the local anatomical reference frame of the umbilicus, perform cross-time alignment between the umbilicus image data at the current acquisition time and the umbilicus image data at the historical acquisition time; Step S105: Extract recovery features based on the aligned umbilical image data. The recovery features include at least root contact interface features, periumbilical fold reconstruction features, and residual surface texture evolution features. Step S106: Perform recovery status identification based on the recovery status features to obtain the neonatal umbilical cord recovery status identification result; Step S107: Based on at least one of the current identification results and historical identification results, output at least one of the following: recovery stability information, trend analysis results, and collected feedback information; The establishment of the umbilical local anatomical reference system includes determining the reference origin, the principal axis direction, and the secondary axis direction. The cross-temporal alignment includes performing differentiated alignment constraints based on the stability differences of different structural regions within the umbilical region.

2. The method for identifying the recovery status of a newborn's umbilical cord based on artificial intelligence according to claim 1, characterized in that, In step S101, acquiring newborn umbilical image data includes acquiring a short-time image sequence of the newborn umbilical region, wherein the acquisition duration of the short-time image sequence is [duration missing]. The number of image frames is Frames, capture frame rate During the acquisition process, the angle between the principal optical axis of the acquisition terminal lens and the local normal of the abdominal wall should not exceed [the specified value]. .

3. The method for identifying the recovery status of a newborn's umbilical cord based on artificial intelligence according to claim 2, characterized in that, In step S102, the preprocessing and quality control of the umbilical image data includes: calculating the sharpness index, exposure balance index, and average displacement index between adjacent frames of each frame; and based on the sharpness index, exposure balance index, and average displacement index, filtering out blurred frames, overexposed frames, underexposed frames, and frames with large drift; and based on the remaining image frames after filtering, determining the target image frames or target image frame set for anatomical anchoring. The sharpness index is calculated based on the Laplacian response or gradient response of the image; the exposure balance index is calculated based on the mean gray level and standard deviation of the gray level in the target area; and the average displacement index is calculated based on the feature point tracking results between adjacent frames.

4. The method for identifying the recovery status of a newborn's umbilical cord based on artificial intelligence according to claim 3, characterized in that, Step S102 further includes constructing a comparability prior scoring model to comprehensively evaluate the overall quality of the image data collected this time. The comparability prior scoring model is configured as follows: a state vector is constructed based on representative features of the available representative frames after filtering. And based on the mean vector obtained in advance from qualified samples under standard collection conditions. With covariance matrix Calculate Mahalanobis distance and generate comparability prior scores. ; when If the image quality falls below a preset threshold, the image acquisition is deemed substandard, a retake prompt is output, and cross-time-series trend analysis of the acquired image data is stopped.

5. The method for identifying the recovery status of a newborn's umbilical cord based on artificial intelligence according to claim 4, characterized in that, Step S103, extracting candidate anatomical anchor points and establishing a local anatomical reference system for the umbilicus, includes: extracting one or more of the following from the umbilicus image: candidate point of the umbilical fossa center, convergence point of the periumbilical folds, endpoint of the contact arc segment of the stump root, candidate line of the local principal axis of symmetry of the abdominal wall, and abrupt change zone of local texture direction; scoring the candidate anatomical anchor points using an anatomical anchor point credibility model; and selecting a target anchor point combination based on the scoring results to establish a local anatomical reference system for the umbilicus. The umbilical local anatomical reference system takes the geometric center of the root contact arc segment or the umbilical fossa stability center as the reference origin, the main symmetry direction of the abdominal wall as the main axis direction, and the direction orthogonal to the main axis direction as the secondary axis direction. The configuration of the anatomical anchor point credibility model is as follows: for the first... A number of candidate anchor points are selected based on the eigenvalues ​​of the Harris matrix formed by the image gradients within their neighborhood windows. and Calculate the corner response of pure shape , To prevent the use of tiny positive numbers with a denominator of zero; Based on the displacement variance of the candidate anchor points in the short-time image sequence Calculate the motion stability penalty term ; Calculate the confidence level of candidate anchor points ,in The parameters used to control the intensity of the penalty for motion stability.

6. The method for identifying the recovery status of a newborn's umbilical cord based on artificial intelligence according to claim 5, characterized in that, In step S104, cross-temporal alignment includes: dividing the umbilicus region into a root contact stability region, a periumbilical fold elastic change region, and a stump free end drift region; applying a strong constraint to the root contact stability region, applying an elastic constraint to the periumbilical fold elastic change region, and not applying an additional geometric constraint to the stump free end drift region. Wherein, the positional consistency constraint strength applied to the root contact stability region is greater than the positional consistency constraint strength applied to the residual free end drift region; Among them, the root contact stability region With reference origin The elastic variation zone of the periumbilical folds is defined by taking the root contact arc segment as the center and the characteristic scale as the radius. For self The outer edge extends outward to a ring-shaped region containing the convergence point of the periumbilical folds, the free end drift region of the stump. Except for the umbilical region and The area where the remaining parts are located.

7. The method for identifying the recovery status of a newborn's umbilical cord based on artificial intelligence according to claim 6, characterized in that, In step S104, cross-time alignment further includes: A partitioned non-rigid alignment model is constructed, and the optimal displacement field is solved by functional energy optimization. The objective function for functional energy optimization is composed of a weighted sum of a data fidelity term, an elastic regularization term, and a rigid regularization term. The data fidelity term is constructed based on the local normalized cross-correlation of the entire map region. The elastic regularization term corresponds to the elastic variation region of the periumbilical fold and is constructed based on the strain tensor. The rigid regularization term corresponds to the root contact stability region and is constructed based on the displacement norm. After alignment is completed, cross-temporal comparability scoring parameters are calculated based on the alignment quality and anatomical pose deviation of the root contact stability zone. ; ; in This represents the normalized cross-correlation coefficient after alignment of the root contact stable region. As a reference to the origin offset, Main axis included angle deviation, and These are the acceptable positional and angular deviation characteristic scales, respectively; When the cross-time series comparability scoring parameter is lower than the preset threshold, the trend analysis results are stopped, and a retake prompt or only a single recovery identification result is output.

8. The method for identifying the recovery status of a newborn's umbilical cord based on artificial intelligence according to claim 7, characterized in that, In step S105, the extraction of recovery status features includes: Extract the root contact interface features and construct structural parameters based on the root contact interface features; The root contact interface features include contact arc length. Average width of contact band Standard deviation of contact strip width and exposed zone area ; The structural parameters include the relative width parameter of the contact strip. Contact zone non-uniformity parameters and relative area parameters of the exposed zone ; Extracting periumbilical fold reconstruction features, the periumbilical fold reconstruction features including the dispersion of the periumbilical fold direction. ; Extracting the surface texture evolution features of the remnant, the surface texture evolution features of the remnant include the proportion of high-frequency texture energy. ; Construct the restored structure state vector at the current moment .

9. The method for identifying the recovery status of a newborn's umbilical cord based on artificial intelligence according to claim 8, characterized in that, Step S106 includes: Calculate the current restored structure state vector Minimum statistical deviation relative to a pre-built set of normal recovery states ; ;in This is a set of normal recovery states. The covariance matrix corresponding to the normally recovered sample; Based on the minimum statistical deviation distance and the cross-time series comparability scoring parameters Generate recovery stability score parameters , Output stability recovery information; The recovery status identification is performed by spatial feature extraction and temporal feature fusion. When performing temporal feature fusion, only the structural state vectors corresponding to historically acquired image data that meet the cross-temporal comparability requirements are included.

10. An artificial intelligence-based neonatal umbilical cord recovery status recognition system, used to implement the artificial intelligence-based neonatal umbilical cord recovery status recognition method as described in any one of claims 1-9, characterized in that, The system includes the following modules: The system includes an image acquisition module, a preprocessing and quality control module, an anatomical anchoring module, a cross-temporal alignment module, a restoration feature recognition module, and a data analysis and feedback module. The image acquisition module is used to acquire umbilical image data of newborns; The preprocessing and quality control module is used to evaluate and filter the umbilical image data to obtain valid image data. The anatomical anchoring module is used to extract candidate anatomical anchor points of the umbilicus and establish a local anatomical reference system for the umbilicus. The cross-temporal alignment module is used to perform cross-temporal alignment between the umbilical image at the current acquisition time and the umbilical image at the historical acquisition time under the local anatomical reference system of the umbilicus, so as to establish a stable anatomical correspondence between different time points. The restoration feature recognition module is used to extract restoration features based on the aligned umbilicus image and perform restoration recognition. The restoration features include at least one or more of the following: root contact interface features, periumbilical fold reconstruction features, and stump surface texture evolution features. The data analysis and feedback module is used to output at least one of the following based on at least one of the current identification results and historical identification results: recovery stability information, trend analysis results, and collected feedback information.

Images