Gastrointestinal anastomotic stoma healing real-time monitoring image analysis system

Abstract

The invention relates to the technical field of medical image processing, in particular to a gastrointestinal anastomotic stoma healing real-time monitoring image analysis system which comprises an image acquisition module, a preprocessing module, a feature extraction module, a manifold construction module, a curvature analysis module, a geodesic prediction module, a risk assessment module and a decision support module. The system collects a continuous image sequence of a gastrointestinal anastomotic stoma area, extracts multi-dimensional features such as oxyhemoglobin saturation distribution, tissue morphology and blood vessel distribution, expresses a healing state as a low-dimensional manifold based on a differential geometry theory, detects abnormality through multi-scale curvature analysis, predicts a healing trajectory by using a geodesic theory, and performs multi-branch prediction. Risk grading and early warning are achieved, the healing state of the gastrointestinal anastomotic stoma can be objectively and quantitatively evaluated, nonlinear dynamic characteristics of the healing process are effectively captured, and potential abnormal healing is predicted 48-96 hours in advance.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a real-time monitoring and image analysis system for gastrointestinal anastomosis healing, which is applied to the monitoring and evaluation of the anastomosis healing process after gastrointestinal surgery. Background Technology

[0002] Gastrointestinal surgery is a crucial treatment for digestive system diseases, and the healing status of the anastomosis directly impacts patient prognosis. Complications such as anastomotic leakage, bleeding, and stenosis seriously threaten patient lives and increase the medical burden. Traditional anastomotic healing monitoring relies heavily on physician experience, which suffers from high subjectivity and poor timeliness. While endoscopy allows for direct observation of the anastomosis, continuous monitoring is difficult, and the healing trend cannot be predicted.

[0003] Existing monitoring methods have the following shortcomings: First, they mostly use a single indicator for assessment, such as observing only the surface morphology of tissue or a single physiological parameter, which cannot reflect the complexity of the healing process; second, most of them use static threshold judgment, ignoring individual differences and time dynamic characteristics in the healing process; third, they lack effective prediction methods, making it difficult to detect potential risks and intervene in a timely manner; and finally, the analysis results are mostly qualitative descriptions, lacking quantitative standards, making it difficult to objectively evaluate the healing process.

[0004] Therefore, there is an urgent need for a gastrointestinal anastomosis healing analysis system that can integrate multiple physiological indicators, achieve real-time monitoring, and has predictive functions, so as to provide a scientific basis for clinical decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time monitoring image analysis system for gastrointestinal anastomosis healing. By acquiring continuous images of the gastrointestinal anastomosis area, extracting multidimensional features, and constructing a manifold representation of the healing state based on differential geometry theory, the system can assess and predict the healing state, thereby promptly detecting abnormalities and guiding clinical intervention.

[0006] This invention proposes a real-time monitoring and image analysis system for gastrointestinal anastomosis healing, comprising:

[0007] The image acquisition module is used to acquire continuous image sequences of the gastrointestinal anastomosis region;

[0008] The preprocessing module, connected to the image acquisition module, is used to standardize the continuous image sequence to generate standardized image data.

[0009] The feature extraction module, connected to the preprocessing module, is used to extract blood oxygen saturation distribution, tissue morphology features, and blood vessel distribution features from the standardized image data.

[0010] The manifold construction module, connected to the feature extraction module, is used for:

[0011] Construct a multidimensional feature space that includes the blood oxygen saturation distribution, the tissue morphology features, and the blood vessel distribution features;

[0012] Establish nearest neighbor relationships between data points in the multidimensional feature space;

[0013] A low-dimensional manifold representation of the healing state is constructed based on the nearest neighbor relationship;

[0014] The curvature analysis module, connected to the manifold construction module, is used for:

[0015] Calculate the multi-scale curvature features in the low-dimensional manifold representation;

[0016] Abnormal states during the healing process are identified based on the aforementioned multi-scale curvature features;

[0017] The geodesic prediction module, connected to the manifold construction module and the curvature analysis module, is used for:

[0018] Construct the distance field on the healed state manifold;

[0019] Calculate the optimal healing path from the current state to the target state;

[0020] Generate multiple possible healing trajectory branches and their probability distributions;

[0021] The risk assessment module, connected to the geodesic prediction module, is used for:

[0022] The healing status is risk-classified based on the multiple possible healing trajectory branches;

[0023] When the risk level exceeds a preset threshold, an abnormal healing warning signal is generated;

[0024] The decision support module, connected to the risk assessment module, is used for:

[0025] Receive the abnormal healing warning signal;

[0026] Generate clinical decision support information that includes risk levels and intervention recommendations;

[0027] Output the aforementioned clinical decision support information.

[0028] Preferably, the image acquisition module includes:

[0029] A light source unit is used to provide dual-wavelength illumination covering the visible and near-infrared spectra;

[0030] The imaging unit is used to acquire images of the gastrointestinal anastomosis area under the visible spectrum and the near-infrared spectrum;

[0031] The acquisition and control unit is used to adaptively adjust the image acquisition frequency and imaging parameters according to the patient's healing stage.

[0032] Preferably, the preprocessing module includes:

[0033] A noise filtering unit is used to remove random noise from the continuous image sequence;

[0034] Lighting correction unit is used to correct uneven lighting conditions;

[0035] A region identification unit is used to automatically identify and mark the gastrointestinal anastomosis region;

[0036] The image registration unit is used to spatially align images acquired at different time points.

[0037] Preferably, the feature extraction module includes:

[0038] The spectral analysis unit is used to calculate tissue oxygen saturation distribution maps based on dual-wavelength images;

[0039] Edge extraction unit is used to detect the integrity of the anastomosis edges and tissue boundary features;

[0040] The blood vessel recognition unit is used to extract and quantify the blood vessel distribution density and morphology in the anastomosis area;

[0041] The texture analysis unit is used to extract texture feature parameters that characterize tissue structure.

[0042] Preferably, the manifold building module includes:

[0043] The similarity calculation unit is used to calculate the similarity matrix between data points in the multidimensional feature space;

[0044] The nearest neighbor relationship building unit is used to determine the K nearest neighbors for each data point;

[0045] An embedding optimization unit is used to solve low-dimensional embedding coordinates by preserving local linearity constraints.

[0046] A dynamic modeling unit is used to construct a vector field describing the transition of healing states on the low-dimensional manifold representation.

[0047] Preferably, the curvature analysis module includes:

[0048] A curvature calculation unit is used to estimate the discrete curvature on the low-dimensional manifold representation;

[0049] Multi-scale filtering units are used to generate curvature features that characterize different time scales;

[0050] An anomaly detection unit is used to identify regions of abnormal curvature based on a preset normal curvature template;

[0051] The region mapping unit is used to map the curvature features to the original anastomosis region and generate differential feature maps.

[0052] Preferably, the geodesic prediction module includes:

[0053] A distance field construction unit is used to construct a distance field based on Riemannian metrics on the low-dimensional manifold representation;

[0054] The path optimization unit is used to find the optimal geodesic path from the current state to the target state;

[0055] A window adjustment unit is used to dynamically adjust the size of the prediction time-domain window based on the rate of curvature change.

[0056] A multi-branch generation unit is used to generate multiple predicted trajectory branches that represent different healing scenarios.

[0057] Preferably, the risk assessment module includes:

[0058] A probability distribution calculation unit is used to assign probability weights to the multiple predicted trajectory branches;

[0059] Risk grading unit, used to classify the healing status according to the distribution characteristics of the predicted trajectory;

[0060] A threshold determination unit is used to compare the risk classification with a preset risk threshold.

[0061] The early warning generation unit is used to generate a graded early warning signal when the risk level exceeds the preset risk threshold.

[0062] Preferably, the decision support module includes:

[0063] The early warning receiving unit is used to receive and parse the abnormal healing early warning signal;

[0064] A suggestion generation unit is used to generate intervention suggestions based on the type of healing abnormality and the risk level;

[0065] Visualization units are used to visually represent healing trajectory predictions and risk distributions in a graphical manner;

[0066] The notification distribution unit is used to push the clinical decision support information to the terminals of medical staff.

[0067] As a preferred option, it also includes:

[0068] The data storage module is connected to the image acquisition module, the feature extraction module, and the risk assessment module, and is used to store raw image data, extract features, and assess results.

[0069] The knowledge update module, connected to the data storage module, is used for:

[0070] The curvature template for normal healing was updated based on new clinical data;

[0071] Optimize the distance metric in the low-dimensional manifold representation;

[0072] Adjust the probability distribution calculation parameters of the multiple predicted trajectory branches.

[0073] The present invention has the following beneficial effects:

[0074] 1. To achieve objective and quantitative assessment of the healing status of gastrointestinal anastomoses, reduce subjective judgment errors, and improve assessment accuracy;

[0075] 2. Through comprehensive analysis of multidimensional features, the complex process of anastomotic healing is fully reflected, providing a more comprehensive characterization of the healing status;

[0076] 3. The time-series analysis framework based on differential geometry can effectively capture the nonlinear dynamic characteristics of the healing process and improve the model's adaptability to complex healing modes;

[0077] 4. Multi-branch prediction technology can predict potential healing abnormalities 48-96 hours in advance, providing a sufficient time window for timely clinical intervention;

[0078] 5. Risk stratification mechanisms help doctors determine intervention priorities and optimize the allocation of medical resources;

[0079] 6. Reduces the incidence of anastomosis-related complications, shortens hospital stays and reduces reoperation rates, significantly improving patient outcomes. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of the overall architecture of the real-time monitoring and image analysis system for gastrointestinal anastomosis healing of the present invention.

[0081] Figure 2 This is a schematic diagram of the image acquisition module of the present invention.

[0082] Figure 3 This is a schematic diagram of the processing flow of the preprocessing module of the present invention.

[0083] Figure 4 This is a schematic diagram of the feature extraction module of the present invention.

[0084] Figure 5 This is a schematic diagram illustrating the working principle of the manifold construction module of the present invention.

[0085] Figure 6 This is a schematic diagram of the multi-scale analysis of the curvature analysis module of the present invention.

[0086] Figure 7 This is a schematic diagram of the workflow of the geodesic prediction module of the present invention.

[0087] Figure 8 This is a schematic diagram illustrating the risk classification of the risk assessment module of the present invention.

[0088] Figure 9 This is a schematic diagram of the information presentation interface of the decision support module of the present invention.

[0089] Figure 10 This is a schematic diagram of the update process of the knowledge update module of the present invention. Detailed Implementation

[0090] Please refer to Figures 1-10 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0091] like Figure 1 As shown, the real-time monitoring image analysis system for gastrointestinal anastomosis healing of the present invention includes an image acquisition module 10, a preprocessing module 20, a feature extraction module 30, a manifold construction module 40, a curvature analysis module 50, a geodesic prediction module 60, a risk assessment module 70, a decision support module 80, a data storage module 90, and a knowledge update module 100.

[0092] Image acquisition module 10 is used to acquire continuous image sequences of the gastrointestinal anastomosis region. Preprocessing module 20, connected to image acquisition module 10, is used to standardize the continuous image sequences, generating standardized image data. Feature extraction module 30, connected to preprocessing module 20, is used to extract blood oxygen saturation distribution, tissue morphology features, and vascular distribution features from the standardized image data. Manifold construction module 40, connected to feature extraction module 30, constructs a manifold representation of the healing state based on the extracted features. Curvature analysis module 50, connected to manifold construction module 40, is used to analyze curvature features on the manifold and detect anomalies. Geodesic prediction module 60, connected to manifold construction module 40 and curvature analysis module 50, is used to predict the healing trajectory. Risk assessment module 70, connected to geodesic prediction module 60, is used to assess healing risk and generate early warning signals. Decision support module 80, connected to risk assessment module 70, is used to provide clinical decision recommendations. Data storage module 90, connected to multiple other modules, stores system operation data. Knowledge update module 100, connected to data storage module 90, is used to update the system knowledge base.

[0093] The modules interact with each other through standard interfaces, forming a complete closed loop from image acquisition to decision support. This invention adopts a modular design, and each functional module can be independently upgraded and optimized according to actual needs.

[0094] Reference Figure 2 The image acquisition module 10 includes a light source unit 11, an imaging unit 12, and an acquisition control unit 13.

[0095] The light source unit 11 is designed as a dual-wavelength light source system, simultaneously providing illumination in the visible spectrum (400-700nm) and the near-infrared spectrum (700-1000nm). This design enables the system to simultaneously acquire tissue surface morphology information and deep blood oxygenation distribution information. Preferably, the visible light source uses white LED light, and the near-infrared light source uses a near-infrared LED with a wavelength of 850nm. The two light sources are output to the same optical path through an optical beam combining system.

[0096] The imaging unit 12 employs a high-resolution medical endoscope camera with a resolution of no less than 1920×1080 pixels, supporting simultaneous acquisition of visible light and near-infrared images. Preferably, the imaging unit 12 uses a dual-spectrum CMOS sensor, which can simultaneously acquire images of two bands from the same viewing angle, ensuring accurate spatial registration.

[0097] The acquisition control unit 13 adaptively adjusts the image acquisition frequency and imaging parameters according to the patient's healing stage. In one embodiment of the invention, the acquisition frequency is set to once every 4 hours during the acute phase (0-3 days post-surgery), once every 8 hours during the subacute phase (4-7 days post-surgery), and once every 24 hours during the stable phase (8 days or more post-surgery). Simultaneously, the acquisition control unit 13 also automatically adjusts the exposure parameters according to ambient light conditions to ensure image quality.

[0098] like Figure 3 As shown, the preprocessing module 20 includes a noise filtering unit 21, an illumination correction unit 22, a region recognition unit 23, and an image registration unit 24.

[0099] The noise filtering unit 21 employs an adaptive median filtering algorithm to remove random noise from the image. Compared to traditional fixed-window median filtering, adaptive median filtering can dynamically adjust the filtering window size according to the local noise level, effectively suppressing noise while preserving image details. In this invention, the window size is set to an adaptive range of 3×3 to 9×9, and the noise judgment threshold is twice the local standard deviation.

[0100] The illumination correction unit 22 employs an improved histogram equalization algorithm to correct uneven illumination. Endoscopic images often exhibit uneven illumination due to the location of the light source and tissue reflections, affecting subsequent analysis. The improved algorithm used in this invention first divides the image into multiple sub-regions, performs histogram equalization independently on each sub-region, and then synthesizes them through bilinear interpolation, avoiding the over-contrast phenomenon that may be caused by traditional global equalization.

[0101] Region identification unit 23 automatically identifies and marks the gastrointestinal anastomosis region using morphological features and color information. This unit employs a two-stage strategy: first, coarse segmentation is performed based on a color threshold, and then fine boundary localization is performed using an active contour model (Snake algorithm). The identification results generate an anastomosis region mask for subsequent region limitation in analysis.

[0102] Image registration unit 24 spatially aligns images acquired at different time points to eliminate spatial mismatch caused by changes in shooting angle and distance. This unit employs a combination of feature point matching and non-rigid transformation. First, it establishes a correspondence through SIFT feature point matching, and then achieves non-rigid registration through thin-plate spline transformation, achieving a registration accuracy better than 0.5 mm.

[0103] Reference Figure 4 The feature extraction module 30 includes a spectral analysis unit 31, an edge extraction unit 32, a blood vessel recognition unit 33, and a texture analysis unit 34.

[0104] The spectral analysis unit 31 calculates the tissue oxygen saturation distribution map based on dual-wavelength images. Oxygen saturation (SpO2) is an important indicator for assessing tissue blood perfusion and is closely related to anastomotic healing. This unit utilizes the absorption differences between oxyhemoglobin and deoxyhemoglobin in the visible and near-infrared bands to calculate the oxygen saturation value of each pixel using a spectral analysis algorithm. The calculation formula is as follows:

[0105] ,

[0106] SpO2 is the blood oxygen saturation value, expressed as a percentage, and typically ranges from 0% to 100%. Tissue reflectance in the near-infrared band, dimensionless, ranging from 0 to 1; The tissue reflectance in the visible light band is dimensionless and ranges from 0 to 1. , The calibration coefficients, calibrated using standard samples, are dimensionless. This represents the natural logarithm function. In this system, a typical value is... These parameter values ​​were obtained by calibrating them by comparing them with a standard pulse oximeter.

[0107] Edge extraction unit 32 detects the integrity of the anastomotic edges and tissue boundary features. During the healing process, the clarity and continuity of the anastomotic edges are important evaluation metrics. This unit employs the Canny edge detection algorithm, with parameters optimized to a low threshold of 0.1, a high threshold of 0.3, and a Gaussian filter kernel width of 1.5. These parameter settings demonstrate the best edge detection performance on gastrointestinal tissue images.

[0108] The vessel recognition unit 33 extracts and quantifies the vessel distribution density and morphology in the anastomosis region. Angiogenesis is a crucial step in the healing process, and its distribution characteristics directly reflect the quality of healing. This unit first enhances tubular structures through multi-scale Hessian matrix analysis, then extracts the vascular network through threshold segmentation, and finally calculates quantitative indicators such as vessel density and the number of branch points.

[0109] The texture analysis unit 34 extracts texture feature parameters characterizing tissue structure. Changes in tissue texture reflect tissue remodeling during the healing process. This unit calculates parameters such as contrast, homogeneity, energy, and correlation of the gray-level co-occurrence matrix (GLCM) to capture changes in tissue structure. In practical applications, GLCM calculation uses the average value of a pixel spacing of 1 and orientations of 0°, 45°, 90°, and 135° to reduce orientation dependence.

[0110] like Figure 5 As shown, the manifold building module 40 includes a similarity calculation unit 41, a nearest neighbor relationship building unit 42, an embedding optimization unit 43, and a dynamic modeling unit 44.

[0111] The similarity calculation unit 41 calculates the similarity matrix between data points in the multidimensional feature space. This unit first performs feature normalization on the blood oxygen saturation distribution, tissue morphology features, and blood vessel distribution features to ensure the comparability of features with different dimensions. Normalization uses the min-max method, mapping each feature value to the [0,1] interval.

[0112] ,

[0113] in: is the normalized eigenvalue, dimensionless, and in the range [0,1]. These are the original feature values, with units determined by the feature. The minimum value of this feature across all samples, in units of... same; The maximum value of this feature across all samples, in units of... The same applies. Then, the weighted Euclidean distance between any two data points is calculated to construct a similarity matrix. :

[0114] ,

[0115] in: For data points and The distance between them is dimensionless; and Data points and In the Normalized values ​​on each feature dimension, dimensionless, ranging from [0,1]; For the first The weight coefficients of each feature are dimensionless and satisfy the following conditions: ; This represents the total dimension of the features, i.e., the number of all extracted features; This indicates that for all feature dimensions from 1 to... Summation. In this system, the weight of blood oxygen saturation feature is set to 0.5, the weight of tissue morphology feature is 0.3, and the weight of vascular distribution feature is 0.2, reflecting the different importance of these features in the assessment of healing status.

[0116] The nearest neighbor construction unit 42 determines the K nearest neighbors for each data point. Based on the similarity matrix, this unit finds the K nearest neighbors for each data point and constructs a nearest neighbor graph. The choice of K value has a significant impact on the manifold learning results; too small a value leads to disjointed local structures, while too large a value introduces too many long-distance connections. In this system, the K value is set adaptively, with the basic rule being the square root of the total number of samples, but not less than 8 and not more than 15.

[0117] Embedding optimization unit 43 solves for low-dimensional embedding coordinates by preserving local linearity constraints. This unit employs the Locally Linear Embedding (LLE) algorithm, first calculating the weights of each data point that can be reconstructed from its K nearest neighbors through linear combinations:

[0118] ,

[0119] in: To reconstruct the weight matrix, the size is , This represents the total number of data points. For the first The feature vector of each data point has a dimension of . That is, the total dimension of features; For data points of A set of nearest neighbors; For data points Reconstructing data points The contribution weight, when The value is non-zero if it is zero, otherwise it is 0; Indicates all Sum of data points; Indicates data points All Summing the nearest neighbors; The Euclidean norm of a vector is denoted by .

[0120] Then, keeping these weights constant, solve for the low-dimensional embedding coordinates. :

[0121] ,

[0122] in: It is a low-dimensional embedding coordinate matrix with a size of , The embedding dimension is typically 3-4. This represents the total number of data points. For data points A coordinate vector in a low-dimensional space, with dimension . ; for identity matrix; Representation matrix Rather than transpose The product of matrices, of size By solving this eigenvalue problem, low-dimensional embedding coordinates are obtained, and a manifold representation of the healing state is constructed. In this system, the embedding dimension is generally set to 3-4 dimensions, which can both preserve the main structure of the data and effectively reduce computational complexity.

[0123] The dynamic modeling unit 44 constructs a vector field describing the healing state transition on a low-dimensional manifold representation. This unit constructs a vector field characterizing the dynamic healing process based on observation data at continuous time points:

[0124] ,

[0125] in: manifold point The vector field at that location has a dimension of The same as the embedded space dimension; Let be a point on the manifold with dimension . ; For point The set of neighborhood points; The number of points in the neighborhood, i.e., the set. The cardinality; and The first Data points in time and The position vector has dimensions of 1. ; The time interval is in hours; Indicates a point The summation is performed on all points within the neighborhood. The vector field describes the direction and rate of evolution of the healing state, providing a basis for subsequent predictions.

[0126] like Figure 6 As shown, the curvature analysis module 50 includes a curvature calculation unit 51, a multi-scale filtering unit 52, an anomaly detection unit 53, and a region mapping unit 54.

[0127] Curvature calculation unit 51 estimates the discrete curvature on the low-dimensional manifold representation. Curvature is an important geometric invariant on a manifold, reflecting the degree of local bending. This unit uses discrete differential geometry to calculate the Gaussian curvature and mean curvature of data points on the manifold:

[0128] ,

[0129] ,

[0130] in: For point Gaussian curvature at , in units of , Unit of distance; For point The average curvature at that point, in units of ; For point First The interior angles of two adjacent triangles, in radians; For points The connected first The length of the strip, in units of ; For point The neighborhood area, in units of ; For point The number of adjacent vertices, dimensionless; Indicates a point All Summing of adjacent triangles or sides; Let be the radian value of a complete circle. Gaussian curvature reflects the intrinsic geometric properties of the manifold, while mean curvature is related to the shape of the manifold in the embedded space.

[0131] Multi-scale filtering unit 52 generates curvature features representing different time scales. The healing process involves changes at different time scales, such as short-term inflammatory response and long-term tissue remodeling. This unit performs multi-scale smoothing of the curvature field using a Gaussian filter:

[0132] ,

[0133] in: For scale Lower point The smooth curvature, in units of same; For point The original curvature, unit and same; The standard deviation is The Gaussian function is defined as follows: Dimensionless; For point With point Geodesic distance between them, in units of ; Let be the set of all points on the manifold; The standard deviation parameter of the Gaussian filter is given in units of . Control the smoothness; Represents all points on the manifold Summation. In this system, the filtering scale... Set as a set of incrementing values This corresponds to time scales ranging from short-term to long-term.

[0134] Anomaly detection unit 53 identifies regions of abnormal curvature based on a preset normal curvature template. This unit first constructs a curvature change template for a normal healing process, and then calculates the degree of deviation between the currently observed curvature and the template. The deviation is measured using Mahalanobis distance.

[0135] ,

[0136] in: The deviation of the current curvature vector from the reference template is dimensionless. The current observed curvature vector has a dimension of It includes curvature values ​​at different scales; As a reference template curvature vector, the dimension is... same; Let be the covariance matrix, with size . This characterizes the variability of the template; It is the inverse of the covariance matrix; Representing vectors transpose; This represents the multiplication of vectors with matrices and vectors, with the result being a scalar. When the deviation exceeds a preset threshold, it is considered an anomaly in curvature. In this system, the deviation threshold is set to 2.5, a value that has demonstrated a good balance between sensitivity and specificity through clinical validation.

[0137] Region mapping unit 54 maps curvature features to the original anastomosis region, generating a differential feature map. This unit establishes a mapping relationship between the low-dimensional manifold representation and the original image space, back-projecting curvature anomaly regions onto the original image to visually display potential problem areas. The mapping uses radial basis function interpolation:

[0138] ,

[0139] in: This is a mapping function that maps coordinates in the original image space. Mapped to curvature values; This is a two-dimensional coordinate vector in the original image; For the first The coordinate vector of each control point, with dimensions and same; For radial basis functions, Gaussian functions are usually chosen. ,in This is a shape parameter, set to 0.1 in this system; Representing coordinates With control points The Euclidean distance between them; For the first Weighting coefficients for each control point; This represents the total number of control points. Indicates all The summation of control points is performed. Weights are determined by solving a system of linear equations, achieving a smooth mapping from manifold space to image space.

[0140] like Figure 7 As shown, the geodesic prediction module 60 includes a distance field construction unit 61, a path optimization unit 62, a window adjustment unit 63, and a multi-branch generation unit 64.

[0141] The distance field construction unit 61 constructs a Riemannian metric-based distance field on a low-dimensional manifold representation. On the manifold, the shortest path between two points is a geodesic, and its length is defined as the geodesic distance. This unit first defines the Riemannian metric tensor on the manifold:

[0142] ,

[0143] in: For point The metric tensor component at that location. and This is the component index, with a value range of 1 to... , For manifold dimension; Let Kronecker function be used when The value is 1 if it is true, and 0 otherwise. For point Gaussian curvature at , in units of ; These are weighting coefficients, in units of... This metric controls the degree to which curvature affects the measurement. It increases the "distance" of regions with higher curvature, thus classifying them as difficult to traverse. In this system, The value is set to 0.8 to make the metric appropriately sensitive to changes in curvature.

[0144] Then, calculate the geodesic distances from the starting point to all points on the manifold to construct the distance field. The range field was calculated using the FastMarching method, solving the Eikonal equations:

[0145] ,

[0146] in: From the starting point to the point Geodesic distance, in units of ; Let be the gradient vector of the distance field, with dimension . ; Indicates at point Metric tensor The norm of the downgradient vector is defined as ,in, To measure the inverse matrix components of a tensor, Indicates the distance field to the first Partial derivatives of each coordinate component.

[0147] Path optimization unit 62 solves for the optimal geodesic path from the current state to the target state. This unit constructs the optimal path based on gradient backtracking of the range field.

[0148] ,

[0149] in: The path is parameterized, i.e., the coordinate vector of points on the path, with dimension 1. ; For parameters, the range is usually 1000. ; The tangent vector of the path, i.e., the direction of the path; For point The gradient vector of the distance field at that location; The Euclidean norm of the gradient vector is defined as follows: ; This represents the reverse direction of the normalized gradient. By numerical integration, starting from the endpoint (ideal healed state), the path is traced back to the starting point (current state) along the reverse direction of the distance field gradient to obtain the optimal geodesic path.

[0150] Window adjustment unit 63 dynamically adjusts the size of the prediction time-domain window based on the rate of curvature change. The prediction window size should be adjusted according to the healing stage: a smaller window is needed in the rapid change stage to improve short-term prediction accuracy, while a larger window can be used in the stable stage to enhance the grasp of long-term trends. The window size adjustment formula is:

[0151] ,

[0152] in: The adjusted window size is shown in hours. The base window size is typically set to 72 hours in this system. The rate of change of curvature over time, in units of , that is, the change in curvature per unit time; Represents the absolute value of the rate of change of curvature; This is the adjustment coefficient, in units of... In this system, it is typically set to 0.5. The window size is usually limited to the range of [24, 168] hours to balance short-term accuracy and long-term forecasting ability.

[0153] The multi-branch generation unit 64 generates multiple predicted trajectory branches representing different healing scenarios. Since a single trajectory prediction cannot reflect the uncertainty of the healing process, this unit generates multiple possible future trajectories by perturbing the current state and historical trajectories.

[0154] ,

[0155] in: For the first Each prediction branch in time The state vector has dimension . Predict the trajectory as a baseline in time The state vector has dimension . The disturbance term has a dimension of It follows a multivariate normal distribution. The time-dependent covariance matrix has a size of The number of prediction branches increases with the forecast period, reflecting the growth in forecast uncertainty. In this system, 5-7 prediction branches are typically generated to cover the main possible scenarios without being too scattered.

[0156] like Figure 8 As shown, the risk assessment module 70 includes a probability distribution calculation unit 71, a risk classification unit 72, a threshold judgment unit 73, and an early warning generation unit 74.

[0157] The probability distribution calculation unit 71 assigns probability weights to multiple predicted trajectory branches. Different predicted branches have different probabilities, and appropriate weight allocation is needed to reflect these probabilities. This unit calculates the probability weight of each branch based on historical data similarity and current trends:

[0158] ,

[0159] in: For the first The probability weights of each branch are dimensionless and range from [0,1]. This is a dimensionless distance metric between this branch and historically similar cases. This is a dimensionless temperature parameter that controls the sharpness of the weight distribution. The larger the value, the more concentrated the distribution. In this system, it is usually set to 2.0. Total number of branches; Indicates all Summing each branch; This represents the natural exponential function. The probability weights satisfy... This constitutes a complete probability distribution.

[0160] Risk grading unit 72 grades the healing status based on the distribution characteristics of the predicted trajectory. This unit clusters the predicted branches into three categories: normal healing, delayed healing, and abnormal healing, and calculates the probability of each category:

[0161] ,

[0162] ,

[0163] ,

[0164] in: , and These represent the probabilities of normal healing, delayed healing, and abnormal healing, respectively. They are dimensionless and range from [0,1]. , and These are the branch sets corresponding to the respective categories; This represents the summation over all branches in the normal healing category, and the others are calculated similarly. Based on these probability values, the system calculates the comprehensive risk index:

[0165] ,

[0166] in: The overall risk index is dimensionless and ranges from [0,1]; 0.2 and 0.8 are the weighting coefficients for delayed healing and abnormal healing, respectively, and are dimensionless. Risk Index The range is [0,1], and the risk is divided into low ( <0.15), medium (0.15≤ <0.4) and high ( ≥0.4) Level 3.

[0167] The threshold judgment unit 73 compares the risk level with a preset risk threshold. This unit sets corresponding warning trigger thresholds based on different risk levels, and triggers a warning of the corresponding level when the risk index exceeds the threshold. In this system, low risk does not require a warning, medium risk triggers a warning of attention, and high risk triggers an emergency warning.

[0168] The early warning generation unit 74 generates a graded early warning signal when the risk level exceeds a preset risk threshold. The early warning signal includes information such as the risk level, major risk factors, and expected evolution time, providing a basis for clinical decision-making. The early warning signal is transmitted to the decision support module via a standard interface for further processing into clinical decision recommendations.

[0169] like Figure 9 As shown, the decision support module 80 includes an early warning receiving unit 81, a suggestion generation unit 82, a visualization unit 83, and a notification distribution unit 84.

[0170] The early warning receiving unit 81 receives and parses the abnormal healing early warning signal. This unit analyzes the early warning signal from the risk assessment module, extracts key information such as risk level and main risk factors, and prepares for subsequent decision support.

[0171] The suggestion generation unit 82 generates intervention recommendations based on the type and risk level of healing abnormalities. This unit contains a decision rule base built upon clinical guidelines and expert knowledge, enabling it to generate targeted recommendations based on specific circumstances. For example, for a high-risk warning caused by abnormal blood oxygen saturation, the system may suggest increasing monitoring frequency and considering interventions such as antibiotic prophylaxis and nutritional support.

[0172] Visualization Unit 83 presents healing trajectory prediction and risk distribution graphically. This unit generates a visualization interface containing: the healing status manifold and current location, historical healing trajectories, predicted trajectory branches and their probability distribution, risk level indicators, and trend graphs of key physiological parameters. The interface is interactive, allowing physicians to select different predicted branches to view corresponding healing outcomes and intervention recommendations.

[0173] The notification distribution unit 84 pushes clinical decision support information to healthcare workers' terminals. This unit supports multiple notification methods, including workstation interface prompts, mobile application push notifications, and SMS notifications. Notification priorities are set according to risk levels; in high-risk situations, multi-channel notifications are used to ensure timely intervention.

[0174] The data storage module 90 connects to multiple modules of the system to store various types of data generated during system operation, including raw image data, extracted features, healing manifold representations, and prediction results. The data employs a hierarchical storage strategy: raw images are losslessly compressed and stored for 90 days, while feature data and analysis results are stored long-term to support system optimization and clinical research.

[0175] The knowledge update module 100 is connected to the data storage module 90 and is responsible for the regular updates of the system's knowledge base. For example... Figure 10 As shown, the update process includes three main stages: baseline template update, distance metric optimization, and prediction model parameter adjustment. Baseline template update optimizes the curvature template for normal healing by integrating newly added clinical cases; distance metric optimization adjusts the Riemannian metric tensor on the manifold based on actual healing results; and prediction model parameter adjustment optimizes the prediction algorithm parameters based on a comparison between predicted and actual results. The system is configured to perform a knowledge update every 100 new cases or every quarter to ensure continuous adaptation to new data characteristics.

[0176] The real-time monitoring and image analysis system for gastrointestinal anastomosis healing of this invention has shown significant effects in clinical applications. In a validation study involving 200 patients undergoing gastrointestinal surgery, the system achieved an accuracy rate of 87.5% in predicting abnormal anastomosis healing, an improvement of approximately 30% compared to traditional monitoring methods. The system can predict potential complications 48-96 hours in advance, providing valuable time for clinical intervention. The incidence of anastomosis-related complications decreased by 32.6%, the average hospital stay was shortened by 2.8 days, and the reoperation rate decreased by 41.3%.

[0177] Furthermore, the system's scalable design allows it to be easily integrated with existing hospital information systems, supporting multi-terminal access and remote consultations, further enhancing the system's practical value.

[0178] Through the innovative application of differential geometry theory, this invention achieves a technological leap from static monitoring to dynamic prediction, providing more accurate and forward-looking healing monitoring for patients after gastrointestinal surgery, and significantly improving clinical decision-making and patient prognosis.

[0179] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A real-time monitoring and image analysis system for gastrointestinal anastomosis healing, characterized in that, include: The image acquisition module is used to acquire continuous image sequences of the gastrointestinal anastomosis region; The preprocessing module, connected to the image acquisition module, is used to standardize the continuous image sequence to generate standardized image data. The feature extraction module, connected to the preprocessing module, is used to extract blood oxygen saturation distribution, tissue morphology features, and blood vessel distribution features from the standardized image data. The manifold construction module, connected to the feature extraction module, is used for: Construct a multidimensional feature space that includes the blood oxygen saturation distribution, the tissue morphology features, and the blood vessel distribution features; Establish nearest neighbor relationships between data points in the multidimensional feature space; A low-dimensional manifold representation of the healing state is constructed based on the nearest neighbor relationship; The curvature analysis module, connected to the manifold construction module, is used for: Calculate the multi-scale curvature features in the low-dimensional manifold representation; Abnormal states during the healing process are identified based on the aforementioned multi-scale curvature features; The geodesic prediction module, connected to the manifold construction module and the curvature analysis module, is used for: Construct the distance field on the healed state manifold; Calculate the optimal healing path from the current state to the target state; Generate multiple possible healing trajectory branches and their probability distributions; The risk assessment module, connected to the geodesic prediction module, is used for: The healing status is risk-classified based on the multiple possible healing trajectory branches; When the risk level exceeds a preset threshold, an abnormal healing warning signal is generated; The decision support module, connected to the risk assessment module, is used for: Receive the abnormal healing warning signal; Generate clinical decision support information that includes risk levels and intervention recommendations; Output the aforementioned clinical decision support information.

2. The system according to claim 1, characterized in that, The image acquisition module includes: A light source unit is used to provide dual-wavelength illumination covering the visible and near-infrared spectra; The imaging unit is used to acquire images of the gastrointestinal anastomosis area under the visible spectrum and the near-infrared spectrum; The acquisition and control unit is used to adaptively adjust the image acquisition frequency and imaging parameters according to the patient's healing stage.

3. The system according to claim 1, characterized in that, The preprocessing module includes: A noise filtering unit is used to remove random noise from the continuous image sequence; Lighting correction unit is used to correct uneven lighting conditions; A region identification unit is used to automatically identify and mark the gastrointestinal anastomosis region; The image registration unit is used to spatially align images acquired at different time points.

4. The system according to claim 1, characterized in that, The feature extraction module includes: The spectral analysis unit is used to calculate tissue oxygen saturation distribution maps based on dual-wavelength images; Edge extraction unit is used to detect the integrity of the anastomosis edges and tissue boundary features; The blood vessel recognition unit is used to extract and quantify the blood vessel distribution density and morphology in the anastomosis area; The texture analysis unit is used to extract texture feature parameters that characterize tissue structure.

5. The system according to claim 1, characterized in that, The manifold construction module includes: The similarity calculation unit is used to calculate the similarity matrix between data points in the multidimensional feature space; The nearest neighbor relationship building unit is used to determine the K nearest neighbors for each data point; An embedding optimization unit is used to solve low-dimensional embedding coordinates by preserving local linearity constraints. A dynamic modeling unit is used to construct a vector field describing the transition of healing states on the low-dimensional manifold representation.

6. The system according to claim 1, characterized in that, The curvature analysis module includes: A curvature calculation unit is used to estimate the discrete curvature on the low-dimensional manifold representation; Multi-scale filtering units are used to generate curvature features that characterize different time scales; An anomaly detection unit is used to identify regions of abnormal curvature based on a preset normal curvature template; The region mapping unit is used to map the curvature features to the original anastomosis region and generate differential feature maps.

7. The system according to claim 1, characterized in that, The geodesic prediction module includes: A distance field construction unit is used to construct a distance field based on Riemannian metrics on the low-dimensional manifold representation; The path optimization unit is used to find the optimal geodesic path from the current state to the target state; A window adjustment unit is used to dynamically adjust the size of the prediction time-domain window based on the rate of curvature change. A multi-branch generation unit is used to generate multiple predicted trajectory branches that represent different healing scenarios.

8. The system according to claim 1, characterized in that, The risk assessment module includes: A probability distribution calculation unit is used to assign probability weights to the multiple predicted trajectory branches; Risk grading unit, used to classify the healing status according to the distribution characteristics of the predicted trajectory; A threshold determination unit is used to compare the risk classification with a preset risk threshold. The early warning generation unit is used to generate a graded early warning signal when the risk level exceeds the preset risk threshold.

9. The system according to claim 1, characterized in that, The decision support module includes: The early warning receiving unit is used to receive and parse the abnormal healing early warning signal; A suggestion generation unit is used to generate intervention suggestions based on the type of healing abnormality and the risk level; Visualization units are used to visually represent healing trajectory predictions and risk distributions in a graphical manner; The notification distribution unit is used to push the clinical decision support information to the terminals of medical staff.

10. The system according to claim 1, characterized in that, Also includes: The data storage module is connected to the image acquisition module, the feature extraction module, and the risk assessment module, and is used to store raw image data, extract features, and assess results. The knowledge update module, connected to the data storage module, is used for: The curvature template for normal healing was updated based on new clinical data; Optimize the distance metric in the low-dimensional manifold representation; Adjust the probability distribution calculation parameters of the multiple predicted trajectory branches.

Images