A spinal cord injury analysis method and system based on medical images

Abstract

The application discloses a kind of spinal cord injury analysis method and system based on medical image, belong to image analysis technical field, obtain the MRI image of target object at multiple time points and carry out registration and normalization processing, construct time series image set;Spinal cord and injury area are extracted using spinal cord structure guided dynamic segmentation network;The time sequence deformation path of injury area is constructed based on morphological tracking algorithm;Extract high-order morphological features including curvature change rate, edge contour jump value and spinal cord compression rate;Features and neural function score are modeled in two domains, and a structure-function coupling model is trained;The model is used to predict the functional status of new patients, identify risk areas, and output rehabilitation recommendations. This method has strong predictive power, high accuracy, and significant clinical assistance value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image analysis, in particular to a spinal cord injury analysis method and system based on medical images. BACKGROUND

[0002] Due to its irreversibility and complexity, spinal cord injury has been a highly concerned problem in neurosurgery and rehabilitation medicine. The existing technology mainly uses static single MRI image for structure observation and interpretation, relies on expert subjective experience to identify lesion location and degree, lacks quantitative standard, and cannot reflect the biomechanical behavior of lesions in the time evolution process.

[0003] Some methods try to extract the spinal cord morphology through image segmentation and three-dimensional reconstruction, but still have the following limitations: first, the image data is limited to a certain time section, ignoring the deformation trajectory of the injury area at different time points; second, it cannot extract the structure-function change pattern implied in the image, and cannot provide accurate decision support for clinical intervention; third, there is no systematic method to integrate multi-time image information to jointly model the dynamic structure response and functional change trend of spinal cord tissue.

[0004] Therefore, there is an urgent need for a spinal cord injury analysis method that combines time series image analysis and structure-function coupling modeling, which can not only automatically extract multi-dimensional features of the injury area, but also capture the deformation evolution behavior of the tissue during the course of the disease, further realize the prediction of the patient's neural function recovery potential and intervention suggestions, and significantly improve the accuracy and timeliness of diagnosis. SUMMARY

[0005] The purpose of the present application is to provide a spinal cord injury analysis method and system based on medical images to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: a spinal cord injury analysis method based on medical images, comprising:

[0007] S100, acquiring a sequence of spinal MRI images of a target object at multiple different time points , t represents the total number of time points of image acquisition, registration and normalization preprocessing is performed, and a time series medical image set I is constructed;

[0008] S200, using a spinal cord structure guided dynamic segmentation network to extract spinal cord tissue regions and suspected injury regions from the image set I frame by frame, and obtaining an injury region sequence ;

[0009] S300, based on a morphology tracking algorithm, constructing a time series deformation path M(t) of the injury region sequence R, wherein the M(t) is used to describe the continuous evolution process of the injury morphology in the time dimension;

[0010] S400, extracting the damaged area at each time point The high-order morphological eigenvectors of , including local curvature change rate, edge contour jump value and spinal cord compression rate;

[0011] S500, collect clinical neurological function scores at corresponding time points , the morphological characteristics and functional score Perform dual-domain mapping and train to obtain the structure-function coupling model G;

[0012] S600, based on model G, performs predictive analysis on the MRI image sequences of new patients and outputs the future evolution trend of their spinal cord injury, potential high-risk areas, and rehabilitation recommendations.

[0013] Preferably, the construction of the time-sequential deformation path M(t) of the damaged area includes:

[0014] Perform non-rigid temporal registration on the damaged area sequences at different time points to align all areas to the coordinate reference system of the initial time point;

[0015] Extracting a set of morphological characteristic parameters of the damaged area of ​​each frame after registration, wherein the parameters include area, volume, boundary curvature, principal axis direction, center of gravity coordinates and geometric moments;

[0016] The Euclidean distance or Mahalanobis distance of morphological features between adjacent time points is used for regional matching to construct a logically coherent damage evolution path;

[0017] Based on the matching results, a continuous morphological change function M(t) is constructed. Polynomial fitting or spline interpolation is used to generate an evolution function f(t) that expresses the damage characteristics over time. Based on this function, the volume growth rate and boundary diffusion rate are calculated.

[0018] The changing trends in different stages are classified and labeled according to the set threshold, and spatially visualized in the image in the form of an evolution trajectory diagram.

[0019] Preferably, the method for extracting the local curvature change rate, edge contour jump value and spinal cord compression rate is:

[0020] Sample the boundary curve of the damaged area and obtain the equidistant point set P on the boundary. At each point Calculate its curvature , find the standard deviation of all curvature values, and use it as the curvature change rate indicator after standardization. The expression is: Where, represents the rate of change of local curvature at time t; is the mean value of all curvature values; m is the total number of boundary points;

[0021] Select the pixel-level contour line of the damage area boundary, and calculate the gray gradient between adjacent boundary points ; the jump value is calculated as the proportion of local mutation points or the change intensity, and the expression is: ; in the formula, is the edge contour jump value at time t; represents the image gray value of the boundary point ; represents the image gray value of the boundary point ;

[0022] Extract the spinal cord cross-sectional area contour; calculate the minimum diameter of the compression site and compare it with the average diameter of the healthy spinal cord ; the expression is: ; represents the spinal cord compression rate at time t.

[0023] Preferably, the construction of the structure-function coupling model G includes:

[0024] During image sequence acquisition, the neurological function scores of the patient at each time point are recorded synchronously ;

[0025] Align the image morphological feature vector at each time point with the corresponding function score on the time axis to form a set of training sample pairs ;

[0026] Based on the training samples, a mapping function is constructed ; wherein: G represents the structure-function mapping model; represents the morphological feature vector at time t; represents the corresponding function score; and ε represents the error term.

[0027] Preferably, the prediction analysis based on the model G includes:

[0028] Receive the MRI image sequence collected at multiple time points of a new patient, and extract the morphological feature vector of the damage area at each time point ;

[0029] Call the trained structure-function coupling model G, and perform forward inference on each feature vector to obtain a function prediction value ;

[0030] Based on the prediction sequence Fitting the functional evolution trend curve, combining M(t) to evaluate the damage area expansion speed and abnormal boundary change, identifying potential high-risk areas and outputting rehabilitation suggestions.

[0031] The application also provides a spinal cord injury analysis system based on medical images, comprising:

[0032] a time-series image preprocessing module, a spinal MRI image sequence of a target object at multiple different time points is acquired , t represents the total number of time points of image acquisition, registration and normalization preprocessing are performed, and a time-series medical image set I is constructed;

[0033] a structure-guided segmentation module, a dynamic segmentation network guided by spinal cord structure is used to extract spinal cord tissue regions and suspected damage regions frame by frame from the image set I, and a damage region sequence R is obtained; ;

[0034] a morphological evolution modeling module, based on a morphological tracking algorithm, a time-series morphological path M(t) of the damage region sequence R is constructed, and the M(t) is used to describe the continuous evolution process of the damage morphology in the time dimension;

[0035] a high-order morphological feature extraction module, a high-order morphological feature vector of each time point damage region is extracted , including local curvature change rate, edge contour jump value and spinal cord compression rate;

[0036] a coupling modeling module, clinical neurological function scores corresponding to the time points are collected , morphological features and function scores are mapped in two domains, and a structure-function coupling model G is trained and obtained;

[0037] a prediction analysis module, based on the model G, a predictive analysis is performed on the MRI image sequence of a new patient, and future evolution trend, potential high-risk area and rehabilitation suggestion of spinal cord injury of the new patient are output.

[0038] In the above technical solution, the application provides technical effects and advantages:

[0039] 1、The application can accurately extract the spinal cord damage area and reconstruct the morphological evolution path by introducing a dynamic segmentation network based on the structure of the spinal cord and a multi-time series image fusion mechanism. Combined with high-order structure features such as local curvature change rate, edge jump value and compression rate, dynamic tracking and quantitative analysis of the damage morphology are realized. Compared with the traditional single-frame image observation method, the application can continuously present the lesion evolution trend, improve the visualization and quantification understanding of the damage evolution process.

[0040] 2、The application realizes the prediction and evaluation of the functional state of the patient by constructing a structure-function coupling model, mapping the image features and the clinical neurological function score in two domains, and intelligently prompting the potential high-risk area in the future. The method not only significantly improves the accuracy of the score prediction (the error is controlled within ±1), but also outputs personalized rehabilitation suggestions and auxiliary diagnosis reports, has the advantages of strong forward-looking, high intelligent degree, and outstanding clinical practical value, and is suitable for multiple medical application scenarios such as neural image analysis and intelligent rehabilitation evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0042] Figure 1 The method mind map of the present application.

[0043] Figure 2 The system module mind map of the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0045] Embodiment 1, please refer to Figure 1 The spinal cord injury analysis method based on medical images described in this embodiment comprises:

[0046] S100, acquiring a sequence of spinal MRI images of a target object at multiple different time points , t represents the total number of time points of image acquisition, registration and normalization preprocessing is performed, and a time sequence medical image set I is constructed;

[0047] S200, using a spinal cord structure guided dynamic segmentation network to extract a spinal cord tissue region and a suspected damage region from the image set I frame by frame, and obtaining a damage region sequence ;

[0048] S300, based on a morphological tracking algorithm, a time sequence morphological change path M(t) of the damage region sequence R is constructed, and the M(t) is used to describe the continuous evolution process of the damage morphology in the time dimension;

[0049] S400, extracting a high-order morphological feature vector of the injury area at each time point , including a local curvature change rate, an edge contour jump value, and a spinal cord compression rate;

[0050] S500, collecting a clinical neurological function score at the corresponding time point , performing two-domain mapping of the morphological features and the function score , and training to obtain a structure-function coupling model G;

[0051] S600, based on the model G, performing predictive analysis on an MRI image sequence of a new patient, and outputting a future evolution trend of spinal cord injury, a potential high-risk area, and rehabilitation recommendations.

[0052] In the embodiment of the present application, step S100 includes acquiring and preprocessing medical image data of the target object to construct a standardized image set for time-series injury analysis.

[0053] First, the spinal MRI image sequence of the target object collected at different time points is acquired, and the time points include but are not limited to the initial visit, the middle of treatment, the later stage of treatment, or the review time point. The image sequence can be represented as , wherein t represents the total number of time points of image acquisition, represents the MRI image data corresponding to the tthtime point.

[0054] The MRI image can include axial, sagittal, or coronal images, and a T2 weighted sequence is preferably used to enhance the contrast of spinal cord soft tissue and injury areas. After image acquisition is completed, the following preprocessing operations are performed on each image:

[0055] Image registration (Registration): A rigid or non-rigid image registration algorithm is used to perform spatial alignment processing on the image sequence. The first time point image is taken as a reference template, and the subsequent time point images are mapped to a unified coordinate system to eliminate spatial deviations caused by patient posture, scanning parameters, or machine errors, and ensure the structural consistency between time-series images.

[0056] Image normalization (Normalization): Intensity normalization processing is performed on the image gray value to make the images at different time points have uniform contrast and brightness standards. Methods such as Z-score normalization or histogram matching are commonly used to enhance the stability of subsequent feature extraction.

[0057] ​Denoising & Resampling: Apply isotropic filtering algorithm or convolution smoothing operation to reduce MRI noise, and standardize the image voxel size to ensure the consistency of spatial resolution, providing the basis for three-dimensional modeling and morphological analysis.

[0058] Constructing the time-series image set I: The image sequence after registration and normalization processing Organize the time-series medical image set I in the order of acquisition time, where each image data is in a unified spatial framework and intensity standard, with time dimension comparability.

[0059] Through the above steps, the image set I constructed provides a standardized and traceable data basis for subsequent damage area extraction, time-series morphological modeling and structure-function coupling analysis. This step significantly improves the accuracy and stability of time-series analysis, especially suitable for damage evolution dynamic monitoring and intervention strategy development.

[0060] In the present application, step S200 aims to automatically extract the spinal cord tissue region and the suspected damage region within it from the time-series medical image set I constructed in step S100 , forming a structured damage region sequence for subsequent dynamic modeling and evaluation analysis.

[0061] Specifically, the extraction process uses a Spinal-Structure-Guided Dynamic Segmentation Network (SSD-Net) based on the anatomical structure features of the spinal cord. This network comprehensively utilizes the spatial context information, texture gradient features and standard anatomical model of the image to realize frame-by-frame segmentation of the target region in the time-series image. The process includes the following steps:

[0062] Network input and structure guidance: input each frame of MRI image in the image set to the dynamic segmentation network, and introduce the spinal cord structure template graph Sref as auxiliary feature map to improve the model's recognition ability of spinal cord tissue boundary and damage features. The spinal cord structure template graph Sref is an anatomical reference atlas constructed based on multiple healthy volunteer MRI image samples (from public datasets or historical clinical data) after registration and averaging processing. Its essence is a standard morphological template graph of the spinal cord region. This graph is used to provide prior structural information during network training, thereby improving the segmentation stability of the model in the case of fuzzy boundaries or signal noise.

[0063] The specific construction method is as follows:

[0064] Select several groups of healthy sample MRI images;

[0065] After registration, voxel-level averaging is performed to form a spinal cord morphology map in a unified coordinate system;

[0066] The template map is embedded in the auxiliary input channel of the network in grayscale form to guide structural perception feature extraction.

[0067] Multi-scale feature extraction: the segmentation network adopts a multi-scale encoder structure to extract local texture features and global spatial information in the image, while fusing long and short-range dependencies to address the challenges of blurred boundaries and diverse injuries in the spinal cord region.

[0068] Spinal cord region segmentation: in the first stage, the network identifies the overall spinal cord tissue region and outputs a spinal cord mask map for spatial restriction.

[0069] Injury region identification: based on the spinal cord mask, the network further detects regions of abnormal signal intensity, morphological mutation or texture disorder in the tissue to generate an injury region candidate map and optimizes the boundary through post-processing operations such as morphological filtering and small region rejection.

[0070] Sequence output: repeat the above operations for each frame in the image set I , extract the corresponding spinal cord region and injury region , and finally obtain a set of time series structured output: injury region sequence .

[0071] The dynamic segmentation method has stronger temporal consistency and spatial robustness compared to traditional static image segmentation methods, and can automatically adapt to changes in tissue morphology caused by disease progression in images at different time points, ensuring the continuity and accuracy of injury region extraction results in time series.

[0072] Through this step, the obtained injury region sequence not only retains the spatial geometric information in the original image, but also has time series indexing characteristics, laying a key data foundation for subsequent injury deformation modeling and functional coupling analysis.

[0073] In the present application, step S300 aims to perform time series analysis on the injury region sequence obtained in step S200, and construct a continuous evolution path M(t) of the injury region in the time dimension based on a morphological tracking algorithm to quantitatively describe the deformation behavior of spinal cord injury as the disease progresses.

[0074] The implementation process of the morphological tracking algorithm includes the following sub-steps:

[0075] S301, to ensure that the injury regions at different time points have a consistent spatial reference system, first perform spatial registration on the injury region sequence Temporal registration is performed. Non-rigid registration algorithms based on B-spline or affine transformation are used to align all to the coordinate system of the initial time point , thus eliminating spatial shifts caused by patient position differences or scanning errors.

[0076] In each registered frame of the lesion region S302, a set of morphological feature parameters is extracted , including but not limited to: area and volume (A / V); principal axis length and orientation; boundary curvature distribution; local shape complexity (such as Fractal Dimension or Shape Irregularity); centroid coordinates and geometric moments, etc. The above features serve as basic data reflecting changes in lesion morphology, which will be used to construct a temporal tracking model.

[0077] S303, using the Euclidean distance or Mahalanobis distance between adjacent time points to calculate the similarity score of morphological features

[0078] and one-to-one matching to ensure the logical coherence of the temporal data chain. In the case of multiple regions or complex branches, the Max-Flow Matching algorithm can be introduced to establish the globally optimal matching path.

[0079] S304, based on the above matching results, a continuous change function M(t) of lesion morphology along the time axis is constructed, which can be modeled in the following two forms:

[0080] Discrete time series form: M(t) = , used to analyze the trend of feature increase or decrease;

[0081] Continuous change curve form: using spline interpolation, polynomial fitting or time series regression to continuously model the change of features over time, generating a morphological evolution function such as , used to predict future trends. f(t) represents the lesion region morphological feature value at time point t, f(t) represents the evolution trend function of a specific feature (such as volume, boundary length, compression rate, etc.) over time; is a constant term, representing the initial feature value of the lesion morphology at the initial time t=0 or t=1; is a linear term coefficient, representing the linear growth rate or downward trend of the lesion morphology feature over time; ​is the coefficient of quadratic term, describing the changing acceleration or curvature, for reflecting the nonlinear dynamics in the process of injury aggravation or alleviation; in addition, the deformation path can also be visually represented as an "evolution trajectory map" in the spinal cord image, for intuitively showing the time sequence change of the injury boundary.

[0082] S305, based on the deformation path M(t), further calculate the morphological change rate (such as volume growth rate, center of gravity drift speed, edge diffusion rate, etc.), and according to the set threshold, classify and label the change trend of different stages (such as "stable period", "acute expansion period", "potential deterioration period"), provide timely risk warning support for clinical decision-making.

[0083] Through the above steps, the morphological tracking algorithm not only realizes the accurate tracking and continuous modeling of the spinal cord injury area in the time dimension, but also improves the quantifiable understanding of the dynamic evolution process of the disease, which is beneficial to the diagnosis, prediction and development of individualized intervention strategies.

[0084] In an embodiment of the present application, step S400 aims to extract high-order morphological features of the injury area for each time point t , for quantifying the geometric complexity and biomechanical response of structural changes. Mainly include the following three key features:

[0085] Local curvature change rate, extraction method: sampling the boundary curve of the injury area, obtaining the equidistant point set P on the boundary; represents the i-th sampling point on the boundary of the injury area, usually a coordinate point in two-dimensional or three-dimensional space; i∈{1,2,…,m} represents the index of the sampling point, which represents the point numbered in order along the boundary; m is the total number of points sampled on the boundary, that is, the boundary sampling resolution. At each point , calculate its curvature , which can be calculated by circle fitting or differential method; find the standard deviation or the difference between the maximum and minimum of all curvature values; after normalization, it is used as the curvature change rate index, the expression is: ; in the formula, represents the local curvature change rate at time t; is the mean of all curvature values; m is the total number of boundary points.

[0086] Edge contour jump value, extraction method: selecting the pixel-level contour line of the boundary of the injury area; statistics of the gray gradient between adjacent boundary points; calculate the proportion of local abrupt points or change intensity as the jump value, the expression (gradient method) is: ; in the formula, is the edge contour jump value at time t; represents the boundary point the image gray value of the boundary point; the image gray value of the boundary point m is the total number of boundary points;

[0087] The spinal cord compression rate is extracted by extracting the contour of the cross-sectional area of the spinal cord, and calculating the minimum diameter of the compression site Compared with the average diameter of the healthy spinal cord The expression is: ; The spinal cord compression rate at the t-th moment is represented by Approaching 1 indicates severe compression, and approaching 0 indicates no compression.

[0088] By jointly extracting the above features, a high-dimensional feature vector is formed, which provides quantitative input for subsequent structure-function coupling modeling. The index comprehensively considers the boundary geometric complexity, tissue continuity and mechanical compression degree, and is significantly better than the static analysis method based on area or signal intensity in the prior art.

[0089] In an embodiment of the present application, step S500 aims to model the correlation between the spinal cord injury morphological features in the image domain and the neurological function scores in the clinical domain, and construct a structure-function coupling model G for revealing the internal mapping relationship between the image structure change and the degree of dysfunction.

[0090] This step specifically includes the following sub-steps:

[0091] S501, while acquiring the image sequence , record the neurological function score value of the patient at each corresponding time point t . The score can be selected from the internationally accepted standardized evaluation scale, such as: ASIA score (American Spinal Injury Association): evaluating motor function (0-100) and sensory function; Frankel classification: five-level classification, used for rough classification of injury severity; JOA score (Japanese Orthopaedic Association): for cervical spinal cord function evaluation; other quantitative neurological function score scales. Each time point t corresponds to a score value, forming a time series score set .

[0092] S502, the morphological feature vector sequence extracted in step S400 is corresponded with the score sequence point by point on the time axis to form a training data pair set: ; wherein each pair of samples describes the structure state and clinical function performance at a certain time point, providing a two-domain input for subsequent modeling.

[0093] S503, based on the training data set , a structure-function coupling model G is constructed for learning the mapping relationship between high-order morphological features and neurological function performance. The optional model structure includes but is not limited to: multivariate nonlinear regression model; support vector regression (SVR); multilayer perception (MLP) or other neural network structure; radial basis function network (RBF) and the like.

[0094] Mapping function form: ; Wherein: G represents a structure-function mapping model; represents the morphological feature vector at the t time; represents the corresponding function score; ε represents an error term for describing the model fitting residual.

[0095] S504, using cross-validation, mean square error (MSE) or mean absolute error (MAE) and the like evaluation indexes, the parameter optimization and generalization ability test of model G are carried out, and the prediction accuracy and robustness are ensured.

[0096] Through the structure-function coupling model, the following technical effects can be realized:

[0097] Multidimensional structure feature fusion: mapping multiple high-order indexes (such as curvature, jump, compression, etc.) in the image domain to the function performance, avoiding the misguidance of a single index;

[0098] Personalized function prediction: can be used to evaluate the current disease state and future function evolution trend of the patient;

[0099] Auxiliary treatment evaluation: assisting doctors to judge the effectiveness of intervention measures or the improvement space of prognosis.

[0100] The coupling modeling process is significantly different from the traditional judgment method which only depends on the subjective experience of doctors, and provides an intelligent solution for the precise evaluation and decision support of spinal cord injury.

[0101] In the embodiment of the application, step S600 calls the aforementioned trained structure-function coupling model G for the MRI image sequence input by the new patient, realizes the prediction analysis of the future evolution trend of spinal cord injury, identifies the potential high-risk area and outputs the personalized rehabilitation intervention suggestion, and provides decision support for the clinic.

[0102] The step includes:

[0103] S601, receiving the spinal MRI image sequence collected by the new patient at different time points , according to the steps of S100~S400, the following processing is carried out:

[0104] Image registration and normalization; dynamic segmentation network extracts the injury area ;

[0105] Calculate the corresponding morphological feature vector .

[0106] This process ensures that the feature extraction method is consistent with the training model, avoiding input domain shift.

[0107] S602, call the trained model G, and input the feature vector of each time point t Forward inference to get the corresponding neural function prediction value : ; If the input is only the initial frame image (such as t=1), the future feature trend can be predicted in combination with the morphological evolution model M(t), and then input into G to realize long-term function estimation.

[0108] S603, based on the time series prediction results output by the model G , trend fitting and change rate analysis are performed to generate the spinal function evolution trend curve. At the same time, combined with the deformation path in the structure model M(t), the expansion direction, rate and morphological abnormality degree of the damaged area are evaluated, and the high-risk area Rrisk that may appear in the future time period is identified.

[0109] S604, according to the prediction results and risk assessment, output the following individualized rehabilitation suggestion information:

[0110] Function level warning: if the predicted score decreases by more than the set threshold, it is automatically marked as a function deterioration risk;

[0111] Key time node prompt: predict the injury aggravation time window to provide timing guidance for clinical intervention;

[0112] Target area labeling: highlight the potential high-risk area in the image to assist image navigation;

[0113] Individualized rehabilitation plan suggestion: based on the structure risk degree and function prediction, recommend intensive treatment or regular review cycle.

[0114] The above suggestions can be displayed in the form of visual reports, including trend curve graph, image heat map and score change table, etc., to facilitate doctors to quickly understand and make decisions.

[0115] In summary, this step realizes individualized structure-function prediction and risk prompt for new patients through model G, with the following innovations and practicality:

[0116] Generalize prediction using historical structure-function coupling patterns to improve forward-looking diagnosis;

[0117] Combine morphological evolution trajectory to realize quantitative and visualized future structure abnormality trend;

[0118] Provide the clinician with quantitative and actionable rehabilitation intervention recommendations, significantly improving the efficiency of diagnosis and treatment.

[0119] The system constitutes an end-to-end intelligent analysis framework, from image acquisition to function prediction to recommendation output, automatically completing the continuous assessment closed loop of spinal cord injury.

[0120] Example 2, to verify the effectiveness of the spinal cord injury analysis method described in the application, real MRI case data from a hospital was selected and the experimental process was constructed as follows:

[0121] Ten patients diagnosed with spinal cord injury were selected, and their spinal T2 weighted MRI images at the pre-treatment, mid-treatment and review stages were collected, with 3 time points for each patient, a total of 30 image data. The image size is 512x512 pixels, the layer thickness is 5mm, and the interval is 1mm. The JOA score and ASIA score of each patient at the three time points were recorded as a reference for the functional state.

[0122] Registration and normalization processing: B-spline registration is performed using the Elastix tool to align the T2, T3 time point images to the T1 image coordinate system; the image gray value is standardized by Z-score.

[0123] Dynamic segmentation: based on the improved U-Net (introducing attention mechanism and multi-scale residual block), the spinal cord and injury area of each image are segmented, and the average Dice coefficient reaches 0.912.

[0124] Feature extraction: the following features are extracted in each frame of the injury area:

[0125] Regional volume change rate (unit mm³);

[0126] Edge curvature change standard deviation (average about 0.32, abnormal threshold 0.5);

[0127] Spinal cord compression rate (maximum value up to 0.36, mild standard <0.15).

[0128] The morphological feature vector of each time point And its corresponding JOA score Paired to form a training set, a total of 30 samples.

[0129] Model structure: three-layer MLP neural network (input dimension = 3, hidden layer 64-32, output = 1), activation function ReLU;

[0130] Training strategy: 80% training, 20% validation, Adam optimizer, learning rate 0.001, batch size 4;

[0131] Loss function: Mean Squared Error (MSE)

[0132] Prediction result: Mean Absolute Error (MAE) = 0.84, R² = 0.91.

[0133] For a patient (No. P09) who did not participate in training, the image was input into the system:

[0134] The system automatically extracted morphological features at three time points;

[0135] The model predicted the JOA score as: [8.6, 10.2, 12.1], which is very close to the actual score [8, 10, 13];

[0136] The system marked the edge expansion segment between T2 and T3 in the C5-6 region, with a risk prompt of "moderate evolution";

[0137] System suggestion: review within three weeks + physical decompression intervention.

[0138] Comparison of traditional method effect

[0139]

[0140] Beneficial effect summary: significantly improve the prediction accuracy of the score (error controlled within ±1); the system can identify potential structural evolution trend in advance, with forward-looking and intervention value; automatically generate auxiliary report, save doctor's judgment time, improve clinical efficiency.

[0141] It should be noted that the purpose of this experiment is to preliminarily verify the feasibility of the structure-function coupling model, not to finally define the performance of the model. In order to ensure the stability of training under limited samples, the following technical measures are taken during the model training process:

[0142] The input dimension of the MLP network used is only 3 (corresponding to three morphological features), and the network size is small, avoiding the risk of overfitting caused by high-dimensional input;

[0143] 80 / 20 training and validation division is adopted, and early stopping mechanism and mean squared error evaluation are introduced to prevent overfitting;

[0144] The extracted volume change rate, curvature change and compression rate are all structural indicators with strong significance and clear association with neural function, which helps to improve the prediction efficiency;

[0145] In the validation set, the mean absolute error (MAE) is 0.84, and R² is 0.91, the prediction result is highly consistent with the actual score, showing good mapping ability.

[0146] In addition, the present application is not limited to the initial sample size, and subsequent clinical data collection can be combined to expand the training set to further enhance the generalization ability.

[0147] Embodiment 3, please refer to Figure 2 As shown in the figure, the spinal cord injury analysis system based on medical images described in this embodiment comprises:

[0148] The time sequence image preprocessing module acquires a sequence of spinal MRI images of the target object at multiple different time points , t represents the total number of time points of image acquisition, registration and normalization preprocessing are performed to construct a time sequence medical image set I;

[0149] The structure-guided segmentation module uses a dynamic segmentation network guided by the spinal cord structure to extract the spinal cord tissue region and the suspected damage region from the image set I frame by frame, and obtains a damage region sequence R; ;

[0150] The morphological evolution modeling module constructs a time sequence morphological deformation path M(t) of the damage region sequence R based on a morphological tracking algorithm, and the M(t) is used to describe the continuous evolution process of the damage morphology in the time dimension;

[0151] The high-order morphological feature extraction module extracts a high-order morphological feature vector of the damage region at each time point , including the local curvature change rate, the edge contour jump value and the spinal cord compression rate;

[0152] The coupling modeling module acquires the clinical neurological function score at the corresponding time point , and performs two-domain mapping of the morphological features and the function score to train a structure-function coupling model G;

[0153] The prediction analysis module performs predictive analysis on the MRI image sequence of a new patient based on the model G, and outputs the future evolution trend of the spinal cord injury, the potential high-risk area and the rehabilitation suggestion.

[0154] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.​

Claims

1. A method for analyzing spinal cord injury based on medical images, characterized by: include: S100: Acquire a spinal MRI image sequence of the target object at multiple different time points , t represents the total number of time points of image acquisition, and registration and normalization preprocessing are performed to construct the time series medical image set I; S200, using the dynamic segmentation network guided by the spinal cord structure, extract the spinal cord tissue area and the suspected injury area from the image set I frame by frame to obtain the injury area sequence ; S300, constructing a temporal deformation path M(t) of the damage region sequence R based on a morphology tracking algorithm, wherein M(t) is used to describe the continuous evolution process of the damage morphology in the time dimension; S400, extracting the damaged area at each time point The high-order morphological eigenvectors of , including local curvature change rate, edge contour jump value and spinal cord compression rate; S500, collect clinical neurological function scores at corresponding time points , the morphological characteristics and functional score Perform dual-domain mapping and train to obtain the structure-function coupling model G; S600, based on model G, performs predictive analysis on the MRI image sequences of new patients and outputs the future evolution trend of their spinal cord injury, potential high-risk areas, and rehabilitation recommendations.

2. The method for analyzing spinal cord injury based on medical images according to claim 1, characterized in that: The temporal deformation path M(t) of constructing the damage area includes: Perform non-rigid temporal registration on the damaged area sequences at different time points to align all areas to the coordinate reference system of the initial time point; Extracting a set of morphological characteristic parameters of the damaged area of ​​each frame after registration, wherein the parameters include area, volume, boundary curvature, principal axis direction, center of gravity coordinates and geometric moments; The Euclidean distance or Mahalanobis distance of morphological features between adjacent time points is used for regional matching to construct a logically coherent damage evolution path; Based on the matching results, a continuous morphological change function M(t) is constructed. Polynomial fitting or spline interpolation is used to generate an evolution function f(t) that expresses the damage characteristics over time. Based on this function, the volume growth rate and boundary diffusion rate are calculated. The changing trends in different stages are classified and labeled according to the set threshold, and spatially visualized in the image in the form of an evolution trajectory diagram.

3. The method for analyzing spinal cord injury based on medical images according to claim 1, characterized in that: The extraction method of the local curvature change rate, edge contour jump value and spinal cord compression rate is: Sample the boundary curve of the damaged area and obtain the equidistant point set P on the boundary. At each point Calculate its curvature , find the standard deviation of all curvature values, and use it as the curvature change rate indicator after standardization. The expression is: Where, represents the rate of change of local curvature at time t; is the mean of all curvature values; m is the total number of boundary points; Select the pixel-level contour line of the damaged area boundary and calculate the grayscale gradient between adjacent boundary points. Conduct statistics; Calculate the local mutation point ratio or change intensity as the jump value, the expression is: Where, is the edge contour jump value at the tth moment; Represents boundary points The grayscale value of the image; Represents boundary points The grayscale value of the image; Extract the contour of the spinal cord cross-section area; calculate the minimum diameter of the compressed area Average diameter of healthy spinal cord Comparison, expression: ; represents the spinal cord compression rate at time t.

4. The method for analyzing spinal cord injury based on medical images according to claim 1, characterized in that: The construction of the structure-function coupling model G includes: During the image sequence acquisition process, the patient's neurological function score at each time point was recorded synchronously ; The image morphological feature vector at each time point Corresponding functional scores Align on the time axis to form a training sample pair set ; Constructing a mapping function based on training samples ; Where: G represents the structure-function mapping model; represents the morphological feature vector at the tth moment; represents the corresponding functional score; ε represents the error term.

5. The method for analyzing spinal cord injury based on medical images according to claim 1, characterized in that: The prediction analysis based on model G includes: Receive MRI image sequences collected from new patients at multiple time points and extract the morphological feature vectors of the damaged area at each time point ; Call the trained structure-function coupling model G, for each feature vector Perform forward reasoning to obtain function prediction values ; Based on the prediction sequence The functional evolution trend curve is fitted and combined with M(t) to evaluate the expansion speed of the damaged area and abnormal boundary changes, identify potential high-risk areas and output rehabilitation suggestions.

6. A medical image-based spinal cord injury analysis system, configured to implement the medical image-based spinal cord injury analysis method according to any one of claims 1 to 5, characterized in that: include: Time series image preprocessing module, obtains the target object's spinal MRI image sequence at multiple different time points , t represents the total number of time points of image acquisition, and registration and normalization preprocessing are performed to construct the time series medical image set I; The structure-guided segmentation module uses a dynamic segmentation network guided by the spinal cord structure to extract the spinal cord tissue area and the suspected injury area from the image set I frame by frame to obtain the injury area sequence ; The morphological evolution modeling module constructs the temporal deformation path M(t) of the damage region sequence R based on the morphological tracking algorithm. The M(t) is used to describe the continuous evolution of the damage morphology in the time dimension. High-order morphological feature extraction module, extracting the damage area at each time point The high-order morphological eigenvectors of , including local curvature change rate, edge contour jump value and spinal cord compression rate; Coupled modeling module, collecting clinical neurological function scores at corresponding time points , the morphological characteristics and functional score Perform dual-domain mapping and train to obtain the structure-function coupling model G; The predictive analysis module, based on model G, performs predictive analysis on the MRI image sequences of new patients and outputs the future evolution trend of their spinal cord injury, potential high-risk areas and rehabilitation suggestions.

Images