Method and system for identifying the freshness of occult fracture based on time-series imageomics

Abstract

This invention belongs to the technical field of medical image analysis, specifically a method and system for identifying the freshness of occult fractures based on temporal radiomics. The method includes: acquiring medical image data from multiple time points for the same examined site and establishing a temporal image sequence; determining bone structure regions using an improved CNN-ViT hybrid network and extracting multimodal features; determining a temporal feature set using a cross-modal contrastive learning mechanism; converting the feature evolution sequence into a parameter sequence using a biomechanical parameter mapping mechanism; performing biological consistency fitting on the feature evolution sequence; and outputting the corresponding fracture freshness level and freshness probability interval based on the results of the biological consistency fitting through a Bayesian inference mechanism. This invention achieves refined identification of areas with minute structural damage by employing an improved CNN-ViT hybrid network for spatial alignment and precise segmentation of bone structure regions.

Description

Technical Field

[0001] This invention belongs to the technical field of medical image analysis, specifically a method and system for identifying the freshness of occult fractures based on time-series radiomics. Background Technology

[0002] Currently, the diagnosis of occult fractures (such as stress fractures, microfractures, and fractures related to bone contusion) relies mainly on X-rays, CT, and MRI. MRI determines the fracture phase by observing changes in bone marrow edema signals, while CT assists in the assessment by observing changes in the fine structure of the bone cortex. However, in clinical practice, the assessment of the freshness of a fracture (acute phase, subacute phase, old phase) mainly depends on the doctor's experience and the characteristics of a single imaging sample, lacking quantitative and objective time-series assessment methods.

[0003] However, existing technologies are mainly limited in accurately determining the freshness of occult fractures due to the lack of multi-temporal image dynamic feature fusion and quantitative evolution modeling mechanisms. This makes it difficult to balance the sensitivity of early identification of small fractures with the accuracy of fracture time staging in practical applications, thus affecting the reliability of clinical decisions (such as whether to restrict weight-bearing or whether to intervene). Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for identifying the freshness of occult fractures based on temporal radiomics, thereby solving the problems of lacking a multi-temporal image dynamic feature fusion and quantitative evolution modeling mechanism in the prior art.

[0005] A method for identifying the freshness of occult fractures based on temporal radiomics includes the following steps:

[0006] Collect medical imaging data and determine the time-series feature set, specifically:

[0007] For the same examined site, medical image data at multiple time points were acquired and a time-series image sequence was established. The time-series image sequence was enhanced. At the same time, the bone structure region was determined by an improved CNN-ViT hybrid network. Based on the determined bone structure region, the suspected fracture region was located by a multi-scale feature pyramid network. Multimodal features were extracted from the suspected fracture region and the time-series feature set was determined by combining a cross-modal contrastive learning mechanism.

[0008] The freshness of occult fractures is identified based on temporal feature sets, specifically as follows:

[0009] A feature evolution sequence is constructed based on a time-series feature set, and the image features are converted into a parameter sequence using a biomechanical parameter mapping mechanism. Based on the constructed parameter sequence, the trend of change is analyzed by calculating the degree of feature drift between adjacent time nodes, and the feature evolution sequence is subjected to biological consistency fitting. Based on the results of the biological consistency fitting, the corresponding fracture freshness level and freshness probability interval are output through a Bayesian inference mechanism.

[0010] Preferably, the establishment of the time-series image sequence is specifically as follows:

[0011] By acquiring raw image data of the same examined area at different time points using medical imaging acquisition equipment, and sorting them according to the acquisition time sequence, a time-series image sequence is obtained. Then:

[0012]

[0013] in, This represents the medical image data corresponding to the t-th time point. This indicates the total number of time points collected. This represents the constructed time-series image sequence.

[0014] Preferably, the image enhancement processing of the time-series image sequence is specifically as follows:

[0015] Image enhancement processing of time-series image sequences is performed based on adaptive histogram equalization and anisotropic diffusion filtering, specifically as follows:

[0016] If adaptive histogram equalization and anisotropic diffusion filtering are performed on each image, then:

[0017]

[0018] in, This represents the anisotropic diffusion filter operator. This represents the adaptive histogram equalization operator. This represents the enhanced image at time point t.

[0019] Preferably, the determination of bone structure regions using an improved CNN-ViT hybrid network is as follows:

[0020] Global structural features and local texture features are extracted by an improved CNN-ViT hybrid network, and spatial transformation parameters between images are calculated to complete registration under a unified reference coordinate system.

[0021] Based on the characteristics of bone tissue density distribution and structural continuity, a segmentation network is used to obtain a bone tissue mask and extract bone structure regions.

[0022] Preferably, the method of locating the suspected fracture area using a multi-scale feature pyramid network is as follows:

[0023] A multi-scale feature pyramid network is used to extract structural response features at different scales, and a sub-pixel feature decoupling mechanism is used to enhance the discriminability of small abnormal structures and obtain anomaly response maps.

[0024] Based on the response intensity threshold and semantic consistency criterion, candidate regions are extracted, and then:

[0025]

[0026] in, This represents the suspected fracture area at time point t. Indicates pixel position, This indicates the response threshold.

[0027] Preferably, the determination of the temporal feature set using the cross-modal contrastive learning mechanism is as follows:

[0028] Calculate the texture response features, structural continuity disruption features, and grayscale statistical distribution features respectively, and construct feature vectors;

[0029] Based on the cross-modal contrastive learning mechanism, image features at each time point are mapped to a unified feature space;

[0030] The features of each node are associated and organized according to the time dimension to determine the time series feature set.

[0031] Preferably, the process of converting image features into parameter sequences using a biomechanical parameter mapping mechanism is as follows:

[0032] If we construct a feature evolution sequence based on a temporal feature set, then we have:

[0033]

[0034] in, This represents a feature evolution sequence, used to describe the changes in fracture-related imaging features over time. This represents the unified feature representation of the t-th time node, where t represents the time index;

[0035] By converting imaging features into a parameter sequence reflecting the mechanical state of bone tissue using a biomechanical parameter mapping function, we have:

[0036]

[0037] in, This represents a biomechanical parameter mapping function used to convert image features into bone tissue biomechanical state parameters. This represents the biomechanical parameter vector at time point t.

[0038] The biomechanical parameter vectors of all time points are constructed into a parameter sequence.

[0039] Preferably, the step of calculating the degree of feature drift between adjacent time points and performing trend analysis is as follows:

[0040] Based on the constructed parameter sequence, trend analysis is performed by calculating the degree of feature drift between adjacent time points, resulting in:

[0041]

[0042] in, This represents the degree of feature drift at time point t relative to the previous time point.

[0043] Based on the calculated characteristic drift degree, the evolutionary stability index is calculated as follows:

[0044]

[0045] in, Indicators of evolutionary stability This represents the average value indicating the degree of drift.

[0046] Preferably, the step of outputting the corresponding fracture freshness level and freshness probability range through the Bayesian inference mechanism is as follows:

[0047] A freshness discrimination model is constructed based on a probabilistic inference framework, and the freshness probability distribution is calculated through a Bayesian inference mechanism. Then:

[0048]

[0049] in, Indicates the freshness level of the fracture. Represents the prior probability. Represents the likelihood function. Represents the posterior probability;

[0050] Based on the posterior probability distribution, the freshness grade of the fracture and its probability interval are determined, then:

[0051]

[0052] in, This indicates the final freshness level.

[0053] Preferably, the system for identifying the freshness of occult fractures based on time-series radiomics includes a data collection and processing module and an occult fracture freshness identification module; the data collection and processing module is used to collect medical imaging data and process the collected medical imaging data; the occult fracture freshness identification module identifies the freshness of occult fractures based on the processed data.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. This invention employs multi-time-node medical image acquisition and time-series image sequence construction techniques, combined with adaptive histogram equalization and anisotropic diffusion filtering enhancement processing mechanisms, to effectively suppress imaging differences between images at different time nodes and enhance structural information. This solves the problems of insufficient contrast in single images and the easy masking of small fracture signals by noise in existing technologies, thereby improving the sensitivity of early identification of occult fractures.

[0056] 2. This invention employs an improved CNN-ViT hybrid network for spatial alignment and precise segmentation of bone structure regions, and combines a multi-scale feature pyramid network with a sub-pixel feature decoupling mechanism to locate suspected fracture areas. This achieves a refined identification effect for areas of microstructural damage, thereby solving the problem of relying on human experience and difficulty in consistently detecting microcortical bone damage in existing technologies, and improving the objectivity and stability of occult fracture detection.

[0057] 3. This invention achieves the technical effect of dynamic feature fusion of multi-temporal images by constructing a unified feature space through a cross-modal contrastive learning mechanism and organizing it into a temporal feature set according to the time dimension. This solves the problem of lacking a temporal feature integration mechanism and being unable to reflect the dynamic change law of fracture in the existing technology, and provides a structured data foundation for subsequent evolution modeling.

[0058] 4. This invention achieves a quantitative expression of the changes in fracture status over time by adopting a biomechanical parameter mapping mechanism to convert image features into bone tissue mechanical state parameters, and by combining feature drift analysis and evolutionary stability index to construct a feature evolution model. This solves the problem of the lack of dynamic evolution modeling mechanism for fracture healing in the existing technology, and improves the objectivity and scientificity of fracture time staging.

[0059] 5. By adopting the technical means of individualized damage tolerance model constraint and biological consistency fitting mechanism, this invention achieves the constraint consistency effect between the judgment result and the actual biological healing trajectory, thereby avoiding the stage jump misjudgment problem that may occur by relying solely on statistical model inference, and improving the rationality and reliability of time stage determination.

[0060] 6. This invention achieves the technical effect of quantitative freshness level output and confidence interval assessment by adopting a freshness discrimination model based on a probabilistic reasoning framework and outputting probability interval and time stage information through Bayesian inference mechanism. This solves the problems of strong subjectivity and lack of probabilistic support in the judgment of fracture time in the prior art, and improves the interpretability and safety of clinical decision-making. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the overall method steps of the present invention for identifying the freshness of occult fractures based on time-series radiomics.

[0062] Figure 2 This is a schematic diagram of the overall structure of the occult fracture freshness identification system based on temporal radiomics of the present invention. Detailed Implementation

[0063] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0064] Example 1

[0065] Reference Figure 1 As an embodiment of the present invention, a method for identifying the freshness of occult fractures based on temporal radiomics is provided, comprising the following steps:

[0066] S1: Collect medical image data and determine the time series feature set.

[0067] Specifically, determining the temporal feature set involves acquiring medical image data from multiple time points for the same examined site, establishing a temporal image sequence based on the acquisition time order, and performing image enhancement processing on the established temporal image sequence based on adaptive histogram equalization and anisotropic diffusion filtering. Simultaneously, an improved CNN-ViT hybrid network is used to perform spatial alignment and feature embedding mapping on images from different time points, and bone structure regions are automatically segmented based on bone tissue density distribution characteristics. Based on the segmentation results, bone structure regions under a unified reference space are determined.

[0068] Based on the identified bone structure regions, a multi-scale feature pyramid network combined with a sub-pixel feature decoupling mechanism is used to detect potential damage signals. Visual-semantic joint embedding analysis is then used to locate suspected fracture regions. From these suspected fracture regions, image texture response features, structural continuity disruption features, and grayscale statistical distribution features are extracted. A cross-modal contrastive learning mechanism is then used to map the image features corresponding to each time point to a unified feature space. Furthermore, these features are correlated and organized according to the time dimension to form a temporal feature set reflecting changes in the fracture tissue state. The specific implementation is as follows:

[0069] Acquire medical image data from multiple time points and establish a time-series image sequence based on the acquisition time order, specifically as follows:

[0070] By acquiring raw image data of the same examined area at different time points using medical imaging acquisition equipment, and sorting them according to the acquisition time sequence, a time-series image sequence is obtained. Then:

[0071]

[0072] in, This represents the medical image data corresponding to the t-th time point. This indicates the total number of time points collected. This represents the constructed time-series image sequence.

[0073] Image enhancement processing of time-series image sequences is performed based on adaptive histogram equalization and anisotropic diffusion filtering, specifically as follows:

[0074] Adaptive histogram equalization and anisotropic diffusion filtering are performed on each image to improve bone tissue boundary contrast and suppress noise interference. Then:

[0075]

[0076] in, This represents the anisotropic diffusion filter operator, which is set by the implementer according to the actual application scenario. This indicates an adaptive histogram equalization operator, which is set by the implementer according to the actual application scenario. This represents the enhanced image at time point t.

[0077] It should be noted that, in order to reduce the changes in the gray-level statistical distribution of the original image caused by preprocessing operations such as CLAHE and anisotropic diffusion filtering, and to avoid adverse effects on the stability of subsequent radiomics features, a feature robustness verification and adaptive normalization mechanism is introduced to enhance feature robustness verification, as detailed below:

[0078] During the model training phase, comparative experiments with and without enhancement were conducted on the same set of original images. Omics features (such as grayscale histogram features and texture features) of the images before and after enhancement were extracted. The consistency correlation coefficient (CCC) or intra-group correlation coefficient (ICC) of each feature was calculated. Only those robust features that changed little before and after enhancement (e.g., ICC > 0.85) were retained for subsequent modeling.

[0079] At the input of the CNN-ViT hybrid network, a learnable adaptive instance normalization (AdaIN) layer is introduced to replace the fixed CLAHE+ filtering process. This normalization layer can adaptively map the gray-level distribution of the input image to a normalized feature space, which can enhance contrast while ensuring that the transformation is differentiable and jointly optimized with subsequent tasks, thereby avoiding information loss or feature distortion that may be caused by manual enhancement methods.

[0080] Based on the enhanced image, cross-temporal spatial alignment processing is performed, specifically as follows:

[0081] By extracting global structural features and local texture features using an improved CNN-ViT hybrid network and calculating spatial transformation parameters between images to complete registration in a unified reference coordinate system, we can then:

[0082]

[0083] in, This represents the spatial transformation parameters at time point t. Represents a spatial transformation function. This indicates the registered image.

[0084] It should be noted that, in order to address the nonlinear deformation of images at different time points caused by patient positional movement, respiratory motion, or equipment differences, this invention uses an improved CNN-ViT hybrid network to calculate preliminary spatial transformation parameters. Following (rigid or affine transformation), a nonlinear deformable registration module and a quality assessment mechanism are further introduced, as follows:

[0085] After completing the initial registration Based on this, a dense displacement field is estimated using a deep learning deformable registration network (such as VoxelMorph). ,right Fine-grained registration is performed to obtain the final registered image. This network uses... Using a reference image as input, the goal is to maximize the local similarity (e.g., local cross-correlation) between the two.

[0086] After registration is completed, an independent quality assessment module is introduced to calculate the structural similarity index (SSIM) and normalized mutual information (NMI) between the final registered image and the reference image. If the registration quality score is lower than the preset threshold (e.g., SSIM < 0.85), the system will trigger a warning, indicating that the confidence of subsequent feature extraction and analysis may be affected, or that poorly registered areas will be given lower weights in subsequent feature analysis, thereby ensuring temporal consistency and the reliability of subsequent analysis.

[0087] Based on the registered images, bone structure regions are segmented, specifically as follows:

[0088] Based on the characteristics of bone tissue density distribution and structural continuity, a segmentation network is used to obtain a bone tissue mask, and bone structure regions are extracted. Then:

[0089]

[0090] in, This represents the bone structure segmentation mask at time point t. This represents a pixel-by-pixel multiplication operation. This indicates the extracted bone structure region.

[0091] Within the defined bone structure region, potential damage signals are detected, specifically as follows:

[0092] By using a multi-scale feature pyramid network to extract structural response features at different scales, and enhancing the discriminability of minute anomalies through a sub-pixel feature decoupling mechanism, an anomaly response map is obtained, which then yields:

[0093]

[0094] in, This represents a function for multi-scale feature extraction and anomaly response calculation. This represents the anomaly response graph at time point t.

[0095] Based on the abnormal response map, the suspected fracture area is located through visual-semantic joint embedding analysis, specifically:

[0096] Based on the response intensity threshold and semantic consistency criterion, candidate regions are extracted, and then:

[0097]

[0098] in, This represents the suspected fracture area at time point t. Indicates pixel position, This indicates the response threshold, which is set by the implementer based on the actual application scenario.

[0099] For areas suspected of fracture, radiomics features were extracted, specifically:

[0100] Calculate the texture response features, structural continuity disruption features, and grayscale statistical distribution features respectively, and construct the feature vectors. Then we have:

[0101]

[0102] in, Represents texture features, Indicates structural damage characteristics. Represents the statistical characteristics of grayscale. Let represent the feature vector at time point t.

[0103] It should be noted that, to ensure the universality and reproducibility of the extracted radiomics features across devices and centers, the feature extraction process of this invention strictly follows the Initiative for Radiomics Standardization (IBSI) guidelines, as follows:

[0104] In extraction (Texture response characteristics) (Characteristics of structural continuity failure) and When calculating grayscale statistical distribution characteristics, all calculation parameters of the features (such as grayscale discretization step size, filter settings, and texture matrix calculation method) are set and recorded in accordance with the IBSI standard to ensure that the feature definitions are clear and reproducible.

[0105] In the feature selection stage, repeated scan data or test-retest datasets from multiple time points are used to perform intra-group correlation coefficient analysis on each extracted omics feature. Only those features that exhibit high stability (e.g., ICC > 0.8) under the same conditions during repeated scans are retained to construct the final time-series feature set. This step effectively filters out unstable features that are greatly affected by noise and scanning parameters, significantly improving the model's generalization ability and reliability on different image data.

[0106] Based on a cross-modal contrastive learning mechanism, image features from different time points are mapped to a unified feature space, resulting in:

[0107]

[0108] in, Represents the feature mapping function. Representation of features in a unified feature space;

[0109] By associating and organizing the features of each node according to the time dimension to form a time-series feature set, we have:

[0110]

[0111] in, This represents a set of temporal features reflecting changes in the state of fractured tissue, used for subsequent fracture freshness discrimination analysis.

[0112] It should be noted that, in order to enhance the robustness of the model under incomplete data, the constructed time-series feature set is preprocessed with time-series data, as follows:

[0113] First, the input time-series image sequence... Perform integrity checks to determine the missing time points. ;

[0114] against Nodes, utilizing existing image data of neighboring nodes (such as...) and The system uses patient-specific parameters (age, bone density, etc.) and pre-trained temporal generative adversarial networks (Time-GAN) or video frame interpolation networks to generate simulated image data with missing nodes.

[0115] In cases where images cannot be directly completed, the extracted temporal feature set and feature evolution sequence can be used to directly deduce the feature representation of the missing time nodes using temporal modeling methods such as Gaussian process regression, and then incorporate them into the feature evolution sequence.

[0116] By preprocessing time-series data, even with incomplete time-series data input, the system can still make reasonable inferences about the freshness of fractures and output results with confidence assessments (e.g., assigning lower confidence weights to the completed data points), thereby meeting the complex and ever-changing needs of clinical practice.

[0117] S2: Identification of the freshness of occult fractures based on temporal feature sets.

[0118] Specifically, the freshness identification of occult fractures based on temporal feature sets involves constructing a feature evolution sequence based on the temporal feature set, and using a biomechanical parameter mapping mechanism to convert image grayscale information into a parameter sequence reflecting the mechanical state of bone tissue. The converted parameter sequence is then subjected to trend analysis and stage pattern recognition, and the degree of feature drift and evolutionary stability index between each time node are calculated. In addition, a fracture evolution constraint relationship is constructed by combining an individualized damage tolerance model, and the feature evolution sequence is subjected to biological consistency fitting.

[0119] Based on the results of the biological consistency fitting process, the fitting results are input into a freshness discrimination model constructed based on a probabilistic inference framework. The model outputs the corresponding fracture freshness level and freshness probability interval through a Bayesian inference mechanism, and generates fracture time stage determination information based on the inference results to assist in determining the occurrence time of occult fractures. The specific implementation is as follows:

[0120] If we construct a feature evolution sequence based on a temporal feature set, then we have:

[0121]

[0122] in, This represents a feature evolution sequence, used to describe the changes in fracture-related imaging features over time. This represents the unified feature representation of the t-th time node, where t represents the time index;

[0123] Based on the feature evolution sequence, biomechanical parameter mapping processing is performed, specifically as follows:

[0124] By converting imaging features into a parameter sequence reflecting the mechanical state of bone tissue using a biomechanical parameter mapping function, we have:

[0125]

[0126] in, This represents a biomechanical parameter mapping function used to convert image features into bone tissue biomechanical state parameters. This represents the biomechanical parameter vector at time point t.

[0127] If we construct a parameter sequence from the biomechanical parameter vectors of all time points, then we have:

[0128]

[0129] in, This represents the biomechanical parameter vector at the first time point. Indicates the first Biomechanical parameter vectors at each time point This represents the sequence of parameters to be constructed.

[0130] It should be noted that the biomechanical parameter mapping function is not a fixed mathematical transformation, but a learnable mapping network based on a physical information neural network, as detailed below:

[0131] A multilayer perceptron (MLP) is used to instantiate the mapping function, with a unified feature representation as input and the output being a corresponding biomechanical parameter vector, which may include bone tissue apparent stiffness, ultimate strength, micro-damage accumulation, etc.

[0132] To ensure that the mapping learned by the biomechanical parameter mapping function conforms to the actual biomechanical evolution, a physical consistency constraint term is introduced into the training loss function. This means that not only must the mapped parameter sequence be accurately used for downstream classification, but its evolution over time must also satisfy the physical equations defined by the damage tolerance model. Specifically:

[0133]

[0134] in, The cross-entropy loss represents the classification of freshness levels. Represents physical constraint loss. The hyperparameter represents the balance between the two losses;

[0135] Mapping Network It can adaptively learn the complex nonlinear relationship between image features and mechanical parameters from data, while being constrained by physical laws, ensuring that its output parameter sequence not only has discriminative power, but also has biological and physical rationality.

[0136] Based on the constructed parameter sequence, trend analysis is performed by calculating the degree of feature drift between adjacent time points, resulting in:

[0137]

[0138] in, This represents the degree of feature drift at time point t relative to the previous time point.

[0139] Based on the calculated characteristic drift degree, an evolutionary stability index is calculated to measure the smoothness of changes in fracture status. Then:

[0140]

[0141] in, Indicators of evolutionary stability This represents the average value indicating the degree of drift.

[0142] Combining an individualized injury tolerance model, a constraint relationship for fracture evolution is constructed. This constraint function is established based on individual patient parameters (such as age, bone mineral density, etc.), resulting in:

[0143]

[0144] in, This represents the damage tolerance constraint function. Let represent the set of individualized parameters, and let represent the constraint state index at time t.

[0145] Based on drift degree, stability index, and constraint state, biological consistency fitting is performed on the feature evolution sequence, specifically as follows:

[0146]

[0147] in, This represents the biological consistency fitting function. This represents the fitted parameter sequence.

[0148] The fitted parameter sequence is then input into the freshness discrimination model, specifically:

[0149] A freshness discrimination model is constructed based on a probabilistic inference framework, and the freshness probability distribution is calculated through a Bayesian inference mechanism. Then:

[0150]

[0151] in, Indicates the freshness level of the fracture. Represents the prior probability. Represents the likelihood function. Represents the posterior probability;

[0152] Based on the posterior probability distribution, the freshness grade and probability interval of the fracture are determined, then:

[0153]

[0154] in, This indicates the final freshness level.

[0155] It should be noted that prior and likelihood modeling is performed before determining the fracture freshness level and probability interval based on the posterior probability distribution, as follows:

[0156] The modeling of prior probabilities is as follows:

[0157] Prior probability represents the distribution of fracture freshness grades before any imaging evidence is available.

[0158] Prior probability can be derived from: large-sample epidemiological statistics: based on a large amount of historical case data, the frequency of fracture occurrence at each stage is statistically analyzed;

[0159] Expert knowledge: Through methods such as the Delphi method, the experience of clinical experts is transformed into probability distributions. For example, in trauma clinics, the prior probability of acute fractures may be higher.

[0160] Likelihood function modeling, as detailed below:

[0161] First, a deep neural network (as part of a kernel function) is used to map the input into a low-dimensional feature space;

[0162] Then, Gaussian process regression is used in this feature space to learn a unique evolutionary trajectory pattern for each level. The Gaussian process can not only give the predicted mean (i.e. the expected value of the parameter sequence) but also the predicted variance, thereby quantifying the model's confidence in the current observation data.

[0163] The final posterior probability distribution not only provides the probability of each level, but by sampling it, we can also obtain the confidence interval for freshness determination. For example, it outputs the acute phase (probability 75%, confidence interval: 2-5 days), providing more comprehensive information for clinical decision-making.

[0164] Based on the judgment results, fracture time stage judgment information is generated to assist clinical judgment of the timing of occult fractures. This involves mapping the freshness level to the corresponding time interval and outputting fracture time stage information, specifically:

[0165] After obtaining the fracture freshness level and corresponding probability distribution results output by the freshness discrimination model, the judgment results are interpreted according to the constructed fracture time stage mapping rule, resulting in:

[0166] In the mapping process, firstly, the candidate time interval range is determined based on the freshness level output by the model. At the same time, the confidence of the time interval is evaluated by combining the corresponding probability score to avoid the uncertainty caused by a single level judgment. If there are multiple possible levels in the model output, the comprehensive time estimation result is calculated by probability weighting to obtain more robust stage judgment information.

[0167] In addition, the consistency of the stage judgment results is verified by combining the trend of temporal feature changes. For example, when the feature evolution trend deviates from the typical fracture healing trajectory, the stage correction mechanism can be triggered to dynamically adjust the time interval to ensure that the judgment results conform to the biological evolution law.

[0168] After generating fracture time stage information, the judgment results are output in a structured form, including fracture stage label, corresponding time range, judgment confidence level, and summary of key features, to provide interpretable auxiliary decision-making information to clinicians. At the same time, the results can be recorded in the case database for subsequent follow-up analysis and continuous model optimization.

[0169] Furthermore, to enhance the correspondence between the model output and routine clinical staging, this invention constructs a mapping mechanism between fracture freshness grades and clinical staging standards. The fracture freshness grade G is divided into a discrete grade set {G1, G2, G3}, corresponding to the acute phase, subacute phase, and old phase, respectively. The specific mapping rules are as follows:

[0170] G1 (acute phase): Fracture occurred ≤7 days ago;

[0171] G2 (subacute phase): Fracture occurred 8–21 days ago;

[0172] G3 (old stage): Fracture occurred >21 days ago;

[0173] When the Bayesian inference module outputs the posterior probability distribution Subsequently, if the probability of a certain level exceeds a preset threshold... If the probability of a time interval is close, it is directly mapped to the corresponding clinical time interval; if multiple levels of probability are close, the expected time value E(T) is calculated by probability weighting, and the stage is classified according to the time interval in which E(T) falls.

[0174] Meanwhile, retrospective verification can be performed using real follow-up case data, and statistical models can determine consistency indicators (such as Kappa value) between staging and actual clinical staging to enhance clinical verifiability.

[0175] It should be noted that the mapping rules can be established based on clinical fracture healing patterns and historical case statistics, and are used to map different freshness levels to specific fracture occurrence time intervals, such as the acute phase, early repair phase, intermediate healing phase, and old phase.

[0176] It should be noted that, in order to clarify the imaging and pathological basis of the 'occult fracture' treated in this invention, the following control relationship was established to guide feature extraction and result interpretation, as shown in the table below:

[0177] Pathological stage Main pathological changes CT imaging features MRI imaging features Corresponding to the feature focus of this invention Acute phase Bone marrow edema, hemorrhage, trabecular fracture There may be no obvious interruption of the cortical bone, or only slight blurring of the trabeculae. Patchy high signal is seen on 2WI / STIR sequences, and low signal is seen on T1WI. Gray-scale statistical characteristics (increased signal in edema areas), structural continuity disruption characteristics subacute phase Inflammation absorption, granulation tissue formation, periosteal reaction Mild periosteal reaction or callus formation may occur, with uneven density. The area of ​​edema signal narrows, and slightly high signal may appear on T1WI. Texture response features (granulation tissue and callus texture), feature drift degree Old period Bone remodeling is complete, or sclerosis and nonunion may occur. Cortical thickening and sclerosis, or clear fracture line and sclerosis of the fracture ends. Bone marrow edema signals disappeared, and fat signals returned to normal. Evolutionary stability indices and biomechanical parameters

[0178] It should be noted that, according to the established control table, it can be found that this control table clarifies the typical manifestations of different pathological stages on CT / MRI, and makes the 'multimodal features' extracted by the model (such as texture, structure, grayscale) correspond to specific pathophysiological processes, providing a traceable basis for clinical validation.

[0179] Example 2

[0180] Reference Figure 2 As an embodiment of the present invention, a system for identifying the freshness of occult fractures based on temporal radiomics is provided, including a data collection and processing module and an occult fracture freshness identification module.

[0181] Specifically, the data collection and processing module is used to collect medical imaging data and process the collected medical imaging data, while the occult fracture freshness identification module identifies the freshness of occult fractures based on the processed data.

[0182] Furthermore, the data collection and processing module is used to complete the acquisition and temporal feature construction of medical image data, as detailed below:

[0183] The system comprises several components: a multi-time-node medical image acquisition unit, which acquires raw image data of the same examined site at different time points and establishes a temporal image sequence; an image enhancement processing unit, which performs adaptive histogram equalization and anisotropic diffusion filtering on the temporal image sequence to enhance bone tissue boundary structure and reduce noise interference; a spatial registration and structural segmentation unit, which uses an improved CNN-ViT hybrid network to achieve cross-time-node image spatial alignment and extracts bone structure regions in a unified reference space based on bone tissue density distribution features; a suspected region localization unit, which uses a multi-scale feature pyramid network and sub-pixel feature decoupling mechanism to detect potential injury signals and locate suspected fracture regions through visual-semantic joint embedding; and a feature construction unit, which extracts texture response features, structural continuity disruption features, and gray-level statistical distribution features from suspected fracture regions and maps them to a unified feature space using a cross-modal contrastive learning mechanism to form a temporal feature set.

[0184] The module for identifying the freshness of occult fractures, based on the time-series feature set output by the data collection and processing module, quantitatively determines the freshness of fractures, as follows:

[0185] The evolutionary modeling unit is used to construct a feature evolution sequence based on a time-series feature set and generate a parameter sequence reflecting the mechanical state of bone tissue through a biomechanical parameter mapping mechanism. The trend analysis unit is used to calculate the degree of feature drift and evolutionary stability index between adjacent time nodes to analyze the trend of fracture status changes. The biological consistency constraint unit is used to construct fracture evolution constraint relationships in combination with an individualized damage tolerance model and perform biological consistency fitting on the feature evolution sequence. The probabilistic inference unit is used to construct a freshness discrimination model based on a probabilistic inference framework and output the fracture freshness level and probability interval through a Bayesian inference mechanism. The time stage generation unit is used to map the discrimination results to specific fracture time stage information and output structured auxiliary diagnostic results.

[0186] It should be noted that through the coordinated operation of the two modules, a complete reasoning chain from image features to time staging results is realized, forming a data-driven system that combines biological laws to determine the freshness of occult fractures.

[0187] Furthermore, if the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0188] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0189] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0190] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying the freshness of occult fractures based on temporal radiomics, characterized in that: Includes the following steps, Collect medical imaging data and determine the time-series feature set, specifically: For the same examined site, medical image data at multiple time points were acquired and a time-series image sequence was established. The time-series image sequence was enhanced. At the same time, the bone structure region was determined by an improved CNN-ViT hybrid network. Based on the determined bone structure region, the suspected fracture region was located by a multi-scale feature pyramid network. Multimodal features were extracted from the suspected fracture region and the time-series feature set was determined by combining a cross-modal contrastive learning mechanism. The freshness of occult fractures is identified based on temporal feature sets, specifically as follows: A feature evolution sequence is constructed based on a time-series feature set, and the image features are converted into a parameter sequence using a biomechanical parameter mapping mechanism. Based on the constructed parameter sequence, the trend of change is analyzed by calculating the degree of feature drift between adjacent time nodes, and the feature evolution sequence is subjected to biological consistency fitting. Based on the results of the biological consistency fitting, the corresponding fracture freshness level and freshness probability interval are output through a Bayesian inference mechanism.

2. The method for identifying the freshness of occult fractures based on temporal radiomics as described in claim 1, characterized in that: The establishment of the time-series image sequence is specifically as follows: By acquiring raw image data of the same examined area at different time points using medical imaging acquisition equipment, and sorting them according to the acquisition time sequence, a time-series image sequence is obtained. Then: in, This represents the medical image data corresponding to the t-th time point. This indicates the total number of time points collected. This represents the constructed time-series image sequence.

3. The method for identifying the freshness of occult fractures based on temporal radiomics as described in claim 2, characterized in that: The image enhancement processing of the time-series image sequence is specifically as follows: Image enhancement processing of time-series image sequences is performed based on adaptive histogram equalization and anisotropic diffusion filtering, specifically as follows: If adaptive histogram equalization and anisotropic diffusion filtering are performed on each image, then: in, This represents the anisotropic diffusion filter operator. This represents the adaptive histogram equalization operator. This represents the enhanced image at time point t.

4. The method for identifying the freshness of occult fractures based on temporal radiomics as described in claim 3, characterized in that: The bone structure region is determined using an improved CNN-ViT hybrid network, as detailed below: Global structural features and local texture features are extracted by an improved CNN-ViT hybrid network, and spatial transformation parameters between images are calculated to complete registration under a unified reference coordinate system. Based on the characteristics of bone tissue density distribution and structural continuity, a segmentation network is used to obtain a bone tissue mask and extract bone structure regions.

5. The method for identifying the freshness of occult fractures based on temporal radiomics as described in claim 4, characterized in that: The method of locating suspected fracture areas using a multi-scale feature pyramid network is as follows: A multi-scale feature pyramid network is used to extract structural response features at different scales, and a sub-pixel feature decoupling mechanism is used to enhance the discriminability of small abnormal structures and obtain anomaly response maps. Based on the response intensity threshold and semantic consistency criterion, candidate regions are extracted, and then: in, This represents the suspected fracture area at time point t. Indicates pixel position, This indicates the response threshold.

6. The method for identifying the freshness of occult fractures based on temporal radiomics as described in claim 5, characterized in that: The determination of the temporal feature set using the cross-modal contrastive learning mechanism is as follows: Calculate the texture response features, structural continuity disruption features, and grayscale statistical distribution features respectively, and construct feature vectors; Based on the cross-modal contrastive learning mechanism, image features at each time point are mapped to a unified feature space; The features of each node are associated and organized according to the time dimension to determine the time series feature set.

7. The method for identifying the freshness of occult fractures based on temporal radiomics as described in claim 6, characterized in that: The process of converting image features into parameter sequences using a biomechanical parameter mapping mechanism is as follows: If we construct a feature evolution sequence based on a temporal feature set, then we have: in, This represents a feature evolution sequence, used to describe the changes in fracture-related imaging features over time. This represents the unified feature representation of the t-th time node, where t represents the time index; By converting imaging features into a parameter sequence reflecting the mechanical state of bone tissue using a biomechanical parameter mapping function, we have: in, This represents a biomechanical parameter mapping function used to convert image features into bone tissue biomechanical state parameters. This represents the biomechanical parameter vector at time point t. The biomechanical parameter vectors of all time points are constructed into a parameter sequence.

8. The method for identifying the freshness of occult fractures based on temporal radiomics as described in claim 7, characterized in that: The analysis of the changing trend of the feature drift between adjacent time points is as follows: Based on the constructed parameter sequence, trend analysis is performed by calculating the degree of feature drift between adjacent time points, resulting in: in, This represents the degree of feature drift at time point t relative to the previous time point. Based on the calculated characteristic drift degree, the evolutionary stability index is calculated as follows: in, Indicators of evolutionary stability This represents the average value indicating the degree of drift.

9. The method for identifying the freshness of occult fractures based on temporal radiomics as described in claim 8, characterized in that: The Bayesian inference mechanism is used to output the corresponding fracture freshness level and freshness probability range, as detailed below: A freshness discrimination model is constructed based on a probabilistic inference framework, and the freshness probability distribution is calculated through a Bayesian inference mechanism. Then: in, Indicates the freshness level of the fracture. Represents the prior probability. Represents the likelihood function. Represents the posterior probability; Based on the posterior probability distribution, the freshness grade of the fracture and its probability interval are determined, then: in, This indicates the final freshness level.

10. A system employing the temporal radiomics-based freshness identification method for occult fractures as described in any one of claims 1 to 9, characterized in that, This includes a data collection and processing module, as well as a module for identifying the freshness of occult fractures; The data collection and processing module is used to collect medical image data and process the collected medical image data. The occult fracture freshness identification module identifies the freshness of occult fractures based on the processed data.

Images