A condylar bone resorption preoperative prediction method and system based on imaging feature absolute difference screening

Abstract

The application discloses a condylar resorption preoperative prediction method and system based on imageomics feature absolute difference screening, and belongs to the field of medical image processing. The method comprises the following steps: collecting preoperative and postoperative condylar CBCT images of a patient, respectively extracting an imageomics feature set and calculating the absolute difference of the corresponding dimension features to generate a feature difference vector; calculating the correlation of the vector and a clinical outcome, and screening out feature dimensions sensitive to bone resorption changes; locking the corresponding core feature subset in the preoperative image through reverse mapping; training a machine learning prediction model using the subset, and constructing a prediction model that only depends on the preoperative image. Compared with a traditional preoperative feature model, the application can effectively screen out core features with "evolution directionality", which helps to improve the prediction performance, feature compression effect and robustness of the model, and provides technical support for preoperative evaluation of condylar resorption risk after orthognathic surgery.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and analysis technology, and more specifically to a method and system for preoperative prediction of condylar bone resorption based on the absolute difference of radiomics features. Background Technology

[0002] Condylar resorption (CR) of the temporomandibular joint is a common complication after orthognathic surgery. Accurately predicting the risk of postoperative bone resorption in patients during the preoperative stage is of great clinical significance for developing individualized surgical plans and postoperative management decisions.

[0003] Currently, existing risk assessment and prediction methods for condylar bone resorption mainly present the following two technical approaches, but both have obvious limitations: 1. Static Feature Prediction Path Based on Preoperative Single-Phase Imaging: This type of method constructs a model by extracting morphological indicators or radiomics features from the patient's preoperative state. While it offers the advantage of timely preoperative warning, condylar bone resorption is a dynamic biological process triggered by surgical stress, and simple preoperative static features often fail to characterize the bone tissue's remodeling potential under surgical intervention. This results in such models lacking clear evolutionary direction in feature selection, and their prediction accuracy and robustness often fall short of the requirements for clinical decision-making.

[0004] 2. Dynamic assessment approach based on preoperative and postoperative two-phase differences: This type of method evaluates the degree of bone resorption by calculating changes in imaging features before and after surgery (such as feature differences or rates of change). Although the difference data can intuitively and accurately reflect the actual changes in bone structure, this type of method relies on postoperative follow-up imaging data and does not establish a mapping relationship between preoperative static indicators and postoperative dynamic remodeling, making it difficult to apply to preoperative prediction scenarios. This "post-operative evaluation" characteristic means that it cannot provide any early warning information before surgery, thus missing the golden window for assisting initial clinical decision-making.

[0005] Therefore, how to provide a method and system for preoperative prediction of condylar bone resorption based on the absolute difference of radiomics features is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomic features, solving the problems of weak correlation between preoperative features and actual bone structure changes and insufficient robustness of prediction models in existing technologies. By introducing feature differences for screening preoperative radiomic features, core features that can predict significant changes in bone structure can be accurately identified from preoperative images, thereby improving the prediction accuracy of condylar bone resorption risk after orthognathic surgery without increasing the data requirements in the prediction stage.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomic features includes the following steps: Obtain a training dataset, which includes preoperative radiomics feature sets, postoperative radiomics feature sets, and corresponding clinical outcome labels for multiple patients; For each patient, the absolute difference between the corresponding dimensional features in the preoperative radiomics feature set and the postoperative radiomics feature set is calculated to generate a feature absolute difference vector. Calculate the correlation coefficient between the absolute difference vector of the features and the clinical outcome label, and sort the features according to the absolute value of the correlation coefficient. Select the top-ranked features to form a set of significant difference features. Based on the set of significant difference features, features of the corresponding dimension are extracted from the set of preoperative radiomics features to form a subset of preoperative core features. The initial prediction model was trained using the aforementioned preoperative core feature subset and its corresponding clinical outcome labels to obtain the preoperative prediction model; The preoperative radiomics features of the patient to be predicted are obtained and input into the preoperative prediction model, and the prediction result of the patient's condylar bone resorption risk is output.

[0008] Furthermore, after selecting the top-ranked features across several dimensions to form a set of significant difference features, the process also includes: A pairwise correlation analysis is performed on the features in the set of significant differences. If the absolute value of the correlation coefficient between two features exceeds a preset threshold, one of the features is removed according to a preset rule to eliminate feature collinearity.

[0009] Furthermore, the initial prediction model is a machine learning model, used to train the machine learning model based on the preoperative core feature subset and the corresponding clinical outcome labels, and to establish a mapping relationship between the preoperative core features and the prediction results of condylar bone resorption risk.

[0010] Furthermore, the machine learning model includes a logistic regression model, a regularized logistic regression model, a support vector machine, a random forest, a gradient boosting model, or a neural network model.

[0011] Furthermore, in the step of calculating the absolute difference between the corresponding dimensional features in the preoperative radiomics feature set and the postoperative radiomics feature set, and generating a feature absolute difference vector, the formula for calculating the absolute difference is as follows: ΔF=|F pre -F post |; Where Fpre F represents the preoperative characteristic value. post These are postoperative characteristic values.

[0012] Furthermore, the feature categories included in the preoperative radiomics feature set and the postoperative radiomics feature set include: first-order statistical features, shape features, and texture features.

[0013] Furthermore, before acquiring the training dataset, the process includes a preprocessing step for the original condylar CBCT images: Histogram-based grayscale normalization was used to process preoperative and postoperative condylar CBCT images to reduce differences in grayscale distribution, brightness, and contrast caused by different equipment or scanning conditions.

[0014] Furthermore, the correlation coefficients include Pearson correlation coefficient, Spearman correlation coefficient, or other statistical indicators used to evaluate the degree of association between the absolute difference vector of features and the clinical outcome label.

[0015] A preoperative prediction system for condylar bone resorption based on absolute difference screening of radiomics features includes: The data acquisition module is used to acquire the training dataset, which includes preoperative radiomics feature sets, postoperative radiomics feature sets, and corresponding clinical outcome labels of multiple patients. The difference calculation module is used to calculate the absolute difference between the corresponding dimensional features in the preoperative radiomics feature set and the postoperative radiomics feature set for each patient, and generate a feature absolute difference vector. The difference-guided screening module is used to calculate the correlation coefficient between the absolute difference vector of the features and the clinical outcome label, and sort the features according to the absolute value of the correlation coefficient, and screen out the top-ranked features to form a significant difference feature set. The feature inverse mapping module is used to extract features of the corresponding dimension from the preoperative radiomics feature set based on the difference significant feature set, and to form a subset of preoperative core features; The model training module is used to train an initial prediction model using the preoperative core feature subset and its corresponding clinical outcome labels to obtain the preoperative prediction model. The prediction application module is used to acquire the preoperative radiomics features of the patient to be predicted, input them into the preoperative prediction model, and output the prediction result of the patient's condylar bone resorption risk.

[0016] Furthermore, the model training module is used to train a machine learning model, which includes, but is not limited to, one or more of the following: logistic regression model, regularized logistic regression model, support vector machine, random forest, gradient boosting model, or neural network model.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method and system for preoperative prediction of condylar bone resorption based on the absolute difference of radiomic features. By constructing the absolute difference between preoperative and postoperative radiomic features, the method is used to characterize the magnitude of bone structure changes. Based on the difference, the preoperative radiomic features are screened, thereby constructing a prediction model that relies only on preoperative data. This improves the performance of the prediction model while ensuring that postoperative information is not introduced during the prediction stage. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 The method flow provided by this invention; Figure 2 This is a schematic diagram of the feature filtering logic provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention discloses a method for preoperative prediction of condylar bone resorption based on absolute difference in radiomics features, comprising the following steps: Obtain a training dataset, which includes preoperative radiomics feature sets, postoperative radiomics feature sets, and corresponding clinical outcome labels for multiple patients; For each patient, the absolute difference between the corresponding dimensional features in the preoperative radiomics feature set and the postoperative radiomics feature set is calculated to generate a feature absolute difference vector. Calculate the correlation coefficient between the absolute difference vector of the features and the clinical outcome label, and sort the features according to the absolute value of the correlation coefficient. Select the top-ranked features to form a set of significant difference features. Based on the set of significant difference features, features of the corresponding dimension are extracted from the set of preoperative radiomics features to form a subset of preoperative core features. The initial prediction model was trained using the aforementioned preoperative core feature subset and its corresponding clinical outcome labels to obtain the preoperative prediction model; The preoperative radiomics features of the patient to be predicted are obtained and input into the preoperative prediction model, and the prediction result of the patient's condylar bone resorption risk is output.

[0022] See Figure 1 This demonstrates the entire process from dual-time image input, feature extraction, difference calculation to final preoperative feature selection and model training.

[0023] Specifically, step one: multi-time period image acquisition and preprocessing: Condylar CBCT images were acquired at fixed follow-up points before and after surgery. Since CBCT does not have the Hounsfield Unit (HU) value of spiral CT, this method uses histogram-based grayscale normalization to reduce differences in grayscale distribution, brightness, and contrast caused by different equipment or scanning conditions.

[0024] Specifically, step two: ROI delineation and feature extraction: Three-dimensional segmentation of the condylar region was performed in preoperative and postoperative images, and the preoperative feature set (F) was extracted respectively. pre ) and postoperative feature set (F post The feature categories include first-order statistical features, shape features, and texture features.

[0025] Specifically, step three: Calculation of absolute feature differences (a core point in the training phase): For the preoperative feature vector F pre With postoperative feature vector F post The features of the corresponding dimensions are matched one by one, and the absolute difference of each feature is calculated to form the feature difference vector ΔF.

[0026] Calculation formula: ΔF = |F pre -F post | This difference is used to quantify the true extent of changes in bone structure before and after surgery.

[0027] Specifically, step four: difference-guided preoperative feature screening: 1. Establish evolutionary correlation criteria: Calculate the correlation coefficient (including but not limited to Pearson correlation coefficient or Spearman correlation coefficient) between the absolute difference vector of features ΔF and the final clinical outcome (e.g., whether bone resorption occurs or quantitative indicators of bone resorption).

[0028] 2. Multi-level feature selection: Preliminary screening (ranking guidance): Sort by absolute value of correlation coefficient from high to low, and select the top N features (e.g., top 10-30) as the candidate set.

[0029] Redundancy filtering (collinearity elimination): Perform pairwise correlation analysis on the features in the candidate set. If the absolute value of the correlation coefficient between two features exceeds a preset threshold (e.g., 0.90), the redundant features are deleted according to the preset rules to reduce the collinearity effect between features and avoid the adverse effects of feature redundancy on model stability and generalization ability.

[0030] 3. Reverse Mapping Locking: The feature difference ΔF serves only as a "guide signal" to reflect the dynamic evolution potential of bone structure; significant terms selected through the difference are reverse-mapped to their corresponding preoperative radiomics features F during the training phase. pre This allows for the precise extraction of a subset of core features with "evolutionary orientation" from a massive amount of preoperative static features.

[0031] Specifically, by establishing a correlation ranking mechanism between ΔF features and clinical outcomes, the feature dimensions most sensitive to changes in bone resorption are prioritized and mapped back to the preoperative radiomics feature space. This allows for the effective compression and screening of key preoperative predictive features without introducing postoperative information as model input, thereby reducing redundant static features and improving the model's stability and generalization ability.

[0032] For details, see Figure 2 This is a schematic diagram of the feature filtering logic, comparing the differences between the traditional filtering path and the difference-guided path of this invention.

[0033] Step 5: Construct a pure preoperative prediction model: 1. Feature Input and Decoupling: Only the preoperative core feature subset locked in step four is used as input to the final prediction model. Postoperative image data is only used during the training phase to construct the absolute difference of features and guide the preoperative feature selection; postoperative image data is not included in the input features of the final prediction model and in the clinical application phase, thus ensuring the independence of the prediction logic in clinical application.

[0034] 2. Model Training: A prediction model is built using a classification or regression algorithm based on machine learning. The model includes, but is not limited to, logistic regression, regularized logistic regression, support vector machine, random forest, gradient boosting model or neural network model.

[0035] 3. Regularization and Feature Sparsity Processing: In a preferred embodiment, a logistic regression model with an L1 regularization term is used for training. L1 regularization sparsifies the model parameters, causing some feature coefficients with low predictive contributions to approach zero or be compressed to zero, thereby achieving secondary automatic feature simplification and outputting condylar bone resorption risk probability or grading information.

[0036] Specifically, step six: model application (prediction phase): In actual clinical prediction, only the patient's preoperative CBCT image features need to be input, without any postoperative image data. Furthermore, during model training or cross-validation, the correlation calculation of absolute feature differences, difference-guided feature selection, and collinearity filtering are all performed based solely on the current training data. Validation data, test data, or data to be predicted do not participate in the above selection process to reduce the risk of information leakage.

[0037] Specifically, taking the prediction of condylar bone resorption after orthognathic surgery as an example: Data preparation: In one embodiment, several patient data may be used for model training and validation.

[0038] Extraction: More than 1,000 radiomics features were extracted from the preoperative and postoperative ROIs using PyRadiomics. The same segmentation strategy and parameters were used for the preoperative and postoperative images to ensure the comparability of the features.

[0039] Calculation: Obtain the characteristic change vector ΔF for each patient.

[0040] Verification: In one specific embodiment, some preoperative shape features, texture features and their corresponding differences are related to postoperative condylar bone resorption and can be used to construct a subset of preoperative core features guided by differences.

[0041] Model training: The classifier is trained based on the selected preoperative features.

[0042] Through the above technical solution, the present invention can achieve the following technical effects: 1. Improved Predictive Performance and Stability: Through the "difference-guided" mechanism, the selected features are not only relevant to clinical outcomes but also reflect the magnitude of changes in bone structure. In one embodiment, compared with traditional preoperative feature models, the method of this invention can achieve better or comparable predictive performance and has advantages in feature compression, evolution directionality, and model stability.

[0043] 2. Strong clinical applicability: During the training phase, this invention uses postoperative change information to guide the screening of preoperative features, but the final prediction model relies only on preoperative image features in the actual application (inference) phase, without the need for postoperative image data, which meets the needs of early clinical warning.

[0044] 3. Enhanced interpretability: The selected preoperative features have been validated by the magnitude of change (ΔF), which improves the correlation between the features and changes in bone structure.

[0045] In one specific embodiment, compared with traditional preoperative feature models, the method of the present invention can achieve better or comparable predictive performance, and has advantages in feature number compression, evolution directionality and model stability.

[0046] On the other hand, the present invention also discloses a preoperative prediction system for condylar bone resorption based on absolute difference screening of radiomics features, comprising: The data acquisition module is used to acquire the training dataset, which includes preoperative radiomics feature sets, postoperative radiomics feature sets, and corresponding clinical outcome labels of multiple patients. The difference calculation module is used to calculate the absolute difference between the corresponding dimensional features in the preoperative radiomics feature set and the postoperative radiomics feature set for each patient, and generate a feature absolute difference vector. The difference-guided screening module is used to calculate the correlation coefficient between the absolute difference vector of the features and the clinical outcome label, and sort the features according to the absolute value of the correlation coefficient, and screen out the top-ranked features to form a significant difference feature set. The feature inverse mapping module is used to extract features of the corresponding dimension from the preoperative radiomics feature set based on the difference significant feature set, and to form a subset of preoperative core features; The model training module is used to train an initial prediction model using the preoperative core feature subset and its corresponding clinical outcome labels to obtain the preoperative prediction model. The prediction application module is used to acquire the preoperative radiomics features of the patient to be predicted, input them into the preoperative prediction model, and output the prediction result of the patient's condylar bone resorption risk.

[0047] In summary, this invention aims to address the shortcomings of existing preoperative static feature models in characterizing bone structure evolution and predicting stability. This invention first acquires preoperative and postoperative condylar CBCT images of patients, extracts radiomics feature sets, and calculates the absolute differences of corresponding dimensional features to generate a feature absolute difference vector. Then, it calculates the correlation between this vector and clinical outcomes, selecting feature dimensions sensitive to bone resorption changes. Through inverse mapping, it identifies the corresponding core feature subset from the preoperative images. Finally, it uses this subset to train a machine learning prediction model, constructing a prediction model that relies solely on preoperative images. L1 regularized logistic regression can be used as a preferred implementation method to achieve sparsity of model parameters and secondary feature simplification. During the training phase, this invention utilizes postoperative change information to construct feature absolute differences, and uses these differences as guiding signals to select core preoperative features. In the final prediction model's input features and actual clinical prediction stages, postoperative image data is not required, thus ensuring the independence and timeliness of clinical application. Compared with traditional preoperative feature models, this invention can effectively screen out core features with "evolutionary orientation", which helps to improve the model's predictive performance, feature compression effect and robustness, and provides technical support for the preoperative assessment of the risk of condylar bone resorption after orthognathic surgery.

[0048] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0049] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomics features, characterized in that, Includes the following steps: Obtain a training dataset, which includes preoperative radiomics feature sets, postoperative radiomics feature sets, and corresponding clinical outcome labels for multiple patients; For each patient, the absolute difference between the corresponding dimensional features in the preoperative radiomics feature set and the postoperative radiomics feature set is calculated to generate a feature absolute difference vector. Calculate the correlation coefficient between the absolute difference vector of the features and the clinical outcome label, and sort the features according to the absolute value of the correlation coefficient. Select the top-ranked features to form a set of significant difference features. Based on the set of significant difference features, features of the corresponding dimension are extracted from the set of preoperative radiomics features to form a subset of preoperative core features. The initial prediction model was trained using the aforementioned preoperative core feature subset and its corresponding clinical outcome labels to obtain the preoperative prediction model; The preoperative radiomics features of the patient to be predicted are obtained and input into the preoperative prediction model, and the prediction result of the patient's condylar bone resorption risk is output.

2. The method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomics features according to claim 1, characterized in that, After selecting the top-ranked features across several dimensions to form a significant difference feature set, the process also includes: A pairwise correlation analysis is performed on the features in the set of significant differences. If the absolute value of the correlation coefficient between two features exceeds a preset threshold, one of the features is removed according to a preset rule to eliminate feature collinearity.

3. The method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomics features according to claim 1, characterized in that, The initial prediction model is a machine learning model, used to train the machine learning model based on the preoperative core feature subset and the corresponding clinical outcome labels, and to establish a mapping relationship between the preoperative core features and the prediction results of condylar bone resorption risk.

4. The method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomics features according to claim 3, characterized in that, The machine learning models include logistic regression models, regularized logistic regression models, support vector machines, random forests, gradient boosting models, or neural network models.

5. The method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomics features according to claim 1, characterized in that, In the step of calculating the absolute difference between the corresponding dimensional features in the preoperative radiomics feature set and the postoperative radiomics feature set, and generating a feature absolute difference vector, the formula for calculating the absolute difference is as follows: ΔF=∣F pre -F post ∣; Where F pre F represents the preoperative characteristic value. post These are postoperative characteristic values.

6. The method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomics features according to claim 1, characterized in that, The feature categories included in the preoperative radiomics feature set and the postoperative radiomics feature set are: first-order statistical features, shape features, and texture features.

7. The method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomics features according to claim 1, characterized in that, Before obtaining the training dataset, the process also includes a preprocessing step for the original condylar CBCT images: Histogram-based grayscale normalization was used to process preoperative and postoperative condylar CBCT images to reduce differences in grayscale distribution, brightness, and contrast caused by different image acquisition devices or scanning conditions.

8. The method for preoperative prediction of condylar bone resorption based on absolute difference screening of radiomics features according to claim 1, characterized in that, The correlation coefficients include Pearson correlation coefficient, Spearman correlation coefficient, or other statistical indicators used to evaluate the degree of association between the absolute difference vector of features and the clinical outcome label.

9. A preoperative prediction system for condylar bone resorption based on absolute difference screening of radiomics features, characterized in that, include: The data acquisition module is used to acquire the training dataset, which includes preoperative radiomics feature sets, postoperative radiomics feature sets, and corresponding clinical outcome labels of multiple patients. The difference calculation module is used to calculate the absolute difference between the corresponding dimensional features in the preoperative radiomics feature set and the postoperative radiomics feature set for each patient, and generate a feature absolute difference vector. The difference-guided screening module is used to calculate the correlation coefficient between the absolute difference vector of the features and the clinical outcome label, and sort the features according to the absolute value of the correlation coefficient, and screen out the top-ranked features to form a significant difference feature set. The feature inverse mapping module is used to extract features of the corresponding dimension from the preoperative radiomics feature set based on the difference significant feature set, and to form a subset of preoperative core features; The model training module is used to train an initial prediction model using the preoperative core feature subset and its corresponding clinical outcome labels to obtain the preoperative prediction model. The prediction application module is used to acquire the preoperative radiomics features of the patient to be predicted, input them into the preoperative prediction model, and output the prediction result of the patient's condylar bone resorption risk.

Images