CT image data acquisition and screening method and system for predicting tooth root absorption

Abstract

The invention discloses a CT image data acquisition and screening method and system for predicting tooth root absorption. The method comprises the following steps: S1, processing a standardized CT image, and segmenting a region of interest; s2, carrying out feature extraction on the VOI; s3, performing three-level feature screening on the radiomics features, firstly performing first-level screening based on statistics, and retaining a plurality of features; performing correlation screening on the plurality of features, and finally performing minimum absolute shrinkage and selection operator regression screening to screen out radiomics features; s4, constructing a radiomics label based on the screened radiomics features; s5, repeating the steps S1-S4, and constructing a clinical tag based on the clinical features; and S6, constructing a radiation-clinical combined line graph based on the radiomics labels and the clinical labels. According to the method, a high-precision and visual solution is provided for early prediction of tooth root absorption in orthodontic treatment, and the oral health prognosis effect of a patient can be improved.

Description

Technical Field

[0001] This invention relates to a method for acquiring and filtering CT image data for predicting tooth root resorption, and also to a system for acquiring and filtering CT image data for predicting tooth root resorption, belonging to the field of image data processing technology. Background Technology

[0002] In the field of orthodontic treatment, orthodontically induced inflammatory root resorption (OIIRR) is a common complication that adversely affects patients' oral health. The risk factors for OIIRR are not yet fully understood, but it is speculated that they may involve both patient-related factors and iatrogenic factors. These factors may act individually or synergistically, increasing the complexity of preventing and controlling OIIRR.

[0003] In the early stages of OIIRR, imaging findings are often subtle, hindering early diagnosis. Furthermore, the CBCT equipment used by different medical institutions varies in parameters, such as voltage and resolution, leading to shifts in image grayscale values ​​and unstable feature extraction. This imaging instability further impacts the accuracy of conventional prediction models, rendering existing models ineffective in predicting OIIRR and failing to meet the high-precision clinical requirements for OIIRR prediction.

[0004] In their paper "Clinical Study on Root Resorption in Clear Aligner Orthodontics" (published in the *Chinese Journal of Orthodontics*, 2018, 25(4): 191-195), Zhang Zhewei et al. conducted a graded study on the degree of root resorption. They selected 150 patients treated solely with clear aligner orthodontics and used panoramic radiographs and cephalometric radiographs to measure the root resorption rate and root movement before and after treatment. They explored the relevant factors affecting the degree of root resorption and found that almost all patients experienced root resorption, with a higher probability of severe root resorption in the mandibular anterior teeth and maxillary molars. Root resorption in anterior teeth was related to the treatment duration and the vertical movement of the roots. However, this study also failed to address the problems of unclear initial imaging findings and inaccurate prediction models due to differences in CBCT equipment parameters. The industry still lacks a solution that can effectively address these challenges and improve the accuracy of OIIRR prediction. Summary of the Invention

[0005] The primary technical problem to be solved by this invention is to provide a method for acquiring and filtering CT image data for predicting root resorption.

[0006] Another technical problem to be solved by the present invention is to provide a CT image data acquisition and screening system for predicting root resorption.

[0007] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0008] According to a first aspect of the present invention, a method for acquiring and filtering CT image data for predicting tooth root resorption is provided, comprising the following steps:

[0009] S1: Process the standardized CT image and segment the region of interest;

[0010] S2: Extract features from VOI;

[0011] S3: Implement a three-level feature screening for radiomics features. First, perform a first-level screening based on statistics to retain multiple features. Then, perform correlation screening on these multiple features. Finally, perform minimum absolute shrinkage and selection operator regression screening to select radiomics features.

[0012] S4: Construct radiomics labels based on the selected radiomics features;

[0013] S5: Repeat steps S1 to S4 to construct clinical labels based on clinical characteristics;

[0014] S6: Construct a combined radiomics-clinical nomogram based on radiomics and clinical labels.

[0015] The nomogram integrates gender, age, treatment time, tooth extraction status, malocclusion classification, type of orthodontic appliance, radiomics characteristics, and severity of root resorption, converting each variable into a corresponding score.

[0016] Preferably, the minimum absolute shrinkage and selection operator regression screening compresses irrelevant feature coefficients to zero through the regularization parameter λ.

[0017] Preferably, the minimum absolute shrinkage and selection operator regression screens out four types of radiomics features, including:

[0018] Category 1: Morphological Features

[0019] Category 2: Texture Features

[0020] Category 3: Wavelet domain features

[0021] Category 4: Statistical Distribution Characteristics

[0022] Preferably, the minimum absolute shrinkage and selection operator regression screens out eight radiomics features, including Angle classification, spatial dependence entropy of gradient images, gray-level contrast / dissimilarity after squared filtering, gray-level non-uniformity of gradient images, average intensity within HLL wavelet decomposition subbands, run entropy of local binary patterns, and intensity skewness of local binary patterns.

[0023] Preferably, step 6 includes the following sub-steps:

[0024] S61: Evaluate diagnostic efficacy using receiver operating characteristic ROC curves in an independent validation queue;

[0025] S62: The calibration performance was evaluated using calibration curves and the Hosmer-Lemeshow goodness-of-fit test;

[0026] S63: Verify the clinical applicability of the model through decision curve analysis.

[0027] Preferably, step 6 employs 7-fold cross-validation to determine the optimal λ value using the minimum error criterion. The λ value is the process value of the minimum absolute shrinkage and selection operator regression method in feature screening.

[0028] Preferably, step 5 includes the following sub-steps:

[0029] S51: Clinical characteristics were screened using univariate analysis, with a p-value threshold of <0.05;

[0030] S52: Apply the same machine learning algorithm as in step 4;

[0031] S53: A fixed 5-fold cross-validation scheme and test set are used to ensure the comparability of results.

[0032] Preferably, in step 1, the window width and window level are standardized, and the voxel values ​​are Z-score standardized to eliminate the influence of device differences on feature extraction.

[0033] According to a second aspect of the present invention, a CT image data acquisition and screening system for predicting tooth root resorption is provided, comprising a processor and a memory, the processor and the memory being coupled together; wherein the memory is used to store a computer program; and the processor is used to run the computer program stored in the memory to perform the CT image data acquisition and screening method for predicting tooth root resorption as described above.

[0034] Compared with existing technologies, this invention uses a three-level feature screening strategy (statistical screening, correlation screening, and LASSO regression screening) to efficiently screen key features from a large number of radiomics features and construct a radio-clinical joint nomogram by combining clinical features. Compared with existing technologies, this invention has the following significant technical advantages: First, by validating the model's cross-institutional applicability through two-center data, it overcomes the limitations of single-center data, achieving an external validation AUC of 0.786, indicating good stability of the model across different devices and cases. Second, by employing a two-stage screening strategy (statistical filtering + LASSO) combined with ensemble algorithms such as random forest, the AUC of the training set is increased to 0.876, while the AUC of the validation set decreases by only 0.095, effectively avoiding overfitting. Third, it transforms abstract AI predictions into a visual scoring system, allowing doctors to quickly assess risks based on factors such as age, gender, and texture features, thus improving clinical applicability. Fourth, it incorporates more readily available clinical variables, lowering the data collection threshold and expanding the model's applicability. Fifth, through rigorous feature screening and cross-validation, it enhances the model's stability and generalization ability in independent validation cohorts, adapting to differences in multi-center data, and providing a high-precision, visualized solution for early prediction of root resorption in orthodontic treatment, contributing to improved oral health outcomes for patients. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall process of the method for acquiring and screening CT image data for predicting tooth root resorption in the first embodiment of the present invention;

[0036] Figure 2 for Figure 1 A flowchart illustrating the steps involved in constructing a radiomics model;

[0037] Figure 3 for Figure 2 Characteristic features and p-values ​​in the Mann-Whitney U test results;

[0038] Figure 4 The 93 features are displayed using t-distributed random neighborhood embeddings.

[0039] Figure 5A A graph of the coefficients for the 7-fold validation obtained from Laszlo regression;

[0040] Figure 5B The plot shows the mean standard error obtained from Lass-So regression;

[0041] Figure 5C This is a schematic diagram illustrating the accuracy of the 7-fold validation obtained from Lass-So regression.

[0042] Figure 6 A Delong graph representing the training set for the performance of the nocline plot;

[0043] Figure 7 A Delong plot for the test set representing the performance of the nodal plot;

[0044] Figure 8 This is a line diagram from the first embodiment of the present invention;

[0045] Figure 9 This is the calibration curve of the nomogram in the first embodiment of the present invention;

[0046] Figure 10 During the testing process of the first embodiment of the present invention, the clinical feature labels, radiomics feature labels, and DCA curves of the radiomics-clinical nomogram are shown.

[0047] Figure 11 This is a schematic diagram of the system structure provided for the second embodiment of the present invention. Detailed Implementation

[0048] The technical content of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0049] First Embodiment

[0050] like Figure 1 As shown, this invention uses a dataset of CT images from 140 eligible patients as the total sample size. Of these, 98 patients from Center 1 constitute the training dataset (dataset A, n = 98), and 42 patients from Center 2 constitute the external validation dataset (dataset B, n = 42). Demographic characteristics, clinical parameters, and outcome endpoints of these two patient cohorts were systematically collected.

[0051] When training with the training group, only the 98 cases in the training group were stratified by risk (high risk / low risk), and three models were independently constructed based on this data: a pure radiomics model, a pure clinical model, and a nomogram model (integrating the pure radiomics model and the pure clinical model).

[0052] During the evaluation phase, the three constructed models (nomogram, radiomics, and clinical) were tested and compared on two populations; internal validation and overfitting checks were then performed using the training group. Finally, an independent external test group (42 cases from different centers, used to rigorously validate the model's generalization ability) was used to compare the efficacy evaluation results of the three models.

[0053] The method for acquiring and filtering CT image data for predicting tooth root resorption provided in the first embodiment of the present invention includes at least the following steps.

[0054] S1: Process the standardized CT image and segment the region of interest (VOI).

[0055] Based on the standardization of CT / MR parameters according to the image scanning protocol, all standard CT images in the training group were converted to 1×1×1mm. 3 Voxel values ​​were Z-score normalized. A window level (WL) of 127 HU and a window width (WW) of 255 HU were used to uniformly calibrate the Henle unit values ​​of each image to eliminate the influence of device differences on feature extraction.

[0056] The VOI (Visible Occlusion) is defined as the assessment area within the visible margin of the tooth, which is verified by two highly qualified dentists (both with over ten years of clinical experience) using Varian Medical Systems' 3Dslicer software (version 5.6.1). Figure 2 As shown, the researchers used the axial view manual annotation tool of ITK-SNAP software (version 3.8; http: / / www.itksnap.org) to perform fine segmentation of the VOI on all two-dimensional slices. Other tissue structures were specifically avoided and excluded during the segmentation process.

[0057] S2: Feature extraction of VOI

[0058] Based on the radiomics feature extraction function of the Pyradiomics toolkit (http: / / pyradiomics.readthedocs.io), multiple feature classes were manually selected: (I) geometric features, (II) intensity features, and (III) texture features. Geometric features describe the three-dimensional morphological attributes of the tooth root, intensity features quantify the first-order statistical distribution of voxel values ​​within the target region, and texture features capture spatial patterns and higher-order correlations between the intensities of adjacent voxels. Texture features were calculated using multiple methods, including the Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRLM), Gray-Level Region Size Matrix (GLSZM), and Neighborhood Gray-Level Difference Matrix (NGTDM). A total of 1550 features across 7 categories were extracted, including 298 first-order features (intensity features), 12 geometric features, and the remainder being texture features.

[0059] S3: Implement three-level feature screening for radiomics features

[0060] See Figure 2 This step first performs a primary screening based on statistics to retain multiple features; then it performs a correlation screening (secondary screening) on ​​these multiple features; and finally it performs a LASSO (minimum absolute contraction and selection operator) regression screening (tertiary screening) to select radiomics features.

[0061] (1) Statistical screening

[0062] Statistical screening involved performing the Mann-Whitney U test on all features, retaining significant features with p-values ​​< 0.05 (93 features were initially screened). Detailed feature information and corresponding p-values ​​are shown below. Figure 3 As shown.

[0063] Figure 3 The graph displays different statistical characteristics of different features. Among them, the distribution characteristics of shape (morphological features) in the graph are: the violin body is tall and slender with a symmetrical shape (narrow width), and the data points are densely distributed near the median line (horizontal axis ≈ 0), with almost no discrete points (no extended "whiskers" on both sides). This indicates that the three-dimensional geometric morphology of the tooth root is highly consistent, suggesting that the patient group has similar anatomical structures in the target area, indicating that this feature is a stable indicator for disease classification.

[0064] Table 1. Distribution characteristics of each feature in the Mann-Whitney U test results.

[0065]

[0066] Using the Mann-Whitney U test to eliminate statistical noise can remove features that are irrelevant to OIIRR (such as more than 90% redundant features), thus avoiding interference from irrelevant variables in the subsequent model. At the same time, because the Mann-Whitney U test does not rely on the assumption of normal distribution of data, it is more robust to outliers in clinical data and has the advantages of nonparametric testing, which is conducive to improving the robustness of the model.

[0067] (2) Correlation screening

[0068] The correlation between features is assessed based on Spearman's rank correlation coefficient. Figure 3 The statistically selected features are then subjected to relevance screening. A greedy recursive deletion strategy is adopted: when the correlation coefficient between any two features is greater than 0.9, the feature with the highest redundancy in the current feature set is deleted each time, ultimately retaining 30 features.

[0069] Iteratively removing the most redundant features from highly correlated features eliminates linear dependencies between features (e.g., GLCM_entropy and GLSZM_entropy are often highly correlated), prevents model coefficient distortion, and resolves multicollinearity. Furthermore, to enhance biological independence, features representing independent pathological mechanisms (e.g., morphology, necrosis, and texture each retain at least one representative feature) are retained from the 30 features after correlation screening. Moreover, the number of features is significantly reduced from 93 to 30, lowering the risk of overfitting in subsequent models.

[0070] (3) LASSO regression screening

[0071] The Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to the discovery dataset, and irrelevant feature coefficients were compressed to zero using a regularization parameter λ. LASSO regression was used to select eight imaging features across four classes (excluding clinical features) as key biomarkers in the radiomics model for predicting OIIRR.

[0072] In this way, not only is intramodal redundancy eliminated before LASSO regression (early stage), but intermodal redundancy is also eliminated. Through a progressive, defensive screening process (statistical testing → correlation resolution → sparsity constraints), over 90% of noisy / redundant features are removed before modeling, ensuring the final model is built on robust, robust features. This is particularly crucial in medical scenarios with limited sample sizes, directly determining the model's generalization ability in real-world clinical settings.

[0073] The aforementioned use of regularization to compress non-critical feature weights not only automatically allocates feature weights, retaining only the features that contribute most to prediction and enabling efficient screening of biomarkers, but also constrains model complexity, avoiding overfitting to noisy features (such as acquisition artifacts or physiological variations in medical images), thereby enhancing stability in independent validation cohorts. The synergistic effect of these two methods allows the model to retain strong predictive features while resisting noise interference, making it particularly suitable for robust modeling of high-dimensional medical data (such as radiomics), balancing model accuracy and generalization ability.

[0074] The aforementioned four categories of eight image features include:

[0075] Category 1: Morphological Features

[0076] original_shape_Elongation (severity of root resorption)

[0077] Category 2: Texture Features

[0078] igradient_glcm_DependenceEntropy (spatial dependency entropy of gradient image)

[0079] is square_glcm_ldn / ldmn (grayscale contrast / dissimilarity after squared filtering)

[0080] igradient_gldm_GrayLevelNonUniformity (gradient image gray-level non-uniformity)

[0081] Category 3: Wavelet domain features

[0082] iwavelet_HLL_firs torder_Mean(Average intensity within HLL wavelet decomposition subbands)

[0083] ilbp_3D_k_glrlm_RunEntropy (Run entropy in local binary mode)

[0084] Category 4: Statistical Distribution Characteristics

[0085] ilbp_3D_k_firs torder_Skewne ss(Intensity skewness of local binary mode)

[0086] Seven-fold cross-validation was used to determine the optimal λ value based on the minimum error criterion. The λ value is the process value for LASSO feature selection. LASSO controls the degree of feature compression through the λ value. We divided the training data into seven equal parts, six parts as training data and one part as test data, and repeated this process seven times. The λ value corresponding to the minimum error was taken as the final λ value. The seven-fold cross-validation selection avoids feature over- or under-selection due to improper λ value selection.

[0087] Table 2. Radiomic features selected by LASSO

[0088]

[0089] As can be seen from Table 2, original_shape_Elongation is the dominant predictor (the largest positive coefficient, significantly larger than other coefficients), which points to biomechanical failure (the elongated shape of the root resorption area around the implant reflects stress shielding effect or fretting wear) and inflammation progression markers (the elongation shape is related to the direction of granulation tissue expansion, indicating progressive root resorption).

[0090] 1) Enhanced stability through a triple defense mechanism: strict single-factor filtering of all features (Mann-Whitney U test) to eliminate statistical noise; 2) Noise interference is eliminated by compressing the coefficients of non-predictive features to 0 (only 8 features are selected); 3) Comprehensive morphological, textural, and wavelet domain features to cover the full-scale pathology at the cellular level (texture) and macroscopic structure (morphology), achieving multi-dimensional synergy; 4) Positive / negative coefficients are used to clearly indicate the activating / inhibiting effect of each feature on OIIRR.

[0091] Firstly, from a statistical perspective, the Mann-Whitney U test of this invention, as a non-parametric test, is insensitive to outliers in clinical data—especially important when the OIRR sample size is small and its distribution is unknown. Directly using optimal subset regression (essentially an exhaustive method) is extremely prone to overfitting in small sample scenarios.

[0092] Secondly, there's the issue of the processing logic for feature interactions. During the research, it was found that when the initial model was tested by incorporating all 1550 radiomics features, the AUC on the training set reached 0.95, but the AUC on the validation set was only 0.68, indicating severe overfitting.

[0093] The study also found that performing LASSO regression on clinical and radiomic features separately and then using optimal subset regression (which is done in a fragmented manner in the early analysis stages and only forcibly combined in the optimal subset regression) can lead to the omission of predictive factors that require cross-modal interaction to be effective (such as specific image texture combined with patient age). Parallel processing of wavelet-transformed features with other original image features weakens the completeness of the algorithm's assessment of relationships between features.

[0094] This invention, through rigorous prior correlation filtering, essentially preserves the "representative features" of each dimension, allowing LASSO to naturally retain valuable interaction effects when compressing features. Furthermore, it introduces cross-validation, using 7-fold cross-validation in LASSO regression to determine the optimal regularization parameter λ, avoiding over-reliance on training data and adapting to the differences in multi-center data.

[0095] S4: Based on the selected radiomics features, construct radiomics signatures.

[0096] Input the features selected by LASSO into the following machine learning algorithms, compare them, and select the best one:

[0097] Logistic Regression

[0098] Support Vector Machine (or Support Vector Machine)

[0099] Random Forest

[0100] XGBoos t

[0101] The formula for calculating radiomics scores (see Equation 1 below) was obtained using the modeling tool Python scikit-learn package. It is a weighted sum of linear combinations that preserve features (the weights are model coefficients).

[0102] S5: Repeat steps S1-S4 to construct clinical labels based on clinical characteristics.

[0103] The process for building clinical labels is the same as that for building radiomics labels:

[0104] S51: Clinical characteristics were screened using univariate analysis through the aforementioned three-level screening method (p-value threshold < 0.05);

[0105] S52: The result of applying multiple machine learning algorithms and selecting the algorithm with the best performance;

[0106] S53: A fixed 5-fold cross-validation scheme and test set are used to ensure the comparability of results.

[0107] The clinical labels selected in step S5 are gender, age, treatment time, tooth extraction status, malocclusion classification (type S), and type of orthodontic appliance (clas sify).

[0108] S6: Constructing a combined radiomics-clinical nomogram based on radiomics and clinical labels

[0109] Integrating radiomics and clinical labels to construct a combined radiomic-clinical nomogram includes the following sub-steps:

[0110] S61: Evaluate diagnostic efficacy using receiver operating characteristic ROC curves in an independent validation queue;

[0111] S62: The calibration performance was evaluated using calibration curves and the Hosmer-Lemeshow goodness-of-fit test;

[0112] S63: Validate the clinical applicability of the model through decision curve analysis (DCA), demonstrating that the model's net benefit is significantly better than the "all-treatment" or "no-treatment" strategies, thereby improving its clinical applicability.

[0113] Result Validation

[0114] To verify the effectiveness of this invention, case studies involving patients from two medical centers were conducted. During statistical analysis, cases were categorized into high-risk (≥2 grade) and low-risk (0–1 grade) subgroups based on the degree of root resorption in the training dataset. ROC curve analysis with area under the curve (AUC) was used to compare the predictive performance between models; Pearson correlation coefficients were used to assess relationships between continuous variables; and Wilcoxon rank-sum tests and t-tests were used to assess differences in non-normally distributed variables. All statistical analyses were performed using IBM SPSS Statistics 25.0 software, and graphs were output in Adobe Photoshop CC 2019.

[0115] Independent samples t-tests were used to compare patients’ clinical characteristics in the clinical baseline analysis. Table 3 shows the patients’ baseline clinical characteristics, in which there were no significant differences between the training group and the test group (P > 0.05).

[0116] Table 3 Baseline characteristics of patients in the cohort

[0117]

[0118] In Table 3, Y=sexual (F=0, M=1), tooth extraction (Yes=1, No=0), trips (fix=1, hidden=0).

[0119] Tag filtering

[0120] (1) Visualization of t-SNE samples

[0121] Ninety-three high-dimensional features, obtained through statistical screening, were visualized using t-distributed random neighborhood embedding (t-SNE) for dimensionality reduction. For example... Figure 4 As shown, samples of different categories exhibit significant distributional differences in the low-dimensional space: red-labeled samples cluster in the lower region, while blue-labeled samples concentrate in the upper right region. This spatial boundary phenomenon demonstrates that the features learned by the model have strong discriminative power for sample categories.

[0122] (2) Lass so feature selection

[0123] To establish the radiomics score (Rad-score), non-zero coefficients were selected, and a logistic regression model with minimum absolute shrinkage and selection operators was used. Figures 5A-5C This shows the coefficient for 70% verification ( Figure 5A ) and mean standard error (MSE) Figure 5B ) and the accuracy of the 70% discount verification ( Figure 5C ).

[0124] Radiomics score label=0.39175257731958585 -0.047609*is quare_glcm_Idmn+0.113146*ioriginal_shape_Elongation-0.018420*igradient_gldm_GrayLevelNonUniformity+0.000522*iwavelet_HLL_firs toorder_Mean-0.061516*ilbp_3D_k_firs toorder_Skewne s-0.000729*is quare_glcm_Idn-0.021544*ilbp_3D_k_glrlm_RunEntropy+0.002031*igradient_gldm_DependenceEntropy

[0125] Each item represents a standardized radiomics feature, and the coefficient reflects its contribution weight to the score. Negative values ​​indicate a negative association.

[0126] Constructing radiomics features

[0127] like Figure 5C As shown, the accuracy and robustness comparison results of multiple models after 7-fold cross-validation are presented. Considering both training and test set performance, among the classifiers constructed using radiomics features—Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, Random Forest, Extra Trees, XGBoost, and LightGBM—Random Forest exhibits the best performance (training set accuracy 0.804, test set accuracy 0.744, see Table 3). The ROC curve shows that its area under the curve (AUC) on the training and test sets reaches 0.876 and 0.771, respectively. Therefore, Random Forest was selected as the base model when constructing radio-clinical nomograms.

[0128] Table 4 Performance Comparison of Each Model

[0129]

[0130] Constructing clinical feature tags

[0131] (1) Correlation between clinical characteristics and outcomes

[0132] The correlation between clinical characteristics and root resorption grading was analyzed using univariate logistic regression and multivariate Cox regression, as shown in Table 4. Correlation was assessed using odds ratios (OR), 95% confidence intervals (95% CI), and p-values ​​in both univariate (_UNI) and multivariate (_MULTI) analyses. Sex and age showed significant correlations in the multivariate analysis. For sex, both the multivariate OR and 95% CI were 2.718. For age, both the multivariate OR and 95% CI were 1.0, both statistically significant (P = 0.00). Other characteristics (such as type, classification, etc.) did not show significant statistical differences, indicating no significant correlation with the predicted outcomes.

[0133] Table 5. Correlation between clinical features and root resorption grading

[0134]

[0135] (2) Clinical characteristic labels

[0136] Using five clinical features (see Table 5), the modeling performance of classifiers such as Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, Random Forest, Extra Trees, XGBoost, AdaBoost, and LightGBM on the five clinical features was compared. The optimal model, AdaBoost algorithm, was determined to be the best, achieving an area under the curve (AUC) of 0.750 and 0.670 in the training and testing queues, respectively.

[0137] (3) Radiology-Clinical Combined Nodal Chart

[0138] In the training cohort, clinical and radiomics labels showed a good fit; in the testing cohort, clinical labels exhibited underfitting, while radiomics labels remained stable. The joint nomogram constructed using the random forest algorithm to integrate clinical and radiomics labels performed optimally (see Delong test comparison results for details). Figure 6 and Figure 7 ).

[0139] like Figure 8 As shown, the nomogram integrates gender (Y), clinical variables (such as age and treatment duration), categorical variables (such as tooth extraction status and malocclusion type), appliance type, radiomics characteristics, and the severity of root resorption, converting each variable into a corresponding score. Ultimately, it predicts the probability of risk based on the total score. Its core function is as a visual tool for clinical prediction, helping physicians or researchers quickly estimate the risk of specific outcomes using multidimensional data, thereby facilitating clinical decision-making and research analysis.

[0140] Figure 9 The calibration curves of the nomograms in the dataset show good agreement between the predicted values ​​and the training and testing cohorts. The Hosmer-Leme show p-values ​​were used to evaluate the clinical feature labels, radiomics feature labels, and nomograms (see Table 6). This indicates that the nomograms fit perfectly in both the training and testing cohorts. Therefore, using this visual score of the nomogram, physicians can quickly assess risk probabilities through the total score. Figure 10 The DCA (Distributed Clinical Analysis) of clinical feature labels, radiomics feature labels, and radiomics-clinical nomograms during testing is shown. Compared to the case without any predictive model (i.e., all-treatment or no-treatment regimen), radiomics nomograms showed a significant benefit to patient intervention at the predicted probability, with a higher net benefit compared to both clinical and radiomics feature labels. Preoperative prediction using radiomics nomograms has been shown to have better clinical efficacy.

[0141] Table 6. P-values ​​of clinical feature labels, radiomics feature labels, and nomograms.

[0142]

[0143] Second Embodiment

[0144] Based on the above methods for acquiring and filtering CT image data for predicting tooth root resorption, such as Figure 11 As shown, a second embodiment of the present invention further provides a CT image data acquisition and screening system for predicting tooth root resorption. The system includes one or more processors and a memory. The memory is coupled to the processors and is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the CT image data acquisition and screening method for predicting tooth root resorption as described in the above embodiment.

[0145] The processor controls the overall operation of the system to complete all or part of the steps in the aforementioned method for acquiring and screening CT image data for predicting tooth root resorption. This processor can be a central processing unit (CPU), graphics processing unit (GPU), field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), digital signal processing (DSP) chip, etc. The memory stores various types of data to support the operation of the system. This data may include, for example, instructions for any application or method operating on the system, as well as application-related data. The memory can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, etc.

[0146] In another exemplary embodiment, the present invention also provides a computer-readable storage medium including program instructions that, when executed by a processor, implement the steps of the method for acquiring and filtering CT image data for predicting tooth root resorption in any of the above embodiments. For example, the computer-readable storage medium may be the memory including the program instructions described above, which can be executed by a system processor to complete the method for acquiring and filtering CT image data for predicting tooth root resorption described above, and achieve the same technical effects as the method described above.

[0147] It should be noted that the above embodiments are merely illustrative examples. The technical solutions of the various embodiments can be combined, and the order of the steps can be changed, all of which are within the scope of protection of this patent.

[0148] The above provides a detailed description of the method and system for acquiring and filtering CT image data for predicting tooth root resorption provided by this invention. Any obvious modifications made by those skilled in the art without departing from the essence of this invention will constitute an infringement of the patent rights of this invention and will incur corresponding legal liability.

Claims

1. A method for acquiring and filtering CT image data for predicting tooth root resorption, characterized in that... Includes the following steps: S1: Process the standardized CT image and segment the region of interest; S2: Extract features from VOI; S3: Implement a three-level feature screening for radiomics features. First, perform a first-level screening based on statistics to retain multiple features. Then, perform correlation screening on these multiple features. Finally, perform minimum absolute shrinkage and selection operator regression screening to select radiomics features. S4: Construct radiomics labels based on the selected radiomics features; S5: Repeat steps S1 to S4 to construct clinical labels based on clinical characteristics; S6: Construct a combined radiomics-clinical nomogram based on radiomics and clinical labels. The nomogram integrates gender, age, treatment time, tooth extraction status, malocclusion classification, type of orthodontic appliance, radiomics characteristics, and severity of root resorption, converting each variable into a corresponding score.

2. The method for acquiring and filtering CT image data for predicting tooth root resorption as described in claim 1, characterized in that: The minimum absolute shrinkage and selection operator regression screening uses a regularization parameter λ to compress irrelevant feature coefficients to zero.

3. The method for acquiring and filtering CT image data for predicting tooth root resorption as described in claim 2, characterized in that: The minimum absolute shrinkage and selection operator regression screened out four types of radiomics features, including: Category 1: Morphological Features Category 2: Texture Features Category 3: Wavelet domain features Category 4: Statistical distribution characteristics.

4. The method for acquiring and filtering CT image data for predicting tooth root resorption as described in claim 3, characterized in that: The minimum absolute shrinkage and selection operator regression screened out eight radiomics features, including root resorption elongation, spatial dependence entropy of gradient image, gray-level contrast / dissimilarity after square filtering, gray-level non-uniformity of gradient image, average intensity within HLL wavelet decomposition subband, run entropy of local binary mode, and intensity skewness of local binary mode.

5. The method for acquiring and filtering CT image data for predicting tooth root resorption as described in claim 4, characterized in that... Step 6 includes the following sub-steps: S61: Evaluate diagnostic efficacy using receiver operating characteristic ROC curves in an independent validation queue; S62: The calibration performance was evaluated using calibration curves and the Hosmer-Lemeshow goodness-of-fit test; S63: Verify the clinical applicability of the model through decision curve analysis.

6. The method for acquiring and filtering CT image data for predicting tooth root resorption as described in claim 5, characterized in that: Step 6 employs 7-fold cross-validation to determine the optimal λ value using the minimum error criterion. The λ value is the process value of the minimum absolute shrinkage and selection operator regression method in feature screening.

7. The method for acquiring and filtering CT image data for predicting tooth root resorption as described in claim 1, characterized in that... Step 5 includes the following sub-steps: S51: Clinical characteristics were screened using univariate analysis, with a p-value threshold of <0.05; S52: Apply the same machine learning algorithm as in step 4; S53: A fixed 5-fold cross-validation scheme and test set are used to ensure the comparability of results.

8. The method for acquiring and filtering CT image data for predicting tooth root resorption as described in claim 1, characterized in that... In step 1, the window width and window level are standardized, and the voxel values ​​are Z-score standardized to eliminate the impact of device differences on feature extraction.

9. A system for acquiring and filtering CT image data to predict tooth root resorption, characterized in that... The device includes a processor and a memory, the processor and the memory being coupled together; wherein the memory is used to store a computer program; the processor is used to run the computer program stored in the memory to implement the method for acquiring and filtering CT image data for predicting tooth root resorption as described in any one of claims 1 to 8.

Images