Establishment method of model for predicting bone and soft tissue sarcoma recurrence

Abstract

The invention provides a method for establishing a model for predicting bone and soft tissue sarcoma recurrence, which comprises the following steps of: 1, collecting and sorting patient data, and selecting clinical variables possibly influencing recurrence risk; 2, pre-processing image data, wherein the pre-processing comprises two parts of tumor ROI manual segmentation and patient MRI image calibration; 3, extracting first-order and second-order radiomics features from the tumor ROI; 4, feature selection and optimization; 6, model construction: combining the obtained clinical variables related to bone and soft tissue sarcoma recurrence with radiomics scores, establishing a clinical radiomics model, and drawing a column graph; and establishing a clinical-MRI radiomics combined model for predicting bone and soft tissue sarcoma recurrence.

Description

Technical Field

[0001] The present invention belongs to the field of radiomics and provides a method for establishing a model for predicting the recurrence of bone and soft tissue sarcoma. Using the model establishment method provided by the present invention, a clinical-MRI radiomics joint model for predicting the recurrence of bone and soft tissue sarcoma can be established. This model can be used to assist doctors in accurately and early predicting the postoperative survival outcomes of patients with bone and soft tissue sarcoma, and can better quantify the risk of recurrence. Background Art

[0002] Sarcoma refers to malignant tumors that occur in mesenchymal tissues, including primary malignant bone tumors and soft tissue sarcomas. The overall incidence of bone and soft tissue sarcomas is low, accounting for about 1% of adult malignant tumors and about 15% of children, but the malignancy is high, most tumors have a poor prognosis, and are prone to recurrence after resection. Surgical resection is the basic treatment for patients with primary sarcoma, and adjuvant radiotherapy and chemotherapy are considered to be important methods to improve the prognosis of patients. However, despite timely surgical treatment, nearly 40% of patients with bone or soft tissue sarcoma still experience local recurrence (LR), distant metastasis (DM), or both during follow-up. Patients with sarcoma recurrence have a poor prognosis, and the long-term survival rate after recurrence is <20%. Sarcoma recurrence is affected by many factors, including tumor site, tumor size, pathological grade, overall health status of the patient, and whether standardized treatment is received. In general, if the sarcoma can be discovered and treated in time after resection and the tumor has not metastasized extensively, the patient's survival will be relatively long. Yang et al. reported that the 10-year local recurrence rate of sarcoma patients who received postoperative radiotherapy was 0%-22%, which was significantly improved. A meta-analysis of 14 randomized trials of 1568 patients with sarcoma treated with anthracycline-based chemotherapy found that 10-year disease-free survival (DFS) was improved by approximately 10% after chemotherapy.

[0003] Sarcomas with high pathological grade, tumors larger than 5 cm, and deep fascia involvement are more likely to recur. However, these factors mainly depend on postoperative pathological histological examination, which is prone to differences between observers and cannot guide preoperative clinical decisions. Therefore, it is urgent to find preoperative non-invasive and reliable evaluation indicators that indicate the prognosis after sarcoma resection to identify patients at high risk of recurrence and optimize personalized treatment decisions.

[0004] Radiomics analysis is fundamentally a high-throughput data-driven processing framework that can objectively and comprehensively quantify tumor characteristics by converting traditional images into advanced features. Many radiomics-based evaluation systems have been used for purposes such as tumor diagnosis, staging, prognosis, gene expression, and treatment response prediction. Chen et al. established a joint model based on radiomics features extracted from contrast-enhanced fat-suppressed T1-weighted (CE FS T1-w) images and whether joint invasion and perivascular involvement were involved to predict osteosarcoma recurrence. Liu et al. constructed a nomogram based on radiomics features extracted from conventional and enhanced MRI scan sequences combined with staging systems and Ki67 to predict tumor recurrence in patients with soft tissue sarcoma (STS) undergoing surgical resection. Wang et al. extracted eight radiomics features to predict the recurrence-free survival (RFS) of patients with subsolid or pure solid stage I non-small cell lung cancer. The results showed that radiomics features could divide all patients into high-risk groups (180 of 378) and low-risk groups (198 of 378) with different RFS, which could guide postoperative treatment strategies in a personalized manner and improve survival outcomes. However, to date, there are limited studies exploring large-sample combined models of radiomics to predict recurrence or metastasis of bone and soft tissue sarcoma after treatment. Summary of the invention

[0005] The purpose of the present invention is to provide a method for establishing a clinical-MRI imaging genomics joint model for predicting the recurrence of bone and soft tissue sarcomas. The technical solution is as follows:

[0006] A method for establishing a model for predicting the recurrence of bone and soft tissue sarcoma comprises the following steps:

[0007] 1. Patient data collection and organization

[0008] Determine the inclusion and exclusion criteria for patients, and collect clinical, pathological, imaging, and follow-up data of patients with osteosarcoma and soft tissue sarcoma who meet the criteria;

[0009] Select clinical variables that may affect the risk of recurrence, including age, sex, tumor size measured on the patient's MRI images, observation of tumor signal heterogeneity, tumor margins, peritumoral edema, and the presence of skip lesions;

[0010] 2. Image data preprocessing, including manual segmentation of tumor ROI and calibration of patient MRI images;

[0011] Third, extract first-order and second-order radiomics features from tumor ROI;

[0012] 4. Feature Selection and Optimization

[0013] Stability verification: The stability of the extracted features was evaluated by the inter-observer ICC coefficient, and the radiomics features with ICC>0.75 were retained;

[0014] Dimensionality reduction processing: includes the following steps:

[0015] (4) For features with ICC ≥ 0.75, SelectKBest based on univariate feature selection was used to analyze the relationship between features and tumor prognosis, and features with P < 0.05 were screened to obtain the initial feature set;

[0016] (5) The maximum correlation-minimum redundancy (mRMR) method was used to eliminate confounding factors, calculate the correlation between features and tumor recurrence, and the redundancy between features, and select a feature subset with high correlation from the initial feature set;

[0017] (6) Using the least absolute shrinkage and selection operator (LASSO) method, the feature subset with high correlation obtained in step (2) is further reduced in dimension, and the feature set with the best λ value is screened out using 5-fold cross validation, excluding the features with a variable coefficient of 0;

[0018] 6. Model Construction

[0019] According to the selected features and the corresponding variable coefficients, the calculation method of radiomics score was constructed;

[0020] The screened radiomics features were used to calculate T1 radiomics, T2FS radiomics, and T1+T2FS combined radiomics scores;

[0021] To select clinical variables associated with recurrence of bone and soft tissue sarcomas;

[0022] The obtained clinical variables related to the recurrence of bone and soft tissue sarcoma were combined with the radiomics score to establish a clinical radiomics model and draw a nomogram;

[0023] To establish a joint clinical-MRI radiomics model for predicting recurrence of bone and soft tissue sarcomas.

[0024] Furthermore, step one also includes: cleaning the collected patient data to remove data with too many missing values ​​or abnormal data.

[0025] Furthermore, the manual segmentation of the tumor ROI in step 2 includes: selecting at least two physicians, each physician outlining the tumor ROI of the patient with bone or soft tissue sarcoma on the T2FS image, and copying the three-dimensional ROI to the patient's T1WI image to ensure the consistency of the tumor ROI; physician B performs manual segmentation of the tumor ROI in the same way.

[0026] Furthermore, in step 2, the patient's MRI images are calibrated to ensure image quality and consistency, including image intensity discretization, intensity normalization, and isotropic resampling processing.

[0027] Furthermore, the first-order features in step three include tumor size, shape, mean, standard deviation, skewness, and kurtosis; the second-order features include gray-level co-occurrence matrix, gray-level size area matrix, gray-level dependency matrix, gray-level dependency matrix, Gaussian Laplace, and wavelet coefficients.

[0028] Furthermore, in step six, univariate Cox regression analysis was used to select clinical variables associated with recurrence of bone and soft tissue sarcoma, including peritumoral edema and skip lesions.

[0029] Furthermore, in step 6, a clinical + T1 radiomics model was established. Table 1 shows the optimal feature distribution of the established model. The Rad score calculation formula based on the optimal feature is:

[0030] F1×411.314+F2×318.092+F3×555.452+F4×430.831+F5×270.128+F6×536.202+F7×280.188+F8×399.036+F9×434.576+F10×274.616+F11×469.686+F12×243.742+F13×319.444+F14×312.036+F15×242.110+F16×664.443+F17×383.167+F18×1134.098+F19×329.060+F20×439.065.

[0031] Table 1 Distribution of optimal radiomics features based on T1 images

[0032]

[0033]

[0034] Furthermore, in step six, clinical + T2FS radiomics was established. Table 2 shows the optimal feature distribution of the established model. The Rad score calculation formula based on the optimal feature is: F1×9.685+F2×9.555+F3×12.335+F4×73.308+F5×11.265+F6×14.440+F7×27.583+F8×14.695+F9× 16.584+F10×38.431+F11×12.981+F12×12.113+F13×28.157+F14×11.446+F15×82.305+F16×13.691+F17×12.263+F18×10.515+F19×29.029+F20×12.143+F21×15.462.

[0035] Table 2 Distribution of optimal radiomics features based on T2FS images

[0036]

[0037]

[0038] Furthermore, in step 6, a clinical + T1 + T2FS combined radiomics model was established. Table 3 shows the optimal feature distribution of the established model. The Rad score calculation formula based on the optimal feature is:

[0039] F1×0.083+F2×0.138+F3×0.100+F4×0.159+F5×0.187+F6×0.091+F7×0.114+F8×0.067+F9×0.065+F10×0.107+F11×0.060+F12×0.155.

[0040] Table 3 Distribution of optimal radiomics features based on T1+T2FS images

[0041]

[0042]

[0043] Using the modeling method provided by the present invention, a clinical-radiological combined model can be established based on the patient's preoperative clinical data and radiomics characteristics. The established model can be used to assist doctors in accurately judging the severity of the patient's condition and tumor characteristics. It has high accuracy and reliability in predicting the recurrence-free survival of sarcoma patients, and indicates the patient's recurrence probability, thereby assisting in formulating a more personalized treatment plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 .C-index of different radiomics models

[0045] Figure 2 .Kaplan-Meier curves showed that the recurrence-free survival rate of the high radiomics score group was significantly worse in both the training set and the validation set. a, c, e, g, i, and k are the training sets of T1, T2FS, T1+T2FS, T1+clinical, T2FS+clinical, and T1+T2FS+clinical, respectively. b, d, f, h, j, and i are the validation sets of T1, T2FS, T1+T2FS, T1+clinical, T2FS+clinical, and T1+T2FS+clinical, respectively.

[0046] Figure 3 A nomogram model was established by combining the radiomic features and clinical characteristics of different sequences to predict 1-, 3-, and 5-year recurrence-free survival rates

[0047] Figure 4Decision curves of various models for predicting recurrence-free survival of patients after resection of bone and soft tissue sarcoma. DETAILED DESCRIPTION

[0048] The steps for building a clinical-radiological genomics model for the risk of recurrence of bone and soft tissue sarcoma include data collection, preprocessing, feature extraction, feature selection and optimization, model construction and validation, etc. The following is a detailed description of the steps:

[0049] 1. Data collection and organization

[0050] Patient selection: Inclusion and exclusion criteria were determined as follows: (a) patients diagnosed with suspected bone or soft tissue sarcoma by a multidisciplinary team; (b) confirmed diagnosis of bone or soft tissue sarcoma by a dedicated sarcoma pathologist after CT-guided puncture biopsy, incisional biopsy or resection; (c) standardized surgical treatment by bone and soft tissue oncologists; (d) standardized magnetic resonance imaging of the tumor before surgery; (e) routine clinical features can be extracted; (f) follow-up for 1 year or more after surgical resection. Exclusion criteria were as follows: (a) patients who had received treatment (chemotherapy or surgical resection) before MRI scanning; (b) patients for whom accurate pathological results were not available; (c) patients who had undergone MRI examinations in other hospitals before treatment and whose MRI data were lost; (d) patients diagnosed with other cancers at the same time; (e) clinical information was not available; (f) patients died due to causes other than bone sarcoma or soft tissue sarcoma; (h) patients with a follow-up time of less than one year. Clinical, pathological, imaging and other follow-up data of patients with bone sarcoma and soft tissue sarcoma who met the criteria were collected.

[0051] Variable selection: Clinical variables that may affect the risk of recurrence were selected, including age, sex, tumor size measured on MRI images, observation of tumor signal heterogeneity, tumor margin, peritumoral edema, and the presence or absence of skip lesions.

[0052] Data cleaning and organization: Clean the collected data and remove data with too many missing values ​​or abnormal data.

[0053] 2. Preprocessing

[0054] ROI (Region of Interest) segmentation: Doctor A manually delineated the tumor ROI of 120 patients with bone or soft tissue sarcoma on T2 FS images and copied the 3D ROI to the patient's T1WI images to ensure the consistency of the tumor ROI. Doctor B performed manual tumor ROI segmentation in the same way.

[0055] Image calibration: Calibrate the patient's MRI images to ensure image quality and consistency, including image intensity discretization, normalization, isotropic resampling and other processes. 1. Image discretization: Discretize the intensity values ​​in the MRI image and use a fixed bin width of 10 for discretization. This step converts the continuous intensity values ​​in the image into discrete values ​​for subsequent processing. 2. Intensity normalization: The discretized image is subjected to intensity normalization so that its mean is 100 and the standard deviation is 1. This adjusts the intensity range of the image to a standardized scale, which helps the stability and comparability of subsequent processing. 3. Isotropic resampling: The image is isotropically resampled and the voxel size of the image is adjusted to 1x1x1 mm. This step is performed by using B-spline interpolation to maintain the spatial resolution of the image and ensure that the voxel size of the image is the same in three directions.

[0056] 3. Feature Extraction

[0057] First-order and second-order radiomics features were extracted from tumor ROIs based on the software system of ResearchPortal V1.0 (Shanghai United Imaging Intelligence, Co., Ltd.) developed by Shanghai United Imaging Intelligence Medical Technology Co., Ltd. These features include first-order features such as tumor size, shape, mean, standard deviation, skewness, and kurtosis, and second-order features such as gray-level co-occurrence matrix, gray-level size area matrix, gray-level dependency matrix, gray-level dependency matrix, Gaussian Laplace, and wavelet.

[0058] 4. Feature Selection and Optimization

[0059] Stability verification: The stability of the extracted features was evaluated by the ICC (intraclass correlation coefficient) method between observers (i.e., physician A and physician B), and the imaging features with ICC>0.75 were retained.

[0060] Dimensionality reduction: Due to the large number of extracted features, dimensionality reduction is required to reduce the amount of calculation and avoid overfitting. For features with ICC ≥ 0.75, SelectKBest based on univariate feature selection is used to analyze the relationship between features and tumor prognosis, and the best features with P < 0.05 are selected. SelectKBest is a feature selection algorithm based on statistical methods. It selects the most relevant or important K features from the original feature set. The core idea is to evaluate the strength of association between each feature and the target variable and select the K features with the strongest association. Subsequently, the maximum correlation-minimum redundancy method (mRMR) is used to eliminate confounding factors, calculate the correlation between features and tumor recurrence, and the redundancy between features, and select a subset of features with higher correlation from the initial feature set. mRMR is a filtering feature selection method that finds a set of features in the original feature set that have the greatest correlation with the final output result, but the correlation between many features is the smallest. Finally, the least absolute shrinkage and selection operator (LASSO) method was used for feature dimensionality reduction, and the feature set with the best λ value was screened using 5-fold cross validation. The features with variable coefficients of 0 were excluded, and the calculation method of radiomics score was constructed based on the selected features and the corresponding variable coefficients. LASSO is a data mining method, that is, in multiple linear regression, a penalty function is added to continuously compress the coefficients, so as to achieve the purpose of streamlining the model to avoid collinearity and overfitting.

[0061] 6. Model Construction

[0062] Radiomics model: The screened radiomics features were used to calculate T1 radiomics, T2FS radiomics and T1+T2FS combined radiomics scores.

[0063] Nomogram: Univariate Cox regression analysis was used to obtain clinical variables associated with the recurrence of bone and soft tissue sarcoma. Clinical variables were combined with the above radiomics scores to establish clinical + T1 radiomics, clinical + T2FS radiomics, and clinical + T1 + T2FS combined radiomics models, and nomograms were drawn.

[0064] 7. Model Validation

[0065] External Validation: Use independent external datasets to validate the model and evaluate its performance in real-world applications.

[0066] 8. Results Analysis and Interpretation

[0067] Results: A nomogram was constructed using the risk factors of Cox regression results. The Harrell consistency index (C-index) was measured to quantify the model discrimination performance. The Kaplan-Meier method and log-rank test were used to plot and compare survival curves. Time-dependent receiver operating characteristic (timeROC) analysis was used to assess the predictive ability of the model over time. Decision curve analysis (DCA) was used to estimate the radiomics nomogram.

[0068] Through the above steps, a radiomics model for determining the risk of recurrence can be established.

[0069] The simulation experiment of the method of the present invention is as follows:

[0070] ROI (Region of Interest) segmentation: Doctor A manually delineated the tumor ROI of 120 patients with bone or soft tissue sarcoma on T2 FS images and copied the 3D ROI to the patient's T1WI images to ensure the consistency of the tumor ROI. Doctor B performed manual tumor ROI segmentation in the same way.

[0071] Image calibration: Calibrate the patient's MRI images to ensure image quality and consistency, including image intensity discretization, normalization, isotropic resampling and other processes. 1. Image discretization: Discretize the intensity values ​​in the MRI image and use a fixed bin width of 10 for discretization. This step converts the continuous intensity values ​​in the image into discrete values ​​for subsequent processing. 2. Intensity normalization: The discretized image is subjected to intensity normalization so that its mean is 100 and the standard deviation is 1. This adjusts the intensity range of the image to a standardized scale, which helps the stability and comparability of subsequent processing. 3. Isotropic resampling: The image is isotropically resampled and the voxel size of the image is adjusted to 1x1x1 mm. This step is performed by using B-spline interpolation to maintain the spatial resolution of the image and ensure that the voxel size of the image is the same in three directions.

[0072] uAI Research Portal software was used for feature selection, and each radiomics feature was z-standardized to obtain a normal distribution. For features with ICC ≥ 0.75, SelectKBest based on univariate feature selection was used to analyze the relationship between features and classification results, and the best features with P < 0.05 were selected. Subsequently, the maximum correlation-minimum redundancy method (mRMR) was used to eliminate confounding factors, calculate the correlation between features and tumor recurrence, and the redundancy between features, and select feature subsets with higher correlation from the initial feature set. Finally, the minimum absolute shrinkage and selection operator (LASSO) method was used for feature dimensionality reduction, and the feature set with the best λ value was screened using 5-fold cross validation. Features with a variable coefficient of 0 were excluded, and radiomics scores were constructed based on the selected features and their corresponding coefficients. All patients were divided into high-risk and low-risk groups using the median radiomics score as the threshold.

[0073] After reducing some radiomic features based on the variability of ROI segmentation, 2086, 1979, and 2172 features were used to develop the joint radiomics models for T1, T2FS, and T1+T2FS, respectively. During the model training process, after the aforementioned feature dimensionality reduction, 20, 21, and 12 features with non-zero coefficients were finally selected on T1, T2FS, and T1+T2FS images, respectively. Among them, 1 first-order statistical feature, 1 shape feature, 1 gray-level co-occurrence matrix feature, 2 gray-level run length matrices, 8 gray-level size region matrices, 4 gray-level dependency matrices, 1 gray-level dependency matrix, 12 Gaussian Laplacian, and 23 wavelet features were selected, and a radiomics score (Rad score) for recurrence-free survival prediction was established. All patients were assigned to high-score (≥ median Rad score) or low-score (< median Rad score) groups, with the median Rad score as the threshold. Tables 1, 2, and 3 show the optimal feature distributions of radiomics models based on T1, T2FS, and T1+T2FS images, respectively. The Rad score calculation formulas based on these features are:

[0074] T1:

[0075] F1×411.314+F2×318.092+F3×555.452+F4×430.831+F5×270.128+F6×536.202+F7×280.188+F8×399.036+F9×434.576+F10×274.616+F11×469.686+F12×243.742+F13×319.444+F14×312.036+F15×242.110+F16×664.443+F17×383.167+F18×1134.098+F19×329.060+F20×439.065;

[0076] T2FS:

[0077] F1×9.685+F2×9.555+F3×12.335+F4×73.308+F5×11.265+F6×14.440+F7×27.583+F8×14.695+F9×16.584+F10×38.431+F11×12.981+F12×12.113+F13×28.157+F14×11.446+F15×82.305+F16×13.691+F17×12.263+F18×10.515+F19×29.029+F20×12.143+F21×15.462;

[0078] T1+T2FS:

[0079] F1×0.083+F2×0.138+F3×0.100+F4×0.159+F5×0.187+F6×0.091+F7×0.114+F8×0.067+F9×0.065+F10×0.107+F11×0.060+F12×0.155.

[0080] Table 1 Distribution of optimal radiomics features based on T1 images

[0081] F1 boxmean_gldm_SmallDependenceLowGrayLevelEmphasis F2 curvatureflow_glszm_SizeZoneNonUniformity F3 boxsigmaimage_glcm_DifferenceVariance F4 boxsigmaimage_glrlm_LongRunHighGrayLevelEmphasis F5 boxsigmaimage_glrlm_ShortRunEmphasis F6 log_glcm_log-sigma-4-0-mm-3D-Imc1 F7 log_glrlm_log-sigma-2-0-mm-3D-LongRunEmphasis F8 log_glrlm_log-sigma-2-0-mm-3D-RunVariance F9 log_glrlm_log-sigma-4-0-mm-3D-RunVariance F10 log_glrlm_log-sigma-4-0-mm-3D-ShortRunEmphasis F11 log_glszm_log-sigma-0-5-mm-3D-LargeAreaLowGrayLevelEmphasis F12 log_glszm_log-sigma-2-0-mm-3D-SizeZoneNonUniformityNormalized F13 wavelet_firstorder_wavelet-LLH-TotalEnergy F14 wavelet_firstorder_wavelet-HLL-Kurtosis F15 wavelet_glcm_wavelet-HLL-SumEntropy F16 wavelet_glrlm_wavelet-HLL-LongRunLowGrayLevelEmphasis F17 wavelet_glszm_wavelet-LHH-SizeZoneNonUniformity F18 wavelet_glszm_wavelet-HLL-LargeAreaLowGrayLevelEmphasis F19 wavelet_glszm_wavelet-HLL-ZoneVariance F20 recursivegaussian_glszm_SizeZoneNonUniformity

[0082] Table 2 Distribution of optimal radiomics features based on T2FS images

[0083]

[0084]

[0085] Table 3 Distribution of optimal radiomics features based on T1+T2FS images

[0086] F1 original_shape_Sphericity F2 log_firstorder_log-sigma-0-5-mm-3D-Skewness F3 log_glszm_log-sigma-0-5-mm-3D-SizeZoneNonUniformity F4 wavelet_firstorder_wavelet-HLL-Kurtosis F5 wavelet_glcm_wavelet-HHH-ClusterShade F6 wavelet_glrlm_wavelet-LHL-LongRunLowGrayLevelEmphasis F7 wavelet_glrlm_wavelet-HLL-LongRunLowGrayLevelEmphasis F8 wavelet_glszm_wavelet-LHH-ZoneEntropy F9 wavelet_ngtdm_wavelet-LHH-Strength F10 wavelet_ngtdm_wavelet-HLH-Busyness F11 normalize_glszm_SmallAreaHighGrayLevelEmphasis F12 laplaciansharpening_gldm_LargeDependenceHighGrayLevelEmphasis

[0087] Univariate Cox analysis showed that peritumoral edema, skip lesions, and radiomics scores were associated with poor RFS and were included in the construction of the nomogram. The C index results showed the discriminatory ability of each model for tumor recurrence-free survival ( Figure 1 The survival curve drawn by the Kaplan-Meier method showed that patients with high radiomics scores had a poorer prognosis ( Figure 2 ). The nomograms and timeROC constructed based on each model show the predictive ability of the model over time ( Figure 3). The AUC values ​​of clinical+T1 imaging omics, clinical+T2FS imaging omics and clinical+T1+T2FS combined imaging omics models for recurrence-free survival at 300 days after sarcoma resection were 0.947, 0.907 and 0.924, respectively; the AUC values ​​of recurrence-free survival at 600 days were 0.913, 0.853 and 0.869, respectively; the AUC values ​​of recurrence-free survival at 900 days were 0.918, 0.866 and 0.889, respectively. Decision curve analysis showed that the combined nomogram model of clinical+T1+T2FS combined with imaging omics score was more accurate in predicting RFS than the single imaging omics model ( Figure 4 ).

[0088] The clinical-radiology model for predicting the recurrence of bone and soft tissue sarcoma established by the model establishment method of the present invention can provide valuable reference information for clinicians and help them formulate more accurate treatment plans. For example, for patients with a predicted short recurrence-free survival period, more active treatment measures may need to be taken, such as strengthening chemotherapy or considering more radical surgical plans. The radiomics model can provide accurate digital prediction results by incorporating the patient's clinical characteristics, pathological characteristics, and imaging characteristics into the model. In addition, the prognostic model constructed based on the clinical-radiology model is visualized in the form of a nomogram, which can intuitively show the impact of various factors on the patient's recurrence-free survival period. Through the nomogram, clinicians can quickly understand the patient's prognosis and make decisions based on it. During the treatment process, doctors can regularly evaluate the patient's recurrence-free survival period and adjust the treatment plan in time according to the evaluation results. If the patient's recurrence-free survival prediction results show a deteriorating trend, the doctor can promptly discover and take corresponding intervention measures to optimize the treatment effect and prolong the patient's survival period.

Claims

1. A method for establishing a model for predicting the recurrence of bone and soft tissue sarcoma, comprising the following steps:

1. Patient data collection and organization Determine the inclusion and exclusion criteria for patients, and collect clinical, pathological, imaging, and follow-up data of patients with osteosarcoma and soft tissue sarcoma who meet the criteria; Select clinical variables that may affect recurrence risk; Clean the collected patient data to remove data with too many missing values ​​or abnormal data; 2. Image data preprocessing, including manual segmentation of tumor ROI and calibration of patient MRI images; Third, extract first-order and second-order radiomics features from tumor ROI; 4. Feature Selection and Optimization Stability verification: The stability of the extracted features was evaluated by the inter-observer ICC coefficient, and the radiomics features with ICC>0.75 were retained; Dimensionality reduction processing: includes the following steps: (1) For features with ICC ≥ 0.75, SelectKBest based on univariate feature selection was used to analyze the relationship between features and tumor prognosis, and features with P < 0.05 were screened to obtain the initial feature set; (2) The maximum relevance-minimum redundancy (mRMR) method was used to eliminate confounding factors, calculate the correlation between features and tumor recurrence, and the redundancy between features, and select a subset of features with high correlation from the initial feature set; (3) Using the least absolute shrinkage and selection operator (LASSO) method, the feature subset with high correlation obtained in step (2) is further reduced in dimension, and the feature set with the best λ value is screened out using 5-fold cross validation, excluding the features with a variable coefficient of 0; 6. Model Construction According to the selected features and the corresponding variable coefficients, the calculation method of radiomics score was constructed; The screened radiomics features were used to calculate T1 radiomics, T2FS radiomics, and T1+T2FS combined radiomics scores; To select clinical variables associated with recurrence of bone and soft tissue sarcomas; The obtained clinical variables related to the recurrence of bone and soft tissue sarcoma were combined with the radiomics score to establish a clinical radiomics model and draw a nomogram; To establish a joint clinical-MRI radiomics model for predicting recurrence of bone and soft tissue sarcomas.

2. The method for establishing a model according to claim 1, characterized in that: In step 1, the clinical variables selected that may affect the risk of recurrence include age, sex, tumor size measured on the patient's MRI images, observation of tumor signal heterogeneity, tumor margin, peritumoral edema, and the presence or absence of skip lesions.

3. The method for establishing a model according to claim 1, characterized in that: Manual segmentation of tumor ROI in step 2 includes: selecting at least two physicians, each of whom outlines the tumor ROI of patients with bone or soft tissue sarcoma on the T2 FS image and copies the three-dimensional ROI to the patient's T1WI image to ensure the consistency of the tumor ROI; physician B performs manual segmentation of the tumor ROI in the same way.

4. The method for establishing a model according to claim 1, characterized in that: In step 2, the patient's MRI images are calibrated to ensure image quality and consistency, including image intensity discretization, intensity normalization, and isotropic resampling processing.

5. The method for establishing a model according to claim 1, characterized in that: The first-order features in step three include tumor size, shape, mean, standard deviation, skewness, and kurtosis; the second-order features include gray-level co-occurrence matrix, gray-level size area matrix, gray-level dependency matrix, gray-level dependency matrix, Gaussian Laplace, and wavelet coefficients.

6. The method for establishing a model according to claim 1, characterized in that: In step 6, univariate Cox regression analysis was used to select clinical variables associated with recurrence of bone and soft tissue sarcoma, including peritumoral edema and skip lesions.

7. The method for establishing a model according to claim 1, characterized in that: In step six, clinical+T1 radiomics, clinical+T2FS radiomics, and clinical+T1+T2FS combined radiomics models were established and nomograms were drawn.

8. The method for establishing a model according to claim 1, characterized in that: In step 6, a clinical + T1 radiomics model was established. Table 1 shows the optimal feature distribution of the established model. The Rad score calculation formula based on the optimal feature is: F1×411.314+F2×318.092+F3×555.452+F4×430.831+F5×270.128+F6×536.202+F7×280.188+F8×399.036+F9×434.576+F10×274.616+F11×469.686+F12×243.742+F13×319.444+F14×312.036+F15×242.110+F16×664.443+F17×383.167+F18×1134.098+F19×329.060+F20×439.

065. Table 1 Distribution of optimal radiomics features based on T1 images 9. The method for establishing a model according to claim 1, characterized in that: In step 6, clinical + T2FS radiomics was established. Table 2 shows the optimal feature distribution of the established model. The Rad score calculation formula based on the optimal feature is: F1×9.685+F2×9.555+F3×12.335+F4×73.308+F5×11.265+F6×14.440+F7×27.583+F8×14.695+F9×16 .584+F10×38.431+F11×12.981+F12×12.113+F13×28.157+F14×11.446+F15×82.305+F16×13.691+F17×12.263+F18×10.515+F19×29.029+F20×12.143+F21×15.

462. Table 2 Distribution of optimal radiomics features based on T2FS images 10. The method for establishing a model according to claim 1, characterized in that: In step 6, a clinical + T1 + T2FS combined radiomics model was established. Table 3 shows the optimal feature distribution of the established model. The Rad score calculation formula based on the optimal feature is: F1×0.083+F2×0.138+F3×0.100+F4×0.159+F5×0.187+F6×0.091+F7×0.114+F8×0.067+F9×0.065+F10×0.107+F11×0.060+F12×0.

155. Table 3 Distribution of optimal radiomics features based on T1+T2FS images

Images