ABSTRACT
Objective: To evaluate the value of a nomogram combined MRI Diffusion Weighted Imaging (DWI) and clinical features to predict the treatment response of Neoadjuvant Chemotherapy (NAC) in patients with osteosarcoma. Methods: A retrospective analysis was conducted on 209 osteosarcoma patients admitted into two bone cancer treatment centers (133 males, 76females; mean age 16.31 ± 11.42 years) from January 2016 to January 2022. Patients were classified as pathological good responders (pGRs) if postoperative histopathological examination revealed ≥90% tumor necrosis, and non-pGRs if <90%. Their clinical features were subjected to univariate and multivariate analysis, and features with statistically significance were utilized to construct a clinical signature using machine learning algorithms. Apparent diffusion coefficient (ADC) values pre-NAC (ADC 0) and post two chemotherapy cycles (ADC 1) were recorded. Regions of interest (ROIs) were delineated from pre-treatment DWI images (b=1000 s/mm²) for radiomic features extraction. Variance thresholding, SelectKBest, and LASSO regression were used to select features with strong relevance, and three machine learning models (Logistic Regression, RandomForest and XGBoost) were used to construct radiomics signatures for predicting treatment response. Finally, the clinical and radiomics signatures were integrated to establish a comprehensive nomogram model. Predictive performance was assessed using ROC curve analysis, with model clinical utility appraised through AUC and decision curve analysis (DCA). Results: Of the 209patients, 51 (24.4%) were pGRs, while 158 (75.6%) were non-pGRs. No significant ADC1 difference was observed between groups (P>0.05), but pGRs had a higher ADC 0 (P<0.01). ROC analysis indicated an AUC of 0.681 (95% CI: 0.482-0.862) for ADC 0 at the threshold of ≥1.37×10-3 mm²/s, achieving 74.7% sensitivity and 75.7% specificity. The clinical and radiomics models reached AUCs of 0.669 (95% CI: 0.401-0.826) and 0.768 (95% CI: 0.681-0.922) respectively in the test set. The combined nomogram displayed superior discrimination with an AUC of 0.848 (95% CI: 0.668-0.951) and 75.8% accuracy. The DCA suggested the clinical utility of the nomogram. Conclusion: The nomogram based on combined radiomics and clinical features outperformed standalone clinical or radiomics model, offering enhanced accuracy in evaluating NAC response in osteosarcoma. It held significant promise for clinical applications.
ABSTRACT
OBJECTIVES: This study aims to evaluate the value of machine learning-based dynamic contrast-enhanced MRI (DCE-MRI) radiomics nomogram in prediction treatment response of neoadjuvant chemotherapy (NAC) in patients with osteosarcoma. METHODS: A total of 102 patients with osteosarcoma and who underwent NAC were enrolled in this study. All patients received a DCE-MRI scan before NAC. The Response Evaluation Criteria in Solid Tumors was used as the standard to evaluate the NAC response with complete remission and partial remission in the effective group, stable disease, and progressive disease in the ineffective group. The following semi-quantitative parameters of DCE-MRI were calculated: early dynamic enhancement wash-in slope (Slope), time to peak (TTP), and enhancement rate (R). The acquired data is randomly divided into 70% for training and 30% for testing. Variance threshold, univariate feature selection, and least absolute shrinkage and selection operator were used to select the optimal features. Three classifiers (K-nearest neighbor, KNN; support vector machine, SVM; and logistic regression, LR) were implemented for model establishment. The performance of different classifiers and conventional semi-quantitative parameters was evaluated by confusion matrix and receiver operating characteristic curves. Furthermore, clinically relevant risk factors including age, tumor size and site, pathological fracture, and surgical staging were collected to evaluate their predictive values for the efficacy of NAC. The selected clinical features and imaging features were combined to establish the model and the nomogram, and then the predictive efficacy was evaluated. RESULTS: The clinical relevance risk factor analysis demonstrates that only surgical stage was an independent predictor of NAC. A total of seven radiomic features were selected, and three machine learning models (KNN, SVM, and LR) were established based on such features. The prediction accuracy (ACC) of these three models was 0.89, 0.84, and 0.84, respectively. The area under the subject curve (AUC) of these three models was 0.86, 0.92, and 0.93, respectively. As for Slope, TTP, and R parameters, the prediction ACC was 0.91, 0.89, and 0.81, respectively, while the AUC was 0.87, 0.85, and 0.83, respectively. In both the training and testing sets, the ACC and AUC of the combined model were higher than those of the radiomics models (ACC = 0.91 and AUC = 0.95), which indicate an outstanding performance of our proposed model. CONCLUSIONS: The radiomics nomogram demonstrates satisfactory predictive results for the treatment response of patients with osteosarcoma before NAC. This finding may provide a new decision basis to improve the treatment plan.
