An MRI radiomics approach to predict the efficacy of chemotherapy for osteosarcoma

Magnetic resonance imaging (MRI) and dynamic contrast enhancement MRI(DCE-MRI) data of osteosarcoma patients prior to neoadjuvant chemotherapy (NAC) were compared in order to investigate the value of imaging histological features in predicting the rate of tumor necrosis in osteosarcoma and patients' postoperative survival.

We enrolled 28 osteosarcoma patients who received three courses of NAC followed by tumor resection. Prior to chemotherapy both conventional MRI and DCE-MRI scans were performed Quantitative analysis of histological features within the tumor region was conducted using T1-weighted, T2-weighted, and DCE-MRI images. Recursive feature elimination guided the selection of relevant features, and a prediction model was constructed using five machine learning algorithms (Random Forest, Logistic Regression, Decision Tree, Gradient Boosting, and Bagging Decision Tree). Model performance was assessed using receiver operating characteristic (ROC) curves, area under the curve (AUC), and accuracy. The five imaging features with the strongest predictive capability were identified, and their association with postoperative survival was explored using Kaplan-Meier survival analysis.

Among the five prediction models, the decision tree, gradient boosting and bagging decision tree models showed high prediction performance, with AUC values of 0.850, 0.856 and 0.872, respectively, and an accuracy of 0.857. Notably, Kaplan-Meier survival analysis highlighted the significance of two features extracted from DCE-MRI images -"Small Dependence Low Gray Level Emphasis" (based on the Gray Level Cooccurrence Matrix, GLCM）and "Run Length Non-Uniformity" (based on the Gray Level Run Length Matrix, GLRLM), were significantly predictive of postoperative survival (P < 0.05).

DCE-MRI-based imaging histology models of osteosarcoma patients prior to NAC can be a useful tool for predicting tumour necrosis rates and patient survival after surgery.