MRI-based radiomics signature for pretreatment prediction of pathological response to neoadjuvant chemotherapy in...

To develop and validate a radiomics signature based on magnetic resonance imaging (MRI) from multicenter datasets for preoperative prediction of pathologic response to neoadjuvant chemotherapy (NAC) in patients with osteosarcoma.

We retrospectively enrolled 102 patients with histologically confirmed osteosarcoma who received chemotherapy before treatment from 4 hospitals (68 in the primary cohort and 34 in the external validation cohort). Quantitative imaging features were extracted from contrast-enhanced fat-suppressed T1-weighted images (CE FS T1WI). Four classification methods, i.e., the least absolute shrinkage and selection operator logistic regression (LASSO-LR), support vector machine (SVM), Gaussian process (GP), and Naive Bayes (NB) algorithm, were compared for feature selection and radiomics signature construction. The predictive performance of the radiomics signatures was assessed with the area under receiver operating characteristics curve (AUC), calibration curve, and decision curve analysis (DCA).

Thirteen radiomics features selected based on the LASSO-LR classifier were adopted to construct the radiomics signature, which was significantly associated with the pathologic response. The prediction model achieved the best performance between good and poor responders with an AUC of 0.882 (95% CI, 0.837-0.918) in the primary cohort. Calibration curves showed good agreement. Similarly, findings were validated in the external validation cohort with good performance (AUC, 0.842 [95% CI, 0.793-0.883]) and good calibration. DCA analysis confirmed the clinical utility of the selected radiomics signature.

The constructed CE FS T1WI-radiomics signature with excellent performance could provide a potential tool to predict pathologic response to NAC in patients with osteosarcoma.

• The radiomics signature based on multicenter contrast-enhanced MRI was useful to predict response to NAC. • The prediction model obtained with the LASSO-LR classifier achieved the best performance. • The baseline clinical characteristics were not associated with response to NAC.