Radiomics signature extracted from diffusion-weighted magnetic resonance imaging predicts outcomes in osteosarcoma

Osteosarcoma often requires multidisciplinary treatment including surgery, chemotherapy and radiotherapy. However, tumor behavior can vary widely among patients and selection of appropriate therapies in any individual patient remains a critical challenge. Radiomics seeks to quantify complex aspects of tumor images under the assumption that this information is related to tumor biology. This study tested the hypothesis that a radiomic signature extracted from Diffusion-weighted magnetic resonance images (DWI-MRI) can improve prediction of overall survival (OS) compared with clinical factors alone in localised osteosarcoma.

Pre-treatment DWI-MRI were collected from 112 patients (9-67 years of age) with histological-proven osteosarcoma that were treated with curative intent. The entire dataset was divided in two subsets: the training and validation cohorts containing 76 and 24% of the data respectively. Clinical data were extracted from our medical record. Two experienced radiotherapists evaluated DWI-MRIs for quality and segmented the tumor. A total of 103 radiomic features were calculated for each image. Least absolute shrinkage and selection operator (LASSO) regression was applied to select features. Association between the radiomics signature and OS was explored. Further validation of the radiomics signature as an independent biomarker was performed by using multivariate Cox regression. The Cox proportional-hazard regression model was also used to analyze the correlation between the prognostic factor and the survival for the clinical (C) model after the univariate analysis. Radiomics (R) model identified radiomics signature, which is the best predictor from the radiomic variable classes based on LASSO regression. Harrell's C-index was used to demonstrate the incremental value of the radiomics signature to the traditional clinical risk factors for the individualized prediction performance.

Cox proportional-hazard regression model shows that: Tumor size, alkaline phosphatase (ALP) status before treatment and number of courses of chemotherapy were proven as the dependent clinical prognostic factors of osteosarcoma's overall survival time. The radiomics signature was significantly associated with OS, independent of clinical risk factors (radiomics signature: HR: 5.11, 95% CI: 2.85, 9.18, P < 0.001). Incorporating the radiomics signature into the coalition (C+R) model resulted in better performance (P < .001) for the estimation of OS (C-index: 0.813; 95% CI: 0.75, 0.89) than with the clinical (C) model (C-index: 0.764; 95% CI: 0.69, 0.85), or the single radiomics (R) model (C-index: 0.712; 95% CI: 0.65, 0.78).

This study shows that the radiomics signature extracted from pre-treatment DWI-MRI improve prediction of OS over clinical features alone. Combination of the radiomics signature and the traditional clinical risk factors performed better for individualized OS estimation in patients with osteosarcoma, which might enable a step forward precise medicine. This method may help better select patients most likely to benefit from intensified multimodality diagnosis and therapies. Future studies will focus on multi-center validation of an optimized model.