Radiomics models predict progression in lung oligometastases after SABR in phase II trial

AIMS: Despite the increasing use of stereotactic ablative radiotherapy (SABR) for oligometastatic cancer, at present, accurate models to predict the time until disease progression are lacking. The study developed radiomics models to predict progression-free survival (PFS) and early polymetastatic progression (PMP) in order to improve upon basic clinical prognostic models in lung oligometastatic patients treated in the population-based single-arm SABR-5 trial. MATERIALS AND METHODS: Among 134 patients treated for lung oligometastases in the SABR-5 trial, pretreatment computed tomography images were available for 126 patients for inclusion in this study. In total, 1116 radiomic features were extracted from the original, wavelet-filtered, and Laplacian of Gaussian (LoG)-filtered images using PyRadiomics. Analyses were developed to predict (1) PFS and (2) early PMP, defined as progression of more than 5 lesions within 6 months of SABR. Clinical-only (ECOG score, primary tumour type, oligoprogression, and gross tumour volume [GTV]), radiomics-only, and combined models were developed. Feature selection was performed using the Pearson correlation coefficient. Cox proportional hazards regression was used to predict PFS and stratify patients into high- and low-risk groups. For the early PMP model, a support vector machine was evaluated using 10-, 5-, and 3-fold cross-validation. RESULTS: The radiomics-only and combined models achieved a concordance index of 0.72 in the test set, versus 0.52 for the clinical-only model. The radiomics model stratified patients into high- and low-risk groups in both the training and test sets (P < 0.001 and 0.041, respectively). In the early PMP model, area under the receiver operating characteristic curve (AUC), true-positive rate, and true-negative rate across all folds (10, 5, and 3) were 0.85, 0.73, and 0.78 for the radiomics-only model and 0.47, 0.45, and 0.62 for the clinical-only model, respectively. CONCLUSION: Radiomics models outperformed clinical models for predicting PFS and early PMP. These radiomics models could potentially assist with optimal patient selection and treatment strategies for patients with pulmonary oligometastases.