ESTRO 2022 - Abstract Book

S1591

Abstract book

ESTRO 2022

Council - Region of Tuscany, G. Monasterio Foundation, Pisa, Italy; 7 IEO European Institute of Oncology IRCCS, Scientific Directorate, Milan, Italy Purpose or Objective In this study, advanced models integrating radiomic features derived from magnetic resonance (MR) images and clinical data were developed for the prediction of local control (LC), distant progression (DP) and overall survival (OS) in patients treated with SRS for BM from non-small cell lung cancer. By doing so, we aimed to investigate the variability in model performance when extracting features with different platforms, and if the empowering of clinical models with radiomics could benefit performance. Materials and Methods A total of 148 patients treated at the same institution, with a total of 276 BM, were retrospectively included. Pre-treatment T1-weighted MR images of the brain were considered. Radiomic features were extracted from the structures of each brain lesion with two different platforms: PyRadiomics (PyR) and SOPHiA Radiomics (SR). A total of 1129 and 192 features were considered for statistical analysis (see Fig1 ), respectively. Clinical data were collected for each patient from the follow- up reports. Five different models were developed for each endpoint: two radiomic models (PyR and SR), one clinical model and two combined models (integrating clinical with PyR and SR information respectively). Performance was asserted in terms of Harrell’s C-index, see Fig2 .

Results In predicting LC, the SR radiomic model outperformed the PyR one. From the clinical model, an increase in the patient's age associated with a slight increase in the probability of LC for his lesions. Cerebellar lesions and concomitant therapy were associated with an increased rate of LC compared to frontal ones and no-therapy, respectively. In predicting DP, 6 radiomic features were significant in the SR model, while only 1 in the PyR radiomic model. Results from the clinical model suggest that parietal and occipital BM are more prone to DP than frontal ones, as are patients with stage IV at diagnosis. In contrast, concomitant therapy resulted in lower DP rate than no-therapy. The combined models showed some differences. In predicting OS, both radiomic models performed equally well. Clinical model’s performance was slightly better than that of the radiomic models. The patient's KPS and prescribed BED were associated with an extended OS, while receiving not- concomitant therapy was associated with lower OS than not receiving therapy at all. The SR combined model performed slightly better than the others. Again, differences exist between the two combined models. Conclusion Overall, the best performing model was the SR radiomic model for LC prediction. DP was the least predictable endpoint in our dataset. This study reveals the choice of radiomic platform may result in differences in performance, and how the merge of both sources of data might not always led to improvement. It provides some important insights into the design of future prospective studies and suggests that individualised assessment of LC and OS probability could be aided by means of such models, for improved decision-making and prognostic assessment.

PO-1784 predicting radiation pneumonitis based on retraining a deep learning feature extraction model

Z. Wang 1 , Z. Zhang 1 , A. Traverso 1 , A. Dekker 1

1 Maastricht University, Department of Radiation Oncology (Maastro), Maastricht, The Netherlands

Purpose or Objective Radiation pneumonitis (RP) is a common adverse effect of thoracic radiotherapy. At present, some researchers use the pre- trained convolutional neural network(CNN) to extract deep learning (DL) features which have been proved that it can achieve better performance than the hand-craft feature extraction method. Although using a pre-trained model is easier to migrate to other tasks, only a model that has been retrained according to a specific data set can be best applied to a