Session Item

In the observational, multi-center "Blue Sky Radiomics” study (NCT04364776), we aim to investigate the prognostic role of radiomic features in predicting progression-free survival (PFS) in a series of stage III, unresectable, PD-L1 positive NSCLC patients undergoing chemoradiotherapy (CRT) and maintenance durvalumab. This is a preliminary report on the first evaluable patients (n=57).

Patients were all affected with stage 3 unresectable NSCLC, and received either sequential or concurrent CRT (Rt dose 60 Gy) followed by durvalumab maintenance (median of 22 doses). Median PFS for the whole cohort was 20.6 months. Median follow-up time 19.3 months. For radiomic analysis we identified the primary lung tumor on computed tomography (CT) images acquired with i.v. contrast medium, and with different scanners. CT images have been collected at 2 time points: at the diagnosis (T0), after CRT (T1). Tumor segmentation was performed by two specialists (thoracic radiologist and radio-oncologist) using the Oncentra Masterplan® software. We extracted 42 radiomic features from the specific ROIs (LIFEx). To harmonize radiomic features we used the Combat tool. We compared Elastic Net (EN), Random Forest (RF) and Support Vector Machine (SVM) models performances for classifying PFS (Leave-One-Out Cross Validation method) at T0 and T1 time points. Moreover, we assessed the change in radiomic features between T0 and T1. We used the Cox and Survival Random Forest models to evaluate the performance of radiomic features, clinical factors, and both, for stratifying for PFS (k-Fold Cross Validation method, k=5).

In the first Table, we reported the accuracy and the AUC of the 3 different models tested, at different time points. The addition of radiomic features to clinical features alone (stage A vs. B-C, histology, durvalumab start <42 days vs. > 42 days, PD-L1 level <50% or > 50%, sequential vs. concurrent CRT) improved the AUC by 4% for EN, and by 2% for both RF and SVM. In the second Table, we reported the performances in predicting PFS of each patient in a time-to-event setting.

Our preliminary study underlines the importance of the development of a robust analysis pipeline for small-datasets. In this preliminary analysis on 57 patients out of 100 (target sample size), we exploited the application of Machine Learning methods for PFS prediction. With the use of radiomic features integrated with clinical factors, the AUC was improved by 4% for the EN, and 2% for both RF and SVM, respectively. A larger sample size should be tested to confirm this favorable trend and to hopefully reach better results.