(Tuesday, 7/14/2020) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Room: Track 3
Purpose: A recent ensemble model of dose to the atria, lung, pericardium and ventricles has been shown to predict overall survival in the randomized controlled RTOG 0617 trial data. Here we further explored the underlying spatial dose-response relationship in this dataset by investigating dose to detailed components of the cardio-pulmonary system.
Methods: DICOM and outcome data were downloaded from TCIA and the 398 patients that received prescription dose cf. their randomization level were included. A previously developed deep learning-based cardiac substructure segmentation framework that included eight structures (aorta, the four chambers, pulmonary artery, and the two vena cavas) was applied to all scans. Dose to these substructures and the mean lung dose (the best lung predictor in the preceding analysis) were investigated for association with overall survival using Cox proportional Hazards. Data were analyzed in a training (n=280) and testing (n=118) approach with bootstrap resampling. In training, the univariate p-value cutoff was corrected for multiple testing (p=0.0002). Stepwise iterated multivariate analysis included only non-strongly correlated variables (|Rs|<0.80), and the performance (c-index and risk group stratification) of the most frequently selected models was explored on validation.
Results: Three predictors were suggested: mean lung dose (MLD), mean of the hottest 5 % of the left ventricle dose (LVMOH5) and the minimum right atrium dose (RAMIN). The three most frequently selected multivariate models were 1. LVMOH5+RAMIN, 2. MLD+RAMIN and 3. LVMOH5+ MLD+RAMIN (frequency: 17-28%). In validation, the c-index was 0.77-0.82) and the largest separation between the low and high-risk groups was obtained for the model that included dose to all three structures (Fig.1).
Conclusion: This update has identified LV and RA in addition to the lung as important cardio-pulmonary components. Introducing dose-volume constraints for the LV and RA and potentially lower used MLD limits may lead to prolonged survival in locally-advanced lung cancer.
Keywords
Taxonomy
Not Applicable / None Entered.
Contact Email
