ESTRO 2021 Abstract Book

S507

ESTRO 2021

We introduce an image-based data mining (IBDM) technique in which a Cox regression is performed for combinations of density and dose in shells around the GTV, allowing assessment of dose-density interactions whilst accounting for clinical variables. We aim to predict different failure types and validate in an independent cohort. Materials and Methods 433 lung cancer patients treated with SABR were available from two centres, 255 for discovery and 178 for validation. In the discovery cohort, events were recorded by failure type (16 LR, 17 RP, 44 DM) but in validation cohort only the presence (Y/N) not location of failure was recorded. Accumulated dose was estimated on 4DCT considering tumour motion and converted to EQD2 (alpha/beta=10). A radial approach was used to sample density (inside lung) on the exhale phase and accumulated dose versus distance from the GTV. For every 1mm shell, mean density and dose standard deviation (SD) were sampled. For each combination of dose distance and density distance, a Cox model was built containing clinical variables plus mean density and dose SD. A separate Cox model was built with an additional interaction term (density*dose) to account for changing risk of dose due to the density biomarker. The change in performance measured by Akaike Information Criterion (AIC) was assessed with an analysis of deviance between the clinical and interaction models for LR, RP and DM. P-value maps were created and p<0.05 indicated statistically significant regions. The p-value for change in AIC due to the interaction term is reported. In the validation cohort, regions were tested against presence of failure. Results Regions were identified for LR and DM (Fig. 1). High tumour border density and dose SD ~1cm from the GTV predict LR. In other words, border density is a predictor for LR, and the risk is greater for patients with non- uniform dose coverage at 1cm.

High peritumoural density predicts DM with greater risk for high dose SD at <1cm and ~3cm - the latter combination was validated in an independent cohort despite patient differences (Fig. 2). The significant region for LR could not be validated due to missing LR data.