An in-house validated technique generated the GTV on the 50% 4DCT phase for 259 patients treated with SABR in 3, 5 or 8 fractions. Dose was converted to EQD2 with α/β=10. Local rigid registration was used to match tumour in each phase to 50% and calculate respiratory motion. Dose distribution was blurred according to respiratory motion to estimate delivered dose.
In-house data-mining was used to sample CT pixel values (inside lung) and dose at radial distance from the GTV. Briefly, we calculated the signed distance transform and created 2D cross-histograms of pixel value/dose versus distance for each patient. For every 1mm, dose standard deviation (SD) and mean pixel value was sampled (Fig.1).
Average cross-histograms of mean pixel value were calculated for patients who did and did not experience DF at 18 months, censored for follow-up. T-statistic describing difference in cross-histograms defined an important region, using permutation testing for statistical significance. Mean pixel value was sampled from this region
.
Dose variability (SD) was sampled from the region (Fig.1A) extending outside the PTV, but within the expected MDE range (5-15mm), where high SD represents higher chance of underdosing MDE.
Patients were split into low and high dose variability based on the median. For each cohort, mean pixel value was assessed in univariable and multivariable Cox regression.