P176] Evaluating automatic contour propagation for estimating tumour shrinkage in small-cell lung cancer patients treated with chemo-radiotherapy

Abstract

Purpose The clinical relevance of tumour shrinkage during thoracic radiotherapy remains unknown with studies showing conflicting results whether shrinkage predicts outcome. Previous work has relied on manual contouring on cone beam CT (CBCT). To enable large scale analysis, automated methods for contour propagation and quantification of tumour shrinkage are essential. Methods 20 small-cell lung cancer patients treated with concurrent chemo-radiotherapy were selected. All patients had a planning CT and, in total, 81 weekly CBCT scans were available. An experienced lung cancer radiation oncologist manually contoured the GTV on all scans. Additionally, the GTV contour of the planning CT was automatically propagated across the CBCTs, using ADMIRE (Research v1.13, Elekta AB, Stockholm, Sweden). To investigate the accuracy of the automated contour propagation, a comparison against the manual contours was performed using median distance to agreement (mDTA). To quantify tumour shrinkage, volumes were calculated for propagated and manual contours. Differences in volumes were assessed using a Bland–Altman plot. Finally, using the day one CBCT as reference volume, the relative percentage change was calculated at each time-point and fitted using linear regression. The slopes for manual versus propagated contours were compared. Results ADMIRE successfully propagated GTV contours all CBCTs. Concordance between the manual and propagated contours showed an unsigned mDTA of 3 mm across all CBCT. Results were worse for GTV contours overlapping with the mediastinum. The Bland–Altman analysis for absolute differences in volume at each time-point resulted a mean of −1.6 cm3. Relative differences provided a mean of −4.8%, with manual contours showing smaller volumes. Finally, manual contours estimated a consistently faster rate of tumour shrinkage (0.018 cm3/day) than automatically propagated contours (0.014 cm3/day). A linear fit between the manual and propagated gradients per patient found a slope of 0.68 (1 for perfect agreement). Conclusions The results show that automatic contour propagation performance is acceptable for use in wider research studies. There are differences between the manual and propagated contours, with the clinician estimated a larger rate of tumour shrinkage, which is suggestive of an observer bias. We will now apply this technique to investigate tumour shrinkage and outcome in the CONVERT trial patients.