Room: AAPM ePoster Library
Purpose: Quantify inter-observer variability in prostate contouring on images from a 1.5T MRI simulator (MRsim) and two 1.5T MRI-Linac (MRL) online adaptive prostate SBRT sequences, including 2-minute (MRL-2min) and 6-min scans (MRL-6min).
Methods: Prostate was contoured on three T2-weighted MRs (MRsim, MRL-2min, MRL-6min), for five patients by seven radiation oncologists (RO). For each image set, the STAPLE contour was generated for comparison. Inter-observer agreement was quantified using: Dice Similarity Coefficient (DSC), Mean Distance-to-Agreement (MDA), Hausdorff distance (HD) and Kappa coefficient (?). Directional differences were quantified by computing 3D distance between centroids. Significance was derived from paired t-tests between image pairs.
Results: Across the cohort, there were no statistically significant differences in mean similarity metrics. Mean ? was 0.902, 0.895 and 0.893 for MRsim, MRL-2min, and MRL-6min, respectively, indicating good agreement. High mean DSC values of 0.905 (MRsim), 0.891 (MRL-2min) and 0.910 (MRL-6min) further indicate good concordance between RO and STAPLE contours regardless of scan (p>0.05 per pairwise-comparison). Since DSC can sometimes be insensitive to local discrepancies, mean (MDA) and maximum (HD) distances from STAPLE were also scored. Though no difference in mean MDA was seen between MRsim and MRL scans, a small improvement in MDA favoring MRL-2min (1.46mm) over MRL-6min (1.87mm) was found (p<0.01). Mean HD was similar between STAPLE and RO contours (6.6mm overall average) but values were variable, ranging from 2.6-25.7mm. Evaluation of directional differences between STAPLE and RO contour centroids revealed statistically larger offsets of: 3.1mm cranio-caudally and 2.4mm anteriorly-posteriorly, respectively, compared to 1.1mm laterally. The majority of differences occurred towards apex and base.
Conclusion: Similarity analyses indicate that a reasonable degree of contouring variability, comparable between MR sets. Larger differences most often seen towards the longitudinal extents of the prostate may have implications on planning margins and could provide a basis for MR-only contouring workshops.