Session Item

Contouring variation is one of the largest systematic uncertainties in modern radiotherapy. Many studies have evaluated inter-observer variation in small patient cohorts, yet its effect on clinical outcomes has never been analysed quantitatively. We propose a novel framework to analyse the effect of contouring variation on clinical outcome for a large cohort of patients.

Planning CTs and contours of 232 intermediate- and high-risk prostate cancer patients, treated with 19x3Gy, were used in this study. For each patient, we compared the manually delineated CTV (prostate gland without seminal vesicles), to an automatically generated CTV contour created using the Deep Learning segmentation tool in ADMIRE v3.0 (Elekta AB, Sweden). This automatic contour can be seen as a consistent (yet imperfect) reference contour created by a virtual observer. Local contour deviation was measured from the reference to the manual contour using spherical coordinates. We sampled the coronal and transverse angles every 3^o, and created contour deviation maps of 60x120 pixels for each patient (Fig. 1). For each pixel, time to biochemical recurrence was modelled using a Cox proportional hazards model accounting for contour deviation, patient age, Gleason score and treated CTV volume (the last three variables were constant for all pixels for a single patient). By assembling the hazard ratios (HR) of the 7200 Cox models in the 60x120 grid, HR maps for each variable were created, and regions of significance were found using permutation testing (permuting contour deviation maps 10^5 times, using the extreme -value as summary statistic).

Fig. 2 shows the HR map for contour deviation and high significance regions (p<0.001). There is a reduced risk of biochemical recurrence of 4-8% in the prostate’s left region, bladder and seminal vesicle interfaces per mm increase of the manual contour relative to the reference. Conversely, there is an increased risk of 8-24% in the prostate''s right and posterior regions and in the anterior region close to the apex per mm increase of the manual contour relative to the reference. HR maps for Gleason scores 6, 7, and 8 relative to >9 were significant for all pixels (p<0.001), i.e. as expected lower scores have reduced risk of biochemical recurrence independent of contour deviation. The HR maps of patient age and manual CTV volume showed no significant regions.

A novel methodology is proposed to analyse the effect of contouring variation on clinical outcome for a large cohort of patients using deviations to a consistent virtual observer. For the studied patient cohort, regions were identified in which contour deviations relate to biochemical recurrence producing some expected and unexpected results: larger contours predicts for better/worse control. This flags the need to include deviation correlations introduced by contouring styles into the analysis. In the future, this methodology could inform contouring protocols based on actual patient outcomes.