ESTRO 2024 - Abstract Book

S3054

Physics - Autosegmentation

ESTRO 2024

1459

Digital Poster

Comparing Delineation Uncertainty between Manual and Auto-contours for Cervical Cancer Radiotherapy

Katherine Mackay 1,2 , David Bernstein 3,2 , Ben Glocker 4 , Daniel Kaye 4 , Sooha Kim 2 , Binnaz Yasar 1,2 , Leslie Cheng 1,2 , Stephen Robinson 5,6 , Alexandra Taylor 1,2 1 The Royal Marsden NHS Foundation Trust, Department of Radiotherapy, London, United Kingdom. 2 The Institute of Cancer Research, Radiotherapy and Imaging, London, United Kingdom. 3 The Royal Marsden NHS Foundation Trust, Department of Physics, London, United Kingdom. 4 Imperial College London, Department of Computing, London, United Kingdom. 5 University Hospitals Sussex NHS Foundation Trust, Sussex Cancer Centre, Brighton, United Kingdom. 6 University of Sussex, Department of Biochemistry, Brighton, United Kingdom

Purpose/Objective:

Auto-contouring has the potential to improve the consistency of target and organ-at-risk delineation in radiotherapy. Despite increasing availability, there is a lack of consensus regarding the most clinically useful methods to assess and validate auto-contouring systems for clinical use. Variation in contouring can be quantified as the delineation uncertainty [1]. This value is used in CTV to PTV margin calculations [2]. Delineation uncertainty can be calculated for manual contours and can also be estimated for an auto-contouring system. Although it provides a clinically meaningful value, delineation uncertainty has been rarely used to assess auto-contouring systems. The aims of this study were to measure the delineation uncertainty from a cohort of clinicians contouring external beam radiotherapy structures for cervical cancer and compare this to the delineation uncertainty for auto-contoured structures.

Material/Methods:

A nnUNet based auto-contouring system [3] was trained to segment CTV-nodes, CTV-primary (uterus, cervix and vagina), parametrium, anorectum and bladder using 100 radiotherapy planning CT scans from a single centre. All contours were peer reviewed prior to inclusion in the training dataset. The resulting system was then applied to produce auto-contours on six radiotherapy planning scans which were independent from the training dataset. An inter-observer manual delineation study was conducted on these same six cases. This involved six clinicians with at least three years’ radiotherapy experience who were individually trained in the contouring protocol including test cases. These clinicians contoured the individual structures within the CTV nodes and CTV primary as well as the parametrium, anorectum and bladder. All contours were reviewed by two experts prior to analysis. Contours that were not protocol compliant were excluded from further analysis at the locations of protocol deviation. The delineation uncertainty was calculated within each of the cardinal axes by taking measurements in the superior, inferior, left, right, anterior and posterior directions at multiple locations within each structure and calculating the overall delineation uncertainty using the pooled standard deviation. The delineation uncertainty for both the auto contouring system and the manual contours were calculated using the BIR methodology described below [1]. For manual contours, the standard deviation of distances to a set reference point for all protocol compliant contours at each landmark was measured, as demonstrated by equation 1. For auto-contours the distance between the auto contour and two expert contours and the distance between two expert contours at each landmark was used as per