ESTRO 2024 - Abstract Book

S3010

Physics - Autosegmentation

ESTRO 2024

507

Digital Poster

Assessing clinical relevance of manual autosegmentation adjustments in head and neck organs at risk

Joëlle E. van Aalst 1,2 , Tomas Janssen 3 , Roel J. H. M. Steenbakkers 2 , Ilse G. van Bruggen 2 , Jelmer M. Wolterink 4 , Stefan Both 2 , Johannes A. Langendijk 2 , Charlotte L. Brouwer 2 1 University of Twente, Technical Medicine, Enschede, Netherlands. 2 University Medical Centre Groningen, Radiation Oncology, Groningen, Netherlands. 3 Netherlands Cancer Institute, Radiation Oncology, Amsterdam, Netherlands. 4 University of Twente, Applied Mathematics, Technical Medical Centre, Enschede, Netherlands

Purpose/Objective:

Automated segmentation using deep learning contouring (DLC) of head and neck organs at risk (OARs) shows promising geometric accuracy and potential for automation of the radiotherapy treatment workflow of head and neck cancer patients. Despite this, in clinical practice, manual adjustment of DLC OARs is common, reducing the potential gain in efficiency. However, not all manual adjustments to OAR segmentation may have a relevant dosimetric effect. The dosimetric impact of OAR segmentation inaccuracies has two aspects that need to be distinguished: it influences the optimised treatment plan and it influences the dose evaluated on the OAR. Suitable measures to assess the clinical relevant impact are the ‘dosimetric impact on planning’ (DIP) and ‘dosimetric impact on evaluation’ (DIE), which can be interpreted as the impact of uncorrected autosegmentation on the optimised treatment plan and the impact on clinical evaluation respectively. The aim of this study was to assess the clinical relevance of manual adjustments of head and neck DLC OARs using the DIP and DIE. A total of 68 oropharyngeal cancer patients from a single centre treated with primary radiotherapy (70 Gy in 35 fractions) between 2018 and 2022, with or without systemic treatment, were included. For all patients, DLC sets of OARs in the head and neck region were available (WorkflowBox 2.0, Mirada Medical Ltd., Oxford, United Kingdom) as well as manually adjusted clinical contour (CC) sets. IMPT treatment plans for both contour sets were automatically created in RayStation Development 10B (RaySearch Laboratories, Stockholm, Sweden) with a U-Net deep learning model. The DIP was defined as the difference between plans optimised using the CC and DLC, evaluated on the CC, while the DIE was defined as the difference between evaluation on the CC and DLC using the plan optimised on the DLC, see Figure 1. The mean dose of 14 OARs and NTCP-values for xerostomia and dysphagia were chosen as dosimetric measures of interest. A difference in mean dose of ≥2.0 Gy or NTCP-value of ≥1.0 percentage point (pp) was defined as clinically relevant. Statistical significance was assessed with a Wilcoxon signed-rank test. Geometric comparison between the DLC and CC was performed using the Sørensen-Dice similarity coefficient (DSC). Material/Methods: