ESTRO 2025 - Abstract Book

S2453

Physics - Autosegmentation

ESTRO 2025

1968

Digital Poster Development of an in-house tool for head and neck cancer organ-at-risk autosegmentation Tom Young 1,2 , Tom Roberts 3 , Michael Woodward 3 , Anil Mistry 3 , Christopher Thomas 1,4 , Sarah Misson 1 , Victoria Butterworth 1,5 , Delali Adjogatse 1,2 , Imran Petkar 1,2 , Miguel Reis Ferreira 1,2 , Anthony Kong 1,2 , Mary Lei 1,2 , Andrew King 4 , Teresa Guerrero Urbano 1,2 1 Radiotherapy, Guy's and St Thomas' NHS Foundation Trust, London, United Kingdom. 2 School of Cancer & Pharmaceutical Sciences, King's College London, London, United Kingdom. 3 Clinical Scientific Computing, Guy's and St Thomas' NHS Foundation Trust, London, United Kingdom. 4 Biomedical Engineering & Imaging Sciences, King's College London, London, United Kingdom. 5 School of Cancer & Pharmaceutical Sciences, Guy's and St Thomas' NHS Foundation Trust, London, United Kingdom Purpose/Objective: Contouring of organs-at-risk (OARs) and target volumes is a key task within the radiotherapy (RT) workflow but is time-consuming and subject to variability 1 . Sub-optimal contouring may affect outcomes 2 . Software using Artificial Intelligence (AI) to autosegment OARs and elective nodal regions has been developed, with state-of-the-art model architectures such as nnU-net available open source. Commercial solutions have been approved for use while data on real-world efficacy and efficiency is collected 3 , however drawbacks include high capital/operational costs, unknown training data provenance, models not trained to a deploying centre’s needs 4 and limitations for users to update and retrain already deployed models. We therefore developed an in-house HNC OAR autosegmentation pipeline (GSTTAutosegHN-OAR) to enable AI lifecycle completion. Material/Methods: 1417 contours for 22 OARs (from 102 HNC-RT patients treated at our centre with radical RT between 1/1/2020 and 31/12/2022) were reviewed and edited following locally-used guidelines by one Radiation Oncologist, before further review by 2 others. Contours were uploaded to XNAT (imaging data repository). A pipeline to extract relevant contours from XNAT to train nnU-net models (3D-full resolution model, 1000 epochs, 0.9 training/0.1 test split) was developed and deployed. 4 sub-models were trained to maximise training data available: • Model 1: Parotids/spinal cord/brainstem/lenses • Model 2: Larynx/oral cavity/pharyngeal constrictor muscles (inferior/middle/superior) • Model 3: Optic chiasm/nerves/retinas/cochleas/pituitary • Model 4: Submandibular glands/mandible Models were evaluated by assessing validation (generated during training to tune hyperparameters) and test (unseen) contours against training dataset contours to generate Volumetric Dice Similarity Coefficient (VDSC) values. Medical Imaging Interaction Toolkit was used to visualise test set contours. Results: 16 structures demonstrated median test cohort VDSC of ≥0.8 (Table 1). Structures with median test cohort VDSC less than 0.8 were smaller structures, with worst performance for optic chiasm (0.60). Median test and validation VDSC values were close for most structures suggesting robust model generalisation. Informal qualitative review (Figure 1) observed good quality contours for all cases reviewed.