ESTRO 2025 - Abstract Book

S1108

Clinical – Head & neck

ESTRO 2025

Sherman, E.J., et al., Radiotherapy and paclitaxel plus pazopanib or placebo in anaplastic thyroid cancer (NRG/RTOG 0912): a randomised, double-blind, placebo-controlled, multicentre, phase 2 trial. Lancet Oncol, 2023. 24(2): p. 175 186.

4004

Digital Poster Evaluation of a deep-learning based auto-segmentation tool for organs at risk in extensively preoperated head and neck cancer patients Christoph Dumke 1 , Rieke Dumke 2 , Tobias Hofmann 1 , Norina Predescu 3 , Stefan Lautenschläger 4 1 Practice for Radiotherapy, Vivantes, Berlin, Germany. 2 Zentrum für Versorgungsforschung, Medizinische Hochschule Brandenburg Theodor Fontane, Rüdersdorf, Germany. 3 MVision, MVision, Helsinki, Finland. 4 Clinic for Radiotherapy, Clinic Lippe, Lemgo, Germany Purpose/Objective: For extensively preoperated head and neck cancer patients (epHNCP) the delineation of organs at risk (OARs) can be challenging and therefore very time consuming. Especially for this situation, deep-learning based auto segmentation may be helpful. Therefore, we analyzed time saving and consistency between manually delineated and autocontoured OARs via an auto-segmentation tool (MVision) in epHNCP. Material/Methods: Planning CT-scans of 10 epHNP with bilateral neck dissection and flap reconstruction were retrospectively obtained at the Vivantes Practice for Radiotherapy in Berlin-Spandau, Germany. Contouring of OARs (parotids, submandibular glands (=SMG), supraglottic larynx (=SLA), glottic area (=GA), pharyngeal constrictor muscle (=PCM), cochlea, brainstem, spinal cord (=SC), mandible) was done manually. For the same OARs auto-segmentation was performed. These autocontours were visually evaluated with a Lickert-scale of 1 to 5 (1 = rejection - 5 = no editing necessary). Subsequently, autocontours were edited until clinically acceptable. For both manual contouring and editing of autocontours the contouring-time was recorded and time saving calculated. The similarity between manual contours and edited autocontours was assessed by Dice-Similarity-Coefficient (DSC), surface-Dice-Similarity Coefficient (s-DSC), and Hausdorff-Distance (HD95). Results: The manual contouring time for OARs was median 20:38 min (range 14:32 – 23:29 min) per patient and this time was reduced to a median of 3:56 min (range 1:31 – 09:06 min) by using the auto-segmentation tool (81% time reduction). The OARs with highest median time saving were mandible 5:54 min (98%), SC 1:05 min (100%), parotids 1:57 min (79%), GA 1:50 min (98%) and brainstem 1:07 min (100%). The lowest was cochlea 0:06 min (19%) and moderate time saving was achieved for SLA 1:05 min (58%), SMG 1:07 min (62%) and PCM 2:09 (73%). OARs with highest time saving were generally rated higher on the Lickert-scale than those with low or moderate time saving (5/4 versus 4/3). None of the structures were rejected and the majority needed only minor editing (see Table 1). In accordance, most of the OARs showed a high DSC and/or s-DSC of around 0.8, except Cochlea and GA with values around 0.6 to 0.7. Furthermore, all OARs showed a low HD95 of mean 2-9 mm (see Table 2).