ESTRO 2025 - Abstract Book

S2434

Physics - Autosegmentation

ESTRO 2025

1358

Digital Poster Artificial Intelligence (AI) Auto-Segmentation in Clinical Target Volume of Elective Nodal Delineation in Head and Neck Cancer Radiotherapy Pathway Matthew Ward 1 , Ruchir Bhandari 2 , Kenneth Oguejiofor 2 , Virgiliu Craciun 1 , Ian S Boon 2,3 1 Department of Radiotherapy Physics, University Hospital Southampton NHS Foundation Trust, Southampton, United Kingdom. 2 Department of Clinical Oncology, University Hospital Southampton NHS Foundation Trust, Southampton, United Kingdom. 3 Clinical and Experimental Sciences, Faculty of Medicine, University of Southampton, Southampton, United Kingdom Purpose/Objective: Radiotherapy for the treatment of head and neck cancers can be complex and the workflow time consuming (i) . Clinical oncologists contour elective neck nodal clinical target volume (CTV) for irradiation according to the risks of recurrence based on accepted guidelines (ii) . An artificial intelligence (AI) software utilising a deep-learning algorithm was employed for auto-segmentation in radiotherapy planning. This study aims to assess the performance and efficiency of edited AI contours compared to clinician drawn CTV. Material/Methods: Curative-intent head and neck radiotherapy cases requiring bilateral elective neck nodal irradiation were retrospectively identified and anonymised. MVision AI commercial auto-segmentation was used. A head and neck specialist contoured level 1b-Va nodes manually (1) and by editing the AI segmentation (2). Contours were checked by a consultant radiologist and timing audits performed. AI contours were also qualitatively assessed if requiring major or minor changes. Dice indices measured the overlap between these contours as well as the unedited AI contours (3). A second set of clinical contours (4) were also included to account for inter-clinician variability. A Dice score of 0 indicates no spatial overlap and 1 indicates complete overlap of contours compared. Results: Fourteen patients were analysed. The average time for manual contouring of bilateral 1b-Va neck nodes (1) was 28 minutes (range 22-35 minutes), compared to 13 minutes (range 7-22 minutes) with AI assistance (2). Qualitative assessment showed major changes were required in 8/14 patients where elective nodal levels were inadequately covered . Minor changes were required in 6/14 patients. Dice indices averaged 0.96 between manual (1) and AI assisted (2) contours, indicating strong agreement. For manual contours (1) and unedited AI contours (3), Dice indices ranged from 0.66 to 0.85, averaging 0.76. For the second set of manual contours (4), overlap with unedited AI contours (3) averaged 0.62. Between the two sets of clinician-drawn contours (1) and (4), the average Dice index was 0.66. Conclusion: Oncologists utilising AI-assisted contouring in head and neck radiotherapy pathway can offer significant time savings. Edited AI contours are almost identical to those drawn without AI assistance, indicating that the AI introduces little-to-no bias in the shape of the final contour. Agreement between clinician-drawn and unedited AI contours is poor, confirming that raw AI cannot yet be used without clinician review and adjustments. Clinicians, physicists, and computer scientists should collaborate for AI quality assurance and to understand the weaknesses of AI for safe adoption into the radiotherapy workflow.

Keywords: artificial intelligence, head and neck cancers

References: i) Huynh, E., Hosny, A., Guthier, C. et al (2020). Artificial intelligence in radiation oncology. Nature Reviews Clinical Oncology, 17(12), 771-781. https://doi.org/10.1038/s41571-020-0417-8 ii) Grégoire V, Evans M, Le QT, et al. Delineation of the primary tumour Clinical Target Volumes (CTV-P) in laryngeal, hypopharyngeal, oropharyngeal and oral cavity squamous cell carcinoma: AIRO, CACA, DAHANCA, EORTC, GEORCC,