ESTRO 2025 - Abstract Book

S4425

Late-breaking abstracts

ESTRO 2025

Material/Methods The study included 85 consecutive patients with oral cavity, oro-/hypo-pharynx or supra-glottic larynx cancers from Sept-2023 to Jan-2025. In our institution, target contours generated by the delineating oncologist are routinely edited and approved at a target conference including a senior oncologist, a radiologist and a nuclear-medicine physician. The magnitude of changes performed to the target structures during target conference was the primary study endpoint. Patients were randomized after informed consent to the control arm (manual contouring) or the intervention arm (AI-assisted contouring). A refined version of a published nnUNet-based model [1] was used for predicting AI-contours, which were imported into the treatment planning system (TPS), using the RadDeploy framework [2]. The target conference was blinded to the origin of contours (control/intervention arm). Changes were measured using Dice score (DSC), surface Dice 1mm tolerance (sDSC), and normalized added path length (APL). Primary tumour (GTV-T) and nodal volumes (GTV-N) were analysed separately. The study aimed for non-inferiority with a margin of 10% for DSC and sDSC. Confidence intervals (CI, 95%) were calculated using bootstrapping (9999 iterations). Mann-Whitney-U test was further performed, using 0.05 as significance threshold.

Results Two patients did not fulfill inclusion criteria, leaving 83 patients for analysis. Table 1 shows patient characteristics.

All AI-contours were successfully imported in the TPS in the intervention arm. For both GTV-T and GTV-N, changes performed at target conference in the intervention (AI-assisted) arm were well within the inferiority margin with respect to the (manual) control arm, confirming the non-inferiority hypothesis. All results are illustrated in figure 1.