S635
Clinical – Head & neck
ESTRO 2026
0.89, p=0.24) (Figure 1). This indicates that oncologists detected involved nodes missed by the AI, while false positives were similar.
Mini-Oral 3772 Qualitative analysis of the results of a prospective randomized clinical trial for AI-assisted GTV delineation in head and neck cancer Jannis Rosenberger 1 , Mathis E Rasmussen 2,3 , Jintao Ren 1,3 , Anne IS Holm 4 , Jasper A Nijkamp 1,3 , Jesper G Eriksen 2 , Stine S Korreman 1,3 1 Department of Clinical Medicine, Aarhus University, Aarhus, Denmark. 2 Department of Experimental Clinical Oncology, Aarhus University Hospital, Aarhus, Denmark. 3 Danish Center for Particle Therapy, Aarhus University Hospital, Aarhus, Denmark. 4 Department of Oncology, Aarhus University Hospital, Aarhus, Denmark Purpose/Objective: The PROSA-GTV prospective randomized blinded clinical trial demonstrated non-inferior performance of AI-assistance compared to manual delineation of GTV in head and neck cancer [1]. Given the trial’s explorative nature, we here examine qualitative patterns in AI performance and use. Material/Methods: Eighty-five consecutive patients with oral cavity, oro- /hypopharyngeal, or supraglottic laryngeal cancer (enrolled Sept2023–Jan2025 in our institution) were randomized 1:1 after informed consent to manual (control) or AI-assisted (intervention) contouring of primary tumor (GTV-T) and involved nodes (GTV-N). A refined nnUNet-based model [2] predicted AI contours, imported into the treatment planning system via RadDeploy [3].In our institution, targets initially contoured by the delineating oncologist are routinely edited and approved at a multidisciplinary conference including a senior oncologist, radiologist and nuclear-medicine physician. The conference was blinded to the origin of contours (control/intervention).With conference-approved structures as ground-truth, true positive (TP), false positive (FP), and false negative (FN) structures were determined for raw AI predictions and oncologist delineations. Recall (TP/[TP+FN]) and precision (TP/[TP+FP]) were calculated per patient. Cohort means are reported separately for GTV-T/GTV-N.We addressed three hypotheses:1) For GTV-N, raw AI predictions exhibit larger difference from approved structures than oncologists’ delineations.2) AI predictions are more accurate for GTV-T than for GTV- N.3) The difference between raw AI predictions and approved structures is independent of study arm.Statistical significance was assessed using Mann– Whitney-U test ( α = 0.05). Results: For GTV-N, raw AI predictions showed significantly lower recall than oncologist delineations (0.58 vs 0.86, p<0.001), while precision was comparable (0.82 vs
For GTV-T, AI recall of 0.89 was significantly higher than for GTV-N (p<0.001). Precision was 0.91, comparable to GTV-N (p=0.05). This indicates high performance in primary tumor detection. However, the AI failed to identify additional lesions in patients with multiple GTV-Ts.For control vs intervention, recall for GTV-T was 0.85 vs 0.92 (p=0.54), and precision 0.88 vs 0.93 (p=0.52). For GTV-N, recall was 0.56 vs 0.55 (p=0.72) and precision 0.76 vs 0.84 (p=0.74), demonstrating no significant differences between arms. Conclusion: The AI model generally performed well, however with underperformance in detection of involved nodes and multi-structure primary tumors. Oncologist oversight mitigated these deficiencies, achieving comparable results to manual delineation. The trial outcomes identify key targets for model refinement and attention points during clinical integration. References: [1] Korreman S, Ren J, Nijkamp JA, …, and Rasmussen ME, “First-in-world demonstration of benefit of AI assisted head and neck cancer target contouring in a prospective blinded randomized clinical trial”, Radiother Oncol 206, pp. S4424-S4427, 2025.[2]
Made with FlippingBook - Share PDF online