KEY POINTS
- This prospective, single-center, randomized clinical evaluation from Aarhus University Hospital, Denmark, enrolled 85 patients with head and neck squamous cell carcinoma between September 2023 and January 2025. Following two post-randomization exclusions, 83 patients were analyzed, including 41 assigned to manual GTV contouring and 42 to AI-assisted contouring. Eligible patients had oral cavity, oropharyngeal, hypopharyngeal, or supraglottic laryngeal cancer treated with curative-intent radiotherapy; oropharyngeal tumors accounted for 71% and 79% of the respective groups.
- AI-generated contours were integrated into the existing clinical planning workflow rather than used autonomously. The intervention used a five-model ensemble of 3D nnU-Net networks trained on 482 historical patients, generating initial GTV contours from multimodal planning imaging. Radiation oncologists could accept, modify, or discard these predictions, after which both study groups underwent the same multidisciplinary target conference involving an oncologist, neuroradiologist, and nuclear medicine physician. Conference reviewers were blinded to whether initial contours were manual or AI-assisted.
- AI assistance did not meaningfully reduce editing of primary tumor contours during peer review. The primary endpoint was normalized Added Path Length (APL), measuring the extent of contour modifications between initial delineation and final multidisciplinary approval. The estimated median difference between AI-assisted and manual contouring was −0.02 (95% CI −0.32 to 0.07) for primary tumor GTV (GTV-T), with negative values favoring AI. However, this exploratory trial did not formally test noninferiority, and the confidence interval does not establish equivalence.
- Nodal GTVs showed a possible increase in editing with AI assistance. The median APL difference for GTV-N was +0.06 (95% CI −0.01 to 0.14). After excluding contours containing uninterpolated or fragmented sections, the difference remained +0.06, with the 95% CI shifting to 0.01–0.14, suggesting more nodal editing in the AI group. Other geometric metrics, including volumetric Dice, surface Dice, and volume differences, did not show corresponding clear between-group differences, so the finding remains exploratory.
- Neither initial contouring time nor multidisciplinary review duration improved with AI. Median target-conference time was 10.0 minutes with manual contouring versus 10.3 minutes with AI assistance. Self-reported initial GTV-T contouring time was 30 minutes in both groups, while GTV-N contouring required 15 minutes in both groups. Perceived task complexity was also broadly comparable, with slightly lower reported complexity for primary tumors but higher complexity for nodal contours in the AI group.
- The primary endpoint was sensitive to incomplete contour preparation. Some initial manual contours contained uninterpolated sections, leading to disproportionately large changes during the target conference. Excluding these cases shifted the estimated GTV-T APL difference from −0.02 to +0.03 (95% CI −0.10 to 0.10), while the nodal editing signal became more pronounced. Post hoc simulations suggested that a future noninferiority trial using an absolute APL margin of 0.10 would require approximately 50 patients per arm for GTV-T and 60 per arm for GTV-N to achieve 90% power.
- The study did not assess the downstream clinical consequences of AI-derived target contours. No study-related adverse events were reported, but treatment-plan dosimetry, geographic miss, tumor control, and radiation toxicity were not evaluated. The single-center design, relatively small sample, self-reported rather than objectively measured contouring times, and dependence on an established multidisciplinary review workflow limit generalizability. The findings also cannot establish how independently generated or unreviewed AI contours would perform.
CLINICAL TAKEAWAY
AI-assisted GTV delineation can be incorporated into a multidisciplinary head and neck radiotherapy workflow without a demonstrated increase in primary tumor contour editing, but this randomized exploratory study found no measurable efficiency benefit. The possible increase in nodal contour modifications warrants further evaluation, and formal noninferiority, dosimetric safety, and clinical effectiveness remain unproven. AI-assisted tumor segmentation should continue to undergo expert review rather than being assumed to perform like established OAR auto-contouring systems.