Extranodal extension (ENE) has long been recognised as an adverse prognostic factor in head and neck cancers, but it is not included in the eighth edition of the American Joint Committee on Cancer (AJCC) staging system for HPV-associated oropharyngeal carcinoma (OPC). Imaging-based ENE (iENE) could provide valuable prognostic information, yet clinical use has been limited by the absence of standardised criteria, the need for specialised radiological expertise, and significant interobserver variability.
A new study from a Canadian tertiary oncology centre suggests that artificial intelligence (AI) may offer a solution. Researchers developed an automated pipeline to segment lymph nodes and classify iENE from pretreatment CT scans in patients with HPV-positive OPC, then assessed its prognostic value.
LISTEN FOR THE LATEST RESEARCH, ANALYSIS AND A GOOD DOSE OF HUMOUR…
Study details
The retrospective cohort included 397 patients with cN+ HPV-positive OPC (mean age 62.3 years; 20% female) treated with up-front chemoradiotherapy or radiotherapy between 2009 and 2020. Follow-up extended through January 2024. Expert radiation oncologists provided ground-truth lymph node segmentations.
For AI modelling, investigators trained an nnU-Net algorithm for automated node segmentation, followed by iENE classification using radiomic and deep learning features. Performance was benchmarked against two expert neuroradiologists.
Performance and outcomes
Classification accuracy: Radiomics-based AI achieved an area under the curve (AUC) of 0.81 for iENE detection.
Survival associations: Patients with AI-predicted iENE had significantly worse 3-year outcomes:
- Overall survival (OS): 83.8% vs 96.8%
- Recurrence-free survival (RFS): 80.7% vs 93.7%
- Distant control (DC): 84.3% vs 97.1%
- Locoregional control (LRC): no significant difference
- Prognostic value: AI-iENE showed higher concordance indices than radiologist-assessed iENE for OS (0.64 vs 0.55), RFS (0.67 vs 0.60), and DC (0.79 vs 0.68).
- Multivariable analysis: AI-predicted iENE remained independently associated with outcomes, even after adjusting for age, T category, N category, and lymph node burden:
- OS: aHR 2.82 (95% CI 1.21–6.57)
- RFS: aHR 4.20 (95% CI 1.93–9.11)
- DC: aHR 12.33 (95% CI 4.15–36.67)
Implications
The findings suggest that an AI-driven approach can reliably identify iENE and may outperform expert readers in predicting survival outcomes for HPV-associated OPC. If validated externally, such tools could help standardise iENE assessment, guide treatment intensification, and potentially extend advanced imaging expertise to centres without subspecialist radiologists.
Conclusion
This single-centre study demonstrates that AI-based iENE classification from routine pre-treatment CT scans is feasible, accurate, and clinically meaningful. AI-predicted iENE was independently associated with inferior overall, recurrence-free, and distant control outcomes in HPV-positive OPC. Future multi-centre validation will be essential before widespread adoption.
Paper: Dayan, G. et al. Artificial Intelligence Model for Imaging-Based Extranodal Extension Detection and Outcome Prediction in Human Papillomavirus−Positive Oropharyngeal Cancer. JAMA Otolaryngol Head Neck Surg. doi:10.1001/jamaoto.2025.3225 Published online September 30, 2025.
