AI Model Forecasts Cancer Outcomes From CT Scans in HPV-positive Throat Cancer

An artificial intelligence (AI) pipeline developed at a Montreal-based cancer center has shown promise in detecting extranodal extension (ENE) from standard CT scans in patients with HPV-positive oropharyngeal cancer—a factor associated with worse long-term outcomes.

Researchers tested the tool in 397 patients with clinically node-positive, HPV-driven oropharyngeal squamous cell carcinoma (OPSCC) who underwent chemoradiation between 2009 and 2020. The pipeline automated the segmentation of lymph nodes on pretreatment CT scans and used radiomics to classify ENE status. Known as imaging-based ENE, or iENE, this feature is not currently part of official staging criteria but has been linked to higher risks of distant failure and lower survival rates.

Detection without a pathologist

The tool identified iENE with an accuracy that matched or surpassed experienced neuroradiologists. Its predictions were associated with overall survival, recurrence-free survival, and distant control, and remained independently significant after adjusting for established clinical factors. Patients with iENE predicted by the AI model had lower three-year survival rates (83.8%) compared to those without (96.8%).

Although the eighth edition of the American Joint Committee on Cancer (AJCC) Staging Manual excludes iENE from HPV-positive cancer staging, the upcoming ninth edition is expected to include it. However, ENE detection remains highly subjective and is often unavailable outside tertiary care settings. This has prompted interest in AI as a reproducible alternative that could support treatment decisions in centers without neuroradiology expertise.

Two-step pipeline for clinical prediction

The researchers developed a two-part system: first, a 3D neural network segmented gross tumor volumes (GTVs) of pathologic lymph nodes. Second, a machine learning classifier identified the presence or absence of iENE using radiomic features. The segmentation model achieved a mean Dice similarity coefficient of 0.74, indicating good overlap with expert contours. For iENE prediction, the best-performing model used LASSO feature selection and XGBoost classification, reaching an AUC of 0.81.

To evaluate clinical impact, the team compared oncologic outcomes between patients with and without AI-predicted iENE. Multivariable analysis showed that AI-iENE remained a significant predictor of worse overall survival (adjusted hazard ratio 2.82), recurrence-free survival (aHR 4.20), and distant control (aHR 12.33). In each case, the model outperformed radiologist assessments when tested against survival data.

Clinical relevance and future applications

While the model was trained on radiologist-determined iENE labels, it demonstrated stronger correlations with survival outcomes than the original human ratings. This suggests the tool may be identifying imaging features linked to nodal aggressiveness that are less apparent to the human eye.

The authors propose that AI-driven iENE classification could help clinicians stratify patients for treatment intensity. For example, those with predicted iENE might not be suitable for treatment de-escalation protocols, while iENE-negative patients could be considered for reduced-intensity options. Importantly, the study’s authors stress that further external validation is needed and that the tool is not intended to alter clinical pathways until more data are available.

Limitations and next steps

This single-center retrospective study used iENE ratings by radiologists as the reference standard rather than pathologic confirmation, which may limit generalizability. The model assessed only the largest node per patient, which may miss smaller foci of ENE. The dataset also lacked demographic diversity, and external validation on multi-institutional cohorts is needed before clinical integration.

Despite these constraints, the researchers suggest that integrating AI tools into clinical workflows could provide consistent and accessible image analysis to improve decision-making and reduce diagnostic variability. Future research will focus on validation, performance across different imaging systems, and incorporation into clinical trials.

Reference: Dayan GS, Hénique G, Bahig H, 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., 2025. doi:10.1001/jamaoto.2025.3225

This content includes text that has been generated with the assistance of AI. Technology Networks' AI policy can be found here.