1045 - Automated Lymph Node Segmentation and iENE Classification Model for HPV-Associated Oropharyngeal Cancer
Presenter(s)
G. S. Dayan1, G. Hénique2, L. Létourneau-Guillon3, K. Nelson3, C. Brodeur2, A. Christopoulos1, E. J. Filion4, F. Nguyen5, B. O'Sullivan4, T. Ayad1, E. Bissada6, P. Tabet7, L. Guertin6, A. Desilets8, S. Kadoury9, and H. Bahig10; 1Division of Otolaryngology-Head & Neck Surgery, Centre Hospitalier de l'Université de Montréal (CHUM), Montreal, QC, Canada, 2Centre Hospitalier de l'Université de Montréal (CHUM) Research Center, Montreal, QC, Canada, 3Department of Radiology, Centre Hospitalier de l'Université de Montréal (CHUM), Montreal, QC, Canada, 4Department of Radiation Oncology, Centre Hospitalier de l'Université de Montréal, Montreal, QC, Canada, 5Department of Radiation Oncology, Centre Hospitalier de l'Université de Montréal (CHUM), Montréal, QC, Canada, 6Division of Otolaryngology-Head and Neck Surgery, Centre Hospitalier de l'Université de Montréal (CHUM), Montreal, QC, Canada, 7Division of Otolaryngology-Head and Neck Surgery, Centre Hospitalier de l’Université de Montréal (CHUM), Montreal, QC, Canada, 8Department of Medical Oncology, Centre Hospitalier de l'Université de Montréal, Montreal, QC, Canada, 9Polytechnique Montreal, Montreal, QC, Canada, 10Department of Radiation Oncology, Centre Hospitalier de l'Université de Montréal (CHUM), Montreal, QC, Canada
Purpose/Objective(s): Though not included in the 8th edition of the AJCC staging system, there is growing evidence suggesting that imaging-based extranodal extension (iENE) is associated with worse outcomes for HPV-associated oropharyngeal cancer (OPC). We hypothesize that an artificial intelligence (AI)-driven pipeline for lymph node segmentation and iENE classification using pre-radiation therapy planning CT scans can accurately identify iENE status and stratify patients based on oncologic risk.
Materials/Methods: From a prospectively maintained OPC database, we analyzed HPV-associated N+ OPC patients treated with (chemo)radiation between 2009-2020. We extracted pretreatment planning CT scans along with lymph node gross tumor volume (GTV-LN) segmentations performed by expert radiation oncologists. Two neuroradiologists consensually assessed iENE (grade 0 to 3) as the primary outcome. We evaluated multiple AI architectures for node segmentation, including CNNs, Vision Transformers, and hybrid models, using Dice and IoU metrics. For iENE classification (dichotomized as grade 0 vs 1,2 or 3), we compared radiomics and deep learning feature extraction methods, using PCA/LASSO feature selection, followed by Random Forest or MLP classification, with five-fold cross validation and SMOTE addressing class imbalance. The prognostic value of predicted iENE was assessed through Kaplan-Meier and multivariable Cox regression analyses.
Results: Among 397 included cases, 126 (31.7%) exhibited iENE based on expert radiological evaluation. The nnUNET segmentation model demonstrated the highest performance for GTV-LN segmentation, achieving a mean Dice Similarity Coefficient (DSC) of 81.0%. The most effective model for classifying iENE used radiomic-based feature extraction with LASSO and MLP, yielding an AUROC of 78.0 ±3.5. In Kaplan-Meier analysis, predicted iENE was associated with significantly worse oncologic outcomes, including 3-year locoregional recurrence-free survival (89.9% vs. 94.8%, P=0.016), distant recurrence-free survival (85.4% vs. 96.0%, P<0.001), disease-free survival (78.6% vs. 89.6%, P<0.001), and overall survival (86.3% vs. 94.1%, P=0.026). On multivariate analysis, predicted iENE remained an independent predictor of disease-free survival (HR 2.16, 95% CI 1.27-3.67, P=0.005), adjusting for age, ECOG performance status, T stage, and N category.
Conclusion: This study demonstrates that an AI-driven pipeline can successfully automate lymph node segmentation and iENE classification from pretreatment CT scans in HPV-associated OPC. The model achieved segmentation and classification performance that meet clinical requirements. Predicted iENE was independently associated with worse oncologic outcomes. Multi-center external validation will be needed to assess generalizability and the potential for implementing this tool to institutions without specialized imaging expertise.