Artificial Intelligence for Endoscopic Surveillance Post‐Treatment for Nasopharyngeal Carcinoma: Detection of Recurrence and Osteoradionecrosis

Radiotherapy is the main treatment for nasopharyngeal carcinoma (NPC) and has significantly improved survival [1]. However, recurrent NPC (rNPC) rates are 10%–20% [2], and skull base osteoradionecrosis (SBORN) remains a serious complication [3]. Early recognition and management of both conditions are critical, making post-treatment endoscopic surveillance essential. Recent advances in vision transformer (ViT) architectures have transformed image analysis. Unlike commonly used convolutional neural networks, ViTs use self-attention mechanisms that capture global context and perform well even on modestly sized datasets [4]. Existing literature on endoscopy image machine-learning model (MLM) analysis has focused primarily on diagnostic applications, rather than on longitudinal surveillance following definitive therapy [5-7]. We developed and evaluated a ViT-based MLM, DINOv2, for automated identification of rNPC and SBORN on post-radiation nasopharyngeal endoscopic images. This study was approved by the Stanford University Institutional Review Board (#70442). Records of patients treated with radiotherapy for NPC between 2013 and 2024 were reviewed. Clinical data were retrieved from the STAnford Research Repository, and corresponding endoscopic images were obtained from VaultstreamGX (Olympus). Images were classified as normal, SBORN, or rNPC based on histopathologic or radiologic confirmation. Three rhinologists independently annotated all images using the Encord platform, resolving discrepancies by consensus. Images obscured by mucus or blood were excluded, while those with crusting or unclear anatomical landmarks were retained to reflect real-world variability. For model development, we utilized the DINOv2-ViT-B/14 architecture implemented in PyTorch [8] (Figure S1). Two models were trained: (1) binary (SBORN vs. non-SBORN) and (2) three-class (normal, SBORN, and rNPC). Training followed a two-phase strategy: Phase 1: the pretrained backbone was frozen while the classification head was optimized; Phase 2: the full network was fine-tuned end-to-end to adapt to domain-specific features. Images were resized to 224 × 224, normalized using ImageNet parameters, and augmented with random rotation (±15°), horizontal flipping, and mild color variation to enhance model generalization and mitigate overfitting. Weighted sampling addressed class imbalance. Optimization used AdamW (learning rate = 1 × 10−⁴, weight decay = 1 × 10−⁵). Learning curves of training and validation loss are presented in Figures S2 and S3. Data were split 8:1:1 into training, validation, and test sets with 10-fold cross-validation. Performance metrics included accuracy, precision, recall, specificity, F1-score, and AUC (scikit-learn, Python). A total of 1483 post-radiation nasopharyngeal endoscopic images from 192 patients were included in the dataset. Note that 866 images were classified as normal, 490 as SBORN, and 127 as rNPC. The model achieved an accuracy of 90.2% in the validation set and 85.3% in the test set. For SBORN, test precision was 87.2%, recall 80.9%, and specificity 94% (Figure 1 and Figure S4) The model achieved an accuracy of 78.0% in the validation set and 73.9% in the test set. Class-specific test accuracies were 81.6% for normal, 84.8% for SBORN, and 82% for rNPC. For SBORN, the precision was 76.0%, recall 75.3%, and specificity 89.1%. Performance for normal mucosa showed precision of 88.3%, recall 75.9%, and specificity 87.4%. For rNPC, performance was lower (precision 40.0%, recall 62.4%, and specificity 85.0%) (Figure 2 and Figure S5). This study demonstrates the feasibility of a ViT-based model for automated endoscopic surveillance following radiotherapy for NPC. The model performed strongly for SBORN and normal mucosa, achieving accuracy of 85.3% and 81.6%, respectively. Precision and recall for SBORN indicate that the model may serve as a useful diagnostic aid. rNPC classification showed moderate accuracy and specificity but reduced precision, limiting its role as a standalone screening tool for rNPC. Misclassification analysis revealed that SBORN was most often confused with rNPC, likely reflecting overlapping post-radiation endoscopic features such as crusting and granulation tissue (Figure S6). These similarities, combined with the limited visualization of submucosal necrotic bone, may contribute to diagnostic delays in clinical practice [2, 9]. Furthermore, misclassification analysis revealed that rNPC was most often confused with normal tissue, which highlights the inherent challenge of identifying submucosal recurrences, which can appear endoscopically normal. Future datasets with a higher proportion of rNPC images and stronger augmentation could improve recurrence detection. Combining AI-assisted endoscopy with complementary imaging modalities, such as MRI, PET/CT, and narrow-band imaging, may further enhance detection sensitivity [5, 7]. Multimodal integration could improve recognition of subtle or submucosal recurrences that are difficult to identify on white-light endoscopy alone. Clinically, such models could potentially serve as real-time decision-support tools during endoscopy, flagging suspicious regions or high-risk areas that would warrant closer follow-up. This study has several limitations. The generalizability may be limited due to the single-center design. Another limitation involves inherent challenges in ViT models, which can lack explainability in the model's decision-making, such as gradient-weighted attention maps. Nonetheless, this study demonstrates the potential of artificial intelligence analyses to augment traditional endoscopic surveillance for NPC, potentially enabling earlier detection of complications and recurrence. Conceptualization: Michael T. Chang, Lirit Levi, and Yi Wang. Data curation: Lirit Levi, Jeffrey D. Huynh, Maxime Fieux, Mahdokht S. Manavi, Axel Renteria, and Shiv K. Sethi. Formal analysis: Yi Wang. Methodology: Yi Wang, Michael T. Chang, and Lirit Levi. Supervision: Michael T. Chang. Validation: Yi Wang. Visualization: Lirit Levi. Writing – original draft: Lirit Levi. Writing – review and editing: Lirit Levi, Michael T. Chang, Maxime Fieux, Mahdokht S. Manavi, Axel Renteria, Shiv K. Sethi, Noel Ayoub, Zara Patel, Jayakar V. Nayak, and Peter Hwang. All authors have read and agreed to the published version of the manuscript. This study was funded by Stanford University Center for Asian Health Research and Education (CARE). The study was approved by an international review board (No. 70442). All data were de-identified, and a waiver for patient consent was obtained. The authors declare no conflicts of interest. Data supporting the reported results are available on reasonable requests. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.