Using Artificial Intelligence to Identify Oropharyngeal Cancer During Fiberoptic Endoscopy

OBJECTIVE: Oropharyngeal tumors with a propensity for submucosal spread can be difficult to detect and delineate during endoscopy. To facilitate their detection, we developed an Artificial Intelligence (AI) algorithm to segment tumors and other anatomically significant structures within the oropharynx. STUDY DESIGN: Model development and validation study. SETTING: Tertiary academic medical center. METHODS: We collected videos from 106 patients undergoing pharyngolaryngoscopy, extracted 942 images, and annotated each pixel as one of 12 classes, such as tumor, uvula, and healthy mucosa. We trained an AI algorithm and evaluated its performance in multiclass and tumor-based segmentation. RESULTS: The AI algorithm achieved a recall statistic of 0.93 and a Dice score of 0.79 for tumor detection. High Dice scores were observed for pharyngeal mucosa (0.90), epiglottis (0.88), and vocal cords (0.82). CONCLUSIONS: This study demonstrates that AI can semantically segment all the structures visualized during endoscopy, paving the way for computer-assisted systems for tumor identification and computer-assisted navigation during endoscopy or intubation.