← Back to List

Program Number: B017
Session Name: Poster Session

Refining the Oropharyngeal Exam using Computer Vision Automated Segmentation of Tumors during Laryngopharyngoscopy

Anita Rau, PhD¹; Nikita Bedi, BS¹; Alberto Paderno, MD, PhD²; Yoon K So³; Masanori Teshima, MD⁴; Hlu Vang, BS¹; Serena Yeung-Levy, PhD¹; F. Christopher Holsinger, MD¹; ¹Stanford University; ²Humanitas University; ³Ilsan Paik Hospital, Inje University College of Medicine; ⁴Kobe University

Background: Oropharyngeal carcinoma (OPC) can be difficult to detect and delineate during endoscopy, due to their submucosal origin in lymphoepithelial crypts. To facilitate the detection of OPC and the understanding of the endoscopic scene, we employ artificial intelligence (AI) to segment tumors and other anatomically significant structures in the oropharynx.

Methods: Videos from 106 patients undergoing routine laryngopharyngoscopy in the clinic were analyzed, extracting 942 images. Each pixel was annotated as one of 12 semantic classes. We trained an AI algorithm and evaluated its segmentation performance. Our AI Pipeline is shown in Figure 1. We used the FCBFormer model for the soft-tissue discrimination task – a method using a transformer backbone with convolutional neural network features.

Results: A deep-learning algorithm could identify tumor within an individual frame of video with true positive rate (sensitivity) of 93% with a corresponding dice score coefficient of 80%. Figure 2. For the detection, corresponding dice score were seen for pharyngeal mucosa (0.90), epiglottis (0.88), and vocal cords (0.82).

Conclusions: This study demonstrates that AI is able to segment endoscopic images of the oropharynx, paving the way for computer-assisted systems that could help identify OPC and enable other applications beyond cancer treatment, such as computer-assisted navigation during endoscopy or intubation.

 

 

← Back to List