The long-term goal of this project is to improve the care of patients with laryngeal disorders through development of automated diagnostic support for in-office flexible laryngoscopy. To accomplish this goal, we propose developing neural network-based algorithms to detect and classify structural laryngeal lesions in laryngoscopy images. An automated diagnostic tool for in-office laryngoscopy such as we propose will have several benefits: (1) It will improve access to care for patients with symptoms of laryngeal dysfunction living in communities with limited otolaryngology resources, (2) It will improve early detection of laryngeal cancers potentially reducing the morbidity of treatment, and (3) It will prove a valuable teaching tool for students and residents first learning to interpret laryngoscopic exams. Flexible laryngoscopy is a common in-office procedure performed by otolaryngologists to evaluate the upper aerodigestive tract in patients with symptoms of laryngeal dysfunction. Subtle differences in the appearance of laryngeal lesions enable otolaryngologists to differentiate benign lesions from suspected malignant ones. The expertise and clinical acumen to correctly interpret laryngoscopic findings requires years of training and therefore laryngoscopy is largely only performed in subspecialty otolaryngology clinics. The primary objective of this project is to develop neural network-based algorithms to detect and classify structural laryngeal lesions. Our hypothesis is that these algorithms can be trained using a large dataset of laryngeal images to accurately detect and classify structural laryngeal lesions on flexible laryngoscopic exam. To test this hypothesis, we propose the following aims: (1) Generate a dataset of high-quality, labeled endoscopic laryngeal images corresponding to normal and structural lesions of the larynx, (2) Develop a location-aware anchor-based reasoning neural network for accurate detection of laryngeal lesions, and (3) Develop an adaptive network model for classification of structural laryngeal pathologies including papilloma, polyp, leukoplakia and suspected malignancy. With expertise in the diagnosis and treatment of laryngeal disorders and computer vision, including object detection and classification, our multidisciplinary team is uniquely qualified to complete this project.
We propose to revolutionize in-office laryngoscopy through development of a deep neural network-based automated detection and classification system for diagnosis of structural diseases of the larynx. Currently, flexible laryngoscopy is only performed by expert subspecialists with years of experience because developing the expertise and clinical acumen to correctly interpret laryngoscopic findings requires years of training. Through development of deep neural network-based algorithms to detect and classify laryngeal lesions on in- office laryngoscopy, we will improve access to care for patients living in communities without subspecialty otolaryngology care and will develop an important teaching tool for clinicians learning to interpret laryngoscopic exams.