Automated Diagnosis of Ear Disease Using Deep Learning with Endoscopic Images of The Tympanic Membrane

Abstract

　 Recent advances in optical imaging technology have enabled obtention of high-definition and magnified images of the tympanic membrane. However, an accurate diagnosis from imaging findings of the tympanic membrane requires much experience and is not always easy in the absence of clinical findings. In this study, we attempted to create a machine-learning program to automatically diagnose middle ear diseases using endoscopic images of the tympanic membrane collected in our hospital. The images of the tympanic membrane were classified into five diagnostic categories (normal, acute otitis media, exudative otitis media, chronic perforating otitis media, and cholesteatoma), based on the diagnosis by a skilled otolaryngologist. Neural Network Console^® provided by SONY was used for the learning, evaluation, and calculation of the accuracy. VGG-11 and InceptionV3 were used as the learning models; a total of 544 images were collected, and the program yielded a 74% accuracy rate for the five diagnostic categories. In our study, the model using VGG-11 without fine-tuning yielded the highest accuracy. Establishment of an automated tympanic membrane diagnostic program is a potential tool to aid non-ENT physicians to diagnose otitis media, but more images would be needed to further improve the accuracy.