Artificial intelligence to detect tympanic membrane perforations

OBJECTIVE: To explore the feasibility of constructing a proof-of-concept artificial intelligence algorithm to detect tympanic membrane perforations, for future application in under-resourced rural settings. METHODS: A retrospective review was conducted of otoscopic images analysed using transfer learning with Google's Inception-V3 convolutional neural network architecture. The 'gold standard' 'ground truth' was defined by otolaryngologists. Perforation size was categorised as less than one-third (small), one-third to two-thirds (medium), or more than two-thirds (large) of the total tympanic membrane diameter. RESULTS: A total of 233 tympanic membrane images were used (183 for training, 50 for testing). The algorithm correctly identified intact and perforated tympanic membranes (overall accuracy = 76.0 per cent, 95 per cent confidence interval = 62.1-86.0 per cent); the area under the curve was 0.867 (95 per cent confidence interval = 0.771-0.963). CONCLUSION: A proof-of-concept image-classification artificial intelligence algorithm can be used to detect tympanic membrane perforations and, with further development, may prove to be a valuable tool for ear disease screening. Future endeavours are warranted to develop a point-of-care tool for healthcare workers in areas distant from otolaryngology.