Deep learning for automatic interpretation of pure-tone audiometry using InceptionV3

Abstract

Abstract Problem Deafness represents a public health issue because of its frequency and its impact on personal and professional life. Its diagnostic and therapeutic management requires an otoscopic examination followed by a pure-tone audiometry. However, there is a lack of ENT specialists capable of interpreting tone audiometry. To overcome this obstacle, an automation of the interpretation of audiometry can solve this problem and thus improve the management of hearing-impaired patients. Objective To classify audiogram images into normal hearing, conductive, sensorineural and mixed hearing loss using a convolutional neural network. Methods Retrospective study of pure-tone audiometries performed in the functional exploration unit of the ENT Department of Charles Nicolle Hospital in Tunis during the year 2021. These pure-tone audiometries were labelled into four types: normal hearing, conductive hearing loss, sensorineural hearing loss and mixed hearing loss. Then, the images were divided into training, validation and test groups and the input format was standardised. A data base augmentation to increase the number of examples in the training group was performed. Finally, a transfer learning was performed using the convolutional neural network InceptionV3. Results One thousand audiograms were collected: 261 normal hearing, 156 conductive hearing, 316 sensorineural hearing and 267 mixed hearing. This database was divided into three groups: 80% ‘training set’, 10% ‘validation set’ and 10% ‘test set’. The data augmentation of the training set provided 5068 images. The convolutional neural network was trained for 20 epochs. This allowed us to obtain an accuracy of 99.17% in the ‘training set’, 96.88% in the ‘validation set’ and 95.24% in the ‘test set’. Conclusion Our study has shown that deep learning is a simple and feasible solution for the automatic interpretation of pure-tone audiometry.