Otitis media (OM) is a term used to describe the inflammation of the middle ear. The clinical inspection of the tympanic membrane is conducted visually by experts. Visual inspection leads to limited variability among the observers and includes human-induced errors. In this study, we sought to solve these problems using a novel diagnostic model based on a faster regional convolutional neural network (Faster R-CNN) for tympanic membrane detection, and pre-trained CNNs for tympanic membrane classification. The experimental study was conducted on a new eardrum dataset. The Faster R-CNN was initially applied to the original images. The number of images in the dataset was subsequently increased using basic image augmentation techniques such as flip and rotation. We also evaluated the success of the model in the presence of various noise effects. The original and automatically extracted tympanic membrane patches were finally input separately to the CNNs. The AlexNet, VGGNets, GoogLeNet, and ResNets models were employed. This resulted in an average precision of 75.85% in the tympanic membrane detection. All CNNs in the classification produced satisfactory results, with the proposed approach achieving an accuracy of 90.48% with the VGG-16 model. This approach can potentially be used in future otological clinical decision support systems to increase the diagnostic accuracy of the physicians and reduce the overall rate of misdiagnosis. Future studies will focus on increasing the number of samples in the eardrum dataset to cover a full range of ontological conditions. This would enable us to realize a multi-class classification in OM diagnosis. (C) 2019 Elsevier Ltd. All rights reserved.