Fusing fine-tuned deep features for recognizing different tympanic membranes

Abstract

Otitis media (OM) refers to a group of inflammatory diseases regarding the middle ear. Although there are a wide variety of disease types regarding OM, the most commonly seen disorders are acute otitis media (AOM), otitis media with effusion (OME) and chronic suppurative otitis media (CSOM). The examination of OM in the clinics is realized subjec tively. This subjective examination is error-prone and leads to a limited variability among specialist. For these reasons, computer-aided systems are in demand. In this study, we focus on recognizing normal, AOM, CSOM, and earwax tympanic membrane (TM) conditions using fused fine-tuned deep features provided by pre-trained deep convolutional neural networks (DCNNs). These features are applied as the input to several networks, such as an artificial neural network (ANN), k-nearest neighbor (k NN), decision tree (DT) and support vector machine (SVM). Moreover, we release a new publicly available TM data set consisting of totally 956 otoscope images. As a result, the DCNNs yielded promising results. Especially, the most efficient results were provided by VGG-16 with an accuracy of 93.05 %. The fused fine tuned deep features improved the overall classification success. Finally, the proposed model yielded promising results with an accuracy of 99.47 %, sensitivity of 99.35 %, and specificity of 99.77 % using the combination of the fused fine-tuned deep features and SVM model. Consequently, this study shows that fused fine-tuned deep features are rather useful in recognizing different TMs and these features can provide a fully automated model with high sensitivity.