Automatic detection of eardrum otoendoscopic images in patients with otitis media using hybrid-based deep models

Abstract

Otitis media with effusion (OME) is fluid accumulation in the middle ear without signs of systemic infection. OME can cause hearing loss, ear fullness, speech retardation, and a decrease in social relations and school success. In the treatment of OME, methods such as medical treatment and placing a tympanostomy tube in the eardrum are used. Correct evaluation of the eardrum is necessary to diagnose OME and follow-up properly in the following period. In this study, we aimed to evaluate the otoendoscopic images of the eardrum in patients with otitis media with effusion using deep models. In the proposed model, Efficientnetb0, Darknet53, and Densenet201 architectures are used as the base. The feature maps obtained using these architectures are combined. In this way, different features of the otoendoscopic images of the eardrum in the data set were extracted and combined using different architectures. neighborhood component analysis optimization technique is used to select important features from the combined feature maps. By using this optimization method, the size of feature maps is reduced. In this way, the model is trained faster, and the testing process is carried out in a shorter time. When these features are classified in different classifiers, the model we propose has the most successful average accuracy value of 94.27% in the SVM classifier.