Optimized fine-tuned ensemble classifier using Bayesian optimization for the detection of ear diseases

Abstract

External and middle ear diseases are common disorders, especially in children, and can be examined using a digital otoscope. Hearing loss can result from delayed diagnosis and treatment which is subjective and error-prone depending on the expertise of the otolaryngologist. For these reasons, deep learning-based automated diagnostic systems are highly needed. In this study, a novel weighted average voting ensemble classifier between MobileNet and DenseNet169 has been developed to diagnose and detect different ear conditions. Bayesian optimization was used to select hyperparameters that gave the best results during the training process. MobileNet and DenseNet169 were fine-tuned by updating the weights of all layers in addition to the newly added layers before fusing them into one ensemble classifier to improve the classification ability of the model and be more specific to our task. This study was performed on a public dataset consisting of 282 otoscopic images. All classes were considered except the Tympanostomy Tubes class for having only two samples. Consequently, the proposed model demonstrated promising results of 99.54 % accuracy and an AUC of 1. Grad-CAM++ saliency maps were employed to highlight the affected area and pertinent features of the otoscopic image. The proposed approach contributes to improving accuracy, decreasing the misdiagnosis rate, and developing an automatic ear disease classification tool.