Automated Diagnosis of Ear Disorders from Otoscopic Images using Hybrid Deep Learning and Multi-Stage Segmentation

Abstract

Effective diagnosis of ear disorders with the help of otoscopic images can enhance the health care of ENT disorders, especially in resourcelimited areas. In this study, deep learning algorithms are applied to five common ear conditions, namely Acute Otitis Media, Chronic Otitis Media, Cerumen Impaction, Myringosclerosis, and Normal, with a particular focus on their multiclass classification. A variety of deep learning architectures, such as MobileNetV2, VGG16, EfficientNetV2B0, NAS Net Mobile, Xception and hybrid models, were examined. Of these, Xception is the most successful model with a validation accuracy of 99.83% and high precision, recall and F1-scores for all classes. EfficientNetV2B0 showed its performance as well, suggesting that state-of-the-art CNN architectures can learn discriminative features from otoscopic images, as well. The findings indicate that the application of deep learning as a computer-aided diagnostic tool for ENT diagnosis is promising. However, the framework can be validated on a broader range of clinical data to assure its generalizability and utility to clinical practice.