Deep learning for otitis media classification using otoscopic image

Abstract

Otitis media (OM) is a globally prevalent health issue that poses significant threats to hearing and overall health. However, traditional diagnostic methods rely heavily on subjective interpretation of otoscopic images, which often leads to inconsistent outcomes and high misdiagnosis rates. This study evaluated 5 deep learning models (ResNet-18, GoogLeNet, AlexNet, MobileNet-V3, and VGGNet-19) for classifying 819 otoscopic images into 4 categories: normal, acute otitis media, otitis media with effusion, and chronic suppurative otitis media. Specifically, the dataset was split into training (60%), validation (20%), and test (20%) sets. Subsequently, model performance was assessed using accuracy, sensitivity, specificity, precision, F1 score, and receiver operating characteristic–area under the curve analysis. The results showed that VGGNet-19 demonstrated superior performance, achieving the highest scores in accuracy (94.51%), sensitivity (94.18%), specificity (98.10%), and precision (93.86%), with particularly outstanding performance in chronic suppurative otitis media identification. While other models delivered competitive results, their metrics were consistently lower in comparison. Therefore, the study confirms VGGNet-19’s potential as an accurate, automated diagnostic tool for OM classification. Future research should address dataset imbalances and validate findings across diverse clinical settings to enhance generalizability.