A One-Stage CNN Object Detection Framework for Automated Classification of Otoscopic Ear Diseases

Accurate diagnosis of ear diseases is essential for preventing complications such as hearing loss and chronic infection, yet traditional otoscopic examination is subjective and prone to variability, particularly in resource-limited settings. This study presents an automated diagnostic system using a one-stage convolutional object detection model trained on an augmented Otoscopy Image Dataset covering five categories: Acute Otitis Media, Chronic Otitis Media, Cerumen Impaction, Myringosclerosis, and Normal. The model achieved strong performance (mAP@0.5 = 0.95, precision = 0.95, recall = 0.93) with near-perfect precision–recall characteristics and stable confidence-based metrics. Qualitative evaluation showed highly reliable detection of visually distinct conditions such as Acute Otitis Media and Cerumen Impaction, while subtle pathologies like Myringosclerosis remained challenging. Deployed via a web-based interface on Hugging Face Spaces, the system enabled real-time classification of both internal and external otoscopic images. These results demonstrate the feasibility of one-stage object detection for automated otoscopic disease diagnosis and highlight its potential to improve diagnostic accuracy, consistency, and accessibility in clinical and remote healthcare environments.