A Deep Learning-based Method for Detection of Severity Stages of Otitis Media by Otoscopic Images

Otitis Media (OM) is a common and potentially serious middle-ear disorder that mostly affects youngsters. Timely and correct OM diagnosis is critical for successful intervention. Often, it is caused by mucosal disorders that affect tympanic membranes. Otitis Media causes effusion, perforation and retraction also in tympanic membrane. This study investigates the use of deep learning techniques on photographs of the tympanic membrane (TM) to detect OM precisely. The dataset was gathered and approved under supervision of a medical expert at AIIMS Raipur, is divided into four categories: acute otitis media (AOM-mild), acute severe otitis media (ASOM-moderate), chronic suppurative otitis media (CSOM-Severe), and normal tympanic membrane. This study compares the performance of the EfficientNet B3 and Inception V3 models in categorizing OM stages using otoscopic pictures, including categories for AOM, ASOM, CSOM, and normal tympanic membranes (TMs). Both models attained an overall accuracy of 78.18%, indicating their promise in OM classification. EfficientNet B3 demonstrated high precision for ASOM but low recall for CSOM (33%). Inception V3 had a higher recall (78% for CSOM) and specificity, indicating a better balance in identifying small differences. These findings demonstrate the potential of deep learning models in assisting OM diagnosis and lay the groundwork for future advancements to enhance clinical practice.