Development and Validation of a CNN-Based Diagnostic Pipeline for the Diagnosis of Otitis Media

Abstract

Background/Objectives: Accurate diagnosis of otitis media (OM) using otoscopic images is often challenging, particularly for non-specialists. Artificial intelligence (AI), especially deep learning-based methods, has shown promising results in supporting the classification of tympanic membrane conditions. This study aimed to develop and validate a multi-step CNN-based AI diagnostic pipeline for the automated classification of tympanic membrane images into four OM categories: normal, acute otitis media (AOM), otitis media with effusion (OME), and chronic otitis media (COM). Methods: A total of 2964 otoscopic images were retrospectively collected and annotated by expert otologists. The proposed pipeline consisted of four sequential stages: image quality assessment, tympanic membrane segmentation, side (left/right) classification, and final disease classification. CNN-based deep learning models including MambaOut, CaraNet, EfficientNet, and ConvNeXt were employed in each stage. Results: The image quality classifier achieved an accuracy of 98.8%, while the laterality classifier reached 99.1%. For disease classification, the ConvNeXt model demonstrated an overall accuracy of 88.7%, with disease-specific F1-scores of 0.78 for AOM, 0.87 for OME, and 0.92 for COM. The system performed reliably across all stages, indicating strong potential for clinical application. Conclusions: The proposed AI pipeline enables automated and accurate classification of tympanic membrane images into common OM subtypes. Its integration into digital otoscopes could support more consistent diagnosis in primary care and underserved settings, while also providing educational support for trainees and general practitioners.