Deep Learning-Based Classification of Otoscopic Images for the Detection of Ear Pathologies

Develioğlu B. Ş.

5th International Ege Congress on Scientific Research, Muğla, Türkiye, 22 - 23 Temmuz 2026, ss.3-15, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Muğla
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.3-15
  • İstanbul Üniversitesi-Cerrahpaşa Adresli: Evet

Özet

This study aims to classify ear pathologies using otoscopic image data. Accordingly, the classification performances of the ResNet50, EfficientNetB0, and MobileNetV2 deep learning architectures pretrained on ImageNet are comparatively evaluated. The dataset used within the scope of the study consists of four otoscopic image classes: chronic otitis media, cerumen impaction, myringosclerosis, and normal ear. The risk of data leakage that may arise from a random splitting approach is eliminated through perceptual hash (pHash)-based image grouping and a splitting strategy that preserves group integrity. In the evaluation conducted on the test set (n=160), ResNet50 exhibits the highest overall performance with 92.5% accuracy and an MCC of 0.9008, whereas EfficientNetB0 demonstrates the highest interclass discriminative performance with a macro AUC of 0.9916. While the cerumen impaction class is classified without error by all models, a systematic confusion is observed between the myringosclerosis and normal ear classes. The regions of the otoscopic images that predominantly contribute to the classification decisions of the models are examined using Grad-CAM (Gradient-weighted Class Activation Mapping) analysis. Thus, visual explainability is introduced into the black-box structure of the models. In this context, the study aims to reveal the capacity of different transfer learning architectures to distinguish ear pathologies in otoscopic images through quantitative performance metrics and Grad-CAM-based visual explainability outputs. The findings indicate that deep learning-based image processing models can be used as a preliminary decision support tool in the otoscopic examination process.