AI-Powered Otoscopic Image Classification for Ear Disease Detection

Abstract

Otoscopy is a method that allows medical personnel to see and assess the external ear canal and eardrum with an otoscope, a handheld instrument with a light source and a lens or a camera. Evaluating the ear through otoscopy is important for indicating the health of the ear and recognizing problems, such as infection, inflammation, wax impaction, foreign bodies, tumors, structural changes, or trauma, particularly if a patient presents with any ear-related symptoms (e.g., pain, discharge, hearing loss). Otoscopy is often complemented by other diagnostic tests, including audiometry or tympanometry, to assess ear function more comprehensively. Even though determining otoscopic images has clinical importance, accurately identifying (or diagnosing) ear conditions from otoscopic images is difficult because of image quality variability and the subtleness of some conditions. In this project present a new deep learning based system for automating the classification of otosocopy images. By collecting a diverse data set of images of normal ears as well images with various conditions such as infections, tumors, and structural defects. By using the Convolutional Neural Networks (CNNs), specifically used the DenseNet and used a pre-trained DenseNet to extract features from images. The model then used categorical cross-entropy loss to optimize the classification. So we improved the robustness of our data set through preprocessing and data augmentation. By assessed the performance using metrics such as accuracy, precision, recall, F1-score and AUC-ROC using hyperparameter tuning to verify results. The results suggest that our approach can reliably classify otoscopic images; and may be a helpful adjunct with clinical diagnosis of ear related conditions.