Otoscopy Image Classification Using Embedded AI

To diagnose different ear diseases and disorders, otoscopy is an essential diagnostic method that looks at the eardrum and external ear canal. This project focuses on preprocessing, augmenting, and implementing Convolutional Neural Network (CNN) architectures for otoscopy image datasets, aiming to classify images such as normal, Acute Otitis Media (AOM), Tympanosclerosis, Acute Otitis Externa (AOE), Foreign Objects in the ear, and other conditions. The primary objective is to identify the architecture that demonstrates superior performance in terms of accuracy across different classes of ear conditions, achieved through fine-tuning and optimizing different layers. Among the various CNN architectures explored, the MobileNetV2 model exhibited notably high accuracy compared to others. Therefore, it was selected for deployment on a Raspberry Pi for real-world testing. A systematic approach to fine-tuning, focusing on optimizing key hyperparameters and architectural components and minor modifications to the base model architecture, the top layers of the base model are unfrozen to allow them to be fine-tuned on the disease identification task. Initially, the pre-trained MobileNetV2 model showed an accuracy of 66%. However, through fine-tuning and modification, the model's accuracy significantly improved to 97%, indicating the effectiveness of the proposed approach in enhancing classification performance. This study contributes to the advancement of automated otoscopy diagnosis by leveraging deep learning techniques, particularly CNN architectures. The successful deployment of the optimized MobileNetV2 model on a low-resource platform like Raspberry Pi underscores its potential for practical clinical applications, facilitating timely and accurate diagnosis of various ear conditions.