A Federated Learning Approach to Accurate Ear Disease Diagnosis with Data Privacy

Abstract

Ear diseases, particularly those affecting the tympanic membrane, pose a significant global health challenge, often leading to hearing loss. Traditional diagnostic methods struggle to balance accuracy with patient privacy concerns. This study introduces OtoFL, a federated learning framework, and Fenet5, a deep learning model, to revolutionize ear disease diagnosis. By leveraging diverse ear imaging datasets including “Ear Imagery Dataset” and “Eardrum Dataset”, our approach simulates real-world collaboration while ensuring patient privacy through differential privacy techniques. Fenet5, a five-block deep convolutional neural network, excels in feature extraction and classification, achieving a remarkable accuracy of 95.13%, precision of 0.96%, recall of 0.90%, and F1 score of 0.92% in diagnosing various ear diseases, even with imbalanced data in an FL environment. Notably, Fenet5 outperforms other state-of-the-art models like DenseNet201, MobileNetV2, and EfficientNetB0, demonstrating superior precision, recall, and F1 scores across different disease classes. Our federated learning approach, using FedProx and the proposed OtoFL, further enhances accuracy and privacy compared to FedAvg and FedSGD. OtoFL's scalability and adaptability are validated through experiments with varying client numbers and communication rounds, showcasing its potential to transform ear disease diagnosis globally.