Federated Learning for Lung Sound Analysis.

Despite the general success of employing artificial intelligence (AI) to help radiologists perform computer-aided patient diagnosis, creating good models with tiny datasets at different sites is still tough. Medical image analysis is crucial for the quick and precise detection of lung disease and helps clinicians treat patients effectively while averting more fatalities. This is why real-time medical data management is becoming essential in the healthcare industry, especially for systems that monitor patients from a distance. To overcome this challenge, we propose a different approach by utilizing a relatively new learning framework. Individual sites may jointly train a global model using this approach, referred to as federated learning. Without explicitly sharing datasets, federated learning combines training results from various sites to produce a global model. This makes sure that patient confidentiality is upheld across all sites. Additionally, the additional supervision gained from partner sites’ results enhances the global model’s overall detection capabilities. This study’s primary goal is to determine how the federated learning (FL) approach may offer a machine learning average model that is robust, accurate, and unbiased in detecting lung disorders. For this aim, we analyze 325 Lung Sound audio recordings collected from https://data.mendeley.com/ and, transform this audio signal into Melspectrograms. Once the labeling and preprocessing steps were carried out, a Convolutional Neural Network (CNN)(FederatedNet) model was used to classify the respiratory sounds into healthy and unhealthy. We achieved the result of almost 88% validation accuracy. Furthermore, this paper discusses the application of FL and its overview. Lastly, we discuss the main challenges to federated learning adoption and potential future benefits.