Towards Federated Learning in Identification of Medical Images: A Case Study
Artificial Intelligence (AI) methods need to learn from adequately large datasets to achieve clinical-grade accuracy and validation, which is vital in the healthcare field. However, sensitive medical data is usually fragmented, and not shared due to security and patient privacy policies. In this context, we aim at classifying abdominal and chest radiographs by applying Federated Learning (FL) with no sharing of the patient data. To perform the analyzes, we use a dataset comprising the abdominal and chest radiographs of patients, derived from Open-i. We implement and evaluate FL framework on distributed data across multiple clients. In the framework, we use multilayer perceptron as a deep learning model for the classification task. FL is a novel approach in which machine learning models are built with the collaboration of multiple clients controlled by a central server or service provider. FL model ensures data privacy and security by retaining the training data decentralized. We compare the performance of FL model against the centralized learning model. According to our results, FL model gives similar results to centralized models. Meanwhile, FL model provides security and privacy for patients by training individual models in distributed clients and sharing merely the model weights. We demonstrate that FL can address classifying private medical data by enabling multiple distributed clients to collaboratively train without the need to exchange or centralize patient records. Consequently, FL is a promising approach to sensitive medical data in terms of security, privacy, and bias.