Federated learning in medical image analysis

• Autor(a) principal: Fabiana Rodrigues da Silva
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/151924
• topic: Outras ciências da engenharia e tecnologias; Other engineering and technologies

Medical image analysis is crucial for the efficient diagnosis of many diseases. Hospitals typically maintain vast repositories of images, which can be leveraged for various purposes, including research. However, access to such image collections is largely restricted to safeguard the privacy of the individuals whose images are being stored, as data protection concerns come into play. Recently, the development of Automated Medical Image Analysis has gained significant attention, with Deep Learning being one solution that has achieved remarkable results in medical image analysis. One promising approach for medical image analysis is Federated Learning (FL), which enables using a set of physically distributed data repositories (the nodes) for analysis, satisfying the restriction that data does not leave the repository. Under these conditions, FL can build high-quality accurate models using a lot of available data wherever it is. This approach can help researchers and clinicians to diagnose diseases and support medical decisions more efficiently and robustly. Detection of pneumonia on chest X-radiography (X-ray) images is proposed in a FL environment using Flower as framework, and FedAvg as strategy. This supervised learning approach uses pre-trained Convolutional Neural Network (CNN) models to leverage transfer learning: VGG-16, Resnet-18 and Resnet-50, and also data augmentation techniques are applied to fine-tune the models. Simulated a FL environment having 8 hospitals sharing their own images, Resnet-18 shows the best result with 98.46\% of accuracy, followed by Resnet-50 with 78.46\% of accuracy, then VGG-16 with 78.46\% of accuracy as well, all evaluated on the server-side after 5 rounds of training. The experiments suggested in the research work exhibited significant computational expense owing to an uneven dataset, a prevalent constraint encountered in the study as well, presented in the current state-of-the-art, which utilizes the identical dataset as this inquiry. Hence, it discusses applications, contributions, limitations, and challenges, and is suitable for those who want to understand how FL can contribute to the medical imaging domain. Furthermore, this solution is applicable as a baseline to solve other binary classification problems using medical images, such as Magnetic resonance imaging (MRI), Computed tomography (CT), X-radiography (X-ray), and histology images.