연합 학습을 통한 딥러닝 기반 대뇌 추출 알고리즘의 성능 향상을 위한 다중기관 학습

Abstract

Brain extraction in magnetic resonance imaging (MRI) is a crucial step in the neuroimage analysis pipelines because this step impacts the following analysis process. In recent years, convolutional neural networks (CNNs) have shown a remarkable performance on brain extraction. However, CNNs usually require large amounts of datasets which are challenging to collect and share with other institutions due to patient privacy issues. To overcome this problem, we propose a multi-institutional collaboration using federated learning (FL) which can utilize datasets at different institutions in a private-preserving manner. However, the performance of the model trained in the FL system can be degraded due to the heterogeneity of the image characteristics caused by differences in imaging devices and protocols across institutions. To deal with this problem, we propose an ensemble model which combines the local model and the global model in the FL system with a probability map ensemble approach to obtain enhanced performance. In our experiments, we evaluate our proposed model with other models obtained in a different strategy in the FL system and other ensemble approaches. We also compare our proposed model with four publicly available algorithms to verify the advantage of the FL model. Experimental results show that the proposed method can be a potential direction to build a robust and consistent deep learning model through multi-institution collaboration without privacy concerns.