Multi-institutional Travelling Model for Tumor Segmentation in MRI Datasets

Abstract

Administrative, ethical, and legal reasons are often preventing the central collection of data and subsequent development of machine learning models for computer-aided diagnosis tools using medical images. The main idea of distributed learning is to train machine learning models locally at each site rather than using centrally collected data, thereby avoiding sharing data between health care centers and model developers. Thus, distributed learning is an alternative that solves many legal and ethical issues and overcomes the need to directly share data. Most previous studies simulated data distribution or used datasets that are acquired in a controlled way, potentially misrepresenting real clinical cases. The 2021 Federated Tumor Segmentation (FeTS) challenge provides clinically acquired multi-institutional magnetic resonance imaging (MRI) scans from patients with brain cancer and aims to compare federated learning models. In this work, we propose a travelling model that visits each collaborator site up to five times with three distinct travelling orders (ascending, descending, and random) between collaborators as a solution to distributed learning. Our results demonstrate that performing more training cycles is effective independent of the order that the models are transferred among the collaborators. Moreover, we show that our model does not suffer from catastrophic forgetting and successfully achieves a similar performance (average Dice score 0.676) compared to standard machine learning implementations (Dice score 0.667) trained using the data from all collaborators hosted at a central location.