Tumor Detection in Distributed Data Silos

Abstract

This thesis focuses on a computer vision problem of tumor detection from magnetic resonance images in federated learning settings and investigates the performance of using different subsets of collaborators in each federated learning round. The benchmark dataset used is from Brats (Brain Tumor Segmentation) initiative. The experiments using two different aggregation algorithms show that using a smaller subset of collaborators, such as 10\%, can achieve comparable performance to using a larger subset, such as 30\%. This finding suggests that it is practical and pragmatic to use a subset of collaborators in each federated learning round in large-scale industrial applications. This approach reduces computational cost while still improving the global model in an iterative fashion. The optimal subset of available collaborators may vary depending on the specific application specially collaborators data. Therefore, it's essential to experiment with different available collaborator subset sizes and evaluate their footprint on the overall performance of the federated learning system. This work contributes to the field of federated learning and provides insights into its practical implementation in real-world scenarios.