Regularized Weight Aggregation in Networked Federated Learning for Glioblastoma Segmentation

TL;DR

This paper presents RegSimAgg, a scalable, cost-efficient federated learning method for glioblastoma segmentation that improves weight aggregation and collaborator selection, achieving competitive results in a federated challenge.

Contribution

The paper introduces a novel regularized similarity weight aggregation method and a collaborator selection technique for federated learning in medical image segmentation.

Findings

Achieved 3rd place in FeTS2022 challenge for weight aggregation.

Scalable approach suitable for heterogeneous non-IID data.

Open-sourced implementation available for reproducibility.

Abstract

In federated learning (FL), the global model at the server requires an efficient mechanism for weight aggregation and a systematic strategy for collaboration selection to manage and optimize communication payload. We introduce a practical and cost-efficient method for regularized weight aggregation and propose a laborsaving technique to select collaborators per round. We illustrate the performance of our method, regularized similarity weight aggregation (RegSimAgg), on the Federated Tumor Segmentation (FeTS) 2022 challenge's federated training (weight aggregation) problem. Our scalable approach is principled, frugal, and suitable for heterogeneous non-IID collaborators. Using FeTS2021 evaluation criterion, our proposed algorithm RegSimAgg stands at 3rd position in the final rankings of FeTS2022 challenge in the weight aggregation task. Our solution is open sourced at: \url{https://github.com/dskhanirfan/FeTS2022}