Distance-Aware Non-IID Federated Learning for Generalization and Personalization in Medical Imaging Segmentation

Julia Alekseenko, Alexandros Karargyris, Nicolas Padoy
Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, PMLR 250:33-47, 2024.

Abstract

Federated learning (FL) in healthcare suffers from non-identically distributed (non-IID) data, impacting model convergence and performance. While existing solutions for the non-IID problem often do not quantify the degree of non-IID nature between clients in the federation, assessing it can improve training experiences and outcomes, particularly in real-world scenarios with unfamiliar datasets. The paper presents a practical non-IID assessment methodology for a medical segmentation problem, highlighting its significance in medical FL. We propose a simple yet effective solution that utilizes distance measurements in the embedding space of medical images and statistical measurements calculated over their metadata. Our method, designed for medical imaging and integrated into federated averaging, improves model generalization by downgrading the contribution from the most distant client, treating it as an outlier. Additionally, it enhances model personalization by introducing distance-based clustering of clients. To the best of our knowledge, this method is the first to use distance-based techniques for providing a practical solution to the non-IID problem within the medical imaging FL domain. Furthermore, we validate our approach on three public FL imaging radiology datasets (FeTS, Prostate, and Fed-KITS2019) to demonstrate its effectiveness across various radiology imaging scenarios.

Cite this Paper

BibTeX
@InProceedings{pmlr-v250-alekseenko24a, title = {Distance-Aware Non-IID Federated Learning for Generalization and Personalization in Medical Imaging Segmentation}, author = {Alekseenko, Julia and Karargyris, Alexandros and Padoy, Nicolas}, booktitle = {Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning}, pages = {33--47}, year = {2024}, editor = {Burgos, Ninon and Petitjean, Caroline and Vakalopoulou, Maria and Christodoulidis, Stergios and Coupe, Pierrick and Delingette, Hervé and Lartizien, Carole and Mateus, Diana}, volume = {250}, series = {Proceedings of Machine Learning Research}, month = {03--05 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v250/main/assets/alekseenko24a/alekseenko24a.pdf}, url = {https://proceedings.mlr.press/v250/alekseenko24a.html}, abstract = {Federated learning (FL) in healthcare suffers from non-identically distributed (non-IID) data, impacting model convergence and performance. While existing solutions for the non-IID problem often do not quantify the degree of non-IID nature between clients in the federation, assessing it can improve training experiences and outcomes, particularly in real-world scenarios with unfamiliar datasets. The paper presents a practical non-IID assessment methodology for a medical segmentation problem, highlighting its significance in medical FL. We propose a simple yet effective solution that utilizes distance measurements in the embedding space of medical images and statistical measurements calculated over their metadata. Our method, designed for medical imaging and integrated into federated averaging, improves model generalization by downgrading the contribution from the most distant client, treating it as an outlier. Additionally, it enhances model personalization by introducing distance-based clustering of clients. To the best of our knowledge, this method is the first to use distance-based techniques for providing a practical solution to the non-IID problem within the medical imaging FL domain. Furthermore, we validate our approach on three public FL imaging radiology datasets (FeTS, Prostate, and Fed-KITS2019) to demonstrate its effectiveness across various radiology imaging scenarios.} }
Endnote
%0 Conference Paper %T Distance-Aware Non-IID Federated Learning for Generalization and Personalization in Medical Imaging Segmentation %A Julia Alekseenko %A Alexandros Karargyris %A Nicolas Padoy %B Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2024 %E Ninon Burgos %E Caroline Petitjean %E Maria Vakalopoulou %E Stergios Christodoulidis %E Pierrick Coupe %E Hervé Delingette %E Carole Lartizien %E Diana Mateus %F pmlr-v250-alekseenko24a %I PMLR %P 33--47 %U https://proceedings.mlr.press/v250/alekseenko24a.html %V 250 %X Federated learning (FL) in healthcare suffers from non-identically distributed (non-IID) data, impacting model convergence and performance. While existing solutions for the non-IID problem often do not quantify the degree of non-IID nature between clients in the federation, assessing it can improve training experiences and outcomes, particularly in real-world scenarios with unfamiliar datasets. The paper presents a practical non-IID assessment methodology for a medical segmentation problem, highlighting its significance in medical FL. We propose a simple yet effective solution that utilizes distance measurements in the embedding space of medical images and statistical measurements calculated over their metadata. Our method, designed for medical imaging and integrated into federated averaging, improves model generalization by downgrading the contribution from the most distant client, treating it as an outlier. Additionally, it enhances model personalization by introducing distance-based clustering of clients. To the best of our knowledge, this method is the first to use distance-based techniques for providing a practical solution to the non-IID problem within the medical imaging FL domain. Furthermore, we validate our approach on three public FL imaging radiology datasets (FeTS, Prostate, and Fed-KITS2019) to demonstrate its effectiveness across various radiology imaging scenarios.
APA
Alekseenko, J., Karargyris, A. & Padoy, N.. (2024). Distance-Aware Non-IID Federated Learning for Generalization and Personalization in Medical Imaging Segmentation. Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 250:33-47 Available from https://proceedings.mlr.press/v250/alekseenko24a.html.

Related Material