Abstract
Brain arteriovenous malformations (bAVMs) are rare and morphologically heterogeneous lesions for which accurate nidus delineation is important for treatment planning and longitudinal assessment. However, the small number of cases available at individual hospitals and restrictions on cross-institutional data sharing limit the development of robust segmentation models, motivating the use of federated learning for bAVM segmentation. We conducted a retrospective real-world multicenter feasibility study of 200 patients from four hospitals across two continents using paired T2-weighted MRI and time-of-flight MRA. Patient images and annotations remained at their institutions of origin. Using the default nnU-Net v2 configuration across sites, we compared local training, cross-site testing without adaptation, and federated training with Federated Averaging (FedAvg). Local models showed limited and asymmetric generalization across hospitals. Federated training increased the macro-averaged Dice coefficient from 0.443 to 0.553 and improved performance at all participating sites, with the largest gains at sites with weaker local models. Training was performed using each hospital’s local computational infrastructure without sharing patient images or annotations. These findings provide real-world evidence that federated learning is feasible for multicenter bAVM segmentation and can improve segmentation performance across heterogeneous clinical sites, although inter-site variability remains a challenge.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DeCaF_017.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=hOEexIFQzJ
BibTex
@InProceedings{DorZol_Federated_MICCAISAT2026,
author = { Dorjsembe, Zolnamar AND Murad, Nazish AND Lin, Ching-Che AND Alsalama, Abdulrhman AND Xiao, Furen AND Chen, Kuo-Wei AND Lin, Chi-Yin AND Yang, Huai-Che AND Liu, Ping-Chuan AND Su, I-Chang AND Li, Baoxin AND Slavin, Konstantin V. AND Alaraj, Ali AND Pao, Hsing-Kuo Kenneth},
title = { { Federated Learning for Brain Arteriovenous Malformation Segmentation: A Feasibility Study Across Four Hospitals and Two Continents } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17252},
month = {pending},
page = {pending}
}