Abstract
Intracranial haemorrhage (ICH) is a life-threatening neurological emergency requiring rapid identification on non-contrast computed tomography (CT). In South Africa, where radiologist scarcity limits timely interpretation, deep learning offers automated detection to support clinical triage. However, deploying such models across multiple hospitals faces two challenges: domain shift from heterogeneous CT scanners degrades accuracy, and the Protection of Personal Information Act (POPIA) prohibits centralised data pooling. We present a framework addressing both challenges through unsupervised domain adaptation and federated learning (FL). Using a DenseNet-121 classifier, we evaluated five adaptation methods and three FL training strategies on public proxy datasets. Deep CORAL was the only adaptation method that did not degrade cross-site performance, showing a directionally positive effect (+1.0 pp, though not statistically significant at n = 322) without requiring target labels. For multi-site training, Federated Averaging (FedAvg) matched centralised performance under realistic non-IID conditions. Adding differential privacy via DP-SGD provided formal guarantees (ε from 4.59 to 0.13) at a substantial accuracy cost that our results attribute largely to the required architectural change rather than the privacy noise. These results suggest a feasible pathway for privacy-preserving ICH detection in resource-constrained settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_013.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=ZPe9Y37F4c
BibTex
@InProceedings{KruJai_Towards_MICCAISAT2026,
author = { Kruger, Jaime Leigh},
title = { { Towards Privacy-Preserving Multi-Site ICH Detection for South African Telestroke Networks } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}