Abstract
Deep learning models for mammographic breast cancer detection are developed and validated predominantly in cohorts from the continental US and Europe, while Native Hawaiian and Pacific Islander women, who experience later-stage diagnoses and poorer outcomes, remain underrepresented. Federated learning (FL) enables collaborative training without sharing patient data, but standard Federated Averaging produces a single global model that may underperform at individual sites differing in population, case mix, and image acquisition. We adapted GMIC, a multiple-instance classifier for mammographic breast cancer detection, for federated fine-tuning with NVIDIA FLARE and evaluated shared-global (FedAvg, FedProx) and personalized (FedBN, Ditto) strategies, plus a module-wise Ditto variant with per-module regularization. We constructed a three-site federation comprising a diverse Pacific Island cohort and cohorts based in the mainland US and Australia from the public RSNA Screening Mammography dataset. FedAvg, FedProx, and FedBN improved the largest site but left smaller sites at or below pretrained performance; FedBN, despite personalizing batch-normalization statistics, was least equitable. Ditto improved the worst-performing site (test AUC 0.87 vs. 0.83 for the others and 0.84 for local-only training) while reducing cross-site AUC spread (0.019 vs. 0.030 for FedAvg) and matching the pooled performance of centralized training, without sharing data. Module-wise Ditto provided no measurable advantage over standard Ditto. To our knowledge, this is the first equity-motivated application of personalized FL to mammographic cancer detection in a Pacific Islander population, suggesting that personalized FL can improve robustness to site heterogeneity by allowing local adaptation while retaining the benefits of collaborative learning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DeCaF_006.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/DeCaF_006_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=syKezaooEp
BibTex
@InProceedings{SolLyd_Personalized_MICCAISAT2026,
author = { Sollis, Lydia J. AND Young, Patricia M. AND Bunnell, Arianna AND Quon, Brandon AND Hernandez, Brenda Y. AND Wolfgruber, Thomas K. AND Shepherd, John A.},
title = { { Personalized Federated Learning for Equitable Breast Cancer Detection in Underrepresented Pacific Islander Populations } },
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}
}