Abstract
Deep Learning models for medical image analysis are only as trustworthy as the data used to train them, motivating automated assessment of images’ compliance with acquisition protocol. Existing compliance-specific methods score each image against a reference distribution estimated over centrally pooled features, an assumption incompatible with federated deployments, in which images must remain within their institution. We introduce DeMICAF, a decentralised framework that distributes a frozen compliance-specific encoder to each client, to rank its images locally by their Mahalanobis distance to a global reference. That reference is aggregated from client-level feature statistics, so that only summary statistics ever leave a client. Because neither encoder training nor client onboarding requires compliance labels, a new institution can participate by receiving the frozen encoder and contributing local feature statistics, avoiding recurrent manual curation costs. To evaluate compliance directly, rather than through a downstream-task proxy, we release compliance annotations for 4,000 chest radiographs drawn from four public datasets, each labelled as compliant or non-compliant together with its causes of non-compliance. DeMICAF outperforms no-reference image quality assessment and out-of-distribution detection baselines, achieving an improvement of at least 22.3 and 18.1 AP points respectively, while remaining comparable to its centralised counterpart. A curated reference requires at least 50 annotated compliant images per client before it surpasses ours. The compliance annotations and associated code are available in our repository.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DeCaF_020.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=AauYZ8PX53
BibTex
@InProceedings{RomHen_DeMICAF_MICCAISAT2026,
author = { Romão, Henrique AND Carvalho Costa, Matilde AND Taiello, Riccardo AND Reis Santos, Diogo AND Sousa, Pedro AND Pedrosa, João AND Serio, Luigi},
title = { { DeMICAF: A Decentralised Medical Imaging Compliance Assessment Framework } },
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}
}