Abstract
A clinic in a resource-constrained setting often cannot retrain a
medical-imaging model; it adopts a frozen pretrained or foundation model and
must decide whether that model is safe on its \emph{local} patient population.
We frame this post-transfer decision as local-audit certification: from a small
labelled audit and a bounded clinical loss, the site obtains a one-sided,
finite-sample, distribution-free upper certificate on the deployment risk of the
fixed predictor. The certificate can be compared with a local acceptable-risk
threshold to decide whether to deploy, collect more labels, or reject the
transfer. We instantiate the protocol with an empirical-Bernstein certificate,
give a variance-adaptive audit-size statement, add a finite-sample Kullback–Leibler distributionally robust optimization
(KL-DRO) margin over one-dimensional loss laws for audit–deployment mismatch, and give a
two-point lower bound showing that the leading variance-dependent slack is
unavoidable up to constants. On real diabetic-retinopathy transfer
(EyePACS$\leftrightarrow$APTOS), the certificates are valid, sharpen with audit
size, and quantify the robustness budget needed to cover a genuine cross-cohort
shift.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_054.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=2bAV6Cwdxu
BibTex
@InProceedings{BisSit_Local_MICCAISAT2026,
author = { Bissu, Sita AND Yadav, Navnit Kumar AND Mehra, Aparna AND Kumar, Sandeep},
title = { { Local Audit Certification of Transferred Medical Imaging Models in Resource-Constrained Settings } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}
