Abstract
Medical-imaging data engineering is increasingly central to building scalable research registries, yet clinical archives often contain heterogeneous studies with incomplete or unreliable metadata. We study a data-engineering pipeline for orthopedic radiograph registry construction that must (i) select in-domain knee radiographs from mixed imaging archives, (ii) assign registry-ready view labels, and (iii) identify cases that should be routed to human review. Rather than proposing a new conformal-prediction algorithm, this work evaluates how standard calibration and conformal tools behave when used as deploy-time data-curation mechanisms under cross-dataset shift. Using multi-source musculoskeletal radiographs for training and held-out target datasets for deployment evaluation, we show that source-trained automation can fail in a way that is directly harmful for data curation: a knee/non-knee domain gate appears well calibrated in-domain (ECE=0.0024) but becomes severely miscalibrated under cross-dataset shift (ECE=0.2654), with knee sensitivity falling to 0.046 while specificity remains 0.991. To make registry ingestion more reliable with minimal target annotation, we evaluate a label-efficient target recalibration protocol. With 50 target examples per class, this protocol restores near-nominal class-conditional coverage at $\alpha=0.10$ while eliminating empty sets. On an independent hospital knee cohort, uncertainty-aware view labeling achieves macro-F1 $=0.937$ and APS coverage $=0.984$. These results position uncertainty-aware and label-efficient recalibration as a practical data-engineering component for portable medical-image registry construction.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DEMI_024.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=G7tKUeho1P
BibTex
@InProceedings{HanYus_LabelEfficient_MICCAISAT2026,
author = { Han, Yushui AND Gao, Fengyi AND Rezvani, Farnaz AND Littlefield, Nickolas AND Kann, Michael R. AND Shah, Palak AND Myers, Nicole AND Plakseychuk, Elizabeth AND Heidinger, Eva M. AND Ewalefo, Samuel AND Hogan, MaCalus AND Plate, Johannes F. AND Tafti, Ahmad P.},
title = { { Label-Efficient Data Engineering for Reliable Orthopedic Radiograph Registry Construction Under Distribution Shift } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}