Abstract
Medical imaging datasets are often curated using global quality-control rules, although case eligibility depends on the predefined requirements of the intended task. AgentQC is a static artifact-grounded agentic framework for task-aware medical dataset curation that separates deterministic evidence extraction, calibrated comparison, controlled reasoning, medical critique, review routing, deterministic final policy, and post-hoc evaluation into typed artifacts with explicit provenance. Reasoning and critique are explanatory and nonbinding; final curation actions remain controlled by deterministic evidence, routing, and policy logic. We instantiate AgentQC for 1,000 pancreas CT cases with visible-pancreas, pancreas-lesion, and lesion-subregion curation requirements. Internal evaluation across task profiles identified 350, 429, and 226 changes in final actions between profile pairs. Nevertheless, lesion-present, confirmed-absent, and unresolved-evidence categories remained consistent across configurations. All cases contained complete reasoning and critique, and deterministic and calibrated modes produced identical final actions. These findings support artifact-backed compliance behavior under configured task requirements, not external medical correctness, expert agreement, or downstream model utility.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_028.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=kXMPfs8X8q
BibTex
@InProceedings{AndGus_AgentQC_MICCAISAT2026,
author = { Andrade-Miranda, Gustavo AND Cao, Yiheng AND Collet, Tiphain AND Vega, Pedro J. Soto},
title = { { AgentQC: Policy-Constrained Agentic Assessment of Task-Aware Reliability in Medical Imaging Datasets } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}