Abstract
Segmentation models deployed for brain-tumor MRI fail silently under domain shift: a confidently missed tumor yields an empty, low-entropy mask that uncertainty ranks as safe. Among tumor-relevant slices, a subset defined by ground truth, slice-mean entropy anti-ranks major failures in every cohort×segmenter cell we test, but not over the unfiltered stream: a property of the evaluation population, not a deployment-time observable. We instead build a small on-device QC agent that perceives a compact slice signature, routes high-risk slices for review, and recalibrates online from case-by-case review, leaving the segmenter frozen. Its durable advantage is needing no upfront labelled target set, not accuracy beyond recalibration: it approaches a full target-domain refit from streamed labels alone; we do not claim fewer labels at a matched budget. Reviewer feedback, the clinician’s per-slice accept/flag decision, is simulated from a ground-truth Dice criterion. Online adaptation substantially improves failure review on an external, out-of-institution African cohort, but at a fixed review budget failure recall barely moves: the discrimination gain is large, the workload gain small. Leave-one-cohort-out across three BraTS cohorts resolves the gain pooled and on that cohort alone. Closing the loop – learning only from slices the agent routes – retains the AUPRC gain at a fraction of the labels, though pooled AUROC does not resolve. The agent runs on a Jetson Orin Nano beside a co-resident 3D segmenter. It is a triage aid, not a patient-level guarantee: at small budgets most failing cases retain an un-reviewed failure slice. Code and calibrated agent will be released.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_023.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MedAgent_023_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=9qSJwz7MHH
BibTex
@InProceedings{LiJac_ALabelEfficient_MICCAISAT2026,
author = { Li, Jacky},
title = { { A Label-Efficient On-Device QC Agent for MRI Segmentation under Domain Shift } },
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}
}