Abstract
The Dice similarity coefficient (DSC) is the default quality gate for cardiac MRI segmentation, yet clinical decisions depend not on overlap but on derived measurements, such as ejection fraction (EF) and ventricular volumes. We ask whether a high Dice actually guarantees that these downstream measurements are clinically reliable. Using a single frozen 2D U-Net trained on ACDC and deployed without refitting on the held-out ACDC test set (n=50) and the external multi-vendor M&Ms cohort (n=320; Siemens/Philips/GE/Canon), we define a silent-failure rate: among cases with conventionally high overlap (DSC ≥ 0.90), the fraction whose derived measurement exceeds a literature-derived repeatability benchmark. Right-ventricular ejection fraction (RVEF) shows a silent-failure rate of 0.52 (95% CI 0.34-0.69) on ACDC and 0.41 (0.28-0.56) on M&Ms, despite mean RV Dice of 0.887 and 0.826, respectively. The effect is observed in a pooled external cohort spanning four scanner vendors and persists across convolutional and transformer segmentation architectures. By contrast, LVEF, LVEDV, and LVSV are protected in-distribution, meaning that their observed silent-failure rates are near zero (≤ 0.02), although LVESV is not; under shift, LV protection degrades noticeably. A mechanism analysis argues against catastrophic cancellation (end-diastolic and end-systolic errors are positively correlated) and instead identifies a systematic, same-sign volume bias to which the unsigned Dice is structurally blind, and which becomes more one-directional under shift. We conclude that a high Dice must not be used as a stand-alone safety gate for RV function, and that segmentation quality assurance should be reported against the clinical endpoint, not overlap alone.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/STACOM2026_010.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/profile?id=~Emmanouil_Gionanidis2
BibTex
@InProceedings{GioEmm_Silent_MICCAISAT2026,
author = { Gionanidis, Emmanouil AND Vrachnos, Panagiotis AND Miltiadous, Andreas AND Tzallas, Alexandros T.},
title = { { Silent Failures of Right-Ventricular Ejection Fraction at High Segmentation Overlap: Evidence from Multi-Vendor Cardiac MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17267},
month = {pending},
page = {pending}
}