Abstract
A useful cardiac imaging world model must do more than delineate anatomy: it must recover a compact, clinically meaningful state from heterogeneous acquisitions and preserve that state under scanner and protocol shifts. We present CardioState, a task-aligned system for the CMR-MULTI setting that targets cine anatomy, left-ventricular ejection fraction (LVEF), late-gadolinium-enhancement (LGE) anatomy, and myocardial scar burden. Instead of a monolithic 3D model, it uses two view-conditioned 2.5D MedNeXt experts. Cine inference combines local spatial and temporal context, whereas LGE inference combines adjacent slices and explicit view codes. A shared anatomical label space, view-valid class masking, exponential moving-average weights, and deep supervision support learning across short- and long-axis views. For LGE, a volume-aware objective aligns voxel segmentation with scar burden, while conservative MRI domain randomization is used to improve acquisition robustness. A physical-volume connected-component rule is selected using strict patient-level out-of-fold (OOF) predictions and two empty-mask scoring conventions. On public validation, CardioState obtains a cine Dice of 0.9070, LVEF Pearson correlation of 0.9245, LGE Dice of 0.7669, scar-mass correlation of 0.8165, and Task 2 score of 0.7663. The local two-task validation estimate is 0.7314. Strict OOF analysis shows that the physical prior improves the robust Task 2 score in all four folds without erasing low-burden positive cases. The deterministic offline Docker pipeline processes 90 cine and 65 LGE test-equivalent volumes in 161 seconds on a 24-GB GPU.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMRSeg_115.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=DrXHf3Gqdi
BibTex
@InProceedings{BaiMen_CardioState_MICCAISAT2026,
author = { Bai, Meng AND Deng, Bo AND Peng, Zilun AND Chen, Yu AND Zhou, Yongle},
title = { { CardioState: Task-Aligned 2.5D MedNeXt for Multi-Sequence Cardiac MRI Segmentation and Quantification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17268},
month = {pending},
page = {pending}
}