Abstract
We describe our submission to the CMR-Multi 2026 chal- lenge, which requires automatic segmentation and functional/quantitative analysis of cardiac MRI across two tasks: (1) cine multi-view segmen- tation (SAX, 2CH, 4CH) with derived LVEF and wall-motion scoring, and (2) LGE scar segmentation and burden quantification across four views (SAX, 2CH, 4CH, RAS). Our pipeline is built entirely on nnU- Net v2 [1], using 2D per-slice models for 2CH/4CH/RAS and a 3D per- cardiac-phase model for SAX cine, trained on the officially released chal- lenge data. For LGE-SAX, where the released training cohort is small (∼40 cases) and myocardial scar is rare and small, five-fold ensembling and validation-tuned test-time augmentation (TTA) decisions stabilize the LGE scar-mass estimate, the most volatile downstream metric. On the official Validation-phase leaderboard, our submission scores overall 0.7153 (Task1 0.6870, Task2 0.7436), driven primarily by a mass-RAE improvement from 1.216 to 0.827 attributable to fold-ensembling sup- pressing a single outlier false-positive scar prediction. We analyze this mechanism in detail and report qualitative best/worst-case results for both tasks on the public validation split.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMRSeg_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=pVmLHzhhXB
BibTex
@InProceedings{AgrHar_nnUNet_MICCAISAT2026,
author = { Agrawal, Harshit},
title = { { nnU-Net for Cardiac MRI Segmentation and Scar Quantification (Cine + LGE): CMR-Multi 2026 } },
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}
}