Abstract
This report describes our submission to the CMR-MULTI challenge on multi-view cardiac magnetic resonance image segmentation. The benchmark includes both functional CINE (steady-state free precession, SSFP) MRI and late gadolinium enhancement (LGE) MRI, with targets distributed across 2-chamber, 4-chamber, short-axis, and right-axis views. Our final system follows a pragmatic design: instead of forcing all subsets into one universal model, we train seven view-specific 2D networks under a shared recipe. The method combines a DINOv3-initialized transformer encoder, multi-level token aggregation, a lightweight upsampling decoder, nnU-Net-based planning and augmentation, and simple connected-component filtering at the end of the pipeline. Models are selected with the official validation protocol rather than overlap scores alone. On our local official validation setup, the final post-processed package achieves a Task1 score of 0.7014, a Task2 score of 0.6382, and a averaged score of 0.6698.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMRSeg_006.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=LSMbtP0tfs
BibTex
@InProceedings{ZhoYih_AMultiscale_MICCAISAT2026,
author = { Zhou, Yihao AND Myrgyyassov, Alisher},
title = { { A Multi-scale DINOv3-based Pipeline for Multi-view Cardiac MRI Segmentation } },
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}
}