Abstract
The MVAA challenge requires mitral valve anatomy segmentation from three highly heterogeneous modalities: cardiac CT volumes, 3D transesophageal echocardiography (TEE) volumes, and intraoperative RGB video frames. These settings differ in dimensionality, image statistics, and annotation availability, making a single shared recipe unattractive for a practical challenge solution. We present a task-specific challenge solution with three separate segmentation pipelines, each tailored to one modality. For Task1 CT, we use a 3D nnU-Net v2 PlainConv 30 GB configuration trained with online Mean Teacher, confidence-filtered pseudo-labels, 3D ClassMix, and foreground largest connected component post-processing. For Task2 TEE, we use a supervised probability-level ensemble of default 3d fullres and residual-encoder nnU-Net v2 models. For Task3 video, we use ImageNet-pretrained U-Net++ models with EfficientNet-B6/B5 encoders, weak-to-strong semi-supervision, a hard-sample B6 branch, weighted checkpoint ensembling, and picture-in-picture cleanup. On the official validation platform, the final pipelines achieve DSC values of 0.8619, 0.8537, and 0.8147 for Task1, Task2, and Task3, respectively.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MVAA_021.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=0IBs0OseUR
BibTex
@InProceedings{XuHui_TaskSpecific_MICCAISAT2026,
author = { Xu, Hui AND Niu, Yichen AND Yin, Hanlong AND Guo, Fangjin AND Zhao, Yue},
title = { { Task-Specific Segmentation Pipelines for the MVAA Challenge } },
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}
}