Abstract
Abstract:
Accurate mitral valve segmentation across CT, 3D transesophageal echocardiography (TEE), and surgical video is challenging because these modalities differ substantially in imaging characteristics and temporal structure. Existing multimodal methods commonly seek shared representations or unified fusion, but the three modalities are unpaired and unregistered and only surgical video carries physical temporal dynamics, making direct cross-modal representation learning poorly defined. We propose StateMV, an anchor-conditioned anatomy world model that instead unifies how anatomical states are refined: it treats a task-specific segmentation posterior as the observed state and learns a bounded residual transition constrained by a modality-specific projection and a held-out trust criterion. A signed-distance CT state evolves in refinement pseudo-time through sparse surface actions and confidence-localized projection; causal latent dynamics predict the subsequent video state under an area-preserving constraint; and for 3D TEE, inconsistent held-out gains lead StateMV to select null dynamics and retain the multiplanar expert consensus. Under strict out-of-fold evaluation, StateMV improves CT Dice from 0.8906 to 0.8961 and ASD from 0.1284 to 0.1199, and video Dice from 0.7829 to 0.7878 and HD from 116.70 to 112.51. Public-validation Dice scores are 0.8603, 0.8575, and 0.8165 for CT, TEE, and video. These results show that heterogeneous modalities can be unified through anatomy-state transitions and evidence-gated refinement without a shared representation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MVAA_106.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MVAA_106_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=OLAJ73Z2Vw
BibTex
@InProceedings{GuoJia_StateMV_MICCAISAT2026,
author = { Guo, Jiarong AND Cheng, Yongzhen AND Hutchinson, Landon Li AND Du, Bodong AND Li, Xiaomeng},
title = { { StateMV: Trust-Gated Anatomy-State Transitions for Multimodal Mitral Valve Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17269},
month = {pending},
page = {pending}
}