Abstract

Mitral valve repair relies on accurate segmentation across preoperative CT, intraoperative three-dimensional transesophageal echocardiography, and surgical video, yet each modality provides only limited labeled data. Existing small-sample pipelines can introduce optimistic calibration bias by tuning thresholds, postprocessing, or fusion with in-sample predictions or non-independent frame splits, which weakens transfer to held-out cases. We present a leakage-aware, calibration-driven framework that couples modality-specific models with held-out prediction calibration and source-geometry-preserving inference for the three MVAA 2026 segmentation tasks. For CT, we use one genuine out-of-fold prediction per labeled case to calibrate the supervised ensemble and to select reliable unlabeled volumes for pseudo supervision. For TEE, a five-fold nnU-Net ensemble segments the anterior and posterior mitral leaflets. We preserve ambiguous pseudo-label boundaries with an ignore label rather than forcing them into either foreground or background, reducing noisy contour supervision during refinement. For surgical video, we use video-disjoint validation and calibrated probability-space fusion of complementary query-based and region-based models. On organizer-provided public evaluation, the CT, TEE, and video systems achieve DSC/HD/ASD of 0.8628/4.509/0.272, 0.7665/20.734/1.560, and 0.8108/73.790/13.051, respectively.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MVAA_087.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=5krSkSHLMg

BibTex

@InProceedings{CheSid_ModalitySpecific_MICCAISAT2026,
        author = { Chen, Sida AND Cao, Chun AND Fang, Xin AND Tian, Shengli AND Tong, Lyuyang AND Du, Bo},
        title = { { Modality-Specific Mitral Valve Segmentation Across CT, 3D TEE, and Surgical Video } },
        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}
}


back to top