Abstract
Few-shot cardiac MRI segmentation is often limited by support-set underutilization and reliance on unlabeled data, degrad- ing performance on unseen domains and compromising the functional measurements (e.g., ejection fraction, chamber volumes) central to cardiac diagnosis. We propose ViPS-FSL, which reformulates few-shot segmentation as support-set densification: from a sparse, pathology-diverse support set, it builds a cross-pathology pairwise support graph with phase-aware edge typing to learn local image-mask co-translation models, then samples a graph-induced Perceptual Feature Space via volume-calibrated latent interpolation to decode fully labeled virtual patients. A segmentation model is then trained on the labeled support set augmented with these virtual patients, using no unlabeled data. We evaluate in two settings: in-distribution on a held-out same-domain ACDC test split, and out-of- distribution on the more challenging M&Ms cohort, unseen and spanning multiple vendors, centers, and pathologies. On held-out ACDC (N=50), ViPS-FSL ranks first at both 7 and 11 support subjects (88.5/89.0 aver- age Dice), outperforming the strongest baseline by +2.6/+1.4 Dice with no unlabeled data. Under single-source transfer to the unseen M&Ms cohort (N=320), it outperforms five state-of-the-art baselines—each using 63/59 unlabeled subjects—reaching 87.4%/3.59 mm and 88.1%/3.29 mm (Dice/HD95), with per-chamber gains up to 5.5 Dice and 39% lower HD95 and generalizing across three unseen vendors despite single-vendor supervision. Thus, densifying labeled support—rather than relying on unlabeled data—yields strong, domain-robust few-shot segmentation. Code: https://github.com/Elbaz-PIC-CVI-Lab/ViPS-FSL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/STACOM2026_036.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/STACOM_036_supp.zip
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Mohammed_S._M._Elbaz1
BibTex
@InProceedings{ElbMoh_ViPSFSL_MICCAISAT2026,
author = { Elbayumi, Mohamed AND Elbaz, Mohammed S. M.},
title = { { ViPS-FSL: Virtual-Patient Densification via Latent Space Co-Translation for Robust Few-Shot Cardiac MRI Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17267},
month = {pending},
page = {pending}
}