Abstract
The integration of automated diagnostic support into clinical practice is often hindered by data scarcity and strict privacy constraints in local centers. While generative models offer a potential solution via data augmentation, standard architectures suffer from mode collapsewhenconditionedonfew-shotlocalsamples.Therefore,weproposeFew-ShotFlowMatching(FSFM).Ourcorecontribution,Stochastic Barycentric Sampling, treats retrieved latent neighbors as vertices of a simplex, dynamically sampling conditioning vectors from the continuous convex hull to prevent mode collapse and preserve structural variance. We apply FSFM to Diabetic Retinopathy (DR) fundus image synthesis, adapting from a large unlabelled source to a constrained target clinic under three protocol-stratified evaluation cohorts. FSFM provides consistent improvements in 5-class DR grading over real-only baselinesintheextreme-scarcityregime(N ≤100),withthelargestabsolutegainsonthehomogeneous-protocoldilatedcohort(QWK0.10→ 0.20 at N=20, 0.29 → 0.37 at N=50). Stochastic Barycentric Sampling produces 26-39% higher synthetic diversity than deterministic centroid conditioning. Code is available at https://anonymous.4open.science/ r/retina_flow_matching-4860.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SASHIMI_011.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=tZ4jquAEEI
BibTex
@InProceedings{KulAle_FewShot_MICCAISAT2026,
author = { Kulbaka, Aleksandra AND Bonnici, John AND Glocker, Ben AND Kainz, Bernhard},
title = { { Few-Shot Flow Matching with Stochastic Barycentric Sampling for Image Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}