Abstract
Self-supervised disentanglement of vessel anatomy from appearance is commonly validated by an encoder-projection cosine between source and swap. This validation is circular: it measures invariance of the very encoder used to generate the swap, not preservation of anatomy. A totally vessel-destroyed image (Gaussian blur, σ=8) still reads as 0.92 anatomy retention while a non-circular vessel-topology metric clDice via an independently-trained FIVES segmenter collapses to 0.003, below an unrelated-pair oor of 0.11; ablations move the cosine by +0.08 and −0.15 while clDice stays at. We propose segmenter-based clDice, read against an unrelated-pair oor, as the evaluation protocol for vessel-bearing medical image synthesis. Applying it to a latent-diusion testbed, a vessel-mask ControlNet, and a MUNIT-style GAN shows that conclusions drawn under it depend on controls that are easy to omit: a clDice margin inverts once resolutions are matched, since resolution dominates the metric, and the mask baseline exhibits a second circularity mode it is conditioned on the evaluating segmenter’s own output. Non-circularity is therefore necessary but not su cient: a topology metric must be paired with matched resolution and compute, an audit of conditioningevaluator independence, and complementary evidence beyond a single scalar.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DGM4MICCAI_014.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/profile?id=~Syed_Abdullah_Basit1
BibTex
@InProceedings{BasSye_Disentangled_MICCAISAT2026,
author = { Basit, Syed Abdullah AND Alam, Tanvir},
title = { { Disentangled Retinal Fundus Synthesis with Non-Circular Vessel-Topology Evaluation } },
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}
}