Abstract
Accurate segmentation of retinal fluids, i.e., intraretinal fluid (IRF), subretinal fluid (SRF), and pigment epithelial detachment (PED), from optical coherence tomography (OCT) is essential for monitoring exudative retinal diseases and guiding anti-VEGF therapy. Conventional deep learning methods often yield topologically fragmented predictions that are clinically implausible. We benchmark UniSegDiff, a published staged conditional diffusion model, on multi-vendor retinal OCT fluid segmentation and integrate a connectivity prior loss inspired by LTSeg to enforce topological coherence during training. to enforce topological coherence during training. On the RETOUCH benchmark (112 volumes: 70 training, 42 test; three vendors), our model achieves a mean Dice of 0.780, outperforming CFENet (0.718), nnUNet (0.716), and base UniSegDiff without the prior (0.702). The connectivity prior yields the largest improvement for SRF (+0.126 DSC), the fluid type most prone to fragmentation, and reduces mean absolute volume difference (AVD) by 57% over CFENet. Zero-shot cross-dataset evaluation on the AROI dataset demonstrates moderate generalization (overall Generalized Dice 0.704), with a clear fluid-type-dependent transferability hierarchy. Our results show that combining diffusion-based iterative refinement with explicit topological priors produces segmentations that are both quantitatively accurate and structurally plausible.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/OMIA_035.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=HeTXIbrqVM
BibTex
@InProceedings{AwaWa’_TopologyAware_MICCAISAT2026,
author = { Awamleh, Wa’ad AND Bogunović, Hrvoje AND Fazekas, Botond},
title = { { Topology-Aware Staged Diffusion for Multi-Vendor Retinal OCT Fluid Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17270},
month = {pending},
page = {pending}
}