Abstract
Ultra-widefield (UWF) retinal imaging captures up to 200° of the retina and contains diagnostically important peripheral and posterior pole features that are often lost when images are aggressively downsampled for vision transformer (ViT) processing. While increasing image resolution could preserve these fine-grained pathological signals, the quadratic dependence of self-attention on token count makes high-resolution training and inference computationally prohibitive. Motivated by the observation that token importance is highly uneven in UWF retinal images, we propose a compute-constrained scaling strategy that reallocates computation from redundant image tokens to increased image resolution through semantic token reduction, enabling higher-resolution UWF image analysis under a fixed compute budget. We instantiate this strategy using four token reduction methods—PiToMe, DiffRate, EViT, and DynamicViT—within a DINOv2 ViT-S/14 backbone constrained to approximately 45 GFLOPs. Across two retinal imaging datasets, compute-matched token reduction generally improves performance over the uncompressed 392×392 baseline, with the magnitude and statistical significance of gains varying across datasets and disease classes. On MMRDR-UWF, per-class AUC gains range from +0.016 to +0.033, with the largest improvements in disease stages characterised by fine lesions such as microaneurysms. On TOP, per-class AUC gains reach +0.022. These results indicate that reallocating compute from redundant tokens to increased image resolution can improve performance across multiple token reduction strategies, with larger and more statistically supported gains on MMRDR-UWF than on TOP. Our findings suggest that semantic token compression provides a practical pathway for scaling UWF image resolution without increasing computational cost and that substantial resolution-driven gains remain available for retinal disease classification, particularly for diseases characterised by fine pathological features.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/OMIA_037.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=XP5SJEb6a4
BibTex
@InProceedings{LopBry_Compress_MICCAISAT2026,
author = { Lopez de Munain, Bryan AND Kerr, Steven AND Pellegrini, Enrico AND Bernabeu, Miguel O.},
title = { { Compress to Scale: Semantic Token Reduction for High-Resolution Retinal Image Analysis } },
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}
}