Compact vision models match domain-specific foundation models for several retinal imaging classification tasks: A systematic benchmark
PLOS One · 18.08.2026 · 21 (8) · e0356202
DOI: 10.1371/journal.pone.0356202 ↗Kurz gefasst
Fachartikel in der Zeitschrift PLOS One, online erschienen am 18.08.2026. Verfasst von Dávid Isztl und 3 weiteren Autor:innen. Beteiligte Einrichtung: Stadtspital Zurich, Department of Ophthalmology, Zurich, Switzerland. Hauptthemen laut MEDLINE: Retina, Image Processing, Computer-Assisted. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Werner H. Spross-Stiftung.
Abstract (Originalsprache)
Large domain-specific foundation models have been widely adopted for retinal image analysis, yet systematic evidence for their advantage over compact general-purpose architectures remains scarce. We benchmarked nine model configurations spanning 22.8M to 303M parameters (vision transformers, hierarchical Swin Transformers, ConvNeXt, and the domain-specific RETFound models) across four tasks: OCT 8-class disease classification, and three fundus photography tasks (DME severity, glaucoma detection, and DR severity grading). All models were evaluated under identical training conditions, with both pretrained (on natural-domain image datasets) and from-scratch initializations compared using Mann-Whitney U tests. Pretraining improved accuracy by 5.18–18.41 percentage points across all tasks (p < 0.05 throughout), with larger benefits for CFP modalities and harder tasks. Compact hierarchical models (27–29M parameters) matched or exceeded larger architectures on three of four tasks. For instance, the SwinV2-tiny architecture ranked first on OCT, DME, and GL classification. The domain-specific RETFound model (303M) achieved the highest accuracy only on the most challenging task (DR severity grading, where the most severe class is underrepresented at 8% of images), where it outperformed the best compact model by 1.54 percentage points. These results indicate that compact general-purpose models may be sufficient for most retinal classification benchmarks, and that domain-specific foundation models may add higher value mainly for severity grading tasks with skewed class distributions.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Autor:innen und Institutionen
- Stadtspital Zurich, Department of Ophthalmology, Zurich, Switzerland.
- Spross Research Institute, Zurich, Switzerland.
ORCID-Profile: Tahm Spitznagel, Rui Santos
Förderung
- Werner H. Spross-Stiftung
Zitieren
APA
Isztl, D., Spitznagel, T., Somfai, G. M., & Santos, R. (2026). Compact vision models match domain-specific foundation models for several retinal imaging classification tasks: A systematic benchmark. PLOS One, 21(8), e0356202. https://doi.org/10.1371/journal.pone.0356202
Vancouver
Isztl D, Spitznagel T, Somfai GM, Santos R. Compact vision models match domain-specific foundation models for several retinal imaging classification tasks: A systematic benchmark. PLOS One. 2026;21(8):e0356202. doi:10.1371/journal.pone.0356202
BibTeX
@article{Isztl2026,
author = {Isztl, Dávid and Spitznagel, Tahm and Somfai, Gábor Márk and Santos, Rui},
title = {Compact vision models match domain-specific foundation models for several retinal imaging classification tasks: A systematic benchmark},
journal = {PLOS One},
year = {2026},
volume = {21},
number = {8},
pages = {e0356202},
doi = {10.1371/journal.pone.0356202}
}
EUVIMED