Compact vision models match domain-specific foundation models for several retinal imaging classification tasks: A systematic benchmark
Dávid Isztl, Tahm Spitznagel, Gábor Márk Somfai, Rui Santos · 2026
Read Compact vision models match domain-specific foundation models for several retinal imaging classification tasks: A systematic benchmark by Dávid Isztl, Tahm Spitznagel, Gábor Márk Somfai, Rui Santos (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0356202.txt.
Description
Dávid Isztl and three colleagues benchmark artificial intelligence models for eye disease classification in this 2026 study, pitting compact general-purpose networks against a large foundation model built specifically for retinal images. Nine model configurations were trained identically across four tasks: sorting retinal scans into disease classes and grading diabetic eye disease, detecting glaucoma, and rating macular edema severity from photographs.
Pretraining on ordinary photographs helped every model. Small hierarchical models with a fraction of the parameters matched or beat the giants on three of the four tasks, with one tiny architecture ranking first on three. The big retinal foundation model won only the hardest grading task, where severe cases were rare in the data. The authors conclude small models suffice for most retinal benchmarks, reserving foundation models for skewed, difficult grading problems.
Catalog record
| CREATOR | Dávid Isztl, Tahm Spitznagel, Gábor Márk Somfai, Rui Santos |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31553 |
| LICENSE | CC_BY |
| RIGHTS | CC BY |
| SOURCE | PLOS |