Related Experiment Video
Updated: May 8, 2026
07:59
Efficient and Consistent Generation of Retinal Pigment Epithelium/Choroid Flatmounts from Human Eyes for Histological Analysis
Published on: October 28, 2022
One Size Fits All? Comparing Foundation and Task-specific Models for Retinal Fluid Segmentation
Xiaoyu Sun1,2, Saiyu You1, Siqi Sun1
1Institute for Informatics, Data Science and Biostatistics, Washington University in St. Louis, St. Louis, MO, USA.
Summary
Task-specific models for retinal fluid segmentation outperform general foundation models in accuracy and consistency. Specialized models like RetiFluidNet are currently more reliable for clinical applications than broad ophthalmic foundation models.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Retinal fluids are crucial biomarkers for diseases like diabetic macular edema and age-related macular degeneration.
- Automated segmentation of these fluids aids clinical research and decision support tools.
- Ophthalmic foundation models show potential but their performance for specialized tasks is unclear.
Purpose of the Study:
- To compare the performance of a task-specific model (RetiFluidNet) against an ophthalmic foundation model (VisionFM) for retinal fluid segmentation.
- To evaluate segmentation accuracy and fluid burden estimation capabilities.
- To determine the current reliability of specialized versus general models for clinical OCT analysis.
Main Methods:
- A standard benchmarking dataset of 4,248 OCT images from 48 patients with retinal diseases was used.
- Models were evaluated using three-fold cross-validation.
- Pixel-level segmentation accuracy and patient-level fluid burden estimation were assessed.
Main Results:
- The task-specific model (RetiFluidNet) demonstrated superior segmentation performance.
- RetiFluidNet provided more consistent fluid quantification across different devices.
- Foundation models showed lower performance compared to the specialized model.
Conclusions:
- Specialized, task-specific models currently offer greater reliability for retinal fluid segmentation than general-purpose ophthalmic foundation models.
- Further targeted adaptation is necessary for foundation models to match or exceed specialized model performance in clinical settings.
- Task-specific models remain the preferred choice for accurate retinal fluid analysis in current clinical research and decision support.
