Frozen Foundation Model Benchmark for Diabetic Retinopathy Grading

A benchmark study comparing the frozen representations of three pretrained encoders—MedSigLIP, RETFound, and EfficientNet-B0—for diabetic retinopathy screening. The models were developed on the APTOS 2019 dataset (3,662 fundus images) and externally validated on MESSIDOR-2 (1,744 images). The study, authored by Mehmet Poyrazer and last updated in April 2026, evaluates binary detection and five-class severity grading using AUC, expected calibration error, and Brier score.

Use Cases

• Benchmarking frozen foundation models for medical image classification based on the described comparison of MedSigLIP, RETFound, and EfficientNet-B0.
• Evaluating model calibration under domain shift based on the reported Brier score and expected calibration error metrics.
• Assessing the generalizability of domain-specific pretraining based on the external validation results on the MESSIDOR-2 dataset.
• Analyzing performance for multi-class diabetic retinopathy severity grading based on the reported macro-F1 scores.

Strengths

• Provides a direct comparison of three distinct pretraining strategies (medical vision-language, retinal self-supervised, ImageNet-supervised) on a specific task.
• Includes external validation on a separate dataset (MESSIDOR-2 with 1,744 images) to assess generalizability.
• Reports co-primary endpoints beyond discrimination, including calibration metrics (Brier score, ECE) and statistical significance (p-values).
• Uses a standardized training protocol with frozen encoders and an identical lightweight head for a controlled comparison.

Limitations

• The dataset is a 497.1 KB DOCX file containing a table and analysis, not the raw image data; the underlying APTOS and MESSIDOR-2 datasets must be sourced separately.
• Column-level documentation for the presented table is absent; field semantics must be inferred from the description.
• Row count for the benchmark results table is unknown, which may limit suitability assessment.

Provenance

Source

figshare

Collection Method

Benchmark analysis of publicly available diabetic retinopathy image datasets (APTOS 2019, MESSIDOR-2).

Time Range

The study uses datasets from 2019 and earlier; the analysis was last updated in 2026.

Freshness

Last updated 2026-04-16 04:19:34

Geography

The geographic origin of the underlying image datasets (APTOS, MESSIDOR-2) is not specified in the provided description.