GPT-Image Fundus Benchmark

This repository contains a processed fundus-image benchmark, training/evaluation code, experiment outputs, and model checkpoints for four ophthalmic classification directions:

• Myopia / Myopic Maculopathy: MMAC, 5 classes
• AMD: ADAM, binary classification
• Glaucoma: AIROGS and PAPILA, binary classification
• Diabetic Retinopathy (DR): IDRiD, APTOS-2019, DeepDRiD, 5 classes

The repository was prepared for experiments comparing RetFound, ResNet-50, and ViT-B/16, including full-data baselines and data-scarcity downsampling experiments.

Repository Layout

code/                         Training, preprocessing, evaluation, and report scripts
data/                         Processed ImageFolder datasets: train/val/test/<label>/
results/                      Metrics, plots, logs, predictions, HTML report outputs (no checkpoint files here)
weights/pretrained/           RETFound pretrained CFP/OCT weights
weights/checkpoints/baseline/ Best checkpoints for 7 datasets x 3 models
weights/checkpoints/downsample/ Best checkpoints for ADAM/AIROGS/PAPILA downsampling experiments
weights/checkpoints/legacy_retfound/ Old-format RetFound checkpoints kept for full provenance
metadata/                     Sample packs, quality-label spreadsheets, and report artifacts
weights_manifest.csv          Manifest of every uploaded weight file
FILES.md                      Directory-level file inventory

Data Format

Each dataset under data/ follows a standard ImageFolder layout:

data/<Disease>/<Dataset>/
  train/<label>/*.jpg
  val/<label>/*.jpg
  test/<label>/*.jpg
  labels.csv

labels.csv contains:

split, filepath, label, class_name, orig_id

Model Checkpoints

All weights are centralized under weights/ for easier reuse on another server.

• weights/pretrained/RETFound_mae_natureCFP.pth: RETFound CFP pretrained checkpoint used for fundus-image fine-tuning.
• weights/pretrained/RETFound_mae_natureOCT.pth: RETFound OCT pretrained checkpoint retained for completeness.
• weights/checkpoints/baseline/: best checkpoints from the full-data benchmark.
• weights/checkpoints/downsample/: best checkpoints from the data-scarcity experiments.
• weights/checkpoints/legacy_retfound/: old-format RetFound outputs retained for provenance; final metrics use the canonical downsample/<dataset>/<pct>/retfound/ paths.

See weights_manifest.csv for the full checkpoint list.

Main Results

The benchmark results are summarized in:

• metadata/report.html and results/report.html: single-file HTML report with dataset backgrounds, class distributions, metrics, confusion matrices, ROC curves, per-class metrics, and data-scarcity analyses.
• results/summary.csv: full-data benchmark summary.
• results/downsample/summary.csv: downsampling experiment summary.

Reproducing / Running

The scripts in code/ expect the project root to contain data/, weights/, and results/ as laid out in this repository.

Typical entry points:

python code/run_all.py --dry_run
python code/run_all.py
python code/downsample_experiment.py --dry_run
python code/downsample_experiment.py
python code/make_report.py

The environment used during experiments was a conda environment with PyTorch 2.5.1 + CUDA 12.1, timm, scikit-learn, pandas, matplotlib, seaborn, and openpyxl.

Data Source and License Notice

This repository aggregates and reformats data derived from multiple public ophthalmology datasets/challenges, including Kaggle, Zenodo, and challenge-hosted sources. Each original dataset may have its own license, terms of use, and citation requirements. Users are responsible for checking and complying with the original licenses before redistribution, publication, or commercial use.

This repository is intended for research and reproducibility. It is not intended for clinical diagnosis.

Quality Labels

Manual image-quality labels for ADAM and MMAC preview subsets are included in metadata/:

• quality_review_AMD_ADAM_MMAC.rechecked_D_standard.csv
• 图像质量类型明细_按数据集分sheet_追加标签.xlsx

Quality categories include light haze, severe haze, low contrast, small-pupil/off-axis appearance, vignetting/peripheral edge-quality decline, reflection/glare, and occlusion.