AI Pipeline for Retinopathy of Prematurity (ROP)

Multi-task deep learning pipeline for ROP severity classification on fundus and fluorescein angiography images.

Tasks

Task	Type	Classes
Plus disease	Binary	plus / no-plus
Stage	Binary	low (0–2) / severe (3+)
Zone	Multi-class	I / II / III

Architecture

• Backbone: EfficientNet-B0/B1 (transfer learning via timm)
• Multi-task heads: shared features → 3 independent classification heads
• Cross-validation: GroupedKFold (patient-level split)
• Explainability: Grad-CAM

Project Structure

ROP/
├── data/
│   ├── raw/              # Original images (organized by patient/eye)
│   ├── labels.xlsx       # Annotations (gold standard + readers)
│   └── processed/        # Preprocessed images
├── src/
│   ├── preprocessing.py  # Retina crop, resize, normalization
│   ├── dataset.py        # PyTorch Dataset
│   ├── model.py          # EfficientNet multi-task model
│   ├── train.py          # Training loop + cross-validation
│   ├── evaluate.py       # Metrics (AUROC, kappa, F1, etc.)
│   ├── interreader.py    # Inter-reader agreement analysis
│   ├── gradcam.py        # Grad-CAM explainability
│   └── run_experiments.py # Experiment orchestrator
├── configs/
│   └── config.yaml       # All hyperparameters
├── notebooks/            # Exploratory & analysis notebooks
├── outputs/              # Models, predictions, figures, reports
└── requirements.txt

Setup

# Create virtual environment
python -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

Labels File Format (data/labels.xlsx)

Column	Description
patient_id	Unique patient identifier
eye	L (left) or R (right)
modality	fundus or FA
image_path	Relative path to image in data/raw/
plus	0 = no plus, 1 = plus
stage_bin	0 = low (stage 0–2), 1 = severe (stage 3+)
stage_raw	Raw stage value (0–5)
zone	0 = zone I, 1 = zone II, 2 = zone III
gs_plus	Gold standard plus
gs_stage_bin	Gold standard stage (binary)
gs_zone	Gold standard zone
reader_*	Individual reader annotations

Running Experiments

# Preprocess all images
python src/preprocessing.py --config configs/config.yaml

# Run all experiments (fundus-only, FA-only, combined)
python src/run_experiments.py --config configs/config.yaml

# Inter-reader analysis
python src/interreader.py --config configs/config.yaml

# Generate Grad-CAM visualizations
python src/gradcam.py --config configs/config.yaml --experiment fundus_only

Metrics

• Plus & Stage: AUROC, Sensitivity, Specificity, F1-score
• Zone: Weighted kappa, Macro F1-score
• Comparison: AI vs Junior readers vs Senior readers (vs Gold Standard)