title	OIRseg
emoji	🔬
colorFrom	yellow
colorTo	red
sdk	streamlit
sdk_version	1.55.0
app_file	app.py
pinned	false

OIRseg

Retinal image segmentation tool for Oxygen-Induced Retinopathy (OIR) research. Segments retinal flatmount images into three zones: neovascular (NV), vaso-obliterated (VO), and retinal tissue, with quantitative area measurements.

Features

• Multi-class segmentation — U-Net with EfficientNet-B4 encoder, trained on retinal flatmounts
• Single & batch processing — upload one image or a batch for bulk analysis
• Test-time augmentation (TTA) — optional multi-transform inference for improved accuracy
• Per-class threshold tuning — adjustable confidence thresholds for NV, VO, and retina masks
• NV post-processing — vessel suppression and VO boundary zone filtering to reduce false positives
• Downloadable results — masks, overlays, and CSV metrics as a ZIP
• PubMed RAG (optional) — AI-powered interpretation of results using retrieved PubMed literature and Claude

Project Structure

.
├── app.py                 # Streamlit web interface
├── src/
│   ├── config.py          # Model & training hyperparameters
│   ├── model.py           # U-Net architecture (segmentation-models-pytorch)
│   └── predict.py         # Inference & post-processing pipeline
├── rag.py                 # PubMed RAG: retrieve abstracts + Claude Q&A
├── ingest.py              # Fetch & embed PubMed abstracts into ChromaDB
├── config.py              # RAG configuration (API keys, search params)
├── theme.py               # Streamlit UI theme
├── requirements.txt       # Pinned Python dependencies
└── .env.example           # Environment variable template

Setup

Requirements

• Python 3.11
• Model checkpoint (best_model.pth) — downloaded automatically from Hugging Face on first run

Installation

python -m venv venv

# Windows
.\venv\Scripts\Activate.ps1
# Linux/macOS
source venv/bin/activate

pip install -r requirements.txt

Run

streamlit run app.py

The app opens at http://localhost:8501. On first launch, the model checkpoint (~227 MB) is downloaded automatically from Hugging Face Hub.

PubMed RAG (Optional)

Enables AI-powered interpretation of segmentation results using PubMed literature.

1. Copy .env.example to .env and add your API keys:

   ANTHROPIC_API_KEY=sk-ant-...
   NCBI_EMAIL=your@email.com
   

2. Ingest PubMed abstracts:

   python ingest.py

3. Restart the app — the RAG tab will become available.

Model

Component	Value
Architecture	U-Net
Encoder	EfficientNet-B4 (ImageNet pretrained)
Decoder attention	scSE
Input size	768 x 768
Output classes	3 (NV, VO, Retina)
Loss	BCE + Focal Tversky (class-weighted)

License

This project was developed for academic research purposes.