README.md

Detection and Segmentation Pipeline

By Scott Kinder scott.kinder@cuanschutz.edu

Overview

This repo is broken down into 3 main parts:

1. detection Detection
2. segmentation Segmentation
3. pipeline Pipeline end-to-end

Architecture

Data

We used 9 public datasets:

• Chákṣu
• Drishti-GS
• G1020
• ORIGA
• REFUGE-1
• RIM-ONE-DL
• RIGA (Composed of Bin Rushed, Magrabi, and Messidor)

Python environment

There are requirements.txt files here. This repo really just uses a couple standard Deep Learning libaries

• Pytorch
• OpenCV (cv2), PIL
• Numpy, Pandas, Matplotlib
• Ultralytics (for YOLO model)
• Albumentations (for image augmentations)

Complete setup

Download the models

See Detection and Segmentation modules for link to downloads

Run from nothing to segmentation output

# Clone repo
git clone https://github.com/QTIM-Lab/fundus_detection_segmentation_pipeline.git
# Change dir
cd fundus_detection_segmentation_pipeline
# Create a venv (recommended)
python3 -m venv venv
# Activate venv
source venv/bin/activate
# Intall packages
pip install -r requirements.txt
# Install repo module
pip install -e .
# Run end to end pipeline with params
python pipeline/scripts/end_to_end.py \
    --detection_model_path /path/to/detect/model.pt \
    --segmentation_model_path /path/to/seg/model.pt \
    --input_dir /path/to/input/images \
    --output_dir /path/to/output/folder \
    --dataset_mean 0.768 0.476 0.290 \
    --dataset_std 0.220 0.198 0.166 \
    --cuda_num 0