DARTv2: Deep Approximation of Retinal Traits

Overview

DARTv2 is a state-of-the-art tool for retinal vascular phenotyping, significantly improving upon the original DART model. It provides highly efficient, robust, and repeatable measurements of Fractal Dimension (FD) and Vessel Density (VD) from retinal colour fundus images. This tool is designed for easy integration and fast analysis, offering both a local GUI for batch processing and an accessible web interface.

Keywords:

• Retinal image analysis
• Deep learning
• Robustness
• Fractal Dimension (FD)
• Vessel Density (VD)

Paper

• Title: Self-consistent deep approximation of retinal traits for robust and highly efficient vascular phenotyping of retinal colour fundus images
• Authors: Lucas Gago, Beatriz Remeseiro, Laura Igual, Amos Storkey, Miguel O. Bernabeu, Justin Engelmann
• Status: Published (available on https://link.springer.com/chapter/10.1007/978-3-031-79103-1_22)

Abstract

Retinal colour fundus images offer a fast, low-cost, non-invasive way of imaging the retinal vasculature, which provides critical insights into both ocular and systemic health. Traditional approaches to retinal vascular phenotyping rely on handcrafted, multi-step pipelines that are computationally intensive and sensitive to image quality issues. DARTv2 overcomes these limitations by leveraging a self-consistent deep learning model that is fast, robust, and repeatable. It enhances the original DART by adding Vessel Density (VD) as a new trait, incorporating additional augmentations, and improving repeatability through a self-consistency loss. DARTv2 demonstrates high agreement with the AutoMorph pipeline (Pearson 0.9392 for FD and 0.9612 for VD), is more robust than both AutoMorph and the original DART, and achieves a significant speed-up in processing.

Features

• Fast and Efficient: 200x faster than AutoMorph and 4x faster than DART.
• Highly Robust: More resilient to image quality issues compared to traditional pipelines.
• Repeatability: Trained with a self-consistency loss for improved trait measurement consistency.
• Batch Processing: Supports drag-and-drop functionality for multiple images.
• Results Export: Outputs Fractal Dimension (FD) and Vessel Density (VD) as CSV, Excel, or TXT files.

Installation and Usage

Local Installation

1. Clone the repository:

   git clone https://github.com/your-repo/dartv2.git

2. Install dependencies:

   pip install -r requirements.txt

3. Run the GUI:

   streamlit run streamlit_dartv2_v3.py

4. Drag and drop multiple retinal images to begin analysis. The results will be saved in the desired format.

Web Interface

DARTv2 is also accessible via a free web interface: DARTv2 Web App. Images are only stored in RAM during the session, ensuring data privacy. Simply upload your images, and download the results in your preferred format.

Model Weights

The model weights are included in the repository for local inference.

Citation

If you use DARTv2 in your research, please cite our publication:

Gago, L., Remeseiro, B., Igual, L., Storkey, A., Bernabeu, M.O., Engelmann, J. (2025). Self-consistent Deep Approximation of Retinal Traits for Robust and Highly Efficient Vascular Phenotyping of Retinal Colour Fundus Images. In: Anazodo, U., et al. Medical Information Computing. MImA EMERGE 2024 2024. Communications in Computer and Information Science, vol 2240. Springer, Cham. https://doi.org/10.1007/978-3-031-79103-1_22

Acknowledgments

We would like to thank the authors and contributors of the original DART model and the AutoMorph pipeline for their foundational work in this field.