Reti-Pioneer

AI Framework for Multidisease Detection via Retinal Imaging

For more information, please visit our website.

News

• 2026/03: version 1.0 released

Hardware and software requirements

• A consumer-grade GPU (~6GB) is recommended for model training and testing.
• The code is tested on Ubuntu 20 and Windows 11.
• Python 3.12 is recommended.

Data

• UK Biobank and tertiary hospital centres in China are used for training and fine-tuning in this study.
• We recommend preprocessing the data to accelerate the training process in the following ways (N, M and D denote the numbuer of samples, clinical variables and diseases):
  • Use RETFound to extract the deep features and save it to UKB_RETF.npz with format
    • "left": N x 1024 numpy array
    • "right": N x 1024 numpy array
  • Save the clinical variables, quality scores, and center ID to UKB_mqd.npz with format
    • "m": N x M clinical variables
    • "mn": N clinical variable names
    • "ql": N x 3 left quality scores
    • "qr": N x 3 right quality scores
    • "center": N center IDs
  • Save diseases information to UKB_y{n}.npz with format
    • "y": N x D numpy array
• Place preprocessed data (UKB_RETF.npz, UKB_mqd.npz and UKB_y{n}.npz) on data/UKBCompressed or specify the path in code.
• A small (dummy) example is provide in data/UKBCompressed.

Installation Guide

• Clone this project.
• Use uv to create the virtual environment and install the dependencies.

git clone https://github.com/lyhyl/Reti-Pioneer.git
cd Reti-Pioneer
uv sync
.venv\bin\activate

Installation will take a few minutes on machines with good internet connection speeds.

Data Preparation

Since data from different sources exhibit distinct characteristics, they require different preprocessing methods. We recommend preprocessing all data independently before training or inference. Currently, only resizing (to 224×224) and center cropping are applied during inference; other preprocessing methods (e.g., padding) should be used as needed.

In our training process, for UKB fundus images, we first extract a central 1400×1400 square and then resize it to 224x224. For fundus images from other sources, we pad them into squares before resizing to 224x224.

Training

• Run the main script.

python main.py

The training will take a few minutes on machines with a consumer-grade GPU. The trained model will be saved to ckpt folder.

Inference

• Run the inference script.

python inference.py

Citation

Zhang, X., Li, Q., Liang, Y. et al. AI framework for multidisease detection via retinal imaging. Nat Med (2026). https://doi.org/10.1038/s41591-026-04359-w