(a)-(e): Retinal images extracted from different datasets: (a) RITE/DRIVE (b) INSPIRE (c) HRF (d) STARE (e) LES-AV . (f)-(j): Corresponding predictions. Note that, in contrast to most existing technques, this model was trained in RITE and never re-trained, see the paper for details.

This is an implementation of a method for Artery/Vein segmentation with uncertainty predictions. If you find this code useful for your research, we would appreciate if you could cite:

Uncertainty-Aware Artery/vein Classification on Retinal Images
A. Galdran,  M. I. Meyer, P. Costa, A. M. Mendonça, A. Campilho
IEEE International Symposium on Biomedical Imaging (ISBI), 2019

• PDF: Follow this link

Assuming you have cloned this repository to a folder F/ and already have a working Anaconda distribution installed in your system, you just need to execute the three lines below:

conda create --name av_uncertain python=3.7
source activate av_uncertain
conda install --file requirements.txt

And you are almost ready to go. After this, just cd into F/ and first download the weights to the models/checkpoints_uncertainty/ subfolder:

wget https://gitlab.com/agaldran/shared_models/raw/9ebce839b046a115ff7ba5defc6251a139eedfda/model_final.pth.tar -P models/checkpoints_uncertainty/

Once finished, run:

python build_predictions.py --path_ims retinal_images --path_out results

where retinal_images is the path containing the images you want to generate predictions for, and results can be replaced by the location where you want your results stored.

Note that this method provides pixel-wise predictions divided into four different classes: Background, Artery, Vein, and Uncertain. The results will come out color-coded:

• Red: Probability of being an Artery pixel
• Blue: Probability of being a Vein pixel
• Green: Probability of being an Uncertain prediction
• Background: 1-Artery-Vein-Uncertain

Running the above line will generate three subfolders of results/: uncertainty, uncertainty_vessels, and pretty_preds.

• Inside the uncertainty/ folder you will find predictions color-coded as above.
• Inside the uncertainty_vessels/ folder, Artery/Vein/Uncertainty have been merged into a single class (vessels), thereby providing a vessel segmentation model.
• Inside the pretty_preds folder, background predictions are reverted, so that pixels predicted as background will show up close to white, which improves predictions legibility, but this is only an aesthetic opinion of mine :)

Code for training your own model will come soon (I hope).