AutoMorphalyzer 2025 👀

This codebase is a significantly adapted version from the original codebase found here. The original codebase was written as part of the publication AutoMorph: Automated Retinal Vascular Morphology Quantification via a Deep Learning Pipeline.

This adaptation was motivated by the need for a faster, more accessible, user-friendly and correct version of AutoMorph at the University of Edinburgh. We hope this may also prove useful to other research institutes.

Please contact Jamie.Burke@ed.ac.uk if you have questions.

Note: This is still in development, and further tests are required to compare the output from AutoMorph and AutoMorphalyzer.

---

Major changes from Automorph in AutoMorphalyzer

• Code is restructured to run from a simple python script automorph/main.py directly on Anaconda Prompt after local installation.

• The end-user is only required to specify file paths in config.txt, requiring two inputs: the path to the data (input_directory) and where to save the results (output_directory). Upon running `python automorph/main.py' while inside the AutoMorphalyzer folder using Anaconda Prompt will trigger the pipeline to run.

• Model weights are packaged as a GitHub release and automatically downloaded upon first running locally. This will not need to be done again.

• CFP images are currently not rejected based on quality (the original M1 module has been removed), and instead are always fed through the pipeline.

• Quality is now computed as a probability of rejection from (QuickQual)[https://github.com/justinengelmann/QuickQual] and saved out in the final results file collating all features.

• At the pre-processing stage, images are downsized to (912,912) as that is the largest dimension used for segmentation and feature measurement. This also help facilitate manual annotation of segmentations using ITK-Snap

• For all modules (pre-processing, segmentation, feature measurement) if an image has previously been analysed, it is skipped to save time. This is based on:

  • Pre-processing: the presence of the image file path in {output_directory}\M0\crop_csv.csv.
  • segmentation: the presence of the segmentation mask in each of the {output_directory}\M2\{model_type}\raw_binary for model_type in binary, artery_vein and optic_disc.
  • feature measurement: if feature_measurements.csv exists in {output_directory}\M3 and the presence of the image file path in the csv file.

• Segmentation is still performed using the original models from the AutoMorph codebase. That is,

  • Vessel segmentation (Binary & Artery-Vein): BF-Net
  • Optic disc segmentation: lwnet

• The feature measurement module has had a major revamp to remove redunancy, improve processing time. Comparisons and execution times to be published soon on this repository.

  • Fractal dimension, vessel density and global vessel calibre remain the same. These are only provided across the whole image, and not across zones B and C.
  • Tortuosity (distance and density) and local calibre are corrected from the original codebase. These are provided across all zones.
    • Extracting individual vessel segments has been sped up significantly through use of Numba and a Depth-first search algorithm.
    • Tortuosity is calculated across individual vessel segments, and the original codebase extracted some vessels segments incorrectly, leading to exaggerated values in tortuosity. This has been corrected.
    • Local calibre was originally measured inefficiently per vessel segment.
  • Tortuosity squared curvature has been removed due to redundancy with other tortuosity measures.
  • We only measure CRAE and CRVE using the Knudtson formula, and remove the Hubbard formula. These are only measured in zones B and C.
  • We include arteriovenous ratio (AVR) in zones B and C.

• If an error is encountered during pre-processing or feature measurement, the file is skipped and a full traceback is both printed out in the terminal and also written to a .txt log file which is saved in {output_directory}\M3. If failure occurs at feature measurement, relevant metrics will be saved as -1s in the output .csv file.

• Extensive use of the os package throughout the codebase ensures that this pipeline will work across different OS systems. It has currently been tested on Windows and MacOS (so is likely to also work in Linux).

• Composite segmentation visualisations of the binary vessels, arteries, veins and optic disc/cup are saved out per file for easy look-up. This can be found in {output_directory}\M3\segmentations.

• AutoMorphalyzer infers laterality based on vessel density either side of the lateral position of the optic cup.

• The only images saved out currently are composite segmentations (as above) and raw binary segmentations of the binary vessels, artery-veins and optic disc/cup. These are saved at the same dimension as for feature measurement (912,912). Thus, we do not save out uncertainty maps (from the ensemble models), or segmentation masks for different zones or at different dimensions.

• Currently we assume no knowledge of pixel resolution, so metrics global/local vessel calibre, disc/cup height and width are left in pixel units. Both the original and new codebases measure metrics using a universal dimension of (912,912) so these features are still comparable across populations analysed using these codebases.

• This pipeline also allows manual editing of binary vessel, artery-vein and optic disc-cup segmentations using ITK-Snap and fed back into pipeline to recompute feature measuremements.

---

Getting started

To get a local copy up, follow the steps in quick_start.txt, or follow the instructions below.

1. You will need a local installation of Python to run AutoMorphalyzer. We recommend a lightweight package management system such as Miniconda. Follow the instructions here to download Miniconda for your desired operating system.

2. After downloading, navigate and open the Anaconda Prompt and clone the AutoMorphalyzer repository.

git clone https://github.com/jaburke166/AutoMorphalyzer.git

3. Create environment and install dependencies to create your own environment in Miniconda.

conda create -n automorph-env python=3.11 -y
conda activate automorph-env
pip install -r requirements.txt

Done! You have successfully set up the software to analyse colour fundus photography data!

See below for how to run AutoMorphalyzer on your own data.

---

Running AutoMorphalyzer

note: While we endeavoured to speed up the codebase, we still recommend the use of GPU acceleration for the segmentation module.

---

Contributors and citing

The contributors to this adapted codebase are:

• Jamie Burke (Jamie.Burke@ed.ac.uk)

The original author of this codebase is:

• Yukun Zhou (yukun.zhou.19@ucl.ac.uk)

If you wish to use this toolkit please consider citing the original work using the following BibText, with additional comment on this being an adapted version.

@article{zhou2022automorph,
  title={AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline},
  author={Zhou, Yukun and Wagner, Siegfried K and Chia, Mark A and Zhao, An and Xu, Moucheng and Struyven, Robbert and Alexander, Daniel C and Keane, Pearse A and others},
  journal={Translational vision science \& technology},
  volume={11},
  number={7},
  pages={12--12},
  year={2022},
  publisher={The Association for Research in Vision and Ophthalmology}
}