autoafids 0.1.4

End-to-end SnakeBIDS workflow for landmark regression and derivative apps

Automatic Anatomical Fiducials (autoAFIDs)

AIMS Lab Research Team at the Robarts Research Institute - 2023-2024

This package is under active development. It should be stable and reproducible, but please let any of the active contributing members know if there are any bugs or unusual behaviour.

This Python package is a standard 3D U-Net (Ronneberger et al. 2015) machine learning model based on Snakemake and SnakeBIDS workflow management tools that leverages the recent release of the anatomical fiducial framework to solve the landmark regression problem on 3D MRI images. It is currently in development phase and contains tunable parameters that are not normally exposed in most other machine learning models; the user is highly advised to get familiar with the above mentioned workflow managaments tools and read docstrings and relevant documentation before using this software. Please see the changelog for more details.

Table of Contents

1. Installation
2. Workflow
3. Train
4. Apply
5. Known issues
6. Roadmap
7. Questions, Issues, Suggestions, and Other Feedback

Installation

Installing Poetry

We use poetry tool for dependency management and to package the python project. You can find step by step instructions on how to install it by visiting it's official website.

Local Installation

After installing poetry, clone this repository via:

git clone https://github.com/afids/autoafids.git

You can then install the python package using one of the following commands, which should be executed within the repository folder (i.e., autoafids/).

poetry install

If you want to install in develop mode, use:

poetry install -e

Workflow

Brief summary of the workflow along with its Directed Acyclic Graph (DAG) can be found below.

Insert DAG image - inspired by khanlab/scattr

Processing landmark data (AFIDs)

1. Extract fiducial points from the landmark files (.fcsv is supported)
2. Generate a landmark Euclidean distance/probability map with each voxel communicating distance to an AFID of interest

Train

Currently, we support generating your own models (i.e., training) in a sperate workflow (i.e., afids-cnn: ). For more details, see Known Issues.

Apply

Known Issues

• Factorize apply workflow to run per landmark of interest

Roadmap

• Model optimization
• Extension to incorporate new modalities (i.e., CT scans)

Questions, Issues, Suggestions, and Other Feedback

Please reach out if you have any questions, suggestions, or other feedback related to this software—either through email ([e-mail]) or the discussions page. Larger issues or feature requests can be posted and tracked via the issues page. Finally, you can also reach out to Alaa Taha, the Science Lead.