marp
true

MorphoDepot

GitHub for Segmentation Projects (In alpha development as of October 2024)

https://MorphoCloud.github.io/MorphoDepotDocs/ (for Slide presentation)

Quick Links

• Instructions for PIs

• Instructions for Segmenters

• Souce repo: https://github.com/MorphoCloud/SlicerMorphoDepot

Big Picture Goals

• Faciliate open science analysis of biological specimens
• Tools and methods for scalable collaboration
• Simplify contributions for anyone (not just lab members)
• Make system as friendly as possible for new users
• Track user contributions (e.g. for co-authorship)
• Ideally usable for many purposes, but SlicerMorph users are first priority

Advantages of GitHub

• Free/inexpensive, widely used, mature
• Traceable, open science
• Already used in MorphoCloud

Use case summary

• Specimen scans that need annotation (e.g. segmentation or landmarks)
• A team of contributors
• A project leader to organize and approve contributions

Example project

• microCT scan of a frog species
• PI interested in anatomical segmentation
• several students in a class assigned to segment different parts (limbs, body, head...)

How this maps to GitHub

• PI and students all create accounts on github.com (free tier is okay, just need email address)
• Segmentation project is stored in a reposotory
  • Scan can be stored at any URL (e.g. s3 or JS2 bucket, or as github asset) in standard formats (nrrd, zarr...)
  • Repository follows MorphoDepot template (source URL, anatomy terminology, resulting segmentation)
  • Students are assigned issues for segmentation tasks
  • Students create pull requests for their assigned issues
• PI reviews and merges pull requests

Nice features of GitHub for MorphoDepot

• We can use the GitHub API to automate and simplify many steps
• GitHub handles user accounts, email integration, sys admin, and many other complex issues
• All GitHub project management features available to organize work
• GitHub repository "topics" can be used to organize, e.g. by species, scan type, or developmental stage
• Extra files can include links, images, experiment descriptions...
• Researchers learn useful open science tools

Big Picture Issues

• GitHub is designed for code, not 3D volumes
  • so text-based tools may not be useful
  • we may need to create custom data-specific replacements
• GitHub is built for code developers, not biologists
  • so tools can be difficult to use
  • we may need to develop automation or custom documentation
• GitHub is a commercial service and terms may change at any time
  • it's been around since 2008, owned by Microsoft since 2018
  • there are some alternatives if needed someday

3D Slicer Integration Plans

• A MorphoDepot module in SlicerMorph to simplify the workflow for contributors
  • List issues assigned to the contributor (or issues avalable to claim)
  • User can select an issue to work on
    • Automatically creates fork/branch and loads the data
    • Allows user to save work in progress to pull request
    • User can mark a contribution as ready for review
• MorphoDepot can be used on a local machine or a Jetstream2 virtual desktop

3D Slicer Integrations to Discuss

• TBD: tools to help PIs create MorphoDepot repositories
  • depends on how hard it turns out to be to do it manually
  • a tool could help ensure a very consistent structure for all repositories
• TBD: a tool to simply comparing and merging segmentations
• TBD: integration with machine learning pipelines (train autosegmentation)
• TBD: deal with large scans (too big for memory)

Discussion Points: Authentication model

• Ensure integrity of repository
  • Accurately track individual contributions
  • Don't allow users to delete or overwrite other contributions
• Respect user accounts
  • Don't give TA or fellow students access to others' accounts
  • Don't expose high-value credentials on shared machines
• Streamline login for new users
  • Simplify/automate but don't obscure underlying security model
  • Use existing GitHub tools when possible

Discussion Points: Repository architecture

• Single specimen / scan or multiples
• Structured metadata about specimen
  • Use of GitHub topics to find repositories
  • JSON Schema for describing segmentation tasks
• Best storage for original scans (GitHub assets, JS2 buckets, etc.)
• Fork vs branch model / are contributors part of the repo

MorphoDepot Status

• Initial prototype available for beta testing
• Feedback from MorphoCloud user community wanted!
  • suggestions for improved documentation or features welcome
  • need segmentation projects with concrete use cases to work out details