Question on suitability of Morpheus for synthetic 3D cell ground-truth generation

[Caspar Amery]

Dear,

I hope you are doing well, and I wish you a happy holiday season and a great start to the new year.

I am a master’s student in Computer Science (AI specialization). I am currently working on my master’s thesis, which focuses on generating realistic synthetic microscopy data to train AI models for cell segmentation.

I am reaching out because I am exploring whether Morpheus would be a good fit for one specific part of my pipeline, and I would greatly value your expert opinion.

Thesis goal (high-level)

My goal is not to study a specific biological process, but to generate realistic 3D ground truth of multiple interacting cells, which I then pass to a separate optics simulator to produce realistic 2D microscopy images. These images are used to train and evaluate AI segmentation models.

What I need from a 3D cell simulator

Concretely, I am looking for a simulator that can:

• Export a clean voxelized 3D grid
• Shape: (Z, Y, X) or (T, Z, Y, X)
• Values as either:
• instance labels (0 = background, 1..N = cell ID), or
• per-voxel density/intensity (fluorophore density), optionally with a parallel instance-ID volume
• Provide voxel size (dz, dy, dx) in µm (or enough metadata to derive it)
• Ideally export to NetCDF / OME-Zarr (xarray-friendly), but NumPy + sidecar metadata is also acceptable

Biological realism needed (focused & minimal)

• Realistic cell clustering and adhesion (this is the most important biological aspect for my use case)
• High diversity of cell morphologies (irregular, elongated, possibly branched shapes, ..; the more diverse the better)
• Temporal sequences with diverse deformations over time, to get diverse training data for models
  (cell division is note necessarily required)

Practical constraints

• ~50–150 cells 
• Typical volume size: 128 × 128 × 128 voxels
• Should run on a good desktop CPU/GPU
• Must be usable in headless / scripted batch mode for dataset generation

Optional but nice-to-have

• Multi-channel structure (e.g. whole cell + nucleus)

My question

Given this very specific goal (synthetic data generation for AI, not hypothesis-driven biology):

1. Would you consider Morpheus a good fit for this use case?
2. Can it naturally produce the kind of voxelized outputs described above, or would this require significant custom development?
3. If not ideal, would you recommend a different (possibly simpler) tool or workflow better suited for this purpose?
4. If it is a good fit: do you have recommendations on how to best get started for this specific use case (e.g. example models, settings, tutorials, or minimal workflows you would suggest focusing on)?

I fully understand if this use case falls somewhat outside the original scope of the tool, and I would greatly appreciate any honest guidance or redirection.

Thank you very much for your time, and again, warm wishes for the holidays and the upcoming new year.