Synthetic Microscopy Timelapse Data Generation for Zero-Annotation AI Mitochondrial Segmentation and Tracking

---

MitoMimicsDemo.mp4

Introduction

MitoMimics is designed to simulate temporal mitochondrial dynamics-based microscopy data. MitoMimics employs a Python-based rendering pipeline to generate synthetic widefield microscopy datasets of mitochondrial dynamics events. Our algorithmic toolset provides researchers a means for testing and validating AI-based models, providing both ease of use and flexibility.

Timeline.2.mp4

Features

• Simulates mitochondrial motility, fission, and fusion in 2D space.
• Generates synthetic time-lapse microscopy data.
• Generates synthetic instance masks along with a JSON file that registers instances between frames, which can be used to validate segmentation and/or tracking models.
• Allows customization of simulation parameters, such as:
  • Mitochondrial Density
  • Mitochondrial Speed
  • Fission and Fusion Rates
  • Simulation Time Scale
  • Dimensions of the Simulation Space
• Allows customization of microscopy parameters, such as:
  • Point Spread Function (2D Gaussian PSF)
  • Noise (Background Offset, Camera Readout Noise, Poisson Noise)

Installation

To install MitoMimics, follow these steps (tested on macOS Sequoia - Tahoe, and Ubunutu 20.04 - 22.04) (Install time 5-15 minutes):

1. Clone the repository: git clone https://github.com/aidanpcquinn/mito_sim_pack.git
2. Navigate to the project directory: cd MitoMimics/envs
3. Install environment: conda env create -f environment.yml
4. Activate the conda environment with conda activate mitodynamicsim

Using the Simulation

To use MitoMimics for dataset generation, follow these steps (Run time: <10 minutes per simulated image stack with default parameters):

1. Navigate to the MitoMimicsGeneration folder and run the simulation script: python sim.py --seed 42
   • This generates the underlying simulation of the mitochondria morphology and dynamics
   • Parameters can be modified in parameters.yaml or with python parametes_gui.py
   • The files are saved in MitoMimicsGeneration/sim_output/42
   • batch scripts for running multiple jobs gen_multi_batch_sim.sh
2. Run the rendering script with python renderer.py --seed 42 --gpu 0
   • This renders the simulation into the synthetic microscopy stacks
   • Parameters can be modified in parameters.yaml or with python parametes_gui.py
   • The files are saved in MitoMimicsGeneration/render_output/42
   • batch scripts for running multiple jobs gen_multi_batch_render.sh
   • Note: on CUDA accelerated systems the cupy library can be installed, the flag --cupy True may speed up rendering.
3. To view the simulated data, run python sim_napari_viewer.py --loc render_output/42
   • The user can vary the timescale indefinitely by changing the save_data or sim_length_seconds variables.

Training With Simulated Data

To use MitoMimics for training on generated synthetic data, follow these steps (Run Time: 24 hours, on 4x NVIDIA H100s)

1. Download a processed synthetic training dataset (Training_Data.tar.gz) from https://zenodo.org/uploads/19603477 REVIEWER TEMP LINK
2. Install Umamba (https://github.com/bowang-lab/U-Mamba) and follow instructions for dataset preperation and training
   • We trained with nnUNetv2_train DATSET_ID 3d_fullres 0 -tr nnUNetTrainerUMambaEnc -num_gpus 4 and otherwise default parameters
3. Note: you will need to process and train on a dataset for both fullmasks and centerlines
4. To use your own tiff stacks, see 2_1_raw_to_intermediary_demo.ipynb and 2_2_intermediary_prepro.ipynb in /RealDataProcessing

Inference and Post Processing

1. Download demo unlabaled photogentle sequences (Unlabelled_Photogentle_Sequences.tar.gz) from https://zenodo.org/uploads/19603477 REVIEWER TEMP LINK
2. Run inference with both trained models on real dataset
   • nnUNetv2_predict -i raw_unlabaled_photogentle_sequences/ -o fullmask_out_location/ -d DATASET_ID_fullmask -c 3d_fullres -f 0 -tr nnUNetTrainerUMambaEnc --disable_tta
   • nnUNetv2_predict -i raw_unlabaled_photogentle_sequences/ -o centerline_out_location/ -d DATASET_ID_centerline -c 3d_fullres -f 0 -tr nnUNetTrainerUMambaEnc --disable_tta
3. Run Post processing scripts 3_1 to 5_3 in /RealDataProcessing. Use python 3_1...py --help to view arguments
   • post processing needs the raw nifti stacks from step 1, and the outputs of both segmentation models in 2

Inference with pretrained model

1. Download demo unlabaled photogentle sequences (Unlabelled_Photogentle_Sequences.tar.gz) from https://zenodo.org/uploads/19603477 REVIEWER TEMP LINK
2. Download the trained model weights ('Trained_UMamba_Models.zip') from https://zenodo.org/uploads/19603477 REVIEWER TEMP LINK
3. Place the unziped folders in your UMamba install at U-Mamba/data/nnUNet_results/
4. Run inference with both trained models on real dataset
   • nnUNetv2_predict -i raw_unlabaled_photogentle_sequences/ -o fullmask_out_location/ -d 901 -c 3d_fullres -f 0 -tr nnUNetTrainerUMambaEnc --disable_tta
   • nnUNetv2_predict -i raw_unlabaled_photogentle_sequences/ -o centerline_out_location/ -d 902 -c 3d_fullres -f 0 -tr nnUNetTrainerUMambaEnc --disable_tta
5. Run Post processing scripts 3_1 to 5_3 in /RealDataProcessing. Use python 3_1...py --help to view arguments
   • post processing needs the raw nifti stacks from step 1, and the outputs of both segmentation models in 2

Citation

License

MitoMimics © 2026 by Aidan Quinn, Volkan Ozcoban & Vijay Rajagopal is licensed under GPL-3.0

Contact

For any questions, feedback, or potential collaborations, please contact the original authors of MitoMimics:

• Aidan Quinn

  📧 aidanpcquinn@gmail.com

  🐙 Github: @aidanpcquinn

• Volkan Ozcoban

  📧 volkanozcoban1@gmail.com

  🐙 Github: @VolkanOzcoban

• Vijay Rajagopal

  📧 vijay.rajagopal@unimelb.edu.au

  🐙 Github: @vraj004

Please include “MitoMimics” in the subject line when contacting us.

CellSMB lab

See more from our lab at our github: https://github.com/CellSMB