MorphoGen

Efficient Unconditional Generation of Long-Range Projection Neuronal Morphology via a Global-to-Local Framework. Code is almost one-click runnable. Below we introduce the environment dependencies, file description, datasets used, and the code execution. Paper is here.

Dependencies

python==3.8.5, pytorch==1.8.2, torchvision==0.9.2, cudatoolkit==11.1

See requirements.txt for detailed environment specifications.

File Description

• sub_process.py: Converts raw SWC files to standardized point cloud data.
• distort.py: Distorts true branches to learn the mapping back to original state.
• DDPM_train.py: Trains the denoising diffusion probabilistic model to predict global structures.
• Auxiliary_train.py: Trains the auxiliary CNN networks to optimize the local structures.
• morphology_gen.py: Generates new morphology point clouds and converts into SWC files.

Dataset

Long-range neuronal data is sourced from this study.

• CT subtypes (45-52): 1,085 neurons (all subtypes)
• PT subtypes (57-64): 1,005 neurons
• IT subtypes (34-44): 985 neurons

Code Execution

train the DDPM：

python DDPM_train.py --dataroot ${dataroot} --model_dir${model_dir} --device ${device}

train the Auxiliary CNN：

python Auxiliary_train.py

generate new neuron morphology:

python morphology_gen.py --dataroot ${dataroot} --model${model} --device ${device} --generate_dir ${generate_dir}

Citation

If you find this repository useful, please cite our paper:

InProceedings{Zhu_2025_ICCV,  
    author    = {Zhu, Tianfang and Zhou, Hongyang and Li, Anan},  
    title     = {MorphoGen: Efficient Unconditional Generation of Long-Range Projection Neuronal Morphology via a Global-to-Local Framework},  
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},  
    month     = {October},  
    year      = {2025},  
    pages     = {13021-13031}  
}

Acknowledgement

Thanks for the wonderful work DiT-3D.