Omnidirectional Volume Slicer - OmniSlicer

Implementation of the omnidirectional volume slicer package in Python and PyTorch for 3D medical image analysis, from our Medical Image Analysis publication "TomoGraphView: 3D Medical Image Classification with Omnidirectional Slice Representations and Graph Neural Networks".

Example Usage

from OmniSlicer import OmniSlicer

volume_path = "path/to/volume.nii.gz" 
mask_path = "path/to/mask.nii.gz" 
output_dir = "path/to/output_dir"
n_views = N

OmniSlicer.encode_volumes(
  volume_list=[volume_path],   # you can pass a list of several image paths here
  mask_list=[mask_path],       # you can pass a list of several segmentation mask paths here
  output_dir=output_dir,
  n_views=n_views,
  model_name="facebook/dinov3-vits16-pretrain-lvd1689m"  # currently, choose from any official facebook/dinov2 or facebook/dinov3 model available on huggingface
)

Tested Dependencies

The functionality of OmniSlicer has been successfully validated using the following dependency versions. These represent the environment in which the package has been developed and tested:

These versions are defined in the project’s installation requirements and are automatically resolved when installing OmniSlicer via pip. While other combinations may work, the dependency set above is the configuration against which all core features have been verified. Please make sure that you install torch with CUDA.

Citation

@article{kiechle2026tomographview,
  title={TomoGraphView: 3D medical image classification with omnidirectional slice representations and graph neural networks},
  author={Kiechle, Johannes and Fischer, Stefan M and Lang, Daniel M and Bercea, Cosmin I and Nyflot, Matthew J and Felsner, Lina and Schnabel, Julia A and Peeken, Jan C},
  journal={Medical Image Analysis},
  pages={104174},
  year={2026},
  publisher={Elsevier}
}