Patch-Based MLP Spleen Segmentation Fusion

This repository provides a GPU-accelerated patch-based multi-layer perceptron (MLP) for voxel-wise fusion of multiple spleen segmentation pipelines.
The model integrates local CT intensity patches, local mask agreement/disagreement, and 3D spatial coordinates to produce an improved spleen segmentation.

A fully reproducible Apptainer/Singularity image is available on Zenodo.

---

⭐ Key Features

• Learned fusion of four pipelines:
  • TotalSegmentator
  • DeepSpleenSeg
  • DeepMultiOrgSeg
  • GennUNet
• CUDA-enabled voxel-wise patch inference
• Dataset-agnostic: accepts any folder structure using --mask_paths
• Includes a working sample test case
• Reproducible: packaged as a .sif container (Zenodo DOI)

---

Structure of expected directory for test run:

🧪 1. Quick Test (Recommended)

A complete working example is included:

test_case/
├── INPUTS/
│ └── image.nii.gz
└── OUTPUTS/
  └──TotalSegmentator.nii.gz
  └──DeepSpleenSeg.nii.gz
  └──DeepMultiOrgSeg.nii.gz
  └──GennUNet_spleen.nii.gz

Run inference:

🚀 3. Run the MLP Fusion on Any Dataset (Singularity)

Use the following command exactly as shown.
This binds your /home-local directory and runs inference on a chosen case:

singularity run --nv  -e --contain  -B /home-local:/home-local   mlp_fusion_v1.0.1.sif   --case_dir  test_case   --mask_paths "OUTPUTS/TotalSegmentator.nii.gz,OUTPUTS/DeepSpleenSeg.nii.gz,OUTPUTS/DeepMultiOrgSeg.nii.gz,OUTPUTS/GennUNet_spleen.nii.gz

singularity run --nv -e --contain \
  -B $PWD:$PWD \
  mlp_fusion_v1.0.1.sif \
  --case_dir $PWD/test_case \
  --mask_paths "OUTPUTS/TotalSegmentator.nii.gz,OUTPUTS/DeepSpleenSeg.nii.gz,OUTPUTS/DeepMultiOrgSeg.nii.gz,OUTPUTS/GennUNet_spleen.nii.gz"

After running, the fused MLP segmentation is written to:

/OUTPUTS/mlp_spleen_torch.nii.gz

Download singularity image sif:

📦 4. Container & Zenodo

Latest version DOI (all versions):

https://doi.org/10.5281/zenodo.17903413