SlicerPSMASegmentation

A 3D Slicer scripted module for automated PSMA PET/CT lesion segmentation using Fine-UNETR — a foundation model trained on 2700 PSMA PET/CT scans.

---

Features

• Automatic CT/PET DICOM series discovery and reading
• SUV conversion from DICOM radiopharmaceutical tags
• CT-to-PET registration and resampling to model training spacing
• Sliding window inference with the Fine-UNETR (UNETR, patch size 8) model
• Segmentation overlay rendered in all Slicer slice views and 3D viewer
• Per-lesion clinical metrics: TTV (mL), SUVmax, SUVmean, centroid coordinates
• Excel export (.xlsx) with Summary and Per-Lesion sheets
• Coronal and sagittal MIP figures (.png) with prediction overlay
• Optional NIfTI export of the binary segmentation mask

---

Installation

1. Install 3D Slicer

Download from https://download.slicer.org (version 5.2 or newer recommended).

2. Add the module to Slicer

1. Clone or download this repository:

   git clone https://github.com/mansour2002/SlicerPSMASegmentation.git

2. In 3D Slicer: Edit → Application Settings → Modules → Additional module paths
3. Click Add and select the SlicerPSMASegmentation folder
4. Click OK and restart Slicer

After restart, the module appears under Segmentation → Nuclear Medicine → Fine-UNETR PSMA Segmentation.

3. Install Python dependencies

On first use, click Install Required Packages in the module panel. This installs:

Package	Version tested
torch	≥ 2.0
monai	≥ 1.2
pydicom	≥ 2.3
SimpleITK	≥ 2.2
scipy	≥ 1.9
nibabel	≥ 4.0
openpyxl	≥ 3.0
pandas	≥ 1.5
einops	≥ 0.6
matplotlib	≥ 3.5

---

Model Weights

The trained Fine-UNETR checkpoint is not included in this repository due to file size.

Download: Fine-UNETR.ckpt — Zenodo / Google Drive / HuggingFace (link will be updated upon public model release)

Place the downloaded Fine-UNETR.ckpt file anywhere accessible (e.g., inside the SlicerPSMASegmentation/ folder or another directory of your choice). The module will prompt you to browse to this file on first use.

---

Usage

DICOM Folder Input (recommended)

1. Set CT DICOM folder — directory containing CT DICOM files
2. Set PET DICOM folder — directory containing PET DICOM files
3. Enable Auto-detect CT/PET subfolders if CT and PET share a common root
4. Browse to the Model checkpoint (.ckpt file)
5. Adjust Threshold (default 0.5) and Min lesion size (default 0 = keep all)
6. Enter a Case ID (used as prefix in output files)
7. Set an Output folder and enable desired export options
8. Click Run Fine-UNETR Segmentation

Output

File	Description
Slicer scene	Orange segmentation overlay in all views
<CaseID>_Fine-UNETR_segmentation.nii.gz	Binary lesion mask (optional)
psma_lesion_metrics.xlsx	Per-lesion and summary metrics (optional)
<CaseID>_MIP_Coronal.png	Coronal PET MIP + prediction (optional)
<CaseID>_MIP_Sagittal.png	Sagittal PET MIP + prediction (optional)

---

Model Architecture

Fine-UNETR is based on the UNETR architecture (Hatamizadeh et al., 2021) adapted for PSMA PET/CT:

• Encoder: Vision Transformer (ViT-B/8), patch size 8³, 12 layers, hidden size 768
• Input: 2-channel [CT, PET] volume, normalized to [0, 1]
• Decoder: CNN with skip connections at ViT layers 3, 6, 9, 12
• Output: Single-channel binary lesion probability map
• Inference: Sliding window, ROI 96³ voxels, stride 64³, 32-voxel overlap

Training data: 2700 PSMA PET/CT scans for self-supervised pre-training; 80 scans with pixel-level lesion annotations for supervised fine-tuning.

---

Repository Structure

SlicerPSMASegmentation/
├── PSMASegmentation.py              # Main Slicer module (GUI + Logic)
├── CMakeLists.txt                   # Slicer extension build file
├── PSMA_Segmentation_User_Guide.html
├── Resources/
│   └── Icons/
│       └── PSMASegmentation.png
└── PSMASegmentationLib/
    ├── __init__.py
    ├── dicom_preprocessor.py        # DICOM reading, SUV conversion, registration
    ├── inference_engine.py          # Model loading and sliding window inference
    ├── sliding_window.py            # Custom sliding window for 2-channel input
    ├── volume_preprocessor.py       # Preprocessor for pre-loaded Slicer volumes
    ├── slicer_output.py             # Segmentation node creation, NIfTI/MIP export
    ├── metrics_reporter.py          # Per-lesion TTV/SUV metrics and Excel export
    └── models/
        ├── __init__.py
        └── unetr8ps_standalone.py   # Standalone Fine-UNETR model (no mmcv)

---

Citation

If you use this module in your research, please cite:

@software{fineunetr_slicer,
  author    = {Abtahi, Mansour},
  title     = {SlicerPSMASegmentation: Fine-UNETR for PSMA PET/CT Lesion Segmentation},
  year      = {2024},
  url       = {https://github.com/mansour2002/SlicerPSMASegmentation}
}

The UNETR architecture:

@inproceedings{hatamizadeh2022unetr,
  title     = {UNETR: Transformers for 3D Medical Image Segmentation},
  author    = {Hatamizadeh, Ali and Tang, Yucheng and Nath, Vishwesh and Yang, Dong
               and Myronenko, Andriy and Landman, Bennett and Roth, Holger R
               and Xu, Daguang},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications
               of Computer Vision (WACV)},
  year      = {2022}
}

---

License

MIT License — see LICENSE for details.

---

Contact

Mansour Abtahi — mansour.abtahi@gmail.com

Issues and pull requests are welcome.