Claritas is a collection of standalone Python tools for denoising PET images using SwinUNETR-based neural networks.
It includes scripts for preprocessing .nrrd PET data, training models with or without tumor masks, and evaluating performance.
The repository is designed for research purposes only.
The developed models can be directly implemented in SlicerPETDenoise Extension which is available from here: https://github.com/4burakfe/SlicerPETDenoise
GUI application for training PET denoising models with tumor masks.
Features:
- Supports SwinUNETR and UNet
- Fully customizable loss functions (MSE, MAE, SSIM, tumor-weighted L1)
- Real-time training plots and logging
- Uses .npy files for training and validation
- The input files shold have be in pairs in seperate folders with exact same names
Same as above but for training without tumor segmentation masks.
Resamples .nrrd PET files to 2×2×2 mm Crops central 256×256 area Saves full volumes as .npy
Extracts non-overlapping patches of size 64×64×64 Saves each patch as a separate .npy file
Models are designed to use with SlicerPETDenoise Extension which is available from here: https://github.com/4burakfe/SlicerPETDenoise
| Filename | Description |
|---|---|
pet_denoiser_std_char.pth |
Trained with standard Charbonnier loss for general denoising. |
pet_denoiser_x19w1_5.pth |
Tumor-weighted loss model (1.5x penalty for underestimation). |
pet_denoiser_x14w3.pth |
Strong tumor-focused model (3x weight for underestimation). |
📦 Requirements Python 3.9+ PyTorch MONAI SimpleITK numpy matplotlib einops
For installation of torch please visit https://pytorch.org For other dependencies you can use command: "pip install monai einops SimpleITK matplotlib numpy"
Disclaimer: This software is for research purposes only and is not certified for clinical use.