Sharp-Diffusion

This repository provides code for the assessment of feature loss for 2D super-resolution of diffusion MRI data as published in: Lohr, D and Werner, R. Layer Selection in Feature-Based Losses Affects Image Quality and Microstructural Consistency in Deep Learning Super-Resolution of Brain Diffusion MRI. ArXiv. DOI: 10.48550/arXiv.2605.15895. 2026

All code is available as jupyter notebooks which can be customized for application to new data. Models are trained using data from n=10 subjects from the 7T HCP young adult cohort.

If you are using this repository in your research, please cite Lohr and Werner

Repository setup

1. in your console move to an appropriate directory and clone the repository

git clone https://github.com/IPMI-ICNS-UKE/Sharp-diffusion

2. create a virtual environment (conda/miniconda) and activate it

conda create -n SHARP python=3.10

conda activate SHARP

3. install required dependencies using the requirement.txt

pip install -r requirements.txt

4. pre-trained models are available via the associated zenodo repository: https://doi.org/10.5281/zenodo.20267632

Quality metrics

MSE and PSNR are tracked during training. Respective metrics could be included into the loss by adjusting the notebook util/Diffusion_Perception_loss.ipynb

Experiments

1. Ablation study: models are trained for 4-fold SR using combinations of VGG16 layers for the loss (Diffusion_SR_Feature_Ablation_desc.ipynb)
2. Isolation study: models are trained for 4-fold SR using isolated VGG16 layers for the loss (Diffusion_SR_Feature_Isolation.ipynb)
3. Resolution assessment study: selected models are fine-tuned for 9-fold SR (Diffusion_SR_x3_Transfer.ipynb)

Data structure of the repository

Download datasets from HCP Servers. Extract HCP files in Sharp-Diffusion/data and run util/nii2png.ipynb to set up Training, Validation, and test splits. Folders x2 and gt build low- and high-resolution pairs.

• 'data/gt/train/volunteerID'
• 'data/x2/train/volunteerID'
• 'data/gt/valid/volunteerID'
• 'data/x2/valid/volunteerID'

Alternatively, set up your own training data following this structure.

Inference

For HCP data, nifti files and original original intensity scaling can be restored using preds2nii.ipynb. Adapting the path_base and testfolders enables application to new data (.pngs)

Citation

@misc{lohr2026layerselectionfeaturebasedlosses,
      title={Layer Selection in Feature-Based Losses Affects Image Quality and Microstructural Consistency in Deep Learning Super-Resolution of Brain Diffusion MRI}, 
      author={David Lohr and Rene Werner},
      year={2026},
      eprint={2605.15895},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2605.15895}, 
}