PRISM: Privacy-preserving Inter-Site MRI Harmonization via Disentangled Representation Learning

• Paper: arXiv preprint (Accepted to ISBI 2025)
• Cite:

  @misc{galada2024prismprivacypreservingintersitemri,
      title={PRISM: Privacy-preserving Inter-Site MRI Harmonization via Disentangled Representation Learning}, 
      author={Sarang Galada and Tanurima Halder and Kunal Deo and Ram P Krish and Kshitij Jadhav},
      year={2024},
      eprint={2411.06513},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2411.06513},
  }

---

Usage:

1. Clone / download this repository, navigate to the code folder and pip install requirements.txt
2. Skull strip your MRI volumes using FreeSurfer / FSL software (aka Brain extraction). [If you don't have MRI data, the openly available IXI Dataset is the easiest option.]
3. Run MRI-Slicer.py on the stripped MRI volumes. [Note: In some cases, where the MRI volumes are stored in different orientations, you may need to manually reorient the data]
4. Run folder2dataset.py to generate separate custom MRI datasets (incl. augmentations) for each site.
5. To train the PRISM model on each site, follow the PRISM-training.ipynb notebook. Alternatively, run train.py. Repeat the training procedure for each participating site.
6. With the models now trained, follow the PRISM-inference.ipynb notebook to harmonize the MRI without any data exchange, as per the PRISM framework. Alternatively, run harmonize.py.
   • reconstruct.py and visualization.py can be run to evaluate PRISM's reconstruction performance and visualize the "style" harmonization.

(More coming soon)

---

Architecture:

---

Results:

--