PRETTIER MRI

Perceptual super-resolution in multiple sclerosis MRI

Our paper has been published in Frontiers in Neuroscience, you can also check out the pre-print version.

This repository contains the script to apply PRETTIER, a framework to perform super-resolution on structural MRI. This framework relies on convolutional neural networks (CNN) that have been fine-tune with data from T2-W FLAIR and T1-W MRIs of people with Multiple Sclerosis.
We have developed and evaluated PRETTIER to increase the through-plane resolution of multi-slice MRI from 6mm to 1mm.

The name PRETTIER comes from Perceptual super-REsoluTion in mulTIple sclERosis. (You can say it is a rather convoluted explanation for a name, and I would agree. But you cannot say it isn't a pretty fitting name for a super-resolution method)

Requirements

This code depends on:

• NiBabel
• PyTorch, this tool has been developed and tested with PyTorch 2.1.1 and CUDA 11.8
• BasicSR is needed to import the generator architecture (RRDBNet) of RealESRGAN.

See full list in requirements.txt

List of fine-tuned models

Model	Info	Fine-tuned weights	# parameters	# FLOP
EDSR	Paper, Repository	EDSR_finetuned.pth	43089947	154.82B
RealESRGAN (generator)	Paper, Repository	RealESRGAN_finetuned.pth	16697987	55.11B
ShuffleMixer	Paper, Repository	ShuffleMixer_finetuned.pth	410579	1.49B

*FLOP (floating point operations) are estimated for a reference input patch of 96 x 16 pixels with 3 channels.

EDSR showed better results than RealESRGAN in our paper. We have recently included the fine-tuned ShuffleMixer model, which is more compact and efficient while achieving quantitative results comparable to RealESRGAN.

Usage

./prettier_mri.py --input <lr_input> --model-name {EDSR,RealESRGAN,ShuffleMixer} --output <output_image> [--gpu-id GPU_ID] [--batch-size BATCH_SIZE] [--no-flip-axes]

Example:

./prettier_mri.py --input demo_data/synth_LR_T1.nii.gz --model-name EDSR --output demo_data/prettier_edsr_synth_T1.nii.gz

Citation

If you use PRETTIER in your research, please cite:

@article{Giraldo2024prettier,
    author = {Giraldo, Diana L. and Khan, Hamza and Pineda, Gustavo and Liang, Zhihua and Lozano-Castillo, Alfonso and Van Wijmeersch, Bart and Woodruff, Henry C.  and Lambin, Philippe and Romero, Eduardo and Peeters, Liesbet M. and Sijbers, Jan},
    title = {Perceptual super-resolution in multiple sclerosis MRI},
    journal = {Frontiers in Neuroscience},
    volume = {18},
    year = {2024},
    doi = {10.3389/fnins.2024.1473132},
    url = {https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1473132},
    issn = {1662-453X}
}

Funding

This project received funding from the Flemish Government under the “Flanders AI Research Program".

Contact

Diana L. Giraldo Franco @diagiraldo