ECLARE: Efficient cross-planar learning for anisotropic resolution enhancement | Paper

ECLARE is a self super-resolution method; no outside training data is required. The only required input is the image itself and the output directory.

The journal manuscript is currently in preparation; the link will be provided soon!

NOTE: ECLARE was developed on top of refactors and improvements to SMORE and as such the SASHIMI 2023 paper points to that URL. However the code and developers are the same.

Quick start

We recommend starting from a fresh Python 3.10 installation. If you use Anaconda or Miniconda, you can do:

conda create -n eclare python=3.10

Although this code may work on version of Python >3.10, it has not been tested so results may vary on such software versions.

Run pip install eclare

To super-resolve an input, run:

run-eclare --in-fpath ${INPUT_FPATH} --out-dir ${OUTPUT_DIR} --gpu-id ${GPU_ID}

This will produce a file ending in {YOUR_INPUT_FILENAME}_eclare.nii.gz in OUTPUT_DIR and will use the GPU you specify.

We also support additional arguments:

run-eclare --in-fpath ${INPUT_FPATH} --out-dir ${OUTPUT_DIR} [--relative-slice-thickness ${SLICE_THICKNESS}] [--relative-slice-profile-fpath ${BLUR_KERNEL_FILE}] [--gpu-id ${GPU_ID}] [--suffix ${SUFFIX}]

where each argument in [] is optional and may be omitted.

• If you have a good guess of your data's slice thickness, you can specify --relative-slice-thickness and provide a number relative to the in-plane. So if in-plane is, say, 0.125mm and you think through-plane is 0.25mm, then you would give --relative-slice-thickness 2. This models the relative slice profile as a Gaussian kernel, but slice separation is still read from the header. Therefore, setting this parameter indirectly specifies slice gap as well.
• The default GPU ID is 0.
• You can change the suffix to be -SR to produce [...]-SR.nii.gz, etc., based on your needs. You specify the delimiter.
• If you know the form of your slice selection profile as a .npy array, you can also specify it with --relative-slice-profile-fpath.

Citations

If this work is useful to you or your project, please consider citing us!

Remedios, S.W., Han, S., Zuo, L., Carass, A., Pham, D.L., 
Prince, J.L. and Dewey, B.E., 2023, October. Self-supervised 
super-resolution for anisotropic MR images with and without 
slice gap. In International Workshop on Simulation and 
Synthesis in Medical Imaging (pp. 118-128). Cham: Springer 
Nature Switzerland.

@inproceedings{remedios2023self,
  title={Self-supervised super-resolution for anisotropic {MR} images with and without slice gap},
  author={Remedios, Samuel W and Han, Shuo and Zuo, Lianrui and Carass, Aaron and Pham, Dzung L and Prince, Jerry L and Dewey, Blake E},
  booktitle={International Workshop on Simulation and Synthesis in Medical Imaging},
  pages={118--128},
  year={2023},
  organization={Springer}
}