Hi all,
I’m writing to ask if a new extension can be interested for the community.
In collaboration with some colleagues of mine, we wrote a python code for fractal dimension computation.
We had this idea because we really struggled to find something “easy to use” (with no success), so now we would like to share our solution.
For the moment it works just with 2D images, but our plan is to support also 3D volumes.
Do you think we can create a new scripted modules and expose it to the slicer users?
@PaoloZaffino for the sake of improved discoverability and reduced maintenance efforts, do you want to consider adding this as a new featureset to pyradiomics? From a quick search, fractal dimension has been used for radiomics studies, e.g., see https://www.ncbi.nlm.nih.gov/pubmed/29230678.
pyradiomics already has a large and thriving community (see mailing list and github repo), and is accompanied by the SlicerRadiomics extension exposing its features to the end user, so you would need to do no Slicer development, and your new features will magically become accessible to the SlicerRadiomics users. As an extra benefit, your features will be accessible in a package via PyPI and Conda, expanding the potential pool of happy users.
@Fernando we use it to quantify vessel complexity.
@fedorov@JoostJM thanks a lot for offering us this possibility…it would be great add our code to this project. @Pierangela00 is the PhD student that wrote the code and that is in charge of deploying it. Do you have any suggestion/guideline/policy about integrating our code into the project?
@fedorov thanks a lot.
Any suggestion about the file/place where we could put our code?
Is there an already defined data structure to get/return images?
There is already functionality to computer texture images, so maybe that is similar. I suggest you create an issue under pyradiomics, describe your code and discuss details there.
If I’m correct, fractal dimension features returns values on a segment level (i.e. where each values describes the entire region of interest), this would make it ideal to implement as a feature class. If it returns values per voxel (and therefore generates a parameter map), it is best implemented as an image type (=filter).
Both signatures are described here.
If you require assistance, let us know by submitting a PR or issue on the github as @fedorov suggested.