[FEA] CUDA accelerated foolfill algorithm with distance return about cucim

[JonasFrey96]

Hello @grlee77,

Thanks a lot for your kind response.
I found an efficient workaround by using the FastGeodis library (BSD-3):
https://github.com/masadcv/FastGeodis
This avoids creating a graph.
Creating python bindings for NPP would be most likely the best solution.

[grlee77]

Thanks @JonasFrey96 for the kind words and for providing a concrete example. It helped me to understand what you are asking for. Let me think about it and see what might be the best option.

There is an implementation of flood fill in NPP, but I don't think it will work for your purposes. As far as I understand, it operates more like a paint bucket tool to fill a region around the seed and does not return distances. That also does not currently have Python bindings.

I don't see how to modify the distance transform kernels we have in cuCIM to do this, but am wondering if we might be able to build a graph to represent the image and then use an appropriate graph algorithm from cuGraph to compute the distance. I think this problem may be called "single source shortest path" (SSSP). I am not too familiar with such graph algorithms, but think this could be a possibility worth exploring. I do see SSSP listed under "Traversal" algorithms on this page in the cuGraph guide.

We would have to figure out how to convert a binary image into the graph format it expects. I know upstream scikit-image does have some graph-related algorithms that utilize NetworkX, but we don't currently have any of the skimage.graph module implemented in cuCIM.