Hi @behxyz , could you please help to check the calculation first. I just tested locally, and will add formal test cases and doc string later.

Compared with the startup code in #1800 , this version enhances:

1. Optimize the calculation via using more numpy based methods, now for a same 2D input probability map, the calculation speed is around 3 times faster when sigma > 0, and 5 times faster when sigma is 0.
2. Remove skimage dependency.
3. Supports any dimension inputs.

@yiheng-wang-nv, maybe we should change the name from RadialNMS to PixelNMS or something. Because although we are receiving the radius as an input, we are suppressing over a square (or a cube) and not a circle (or a sphere). Unless, we use a radial base to suppress the pixels within the given radius.

@yiheng-wang-nv, maybe we should change the name from RadialNMS to PixelNMS or something. Because although we are receiving the radius as an input, we are suppressing over a square (or a cube) and not a circle (or a sphere). Unless, we use a radial base to suppress the pixels within the given radius.

Agree, do you think we should also change the name of radius?

@yiheng-wang-nv
Could you please finalize this PR without resolution and add another pathology specific one with resolution in monai/apps/pathology/?