Ilkke's approach should work for images styled like the first one in the example. There's one big difference though: the example image has a higher resolution, and in particular, the horizontal resolution is greater than the vertical resolution (i.e. the source "pixels" aren't when they go into the polarization filter). This can probably be achieved by scaling the smart object appropriately. A minor difference is that the example image appears to have "pixels" of varying sizes, this can be achieved by having a very high horizontal resolution (so that the smallest unit can be the smallest radial pixel you want), and having the colours repeat as needed to create the larger "pixels". This can also be used to achieve the look of having the same size "pixels" in the middle as towards the outside - the inner "pixels" would all just consist of several actual "pixels".
The other images in the post aren't made of radial "pixels", in that case the effect uses a radial distortion only as a base to work on.