There is certainly nothing revolutionary about nrrd, it just does what it is designed for - storing 2D/3D/4D images in 3D space with various scalar types and bit depth, with known origin, spacing, and axis directions.
We try to steer users away from using png and other consumer file formats because they are not designed for storing medical images and they do a very bad job at that: they cannot store 3D information, physical size, position, orientation, only support a few pixel types, etc.
Have a look at torchio and monai for good examples of preprocessing 3D medical images to be used as training data for deep learning.