Personally I have always favoured sphere collision detection above rectangular.
Less calculations and it could also be used as a rough detection first. i.e. make the sphere’s diameter as large as your longest side, then if you collide, go for the detailed & calc intensive checks.
Advantage is also a sphere can work in 2d and 3d without much change.
You also have to consider how seldom ‘pure’ rectangular or sphere collection detections are. Chances are high you might even have to go all the way down to pixel or 3d level detection (forgot the name) to make it ‘look right’ anyway.
Also when your set of collide-able objects grows, you might also consider using space partitioning patterns, such as BSP, QSP or OSP.