Reinventing 2D Convolutions for 3D Medical Images