Four Paths, One Output

$224$ — pooling has no filters, so it contributes no channels of its own; the three convolutional paths are what set the output depth, and the block hands on less depth than it received.

$416$ — path D passes all $192$ input channels on, so the block returns more depth than it received, and each block stacked on top starts deeper and costs more than the last.

$192$ — concatenation lays the four paths over the same $192$ channels, so the depth coming out is the depth that went in, and a stack of these blocks holds its width steady.

$416$ — and that is the intention: an Inception network is meant to grow wider with every block, which is exactly what the parallel paths are there to do.