On the surface it does not seem that odd that the visual information from the retina stops on the way to the cortex for processing. Why not, most other sensory information does, the brain needs to process information, visual information is complex, etc. However, I am not convinced that these arguments for me make much sense. Synapses take time. There needs to be a sufficient gain in processing so that the delay in the signals is worth the time. However, the receptive fields at the retina and the LGN are essentially of the same types. The shapes of the receptive fields are center surround and the two basic types of responding are maintained. The parvo/magno cell types pick up where the X and Y cells left off. The responses of the cells in the ganglion cells and the LGN are not be identical but is enough gained to justify the delay? There is not even binocular responses in the LGN cells so even though it might anatomically be possible to have cells now respond to both eyes, the cells receiving input from the two eyes seem to be kept chastely separate in separate layers.
I have used a question mark in the title to indicate that I have no clear ideas on this matter at this time. However, one idea suggests itself to me recently. This is the idea of the attractor (Hofstader, 1985c). The attractor is a mathematical phenomenon that for functions, like the parabola below, if you take the output of the function from arbitrary start values and feed it back into itself, and repeat this process, eventually the iterations will constantly give the same output value. As seen below, this happens for the parabola when x=y crosses the parabola. Thus this value can be said to attract the iterations of the functions. What if the retina/LGN serve as a couple of iterations towards an attractor. In a sensne the LGN receives as an input, the output of the same DOG functions of the retina. What if this iteration leads to a more stable output region, the attractor in the multidimensional visual space, before the elaboration in the cortex. In other words, the output of the LGN may be less dependent upon vagaries of the retina stimulation than the output of the ganglion cells simply by virtue of the iteration of the LGN? I have no clue if this makes real quantitative sense but it does suggest some role of the LGN that might be very useful. To be useful, there would have to be a relationship between the retinal input and the retinal output and the LGN such that stability could be obtained in a few steps. Playing with the seed below and the parabola values you can see that some settings stabilize in a few steps and others in several.