The difference is the number of transmitters / receivers. TSA uses an array of them (from top to bottom) with narrow beams and they also rotate them and ask you to not move, so that effectively their number is considerably higher (ten of thousands), when reconstructing the image. Additionally, they use a higher frequency (160-400GHz), which additionally helps with resolution. See [1] and [2].
If we limit the number of transmitters / receivers to 3/4, the image reconstruction becomes impossible, similar how you would not really make a real photo from a few color sensors stationary looking at a scene, while a linear array of them that moves around an object would make a perfect scanner. It's a non-ideal analogy, but the best I can offer.
This is true in the sense that NNs offer a convenient way to estimate a distribution of hidden values conditionalized upon visible values. That is an approximation of what is usually understood by "a reconstruction". I would argue that a reconstruction per se would only include parts for which the credible interval could be systematically bounded within a suitable tolerance threshold.
