Noisy Segmentation Data Ideas

I’m working on a medical segmentation task and currently experiment with synthetic noise on the training and validation data. I noticed that the predictions show higher accuracy than the noisy ground truth data it learned from. I found that kinda cool.

Anyway, now i’m imagining a workflow that goes something like this:

  1. set up training and validation data using the same samples, but add noise to the training labels
  2. train network on noisy labels until performance on validation is more accurate than the noise on the training labels
  3. predict labels for the training set and use them as new training labels
  4. start all over

and then i thought: hey, doesnt that kinda sound like a GAN? so, now im thinking how would i set up something like this using a GAN? and also: does any of this make sense?