Hi guys. I’m trying to replicate a study of using low dose CT and up-sampling to synthetic full dose using fast.ai. I’m wondering if there was a way to do this using full resolution ct images so that I don’t sacrifice quality and keep the images diagnostic (for evaluation later). Any tips on what to use and how to modify to keep the original resolution. GAN or autoencoder would be best?
I’m dashing off a quick response, without having thought very deeply about it.
This type of application raises warnings for me. In essence, a neural net hallucinates the most probable details into the upsampled image. This seems fine to do for an old photgraph, but not so great for a diagnostic medical image.
Suppose I get a low dose CAT scan to assess my small cancer. The ML process may well hallucinate a bigger cancer than is real, or worse, a smaller cancer, according to what that low res image is most similar to in the training set. That’s not something I would want for deciding my own treatment.
Maybe your loss funtion could heavily penalize false negatives, but then you’d have to know what a negative means for a training image. Or provide a confidence measure for new “objects” placed into the reconstructed image, whatever objects are.
Just musing here, but I wonder whether anyone has deeply considered these issues.