[new299]

It's a first step I guess. At least if you can get the classification to work with controlled data collection you have reasonable evidence that the approach can work.

If all other parameters area randomized between case and control, this is also fine. I'd guess you can also add illumination and other artifacts to the datasets to make the training robust to this.

But ultimately in the extreme case (like wanting to submit images from a point and shoot camera) I suspect you'll have a hard time building a system that's robust to that... or robust enough to be used in a diagnostic context...

I personally don't think I'd be comfortable working without a standard configuration used by the radiologists and a controlled protocol.