I've think I've noticed a correlation between images which have all sorts of bad Machine selected craters and that the indicated light source direction is wrong.
Here's an example, I just had another one much like it and decided not to post it, but when I hit another I had to post it.
The light here is indicated as being from the left bottom, but the source is really the upper left side. Not a single crater marked there is legit. So I wonder if the lighting is just wrong but not affecting the algorithm, the lighting is wrong and feeding into the algorithm causing problems, or some other problem such as the image isn't really for that set of data, etc.
I've been removing all of the bad blues and selecting the craters anyway.