tcav diabetic retinopathy

[FilipUniBe]

I'm unable to find any guidelines for TCAV in medical applications. The only information I found was tensorflow/tcav#115, where Kim Been confirms that we center-crop on the medical concepts.

So I'd like to know more details:

• Do you crop in detail as you would in a segmentation or does it have more leeway?
• What does work as "concept"? Does it have to be a specific pattern, such as a micro-aneurysm? Or could it be something less tangible like an angle or a size in relation to another size?
• Do the random medical concept have to similar to the actual concepts to be valid for statistical testing? Could I for example pair center-cropped micro-aneuryms with whole-image healthy DR-images?