Inference / Training on less than 4 views
Linardos opened this issue · 5 comments
I wanted to do some experiments with the DDSM dataset using your code but your neural network seems to expect four inputs per patient corresponding to left / right + cc / mlo. In some patients of DDSM unfortunately not all views are present. Some have 2 some 3 some have all 4.
Not sure why that's the case with DDSM, It's a pretty 'dirty' dataset in many respects though so I'm not surprised.
Is there a way to use your model on patients that don't have all 4 views?
Hi Akis,
I think the easiest thing which is a little wrong but also pretty easy, would be to duplicate one of the available views (probably doesn't matter which one) so that there are 4 views in total. I expect that this should work reasonably well.
Hmm I see. That was my first thought as well. I guess It wouldn't hurt to try. I'll probably make some comparative analysis later on using just the 4 view samples and see whether adding the incomplete samples severely impacts performance.
Thank you so much for making this open source and well documented by the way!
One more question regarding this if I may: We have received some feedback that there asymmetries between breasts occur so that the left's density is different to the right's which is indicative of cancer / cause for alarm. It seems this information is lost if we concatenate them.
Was it a radiologist's recommendation to use 4 different views towards a single density classification or is there some practical reason for it?
This is how radiologists report it (at least at NYU). The entire exam has only one label characterizing breast density.
That's really useful to know, thanks!