From a quick glance at the pilot data and the data description it seems that training data is annotated at the "whole image" level. I did not see any mention of localization information (e.g. tumor centers or contours). I have not worked with mammography data before, but I suspect this additional information would be of significant utility when training machine learning algorithms. Does such information exist and, if so, could it also be made available?

Created by zzgo
This information is not available to us and it's unrealistic to obtained it considering the size of the dataset (640905 images). Most importantly, we hope that learning algorithms will be able to identify different, novel markers than the ones currently used by radiologists to detect cancer.
I would expect that kind of information to have significant utility as well, but I think it unlikely to be included with the training/test data. Your model should be able to discover such localization information on its own.

localization ground truth for training data? page is loading…