Training lesion detectors from noisy annotations: an empirical study in mammography

Machine learning algorithms need carefully-annotated datasets, these are not always available for medical images. When it is possible the annotations, such in mammography, need the presence of experts and it is much more costly than general object detection task. Moreover, the opinion of the experts may not be unanimous. As a consequence, in many cases may happen that it is not feasible having scrupulously-annotated datasets. One alternative solution is to use the annotations that are already available in clinics or hospitals, which may not be particularly made for research. These annotations are usually bigger than the actual size of the lesions. This enlargement can be considered as noise, which is been modeled and injected in a publicly available dataset. By further exploring the behavior of the Faster R-CNN, it has been observed that the matching criterion used for labeling anchor/bounding boxes plays an important role. It has been observed this model tends to overfit with small datasets. For this reason, an alternative samples selection has been proposed, which shows a significant improvement. The noise injected produces a decrease in the quality of the detection which proves the lack of robustness of the Faster R-CNN.