Refinement of Weak Annotations for the Segmentation of Bone Marrow Leukocytes

Philipp Gräbel, Martina Crysandt, Reinhild Herwartz, Melanie Baumann, Barbara M. Klinkhammer, Peter Boor, Tim H. Brümmendorf, Dorit Merhof

Keywords: Leukocyte Segmentation, Weak Annotations, Annotation Refinement, Bone Marrow Microscopy

TL;DR: Evaluation of several refinement methods and their impact on improving weak annotations for training a segmentation network on bone marrow microscope images. Our novel method 
yields segmentation results 
close to manual training data.

Abstract:

Deep neural networks are well suited to address medical problems such as the automated analysis of leukocytes in bone marrow images. However, their training requires large annotated datasets.
The shortage of annotations is one of the most prevalent problems in biomedical image analysis. Particularly with polygonal contours as segmentation training data, creating high quality annotations is infeasible. Weak annotations, e.g. bounding boxes, can be obtained more easily.
This paper investigates several approaches that aim at refining weak annotations. The resulting refined contours are used to train a semantic segmentation network. Our evaluation shows that it is possible to achieve precision close to training with ground truth data, with a novel U-net method, presented in this paper.