Morphology-based losses for weakly supervised segmentation of mammograms

Mickael Tardy, Diana Mateus

Keywords: Mammography, Weakly supervised learning, Segmentation, Morphology

TL;DR: Morphological operators such as top-hat and closing allow to improve the precision of segmentation in a weakly supervised setting.

Abstract: Segmentation is one of the most common tasks in medical imaging, but it often requires expensive ground truth for training. Weakly supervised methods cope with the lack of annotations, however, they often fall short compared to fully supervised ones. In this work, we propose to constrain the segmentation output with morphological operations, leading to an increase in the overall performance. In particular, we use top-hat and closing operations. We evaluate the method on high-resolution images from INBreast dataset and achieve an increase in F$_1$ of $\approx 0.14$ and in recall of $\approx 0.22$ compared to the training without morphology loss.

Paper Type: methodological development

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Radiology

Paper Status: based on accepted/submitted journal paper

Source Code Url: The source code is protected by the IP of company that funded this research.

Data Set Url: https://pubmed.ncbi.nlm.nih.gov/22078258/

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.