Superpixel-Guided Label Softening for Medical Image Segmentation

• DOI: 10.48550/arxiv.2007.08897
• Publication Date: 2020-01-01
• Authors: Hang Li; Dong Wei; Shilei Cao; Kai Ma; Liansheng Wang; Yefeng Zheng
• Keywords: Ground truth; Intuition

Segmentation of objects interest is one the central tasks in medical image analysis, which indispensable for quantitative analysis. When developing machine-learning based methods automated segmentation, manual annotations are usually used as ground truth toward models learn to mimic. While bulky parts segmentation targets relatively easy label, peripheral areas often difficult handle due ambiguous boundaries and partial volume effect, etc., likely be labeled with uncertainty. This uncertainty labeling may, turn, result unsatisfactory performance trained models. In this paper, we propose superpixel-based label softening tackle above issue. Generated by unsupervised over-segmentation, each superpixel expected represent a locally homogeneous area. If intersects annotation boundary, consider high probability uncertain within Driven intuition, soften labels area on signed distances boundary assign values [0, 1] them, comparison original "hard", binary either 0 or 1. The softened then train together hard labels. Experimental results brain MRI dataset an optical coherence tomography demonstrate that conceptually simple implementation-wise method achieves overall superior performances baseline both 3D 2D images.