Example-based Segmentation for Breast Mass Images

Qingying Huang,Songhua Xu,Xiaonan Luo

DOI: https://doi.org/10.1117/12.2008022

2013-01-01

Abstract:A new example-based mass segmentation algorithm is proposed for breast mass images. The training examples used in the new algorithm are prepared by three medical imaging professionals who manually outlined mass contours of 45 sample breast mass images. These manually segmented mass images are then partitioned into small regular grid cells, which are used as reference samples by the algorithm. Each time when the algorithm is applied to segment a previously unseen breast mass image, it first detects grid cell regions in the image that likely overlap with the underlying mass region. Upon identifying such candidate regions, the algorithm then locates the exact mass contour through an example based segmentation procedure where the algorithm retrieves, transfers, and re-applies the human expert knowledge regarding mass segmentation as encoded in the reference samples. The key advantage of our approach lies in its adaptability in tailoring to the skills and preferences of multiple experts through simply switching to a different corpus of human segmentation samples. To explore the effectiveness of the new approach, we comparatively evaluated the accuracy of the algorithm for mass segmentation against segmentation results both manually produced by several medical imaging professionals and automatically by a state-of-the-art level set based method. The comparison results demonstrate that the new algorithm achieves a higher accuracy than the level set based peer method with statistical significance.2