An atlas based approach to segmenting MLO view mammograms

摘要

A method for segmenting mediolateral oblique view mammograms is presented here. While most existing algorithms rely on histogram information to segment the breast region, we propose to use the shape and profile information to achieve more accurate and reliable segmentation. Here, we suggest a two step procedure consisting of feature based registration followed by Demon's implementation of deformable registration. This paper also addresses the issue of automatically detecting suitable landmarks such as nipple and pectoral muscle edge using existing algorithms. The results were evaluated against contours drawn by a radiologist on 18 mammograms selected randomly from the mini-MIAS dataset. Finally, a quantitative comparison using the criteria of false positive rate and false negative rate show that this algorithm clearly outperforms existing methods.