C-V MODEL WITH REGION CONSTRAINT FOR MAMMOGRAPHIC MASS SEGMENTATION

YIHUA LAN*,**, ZHENSHUANG LI**, XIAO SONG*, JINJIANG LIU**, ENMING SONG*, JINXIN WAN***, KUNXIAO SHEN****

* Center for Biomedical Imaging and Bioinformatics, School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, Hubei 430074, China - **School of Computer and Information Technology, Nanyang Normal University, Nanyang 473061, China - ***Medical imaging department, Lianyungang second people’s hospital, Lianyungang 222005, China - ****School of Biological Engineering, Nanyang Normal University, Nanyang, 473061, China

Introduction: Lesion segmentation in mammograms, which affects the subsequent feature extraction and classification, plays a critical step in most of computer-aided detecting and diagnosis (CADx) systems. However, accurate segmentation of suspicious masses from surrounding breast tissues is a challenging and difficult task as lesion boundaries are usually overlapping or touching normal structures, obscured or irregular or low contrast. The aim of this paper is to propose an accurate and robust algorithm for mammographic mass segmentation.

Material and methods: In this paper, we present new scheme for accurately segmenting breast lesions in mammograms. To achieve this task, two stages are included in this scheme. In the first, a hyperbolic secant template was applied to locate a rough rectangular region of breast lesions. Then, based on the rough region, C-V (Chan-Vese) model was employed to obtain the contour of the being segmented lesions in this region. 483 regions of interest (ROIs) extracted from 328 patients were taken from a publicly available database, the Digital Database for Screening Mammography (DDSM) and formed as the test data set. Area based and boundary distance based similarity measures based on radiologists’ manually marked signs were used to evaluate the performance of the proposed approach and other four algorithms. This paper also provided a comparison between the proposed approach and other four different segmentation algorithms (i.e., Geodesic Active Contours, C-V model, Timp and Karssemeijer’s dynamic programming method and marker-controlled watershed segmentation method).

Results: Experimental results show that our scheme are better than those of other algorithms, the mean overlap percentage and combined measure were 0.7198 0.1149, 0.8056 0.08 respectively. Form experimental results, we also notice that values for all the measures distribution of the proposed method were.

Discussion: Through the comparison, our scheme has better performance to improve the accuracy of lesions segmentation than other four algorithms.