• Weijie Chen, Maryellen L Giger, and Ulrich Bick.
• University of Chicago, Radiology, 584 South Maryland, MC Chicago, IL , USA. weije@uchicago.edu
• Acad Radiol. 2006 Jan 1; 13 (1): 63-72.

Rationale And ObjectivesAccurate quantification of the shape and extent of breast tumors has a vital role in nearly all applications of breast magnetic resonance (MR) imaging (MRI). Specifically, tumor segmentation is a key component in the computerized assessment of likelihood of malignancy. However, manual delineation of lesions in four-dimensional MR images is labor intensive and subject to interobserver and intraobserver variations. We developed a computerized lesion segmentation method that has the advantage of being automatic, efficient, and objective.Materials And MethodsWe present a fuzzy c-means (FCM) clustering-based method for the segmentation of breast lesions in three dimensions from contrast-enhanced MR images. The proposed lesion segmentation algorithm consists of six consecutive stages: region of interest (ROI) selection by a human operator, lesion enhancement within the selected ROI, application of FCM on the enhanced ROI, binarization of the lesion membership map, connected-component labeling and object selection, and hole-filling on the selected object. We applied the algorithm to a clinical MR database consisting of 121 primary mass lesions. Manual segmentation of the lesions by an expert MR radiologist served as a reference in the evaluation of the computerized segmentation method. We also compared the proposed algorithm with a previously developed volume-growing (VG) method.ResultsFor the 121 mass lesions in our database, 97% of lesions were segmented correctly by means of the proposed FCM-based method at an overlap threshold of 0.4, whereas 84% of lesions were correctly segmented by means of the VG method.ConclusionOur proposed algorithm for breast-lesion segmentation in dynamic contrast-enhanced MRI was shown to be effective and efficient.

13 keywords...

hide…