Glandular segmentation of cone beam breast CT volume images

Cone beam breast CT (CBBCT) has potential as an alternative to mammography for screening breast cancer while limiting the radiation dose to that of a two-view mammogram. A clinical trial of CBBCT has been underway and volumetric breast images have been obtained. Although these images clearly show the 3D structure of the breast, they are limited by quantum noise due to dose limitations. Noise from these images adds to the challenges of glandular/adipose tissue segmentation. In response to this, an automated method for reducing noise and segmenting glandular tissue in CBBCT images was developed. A histogram based 2-means clustering algorithm was used in conjunction with a seven-point 3D median filter to reduce quantum noise. Following this, a 2D parabolic correction was applied to flatten the adipose tissue in each slice to reduce system inhomogeneities. Finally, a median smoothing algorithm was applied to further reduce noise for optimal segmentation. The algorithm was tested on actual breast scan volume data sets for subjective analysis and on a 3D mathematical phantom to test the algorithm. Subjective comparison of the actual breast scans with the denoised and segmented volumes showed good segmentation with little to no noticeable degradation. The mathematical phantom, after denoising and segmentation, was found to accurately measure the percent glandularity within 0.03% of the actual value for the phantom containing larger spherical shapes, but was only able to preserve small micro-calcification sized spheres of 0.8 and 1.0 mm, and small fibers with diameters of 1.2 and 1.4 mm.