Size-Dependent Computed Tomography Histogram Analysis: Towards Breast Tissue Segmentation
SD Mann*, JP Shah, MP Tornai, Duke University, Durham, North CarolinaSU-E-I-2 Sunday 3:00:00 PM - 6:00:00 PM Room: Exhibit Hall
Purpose: To examine the effects of object size on scatter-corrected CT histograms to be used in threshold-based tissue segmentation.
Methods: A polyethylene cone filled with various concentrations of water and methanol mixtures, simulating glandular and adipose tissues, were imaged with our quasi-monochromatic dedicated breast CT, and scatter corrected using beam-stop array measurements. Images were reconstructed using iterative OSC, and individual coronal slices along the central axis of the cone were analyzed, with radii ranging from 3.25-6.25cm. Histograms from each slice were fit with two Gaussians (fluid filling and cone material) using nonlinear least squares methods, and the corresponding standard deviation and peak centroid of each filled material were evaluated as a function of object radius. Identical methods were applied to dedicated breast CT images of four patients for clinical comparison.
Results: Analysis of phantoms and breast data indicates low correlation between the standard deviation and object diameter. The centroids of the Gaussian peaks demonstrate an inverse linear relationship with increasing object size, independent of object material. The clinical datasets show a similar linear relationship between centroids and breast radius.
Conclusions: Data indicate that using a linear combination of Gaussian distribution functions to segment breast tissue in scatter corrected, quasi-monochromatic cone beam dedicated breast CT is possible. An object size-independent variation of attenuation values may allow for consistent restraints on initial fit parameters, resulting in improved confidence using Gaussian curve fitting. Appropriate scaling of the size-dependent volume slices, or independent slice analysis, is necessary to minimize binning variability for accurate tissue segmentation using Gaussian curve fitting.