Experience using Bayesian segmentation on low-noise ultrasound images of residual limbs
M.C. Baker, Martin R. Stytz · 2002
To create a 3D rendering of an amputee's residual limb for a Computer-Aided Socket Design software, the image data must be accurately segmented to extract the significant fit points for the limb. This paper reports on the use of a Bayesian segmentation approach created to be robust enough for speckle-laden or noisy images on relatively noise-free ultrasound images of a residual limb. The authors wanted to see if the algorithm would be able to produce a segmented image with improved detail given that the original image was free of noise. The algorithm provides a multi-resolution segmentation using the maximum a posteriori probability (MAP) densities based on modeling the a priori probability distribution with a Gibbs random field (GRF). The authors summarize the segmentation process which consists mainly of a filter-and-decimate procedure and an adaptive clustering algorithm. Their results on low noise ultrasound data indicate that the Bayesian approach does not improve noticeably better segmentation of detail in the image data.