Bias field is a common phenomenon in a breast sonogram. Although artifacts caused by bias filed may carry important
information, e.g., shadowing behind a lesion, they are generally disturbing in the process of automatic boundary
delineation for sonographic breast lesions. This paper presents a new segmentation algorithm aiming to decompose the
region of interest (ROI) into prominent components while estimating the bias field in the ROI. A prominent component is
a contiguous region with a visually perceivable boundary, which might be a noise, an artifact, a substructure of a tissue or
a part of breast lesion. The prominent components may be used as the basic constructs for a higher level segmentation
algorithm to identify the lesion boundary.
The bias field in an ROI is modeled as a spatially-variant Gaussian distribution with a constant variance and
spatially-variant means, which is a polynomial surface of order n. The true gray levels of the pixels in a prominent
component are assumed to be Gaussian-distributed. The proposed algorithm is formulated as an EM-algorithm composed
of two major steps. In the E-step, the ROI is decomposed into prominent components using a new fuzzy cell-competition
algorithm based on the bias field and model parameters estimated in the previous M-step. In the M-step, the bias field
and model parameters are estimated based on the prominent components derived in the E-step using a least squared
approach. The results show that the effect of bias field on segmentation has been reduced and better segmentation results
have been attained.