Fat Segmentation in Abdominal MR-scans

This thesis describes a method for automatically segmenting abdominal adipose tissue from 3-dimensional magnetic resonance images. The segmentation distinguishes between three types of adipose tissue; visceral adipose tissue, deep subcutaneous adipose tissue, and superficial subcutaneous adipose tissue. Prior to the segmentation, the image data is preprocessesed to remove withinclass image intensity inhomogeneities caused by the so-called bias field effect. The field is sampled as two classes of intensity points and the effect is estimated using an extension of thin plate splines. The adipose tissue is labelled across the abdomen by unsupervised classification using fuzzy c-means clustering and locally determined thresholds. The abdomen boundary is segmented, and the visceral adipose tissue is separated from the subcutaneous adipose tissue by means of active contours; incorporating intensity information derived through the unsupervised classification. The subcutaneous adipose tissue layer is subdivided into a deep and superficial part by dynamic programming and a polar transformation of the image data. In the absence of ground truth segmentations, the results are subject to a visual validation; good results are obtained across the broad spectrum of images present in the data set.