Segmentation of male abdominal fat using MRI

This thesis describes the methods used to construct a pipeline for the automatic and robust segmentation of adipose tissue in the abdominal region of human men. The segmentation is done into 3 classes: subcutaneous adipose tissue, visceral adipose tissue and other tissue. The MRI data is preprocessed to remove the field of non-uniformity in intensity levels that are present on MR images. A novel way of sampling the field is introduced and the field is estimated using Thin Plate Splines. The initial clustering of the data is done on the preprocessed data using Fuzzy c-mean clustering. The results of the clustering are accurate partly due to a successful preprocessing. The segmentation of adipose tissue into the subcutaneous adipose tissue and visceral adipose tissue classes is done using a combination of Active Shape Models and Dynamic Programming. This hybrid approach of combining the two methods makes for a both robust and accurate segmentation. No ground truth is available to verify the accuracy of the results against. The results have however been found accurate by visual inspection of the results on a large number of patients.