Machine learning techniques have shown to be a viable means of analyzing medical images for tumor segmentation. Implementation of classi ers... Show moreMachine learning techniques have shown to be a viable means of analyzing medical images for tumor segmentation. Implementation of classi ers used for tu- mor segmentation requires training data constructed from known data sets. In most applications, obtaining the training dataset requires image registration. This thesis proposes a new combined image registration-segmentation framework to be applied to prostate tumor segmentation in MR images in the future. In order to construct training sets, known pathology from histological slides must be transferred to the MR images. Both manual and automated registration techniques have been previ- ously developed to accomplish this information transfer. Automated methods provide a reproducible, measurable result; however there is still room for improvement. The focus of this thesis is on improvement of the accuracy of the training data set through an iterative approach combining automated registration and segmenta- tion. The performance of the automated method of registration improves with an increase in the number of landmarks for matching. These extra landmarks used for registration are obtained from tumor segmentation; hence the name iterative segmentation-registration. This improvement leads to more accurate data transfer between histology and MR images and hence more accurate training data. Through numerical simulations, we show that by increasing the number of landmarks, using not only anatomical features but obtained segmentation results, the quality of the train- ing data can be improved. The proposed method results in a measurable increase in registration performance which leads to improved training data and consequently improved tumor segmentation performance. M.S. in Electrical Engineering, May 2012 Show less
