Analysis of Medical Images : Registration, Segmentation and Classification

Abstract: A large number of medical examinations involve images in some way. Images can be used for diagnostics, follow-up studies and treatment planning. In this thesis mathematical methods have been developed and adapted in order to analyze medical images. Several applications for different imaging modalities have been studied and the usefulness of such methods is demonstrated.A complete system for detection and diagnosis of kidney lesions in scintigraphy images has been developed. We segment the kidneys with the use of an active shape model. The uptake of a biological molecule is then compared to the uptake in a healthy kidney and potential lesions are detected. A number of properties of the potential lesions are gathered and the lesions are classified as healthy or unhealthy with a linear classifier. We are able to correctly classify 86 % of the lesions.Ultrasound images have also been studied. In the first case for the purpose of segmenting the left heart ventricle, which can be used for computing the ejection fraction. This was done using a region based snake with anchor points at each side of the cardiac valve. The second application in ultrasound images is also of the heart but with patients that, due to heart failure, have had a mechanical pump implanted. The septum wall between the ventricles is segmented using a shortest path approach and a measure of how much the septum bulges towards either of the ventricles is obtained. By studying this measure a more objective indication is given on whether the speed of the pump is correct for a patient than by only visually study the images.In computed tomography (CT) whole-body images, several organs have been segmented using a multi-atlas approach. The fused labels are refined with a random forest classifier and a final graph cut segmentation. This method was evaluated in the VISCERAL Grand Anatomy Challenge and achieved the highest Dice score for 13 out of 20 organs. A development of this approach was done in order to achieve qualitatively better segmentations of the organs. Instead of fusing organ labels, a map of corresponding landmarks is obtained and the segmentation is given by the robust average of these with similar refinement steps as in the origin work. The segmentation results using this method is on par with or better than state-of-the-art. Segmentation of organs is important in e.g. radiotherapy planning.In another project with CT images, vertebrae have been detected and identified. This is useful in for instance surgical planning. The detection is done using convolutional neural networks. A shape model of the spine is fitted to the detections in order to correctly identify them. The task is difficult because, in general, only a limited part of the spine is visible. We are able to correctly identify 63 % of the vertebrae.