Dynamic Programming for Ultrasound Image Analysis of Atherosclerosis
Cardiovascular diseases, due to atherosclerosis, are among the most common causes of death in the industrialized world. Atherosclerotic plaque composition as well as arterial diameter and wall thickness can be used to predict the risk of cardiovascular disease. These risk factors can be assessed non-invasively by ultrasound.
Both diameter and wall thickness measurements require detections of boundaries between tissue layers in the blood vessel. Dynamic programming has previously been used for automatic boundary detection in ultrasound artery images. In this thesis, an extension to a previously reported dynamic programming procedure applied to ultrasound image sequences is proposed. The new method is a two step procedure that takes advantage of the correlation between frames in a sequence, i.e. the location of a boundary should be at approximately the same position in two consecutive frames. First, several "candidate boundaries" are detected in each frame. Then, one boundary from each frame in the sequence is selected so that a certain cost function is minimized. This function includes the cost of individual boundaries as well as a movement cost. The method is suboptimal in the sense that only a small subset of all possible boundaries are considered in step two, but among these candidates the detected boundary sequence is optimal.
In order to evaluate and compare, another method based on graph cuts is also presented. The previously mentioned dynamic programming procedures and the graph cuts method are tested on both synthetic data and real ultrasound image sequences. The result is that the new dynamic programming procedure outperforms the previous one based on individual frame by frame detection. Also, it is shown to have similar detection performance at a significantly lower computational complexity than the graph cut method.
This thesis also describes an automated method for classification of ultrasound carotid plaque images as being echolucent or echogenic. To compensate for variability among the images, the method uses an adaptive thresholding technique. The results show a good agrement between the automated method and visual Gray-Weale classification.