* Faculty       * Staff       * Contact       * Institute Directory
* Research Groups      
* Undergraduate       * Graduate       * Institute Admissions: Undergraduate | Graduate      
* Events       * Institute Events      
* Lab Manual       * Institute Computing      
No Menu Selected

* Research

Ph.D. Theses

Computer Vision Algorithms for Retinal Vessel Detection and Width Change Detection

By Kenneth Fritzsche
Advisor: Charles V. Stewart, Badrinath Roysam
April 8, 2004

Detection of width changes in blood vessels of the retina may be indicative of eye or systemic disease. However, fundus images taken at different times have different scales and are difficult to compare side by side. This research presents automated techniques for detection of vessel width change from two images acquired at different points in time.

In order to detect vessel width change, vessels must first be identified in available images. This research starts by making numerous improvements to an existing vessel tracing algorithm. However, even the improved results exhibit too much variance, primarily attributable to the discrete nature of the tracing algorithm. Thus new methods for estimating vessels are explored to address the limitations of the vessel tracing algorithm. These methods are designed to provide smooth, continuous boundaries. Five ribbon-like objects are put into an active contour framework, all initialized using the results from tracing. These ribbons are shown to provide more repeatable vessel boundaries than tracing with an innovative technique named in this research as cross section snakes selected as best.

In addition to estimating vessels from a single image, a technique is explored that uses information from multiple images acquired in a single sitting to estimate vessel boundaries in a single image with greater accuracy. This technique is subjectively evaluated and is chosen at a rate that is double the sum of two other single image estimation techniques.

In detecting change, two approaches are tried. One that alternately tests vessel boundaries in the other image before determining change and one that uses a statistical hypothesis test as the basis for the determination of change. The first method is the better of the two based on three criteria. First, method 1 achieves a positive predictive power of 78% and method 2, 57%. Second, method 1 correctly identifies 15 out of 20 vessel segments as changed compared with only 7 in method 2. Finally, method 1 is shown to be superior at correctly identifying no change (i.e. less false positives) over method 2.

* Return to main PhD Theses page