Assessing relative performance of edge detection algorithms

In the field of medical imaging there is a growing need for automatic procedures for processing visual recordings. Image segmentation based on edge detection represents one of the mainly used techniques of analysis and has an important role in medical diagnostics, researches and in treatment planning. Edge detection procedures must be as reliable and accurate as possible. This represents a problem since medical images are mostly unclear and contain noise which causes difficulties in detecting true contours of objects. In this thesis we compared two edge detection algorithms, the Canny edge detector and the Marr-Hildreth edge detector, which are based on the use of the first and second derivative. Our first task was to implement the algorithms, then we compared them in the sense of performance. The testing dataset was composed from computed tomography (CT) and magnetic resonance imaging (MRI) images that are a part of the CTMRI DB database. We used qualitative and quantitative approach to compare the performance of both detectors. First we compared detection results visually, based on a subjective assessment of the presence of the edges. Quantitative comparison consisted of generating a reference image which was then compared to the detection results. For this purpose we developed a procedure for reference image generation and established new performance metrics for comparison of performance of two detectors. The developed performance metrics can be generalized to any number of edge detectors and offer new opportunities in the field of algorithm performance evaluation