Supervisors: Firat Özdemir, Ece Özkan, Dr. Christine Tanner, Prof. Dr. Orçun Göksel
Finding spatial correspondences in a set of images is an important topic in medical imaging, as they can be used for image registration, motion tracking and many other applications. Point correspondences are often annotated manually by specialists which is a time consuming task. Automatic detection of point correspondences is preferred as it is more scalable to large datasets, providing opportunities for data analysis techniques which rely on large amounts of data. In this project a method for both automatic detection and matching of anatomical landmarks in medical images was developed. The Rohr corner detector in combination with the Histograms of Oriented Gradients descriptor was used to generate keypoints and corresponding descriptors, after initial comparisons with other approaches. The use of a greedy matching method that assigns matches individually as well as a global matching method that finds a non-conflicting global solution was investigated. The performance was tested on non-linearly transformed 3D shoulder MR images and compared to the nSIFT algorithm. Also linear regression was used to predict the error of each candidate match. This was used to extract a confidence score for each match and select the best 10 matches according.