Automated Feature Detection in Multidimensional images

Manual identification of structures and features in multidimensional images is at best time consuming and operator dependent. Feature identification need to be accurate, repeatable and quantitative. This thesis presents novel methods for automated feature detection in multidimensional images that are independent on imaging modality. Feature detection is described at two abstraction levels. At the first low level the image is regionally processed to find local or regional features. In the second medium level results are taken from the low level feature detection and grouped into objects or parts that can be quantified. A key to quantification of cardiac function is delineation of the cardiac walls which is a difficult task. Two different methods are described and evaluated for delineation of the left ventricular wall from anatomical images. The results show that semi-automatic delineation is a huge time saver compared to manual delineation. To obtain a robust results as much a priori and image information as possible should be used in the delineation process. Regional cardiac wall function is further studied by deriving and analyzing strain-rate tensors from velocity encoded images. For flow encoded images novel methods to find regional flow structures such as vortex cores, flow based delineation, and flow quantification are proposed. These methods are applied to study blood flow in the human heart, but the techniques outlined are general and can be applied to a wide array of flow conditions.