Research project

Segmentation of Multi-Modal Volume Data
The conceptual formulations are the segmentation and registration of multi-modal volume data. Within the scope of the segmentation project, the goal is to classify small anatomical parts (e.g. thyroids) in 3D ultrasound (US) images. As the image quality of US usually is very bad due to speckle and noise contained in the image signals, this approach cannot succeed without appropriate pre-processing techniques. So far, various heuristic and numerical techniques for image enhancement have been covered, such as variants of anisotropic diffusion, Mumford-Shah approaches and morphological filters. Some of these filters have been extensively evaluated in the article by Kollorz et.al.: "Quantification of Thyroid Volume Using 3-D Ultrasound Imaging".
For the segmentation based on level sets, it is very important to conserve or enhance the edge information in the images. A quantitative analysis of different pre-processing filters resulted in a clear advantage of anisotropic, edge enhancing diffusion filters for this specific task.
In addition, alternative approaches to level sets have been explored. The Random Walk algorithm has been implemented for the segmentation of kidneys and cysts of patients with Autosomal Dominant Polycystic Kidney Disease (ADPKD). Approaches using prior knowledge of training shapes are currently evaluated for the segmentation of organs, such as kidneys. A corresponding article has been submitted and is currently under review by the journal board. Experiments showed that using active shape models helps to increase the robustness of automatic segmentation algorithms and reduces manual interactions.

Multi-Modal Image Registration
The information that is obtained by the segmentation is used in the second project: the multi-modal registration of CT (Computed Tomography), MRI (Magnetic Resonance Imaging), SPECT (Single Photon Emission Computed Tomography) and US. Compared to CT or SPECT, 3D US imaging is a relatively cheap acquisition method, which is increasingly used in the clinics today. Nowadays, the main field of application for this technique is situated in the field of prenatal diagnoses. However, it is not restricted to this application as the method is very suitable for the acquisition of morphologically small and bounded regions. The DFG proposal "Three-dimensional subtraction and processing of US for improved diagnoses of thyroid diseases with the focus on thyroid cancer" formulates a project that allows for a period of processing of three years.
Registration is a vital part in analyzing multi-modal images. Several registration algorithms are currently under development, both rigid and non-rigid. The rigid registration is focusing on speed, accuracy and robustness with respect to transforms that contain only rotations and translations. In order to increase the speed of state-of-the-art rigid registration approaches, a new projection based algorithm has been developed that allows for completely disjoint optimizations of the parameter dimensions. Non-rigid registration techniques have been implemented and successfully applied for the analysis of differences between inter- and intra-ictal SPECT images of Epilepsy patients.
In future, one focus of the project will be to incorporate segmentation results as prior knowledge within the registration algorithms. Especially for the non-rigid registration, this acts as an additional regularization for the highly ill-conditioned problem. Another important aspect will be the evaluation of the registration accuracy.
Very important for this project is the cooperation between the Institute of Pattern Recognition (Prof. Dr.-Ing. J. Hornegger) with clinics in Erlangen-Nuremberg: the Clinic for Nuclear Medicine with Policlinic (Prof. Dr. med. T. Kuwert) and the Special Ambulance for Prenatal Diagnostic with Ultrasound (Prof. Dr. med. R. Schild) of the Gynaecological Hospital of the University of Erlangen - Nuremberg (Prof. Dr. med. W. Beckmann).