Organ motion occuring during acquisition of medical images can cause motion blur artifacts, thus posing a major problem for many commonly employed modalities. Therefore, compensating for that motion during image reconstruction has been a focus of research for several years. However, objectively comparing the quality of different motion compensated reconstructions is no easy task. Often, intensity profiles across image edges are utilized to compare their sharpness. Manually positioning such a profile line is highly subjective and prone to bias. Expanding on this notion, we propose a robust, semi-automatic scheme for comparing edge sharpness using an ensemble of profiles. We study the behavior of our approach, which was implemented as an open-source tool, for synthetic data in the presence of noise and artifacts and demonstrate its practical use in respiratory motion-compensated MRI as well as cardiac motion-compensated C-arm CT.

In the field of image-guided surgery, Time-of-Flight (ToF) sensors are of interest due to their fast acquisition of 3-D surfaces. However, the poor signal-to-noise ratio and low spatial resolution of today's ToF sensors require preprocessing of the acquired range data. Super-resolution is a technique for image restoration and resolution enhancement by utilizing information from successive raw frames of an image sequence. We propose a super-resolution framework using the graphics processing unit. Our framework enables interactive frame rates, computing an upsampled image from 10 noisy frames of 200×200 px with an upsampling factor of 2 in 109 ms. The root-mean-square error of the super-resolved surface with respect to ground truth data is improved by more than 20% relative to a single raw frame.

In the course of this project, we developed a library for non-linear optimization using L-BFGS (a Quasi-Newton method) on the GPU using CUDA. The super-resolution framework that builds on it is also freely available.

Simple algorithms for segmenting healthy lung parenchyma in CT are unable to deal with high density tissue common in pulmonary diseases. To overcome this problem, we propose a multi-stage learning-based approach that combines anatomical information to predict an initialization of a statistical shape model of the lungs. The initialization first detects the carina of the trachea, and uses this to detect a set of automatically selected stable landmarks on regions near the lung (e.g., ribs, spine). These landmarks are used to align the shape model, which is then refined through boundary detection to obtain fine-grained segmentation. Robustness is obtained through hierarchical use of discriminative classifiers that are trained on a range of manually annotated data of diseased and healthy lungs. We demonstrate fast detection (35s per volume on average) and segmentation of 2 mm accuracy on challenging data.