Doctoral Defense by Athanasios Karamalis

Date: 25.06.2013 Time: 16:00 Location: Room 00.12.019, FMI-Building, Garching, Boltzmannstr. 3

Ultrasound Confidence Maps and Applications in Medical Image Processing

Short Abstract:
This dissertation introduces a novel method for confidence estimation from ultrasound images, denoted as ultrasound confidence map. The estimation problem is formulated as a random walk on a graph, taking into account ultrasound specific constrains. The method was successfully introduced into different ultrasound image processing applications and improved their results.

Full Abstract:
Advancements in medical ultrasound imaging have been driven by the continuous improvement of imaging systems and novel approaches for image processing applications. The modeling and simulation of ultrasound contributes to our understanding of complex sound propagation in human tissue and also to the improvement and development of ultrasound systems. Furthermore, modeling the ultrasound image formation process provides valuable prior information for image processing algorithms.

In terms of simulation this thesis contributes a new framework for fast and realistic ultrasound image simulation. The nonlinear full-wave equation is used for the modeling of sound propagation and solved with a Finite-Difference Time-Domain scheme on GPU hardware to reduce simulation time. The essential image formation pipeline is simulated including pulse transmission, beamforming and signal processing. Realistic images are simulated from computer phantoms and phantoms generated from histology images. In the later case histology and intravascular ultrasound images were acquired from the same cross-section of vessels. Consequently, simulated images based on histology could be compared to real images in the same reference coordinate system.

The main contribution of this thesis is a novel method for estimating confidence in ultrasound images, denoted ultrasound confidence maps. The novelty of the method lies in the formulation of the confidence estimation problem as a random walk on a graph, taking into account essential ultrasound physics and the image formation process. The fast and closed-form solution to this problem comes with an inherent robustness to image acquisition parameters, which is demonstrated by obtaining consisting confidence maps for images acquired with different transducers, systems and settings. In addition, the benefit of using confidence maps for image processing applications is demonstrated by introducing them into algorithms for shadow detection, 3D freehand reconstruction, classification in intravascular ultrasound and mono- and multi-modal registration. In all cases algorithms benefited from confidence map information and produced improved results.

In the process of this thesis an ultrasound-guided navigation system was developed for computer-aided orthopedic surgery. For this, a novel feature descriptor is introduced for automatic bone detection in ultrasound, based on the previously introduced confidence maps. The system performed automatic bone surface detection in ultrasound images in real-time and subsequent rigid registration to the bone surface extracted from pre-operative Computer Tomography data set. This would allow fusion of pre-operative imaging and planning data with intra-operative imaging by bringing this information into the same reference coordinate system. The algorithms and system were evaluated on data sets acquired in-vivo and in a cadaver study, demonstrating high accuracy, robustness and the potential for intra-operative application.