Statistical Segmentation and

Registration of Medical Ultrasound Data

The interpretation of ultrasonic imagery is typically not straightforward and of quite subjective nature 
and therefore strongly dependent on the expertise of its users. Thus the development of algorithms 
which aid in the interpretation of ultrasonic data is a highly relevant topic. This thesis examines aspects 
of segmentation and registration of ultrasonic data, utilizing the fact that the ultrasound signal can be 
modeled statistically. The object of segmentation is the endocardium in the left-ventricular long-axis 
view of the human heart in clinical B-mode ultrasound (US) image sequences, while similarity measures 
for feature descriptors and registration are applied to the envelope-detected radio frequency US data 
of the human neck and brain. Locally and globally optimal variational active contour methods and a 
Bayesian Markov Chain Monte Carlo sampling method are applied to the segmentation problem, utilizing
prior formulations for shape and regularization. A feature descriptor is proposed which combines global 
data statistics, by a maximum-likelihood-estimated distribution, with local pattern characteristics, 
employing Markov Random Field interaction parameters. For registration we propose two approaches. 
Firstly, a hybrid procedure incorporating global statistics, by Hellinger distance between distribution in 
images, and local textural features by a statistics-based extension of Fuzzy Local Binary Patterns. 
Secondly, we explore the registration of 3D freehand US data, where view dependency of ultrasound 
is addressed by modeling speckle statistics, using a finite mixture model. The proposed methods for 
segmentation, feature description and registration are evaluated through experiments and/or 
comparative experiments to state-of-the-art models.