Tenbrinck D (2013)
Publication Type: Thesis
Publication year: 2013
This thesis is focused on variational methods for fully-automatic processing and analysis
of medical ultrasound images. In particular, the e↵ect of appropriate data modeling in
the presence of non-Gaussian noise is investigated for typical computer vision tasks.
Novel methods for segmentation and motion estimation of medical ultrasound images
are developed and evaluated qualitatively and quantitatively on both synthetic and real
The first part of the thesis is dedicated to the problem of low-level segmentation. Two
di↵erent segmentation concepts are introduced. On the one hand, segmentation is formulated
as a statistically motivated inverse problem based on Bayesian modeling. Using
recent results from global convex relaxation, a variational region-based segmentation
framework is proposed. This framework generalizes popular approaches from the literature
and o↵ers great flexibility for segmentation of medical images. On the other hand,
the concept of level set methods is elaborated to perform segmentation based on the
results of a discriminant analysis of medical ultrasound images. The proposed method
is compared to the popular Chan-Vese segmentation method.
In the second part of the thesis, the concept of shape modeling and shape analysis is
described to perform high-level segmentation of medical ultrasound images. Motivated
by structural artifacts in the data, e.g., shadowing e↵ects, the latter two segmentation
methods are extended by a shape prior based on Legendre moments. Ecient
schemes for encoding and reconstruction of shapes are discussed and the proposed highlevel
segmentation methods are compared to their respective low-level variants.
The last part of the thesis deals with the challenge of motion estimation in medical
ultrasound imaging. A broad overview on optical flow methods is given and typical
assumptions and models are discussed. The inapplicability of the popular intensity constancy
constraint is shown for the special case of images perturbed by multiplicative
noise both mathematically and experimentally. Based on the idea of modeling image intensities
as random variables, a novel data constraint based on local statistics is proposed
and their validity is proven. The incorporation of this constraint into a variational model
for optical flow estimation leads to a novel method which outperforms state-of-the-art
methods from the literature on medical ultrasound images.
This thesis aims to give a balanced view on the di↵erent stages involved in solving
computer vision tasks in medical imaging: Starting from modeling problems, to their
analysis and ecient
numerical realization, to their final application and adaption to
real world conditions.
Tenbrinck, D. (2013). Variational Methods for Medical Ultrasound Imaging (Dissertation).
Tenbrinck, Daniel. Variational Methods for Medical Ultrasound Imaging. Dissertation, 2013.