| Περίληψη: | Ultrasound is a well-established medical imaging modality competing more advanced methods such as MRI
and CT, while the recent advancements in super-resolution imaging have increased its usefulness even more. Super
resolution ultrasound is frequently used in tumor detection applications, since micro-bubble tracking provides
information of tissue perfusion and local haemodynamic characteristics. The particles are inserted into the blood
stream via an intravenous injection and due to their mm dimensions they can reach even the smallest capillaries.
Usually, they are fabricated with an gaseous inner layer and an outer shell composed by various materials such as
lipids, polymers and proteins. Replacing the inner layer with chemotherapeutic drugs, targeted tumor treatment
is possible.
The super-resolution imaging approach is based upon the tracking of the particles as they travel in the bloodstream.
To achieve that a two-step algorithm is employed which includes the segmentation and localization of
the bubble and its tracing in the temporal domain. Since in the tracking the particle is considered as a single point
object, replaced by its center of mass, the accuracy of the segmentation and localization determines the success
of the entire algorithm. The current method includes a simple averaging filter to remove any artefacts from the
image and through a probabilistic approach distinguish the pixels to those that belong to a particle and those that
belong to the background.
This thesis addresses some of the misconceptions of the already developed algorithm and presents a comparative
study between the above segmentation and two new approaches. Firstly, the averaging filter is replaced
with more proper enhancing methods such as adaptive wiener filter, non-local means or bilateral filter, replacing a
simple smoothing filter with edge-preserving filters. The analysis of the original algorithm revealed an ill-defined
step using the Haar-Like features, which in the new approaches is removed, while an entirely new method based
on the maximization of the cross correlation between an investigated bubble and a theoretical template is used.
After describing thoroughly the principles of each filter and each method, this study presents a detailed
evaluation of each proposed algorithm via a synthetic data set and several feature maps produced from real US
videos. The utilization of synthetic data is critical since they provide information of ground truth events and thus
measurements of wrong or missed detections can be made. However, due to specific limitations of the synthetic
data. an extensive assessment on real data, on both sheep ovaries and patients with prostate cancer, is necessary.
Considering the same linking algorithm, any difference in the feature maps is attributed to the segmentation
algorithm.
|
|---|
