Faculty of Informatics Vienna University of Technology Institute of Computer Aided Automation PRIP Home PRIP Home
Personal tools
You are here: Home Teaching PRIP-Talks Lukas Fischer: Optimizing Shape Particle Filters for the Detection and Segmentation of Medical Images
Navigation
 

Lukas Fischer: Optimizing Shape Particle Filters for the Detection and Segmentation of Medical Images

Abschlussvortrag DA

What
  • Presentation
When Mar 15, 2010
from 04:15 pm to 04:25 pm
Where Sem 183/2
Add event to calendar vCal
iCal
In recent years segmentation approaches based on sequential Monte Carlo Methods delivered promising results for the localization and delineation of anatomical structures in medical images. Also known as Shape Particle Filters, they are used for the segmentation of human vertebrae, lungs and hearts, being especially well suited to cope with the high levels of noise encountered in MR data and overlapping structures with ambiguous appearance in radiographs. Shape Particle Filters rely on a region template or map based on a shape models' mean shape, which is defined manually in existing approaches. During search, a classification step based on appearance features yields the probabilities for each pixel to belong to a certain region within the template. This forms the basis for the actual segmentation process.

This thesis aims to optimize Shape Particle Filters in terms of computational performance as well as segmentation accuracy. Two novel approaches for the generation of the region map are proposed, namely automatic region maps and per-pixel region maps. The automatic region map approach, where the optimal distribution and number of template regions is derived from a set of training images, adapts to complex data and finds consistent features in the training examples without manual interaction. Using appearance features based on the Monogenic Signal the per-pixel region map approach eliminates both the need for the region estimation as well as the classification step, resulting in considerably faster segmentation while retaining the same level of accuracy.

The proposed methods are evaluated on four different data sets, synthetic rectangles, metacarpal bone radiographs, MRI slices of the heart and CT slices of the lung. Experimental results show a major gain in computational performance as well as better or at least equal segmentation results when compared to current approaches.
Document Actions