Nov 10, 2016
Jolley Hall, Room 309
"Guided Medical Data Segmentation Using Structure-Aligned Planar Contours"
Advisor: Tao Ju
Segmentation of 3D/4D biological images is a critical step for a wide range of applications such as treatment planning, quantitative analysis, virtual simulations, and rendering visualizations. Automatic segmentation methods are becoming more reliable, but many experts still rely on manual intervention which makes segmentation a time and resource intensive bottleneck. Marking boundary contours in 3D images can be difficult when images are often noisy or the delineation of biological tissue is unclear. Non-parallel contours can be more accurate and reduce the amount of marking necessary, but require extra effort to ensure boundary consistency and maintain spatial orientation. This dissertation focuses three problems that pertain to drawing non-parallel contour networks and generating a segmentation surface from those networks.
First a guided structure-aligned segmentation system is detailed that utilizes prior structure knowledge from past segmentations of similar data. It employs a contouring protocol to aid in navigating the volume data and support using arbitrarily-oriented contouring planes placed to capture or follow the global structure shape. A user study is provided to test how well novices perform segmentation using this system. The following two problems then aim to improve different aspects of this system. A new deformation approach to reconstruction is discussed which deforms previous segmentation meshes to fit protocol drawn contours from new data instances in order to obtain accurate segmentations that have the correct topology and general shape and preserves fine details. The focus is on the problem of finding a correspondence between a mesh and a set of contours describing a similar shape. And finally, a new robust algorithm that resolves inconsistencies in contour networks is detailed. Inconsistent contours are faster and less demanding to draw, and they allow the segmenter to focus on drawing boundaries and not maintaining consistency. However, inconsistency is detrimental to most reconstruction algorithms, so the network must be fixed as a post process after drawing.