Dana Brooks, Jennifer Dy, Armen Stepanyants
Date of Award
Doctor of Philosophy
Department or Academic Unit
College of Engineering, Department of Electrical and Computer Engineering
electrical engineering, bioinformatics, medical imaging and radiology, diadem, image processing, kernel regression/interpolation, manifold clustering/deviation/extraction, morphological reconstruction, pattern recognition
Electrical and Computer Engineering
Robust techniques for structural analysis of curvilinear objects are essential in various biomedical image processing applications. Conventional approaches in the literature rely on an accurate initial segmentation of the data and problem specific formulations with user defined meta-parameters to estimate the morphology of the curvilinear structures. Outliers and variations in the data make such methods ill-suited for generalization to a broader range of applications. In that regard, this thesis provides novel tools to extract accurate structural information, i.e. morphology, from complex curvilinear networks of objects by providing necessary mathematical background and implementation details.
We start with a shape driven approach that builds a compartment model using the color information, where compartments are constrained to the edges of tubular structures in space. Similar to the conventional flux based tubular filtering systems, the proposed approach spans the structure by fitting the compartment model in an iterative fashion. Color information is incorporated to the local optimization procedure that enables us to use the method in dense regions where multiple curvilinear objects propagate together as a fiber bundle. Qualitative and quantitative results obtained from synthetic and Brainbow confocal images are reported to evaluate the performance of the approach.
In general, biological curvilinear structures have arbitrary crosssections and the data is not standardized, having variations in the appearance and shape. Moreover, the morphology estimation process is hampered by the presence of noise or artifacts and is not trivial to accomplish with model-based approaches. In order to tackle these problems, a geometric approach that is based on the definition of principal curves and subspace local maxima using first two spatial derivatives of a distribution function is provided. We call this method recursive principal curve tracing (RPCT) approach, which was a finalist in the 2010 Diadem Challenge. Quantitative results obtained from the simulations and experiments conducted on Brainbow confocal microscopy images validated the effectiveness of the approach. The generalization capability of the approach for connectivity analysis is verified on neuronal arbor datasets from confocal and brightfield microscopy images, retinal scans, and computer tomography (CT) images. We reported the Diadem measure for the neuronal arbor reconstructions obtained from the confocal images, where groundtruth validation data has been made available.
Last but not least, we present a novel approach, principal spanning forest (PSF), based on a new, purely data driven measure that becomes zero between pairs of points on the same principal curve segment - this will allow us to extract morphology reliably by identifying samples from the ridge of an intensity distribution and then determining their connectivity using an affinity measure based on this measure. The proposed affinity measure will be useful for (i) clustering curvilinear manifold segments with smooth shapes (ii) downsample a selected principal curve segment with a preset accuracy level sparsely. We will demonstrate these concepts in neural morphology extraction and approximation and validate the approach on confocal microscopy and CT images.
Baş, Erhan, "Extracting structural information on manifolds from high dimensional data and connectivity analysis of curvilinear structures in 3D biomedical images" (2011). Electrical Engineering Dissertations. Paper 41. http://hdl.handle.net/2047/d20002052
Click button above to open, or right-click to save.