Segmentation and Enhancement of Filamentous Objects for Biological Image Analysis

The neurome is an atlas of neurons for a given organism that includes description of individual neuron morphology and of neuron variability. Construction of such an atlas will be critical to understanding the complex neural system of an organism, eventually providing clues to how animals think and function. As the organisms under investigation scale from the worm to the human, the number of neurons range from hundreds to trillions.

Image analysis is a key enabling tool for building the neurome for complex organisms. From an image processing engineer’s perspective, an essential aspect of this study is development of segmentation algorithms that obtain an automated digital reconstruction of the neural morphology. The primary objective of this thesis is to develop novel image analysis methods for
creating a database of neurons of an animal. We investigate automated algorithms for segmenting single neuron cells from noisy and cluttered confocal microscopy images. The end goal is to embed the shape and geometry of the segmented neurons in a mathematically representable, graph theoretic tree. This provides a digital reconstruction that could be used for further neurological studies such as shape comparison, structural categorization and disease identification.

In this regard, we propose three automated segmentation algorithms. The first method, Tree2Tree-2 is a graph based tracer that uses graph theory in conjunction with image analysis to perform tracing. Tree2Tree-2 uses a graph based geometry analysis procedure to create a global neuron tree from a set of disconnected objects. Tree2Tree-2 improves on its predecessor Tree2Tree by establishing a variational formulation for initial segmentation and a shortest path based path refinement policy to accurately link the neuron branches.
Experimental results suggest improvement over Tree2Tree, where the average tracing error is reduced by 40% (on average) by virtue of precise inter-branch path assessment and robust initial segmentation..

While the above mentioned solution is effective for relatively simple structures, we hypothesize that a more efficient solution can be achieved by introducing a methodology which is adaptive to the object topology. One approach to solve this problem is by using geometric deformable models which are topology-adaptive. This motivates our next solution, Legendre Level Sets (L2S), which is designed to perform region based segmentation of 2D images in presence of heterogeneous intensity. It extends the Mumford-Shah model
to applications where object and background illumination is inconsistent, and the contrast is poor. Although L2S is not specifically designed for tracing filamentous objects only, we have observed encouraging performance in tracing 2D neurites. L2S improves segmentation performance on a broad category of 2D images, with more than 28% improvement in average Dice score over contemporary region based techniques Finally, a third algorithm is discussed that leverages the clutter rejection capability of Tree2Tree-2 and the topology-adaptive property of L2S. While L2S is a generalized method, Tubularity Flow Field (TuFF) is geometric deformable model which is specifically designed for segmenting filamentous objects. TuFF handles discontinuities in the
filament appearance in a natural framework, by using a local attractive force in the geometric set up, and is more robust against low contrast and structural complexity.

The efficacy of TuFF over contemporary automated and semi-automated neuron tracers is demonstrated via 2D and 3D segmentation results, with significant improvement in quantitative performance (> 70%) as well as qualitative tracing evaluation. To improve robustness to contour initialization and extend its general applicability, two improvements to the TuFF framework are proposed. First, we design a novel 2D enhancement procedure
for filament enhancement. The proposed algorithm, Local Directional Evidence (LDE), uses multiple ridge detector filters to incorporate evidence of filamentous objects in a local neighborhood of a point, which allows it to identify structures when object brightness is inconsistent. Second, we integrate an edge dependent term in TuFF’s mathematical formulation to generalize its use in cases where the object edges are prominent. The Edge Assisted TuFF (EATuFF) model is used in conjunction with the LDE filter for segmentation applications in medical imaging (blood vessel tracing) and civil engineering
(automated detection of cracks on concretes).