Sparsity-Driven Methods for Tracing of Tubular Structures in 3-D Confocal and SD-OCT Images: Application to Reconstruction of Astrocyte Arbors and Blood Vessels
Kulkarni, Prathamesh M
MetadataShow full item record
Tubular structures with complex morphologies occur frequently in biomedical research, often at a scale which necessitates their large-scale comprehensive and computational analysis. In this thesis, we focus on two such structures, astrocytes and blood vessels. The majority of the glial cells present in the brain are astrocytes which play critical roles in brain development, physiology and pathology. Astrocytes can be imaged as three-dimensional (3-D) objects in the brain tissue using fluorescence confocal microscopy. However, their vast number and the complexity of the patterns and mechanisms associated with their dynamics hinder an objective quantitative study. Therefore it is critical to develop automated computational methods facilitating comprehensive analysis of the astroglial networks. Congenital cardiovascular defects are one of the most common diseases responsible for infant mortality. These defects are closely associated with development of embryonic yolk sac vasculature. Recently developed imaging techniques such as optical coherence tomography (OCT) allow non-invasive imaging of embryonic structures including blood vessels. However, the reconstruction of blood vessels as such is a non-trivial task, mainly due to low signal-to-noise ratio (SNR). Therefore, it is critical to develop automated methods for longitudinal quantification of embryonic vascular networks as imaged using OCT. Understanding the commonalities between these two diverse problems, we propose to investigate the use of sparse representations for reconstruction of tubular structures from biological data. For astrocyte quantification, we propose a novel two-step approach. In the first step, we use a machine-learning method for detecting astrocyte root points while the second step is responsible for efficient tracing of astrocyte arbors. For OCT blood vessel reconstruction, we propose the use of anomaly detection from hyper-spectral imaging to handle the low SNR problem, followed by a smooth reconstruction using a parametric dictionary. Finally, we propose an integrated software framework for tracing, visualization, editing and feature-computation. To the best of our knowledge, this is the first reported comprehensive framework for reconstruction of astrocyte arbors from confocal data. The proposed approach includes a robust method for astrocyte nuclei detection which has not been effectively addressed by the prior work. Additionally, we propose the parallel arbor reconstruction algorithm which has been specifically designed to address the challenges involved in tracing astrocyte arbors. With regards to OCT blood vessel reconstruction, the application of anomaly detection improves upon the reconstruction quality compared to the prior methods such as speckle variance. Additionally, this work also introduces the application of sparsity-based methods for analysis of tubular biological objects. Results demonstrate that the proposed methods can facilitate efficient large-scale automated analysis of these important biological structures. Validation of the results provides convincing evidence to substantiate this claim. To this end, the error rate for the proposed reconstruction method was found to be 3.2%, compared to the fast-marching method (FMM) which had an error rate of 9%. With regards to OCT vessel reconstruction, the proposed method resulted in reconstructions with overall higher vesselness (0.8±0.09) compared to the standard speckle variance (SV) method which resulted in a vesselness of (0.6±0.2). The proposed methods, being integrated and distributed through FARSIGHT, an open source image analysis toolkit, can potentially have major contributions in two broad areas. Firstly, in advancing the state of the art understanding of the role of astrocytes in brain development, physiology and pathology and secondly, in advancing the longitudinal image-based quantification of embryonic yolk sac vascular development.