Associative image analysis: A method for automated quantification of 3D multi-parameter images of brain tissue

https://doi.org/10.1016/j.jneumeth.2007.12.024Get rights and content

Abstract

Brain structural complexity has confounded prior efforts to extract quantitative image-based measurements. We present a systematic ‘divide and conquer’ methodology for analyzing three-dimensional (3D) multi-parameter images of brain tissue to delineate and classify key structures, and compute quantitative associations among them. To demonstrate the method, thick (∼100 μm) slices of rat brain tissue were labeled using three to five fluorescent signals, and imaged using spectral confocal microscopy and unmixing algorithms. Automated 3D segmentation and tracing algorithms were used to delineate cell nuclei, vasculature, and cell processes. From these segmentations, a set of 23 intrinsic and 8 associative image-based measurements was computed for each cell. These features were used to classify astrocytes, microglia, neurons, and endothelial cells. Associations among cells and between cells and vasculature were computed and represented as graphical networks to enable further analysis. The automated results were validated using a graphical interface that permits investigator inspection and corrective editing of each cell in 3D. Nuclear counting accuracy was >89%, and cell classification accuracy ranged from 81 to 92% depending on cell type. We present a software system named FARSIGHT implementing our methodology. Its output is a detailed XML file containing measurements that may be used for diverse quantitative hypothesis-driven and exploratory studies of the central nervous system.

Introduction

Brain tissue has a complex three-dimensional (3D) architecture containing multiple neuronal and glial cell types, immune system cells, and vascular elements (Bear, 2006). This complexity presents a fundamental challenge to studies that require quantitative image-based measurements at the tissue scale. Examples of such studies include quantifying the location and distribution of cells in the neural stem-cell niche (Fuchs et al., 2004) and the cancer stem cell microenvironment (Calabrese et al., 2007), cell relationships associated with the blood–brain barrier (Iadecola, 2004), and changes in cell associations with injury or disease (Spataro et al., 2005).

Our methodology for describing the cellular organization in the brain is based on 3D spectral confocal microscopy, and computational image analysis (Pawley, 2006). By careful choice of fluorescent probes, it is now possible to image multiple molecular species simultaneously while preserving their spatial relationships. The advent of spectrally resolved confocal microscopes, such as the Zeiss LSM META, allows a 32-point emission spectrum to be recorded at each point in the image. These data can now be processed with linear unmixing software provided by the microscope manufacturer to compute a smaller number of essentially ‘pure’ channels containing negligible crosstalk from other fluorophores (Dickinson et al., 2001, Garini et al., 2006, Zimmerman, 2005). In summary, it is now possible to acquire (i) sufficiently large 3D images to permit extensive measurements of cell-to-cell relationships and (ii) sufficiently large number of imaging channels to provide independent labeling of all the components under study (Lin et al., 2005a, Seymour and Kipke, 2007, Spataro et al., 2005). In this report, we present 3D images with five channels corresponding to cell nuclei, neurons, microglia, astrocytes, and blood vessels (see Fig. 1).

Once such large 3D multi-channel images of brain tissue are acquired, there is a need to extract meaningful quantitative measurements. Traditionally, this is performed by computer-assisted manual analysis, relying on human pattern recognition and computerized data recording using software tools such as Metamorph (Molecular devices, Sunnyvale, CA), Volocity (Improvision Inc., Waltham, MA), ImageJ (NIH), and Neurolucida (MBF Biosciences). Among manual methods, two important approaches exist: object-based, and fluorescence intensity-based. An example of the former is stereology, as implemented in tools like Stereo Investigator (MBF Biosciences, Williston, VT). In this method, the user counts and/or delineates a random subset of objects guided by unbiased stereological principles. Fluorescence intensity-based methods utilize the overall signal intensity for quantification without explicitly delineating structures (e.g., Biran et al., 2007, Zhong and Bellamkonda, 2007). Stereological methods suffer from three major limitations: (i) extensive manual sampling of objects is required to reduce the variance, obviating the advantages of sub-sampling; (ii) they assume that the tissue is homogeneous, but the brain histology is not homogeneous; and (iii) they are inherently limited in their ability to cope with multi-dimensional data. Fluorescence intensity-based methods suffer from lack of structural information, and the fact that the fluorescence signal is an unreliable indicator of molecular concentration since it is affected by specimen preparation, and imaging artifacts. Overall, manual methods suffer from slowness, cost, tedium, subjectivity, lack of steadiness in tracing structures, limited attention span, limitations of the human visual system that can only perform binocular stereo viewing rather than volumetric viewing, and that scoring methods are limited to the two-dimensional (2D) computer screen. Finally, manual methods are inherently limited in their ability to analyze associations among structures—their greatest limitation.

Automated algorithms have been developed to delineate individual elements of neuroanatomy, e.g., counting of nuclei, tracing neuronal processes, branching pattern of dendrites for a single neuron, and vascular tracing (Al-Kofahi et al., 2003, Al-Kofahi et al., 2008, Fernandez-Gonzalez et al., 2002, Lin et al., 2005a, Lin et al., 2007, Tyrrell et al., 2007, Weaver et al., 2004, Zhang et al., 2007). These tools are able to delineate a specific class of structures from single channel images with error rates in the range of 5–10%, and have progressed greatly in their ability to handle morphological variability. Since 100% automation is beyond the current state of the art, we accept these tools to be automated, and introduce an efficient inspection and editing method that can allow the investigator to further reduce the error as necessary. Next, considering that brain tissue has a diverse architecture with complex networks of relationships, two major needs exist. First, there is a need to apply multiple segmentation tools simultaneously to analyze multi-parameter datasets. Second, there is a need to develop comprehensive and broadly applicable methods to relate the measurements from multiple segmentations in a biologically meaningful manner.

We propose a ‘divide and conquer’ strategy to tackle the complexity of multi-channel, 3D image data (see Fig. 2). This strategy uses commercially available linear unmixing software to divide the 32-channel spectral image data into a small number of non-overlapping channels (for a thorough discussion of this method see Dickinson et al., 2001, Garini et al., 2006, Zimmerman, 2005). Each channel has a much lower morphological diversity compared to the overall tissue, making it feasible to segment (delineate) the structures in each channel using separate previously developed algorithms, each specialized to one morphological category. Once the main structures in each channel are segmented, inspected, and edited to achieve acceptable accuracy, a rich set of measurements can be computed.

We consider two classes of measurements: intrinsic and associative. Intrinsic measurements quantify aspects of objects within an individual channel, including the spatial locations, sizes, morphologies of cellular compartments, lengths and branching patterns of processes, width variations of vessels, and areas of membranous surfaces. Methods for computing intrinsic measurements are widely described (see Sonka et al., 2008). Associative measurements a central contribution of this work. They quantify relationships among two or more structures identified by segmentation. The importance of associative measurements cannot be overstated—they are essential to developing a system-level understanding of brain tissue. The objects being associated may arise from a single channel (e.g., two cell nuclei), or different channels (e.g., nuclei and vessels). Precursors to our concept of associative measurements include 3D-catFISH for quantifying immediate-early gene transcription (Guzowski et al., 2005), and 3D quantification of the neurovascular unit (Lin et al., 2005a). Given the large number of constituents of brain tissue, a combinatorial number of associations are possible. However, a much smaller set of associations are biologically relevant, and immediately useful. In this paper, we focus on associations that are based on the broadly relevant notions of spatial proximity and adjacency. Other important associations, such as synaptic connectivity and neuronal networks, are not addressed here.

Once a set of associative measurements is computed, an important issue is how to represent them. Traditional two-dimensional representations such as tables and charts are often inadequate to capture the information in the network of associative measurements describing the 3D multi-channel data. We have adopted attributed graphs as an effective representation due to their many advantages. First, they are supported by graph theory, a well-studied mathematical discipline that is specifically intended for abstract analysis of relationships (Gross and Yellen, 2005). Second, general-purpose software tools are available for visualizing and analyzing graphs (e.g., Graphviz, www.graphviz.org). An attributed graph is composed of nodes representing objects, and links representing associations among objects. Attached to each node is a list of intrinsic measurements (attributes) of the object, and attached to each link is a list of associative measurements quantifying the relationship. In our work, the graph data structure is represented as an extensible markup language (XML) file. Once an attributed graph is constructed, it can be queried to answer specific questions. The main advantage of our strategy is flexibility, generality, and the potential for future expansion. We present examples illustrating the power of this strategy.

This paper introduces a divide-and-conquer segmentation strategy, and the concepts of associative image analysis for 3D multi-parameter images. It presents FARSIGHT, a software system that implements these strategies. There are several features that make FARSIGHT a versatile tool for multi-parameter image analysis: (i) it permits efficient investigator validation of the results; (ii) the code is able to handle morphological diversity; and (iii) its modular design will enable adaptation to other applications, and future extensions. Thus FARSIGHT provides a systematic and methodical strategy for making quantitative measurements of complex tissue-level image data. It is broadly applicable for quantifying the organization of brain tissue, and changes in organization during development, or following disease, injury, application of a pharmaceutical agent, or aging.

Section snippets

Specimen preparation and imaging

The Wadsworth Center Institutional Animal Care and Use Committee (IACUC) approved all animal procedures. Three adult male Sprague–Dawley rats were anesthetized with a ketamine/xylazine mixture, and transcardially perfused with 200 mL warm (37 °C) phosphate buffered saline (PBS) followed by 200 mL 4% paraformaldehyde in PBS using a constant-pressure system (Olsen, 1985). Brains were removed and immersion fixed in 4% paraformaldehyde for an additional 24 h, then washed in HEPES-buffered Hanks’ saline

Results

Three datasets were used to test our segmentation and classification analyses. Two of these were from the hippocampus; the third was from cingulate cortex. One of the hippocampal datasets was used for our primary example because of its clear laminar organization. In addition, a large majority of hippocampal astrocytes express sufficient amounts of glial fibrillary acidic protein (GFAP) to enable immunohistochemical identification of this cell population (Kimelberg, 2004, Nixdorf-Bergweiler et

Discussion

This study demonstrates the power and utility of associative image analysis for computing biologically meaningful measurements of complex brain tissue from multi-channel 3D images. The measurement data generated from each image are very rich. They can be tapped for diverse hypothesis-driven and/or discovery-oriented studies. To enable widespread usage, we have chosen to cast all the intrinsic and associative measurements, as well as the attributed graphs in the widely used XML format. Since

References (46)

  • D.H. Szarowski et al.

    Brain responses to micro-machined silicon devices

    Brain Res

    (2003)
  • G.M. Treece et al.

    Regularized marching tetrahedra: improved isosurface extraction

    Comput Graph

    (1999)
  • J.N. Turner et al.

    Cerebral astrocyte response to micromachined silicon implants

    Exp Neurol

    (1999)
  • Y. Zhang et al.

    A novel tracing algorithm for high throughput imaging. Screening of neuron-based assays

    J Neurosci Methods

    (2007)
  • Y. Zhong et al.

    Dexamethasone-coated neural probes elicit attenuated inflammatory response and neuronal loss compared to uncoated neural probes

    Brain Res

    (2007)
  • K. Al-Kofahi et al.

    Median-based robust algorithms for tracing neurons from noisy confocal microscope images

    IEEE Trans Inf Technol Biomed

    (2003)
  • Y. Al-Kofahi et al.

    Improved detection of branching points in algorithms for automated neuron tracing from 3D confocal images

    Cytometry A

    (2008)
  • M. Bear

    Neuroscience: exploring the brain

    (2006)
  • J. Bezdek

    Pattern recognition with fuzzy objective function algorithms

    (1981)
  • R. Biran et al.

    The brain tissue response to implanted silicon microelectrode arrays is increased when the device is tethered to the skull

    J Biomed Mater Res A

    (2007)
  • Y. Chikuse

    Statistics on special manifolds. Lecture notes in statistics

    (2003)
  • M.E.B.G. Dickinson et al.

    Multi-spectral imaging and linear unmixing add a whole new dimension to laser scanning fluorescence microscopy

    Biotechniques

    (2001)
  • R.E. Fan et al.

    Working set selection using the second order information for training SVM

    J Mach Learn Res

    (2005)
  • Cited by (0)

    1

    These authors contributed equally to this work.

    View full text