2006 Scientific Report Structural Summary

Brain Tissue Classification and Subparcellation of Brain Structures

Structural Image Analysis

Overview

Structural image analysis is concerned with the morphology of anatomic structures, which includes identification and characterization of volume and shape. The processing comprises image filtering, tissue segmentation, anatomically true subdivision of large structures into relevant substructures, and extraction of features to be used for population-based statistical analysis. Linear and nonlinear volumetric registration tools have multiple uses, i.e., mapping individual image datasets into an anatomic reference database, fusing data by combining multiple modalities, and linking the individual anatomy to a statistical prior (atlas). These tools are also critically important to structural image analysis. NA-MIC is developing a new set of tools driven by the needs of its clinical partners (DBPs), using the existing image databases for testing and validation. The development of methods and tools for structural image analysis is a collaborative effort betweeen the computer science partners, clinical Core 3 counterparts, and engineering Core 2 partners. This collaborative approach ensures the development of state-of-the-art methods that are driven by the requirements of clinical imaging research and implemented with industry-standard programming style and design.

Image Segmentation

Brain tissue classification, subparcellation into lobes and cortical areas, segmentation of subcortical structures

• Utah: Multi-spectral MRI tissue classification with a novel spatial filtering scheme using Markov random fields. Comparison of new segmentation method to EM-segmentation (Leemput et al.) based on simulated data (MNI) with varying levels of noise.
• GATech: Development of rule-based semi-automatic segmentation tools, which include a “thumb extractor” using energy-based minimization for efficiency and a Bayesian tissue classifier. Both tools are combined with rules developed by Jim Fallon (UCI) and Jim Levitt (BWH) and integrated into a GUI to efficiently segment the dorsolateral and dorsomedial prefrontal cortex (DLPFC and DMPFC), putamen, and striatum. Rule-based striatum segmentation is integrated into Slicer and thus already part of the NA-MIC toolkit.
• GATech: Statistical PDE methods: Statistical measures are added into the PDE method (curvature driven flow), which allows incorporation of global image information into locally defined PDE framework.
• MIT: Shape-driven segmentation: EM framework used for statistical segmentation. The tool uses registration of atlas information (shape prior) to guide subtle segmentation and incorporates additional statistical shape information to refine the segmentation of substructures. Preliminary validation on 22 scans where expert segmentations are available. The development of atlas priors and statistical shape models was done in close collaboration with the clinical partner group of M. Shenton, Harvard. The segmentation system is currently integrated into the Slicer platform.
• Dartmouth: Volumes of medial temporal lobe structures in patients with schizophrenia, including the hippocampus, amygdala and entorhinal cortex, are measured with Slicer using manual segmentation.

Image Registration

• GaTech: Registration of pairs of image data sets via registration of segmented surfaces. Using sulci as landmarks, surfaces are flattened to an annulus, and mass-preserving mapping is used to register the two annuli.
• Isomics: Slicer Registration Framework (Registration Tools): The registration framework is designed to support translation, rigid, affine, and deformable forms of registration between volumetric image data in Slicer. These tools will form a most crucial component of the NAMIC toolkit, since linear and nonlinear image registration is a key component for segmentation, mophometric analysis, and data fusion.

Morphometric Measurements and Shape Analysis

• UNC: Toolkit for population-based statistical shape analysis, which runs as a dataflow pipeline. ITK modules include surface parametrization [(spherical, harmonics, and point distribution models (PDMs)], surface correspondence, shape alignment, statistical analysis with correction for multiple comparison, and visualization of complex statistics for verification and interpretation. Application to Harvard VA SPD clinical study of caudates (20 SZ and 20 CNTL), second BWH SPD caudate study, and Dartmouth hippocampus study (20 SZ and 20 CNTL) in close collaboration with Core 3 partners.
• UNC: Loni shape pipeline prototype for automatic population-based shape analysis in collaboration with UCLA (Loni pipeline) and with GE (Dart 2 based data storage).
• GATech: Localized shape analysis via spherical wavelet basis functions, which encode local shape in space at different scales. Application to BWH caudate dataset and comparison of wavelet prior estimation with PCA estimation.
• MIT: Shape modeling of anatomical structures of interest to be used as a shape prior for segmentation. Shapes are coded via distance transforms and represented as mean shapes plus major eigenmodes of deformations. Result is a mult-object hierarchical representation of brain structures.
• Utah: Automated shape model construction: Tool to automatically construct shape models from an input of several binary segmentations. This method finds boundary correspondences by creating the most efficient probability distribution for the population of input shapes, and can be used for statistical group comparison. Hippocampal segmentations from Harvard are used as a driving clinical application.

Algorithms and Software Infrastructure

Most of the new tools listed above are either already integrated into Slicer or are available as ITK modules of the NA-MIC toolkit. Complex sequences of tools, such as rule-based segmentations, shape-driven EM segmentation, or population-based shape analysis, are developed as sets of ITK modules, which, in turn, are integrated into Slicer, into a separate GUI for testing and validation, or combined to automatic pipelines via the Loni pipeline architecture.

Key Investigators

• MIT: Kilian Pohl, Sandy Wells, Eric Grimson
• UNC: Martin Styner, Ipek Oguz, Guido Gerig
• Utah: Ross Whitaker, Suyash Awate, Tolga Tasdizen, Tom Fletcher, Joshua Cates, Miriah Meyer
• GaTech: Allen Tannenbaum, John Melonakos, Tauseef ur Rehman, Shawn Lankton, Ramsey Al-Hakim, Eric Pichon, Delphine Nain, Oleg Michailovich, Yogesh Rathi, James Malcolm

Key co-investigators Core 3

• Harvard PNL: Sylvain Bouix, Motoaki Nakamura, Min-Seong Koo, Martha Shenton, Marc Niethammer, Jim Levitt
• Dartmouth: Andrew Saykin
• UCI: Jim Fallon

Key co-investigators Core 2

• Steve Pieper, Bill Lorensen, Luis Ibanez, Karthik Krishnan, Michael J. Pan, Jagadeeswaran Rajendiran, Jim Miller, Karthik Krishnan, Luis Ibanez