Software @ CIS : LDDMM : Clinical Background

The discipline of Computational Anatomy focuses on shape analysis of anatomical structures obtained in biomedical imaging. Under the auspices of the Brain Morphometry Biomedical Informatics Research Network (mBIRN), a processing pipeline (SASHA - Semi-Automated Shape Analysis) is being developed to enable seamless processing of brain morphometry data for subcortical structures through the integration of multiple site applications.

As a testbed, the tools are being integrated to perform semi-automated shape analysis of hippocampus volumes in a study of Alzheimer's Disease. Brain structural MRI data from Washington University was made available to researchers at Massachusetts General Hospital (MGH) and Johns Hopkins University (JHU). The initial data consisted of 45 subjects (21 controls, 18 Alzheimer's, 6 from a rare form of dementia, called semantic dementia, a disorder of language in which patients demonstrate a progressive deterioration of understanding and recognizing words while other cognitive faculties remain spared).

The subjects were scanned using high resolution (1.5 Tesla) T1-weighted structural MRI at Washington University. These scans were first anonymized and then automatically segmented at MGH's Martinos Center using Freesurfer, yielding segmented data sets. These data sets were aligned and processed at JHU's Center for Imaging Science (CIS) using the Large Deformaion Diffeomorphic Metric Mapping (LDDMM) tool and visualized with 3DSlicer from the Surgical Planning Lab (SPL) at Brigham Womans Hospital (BWH).

Velocity Vectors and Metric Distance

LDDMM computes the velocity vectors that transform one binary image to another, giving the metric distance between the two images. The velocity vector fields are generated by the group of infinite dimensional diffeomorphisms (the generalization of rotations, translations and scale group), the necessary group for studying shape.

The metric distances give a precise mathematical description of what shapes are similar and different.

Statistical Analysis

The resultant data was then uploaded into the BIRN database (Storage Resource Broker) for sharing among institutions and for further analysis. From the 2050 LDDMM comparisons of the left hippocampus data, a preliminary statistical analysis of the 45x45 matrix of metric distances was performed. The figure below illustrates a non-linear mapping to the two dimensional plane of a set of high-dimensional Euclidean points which minimize the interpoint distance distortion from the LDDMM matrix.

The figure suggests that there is class-specific information in the LDDMM matrix with upper-left observations predominantly controls (black), and the other observations for dementia (red) and semantic dementia (green) elsewhere.

Conclusion

Researchers were able to perform shape analysis on the hippocampi of 45 subjects. Ongoing statistical analysis of the data is being performed. The tools (Freesurfer, LDDMM and 3DSlicer) are expected to enable biomedical scientists to perform shape analysis of anatomical structures leading to a better understanding of diseases and disorders with greater statistical power.

Poster abstract at 10th Organization for Human Brain Mapping, Budapest, Hungary, 2004:

Biomedical Informatics Research Network: Multi-Site Processing Pipeline for Shape Analysis of Brain Structures: MF Beg, C Ceritoglu, AE Kolasny, CE Priebe, JT Ratnanather, R Yashinski, L Younes, P Yu, J Jovicich, RL Buckner, S Pieper, B Fischl, MI Miller.