Download presentation
Presentation is loading. Please wait.
Published byApril Barton Modified over 8 years ago
1
Statistical Shape Analysis of Multi-object Complexes June 2007, CVPR 2007 Funding provided by NIH NIBIB grant P01EB002779 and NIH Conte Center MH064065. MRI images and expert manual segmentations funded by NIH RO1 MH61696 and NIMH MH64580. REFERENCES [1] P. Golland, W. E. L. Grimson, M. E. Shenton, and R. Kikinis, “Detection and analysis of statistical differences in anatomical shape,” Medical Image Analysis, vol. 9, pp. 69–86, 2005. [2] M. Styner, A. Lieberman, R. K. McClure, D. R. Weinberger, D. W. Jones, and G. Gerig, “Morphometric analysis of lateral ventricles in schizophrenia and health controls regarding genetic and disease-specific factors,” Proc. of the National Academy of Sciences, vol. 102, no. 13, pp. 4872–4877, March 2005. [3] J. Juranek, P. A. Filipek, G. R. Berenji, C. Modahl, K. Osann, and M. A. Spence, “Association between amygdala volume and anxiety level: magnetic resonance imaging (mri) study in autistic children,” Journal of Child Neurology, vol. 21, no. 12, pp. 1051–8, Dec. 2006. [4] M. Langen, S. Durston, W. G. Staal, S. J. M. C. Palmen, and H. van Engeland, “Caudate nucleus is enlarged in high-functioning medicationnaive subjects with autism,” Biol Psychiatry, 2007. [5] J. Marron and M. Todd, “Distance weighted discrimination,” Operations Research and Industrial Engineering, Cornell University,” Technical Report 1339, 2002, available at http://www.stat.unc.edu/postscript/papers/marron/HDD/DWD/DWD2.pdf. Data Fig. 2: Left: Object set after global alignment without scaling. Center: Global alignment with scaling, local pose differences remain. Right: Object set after global and local alignment. Summary: We studied the ability to discriminate between two populations based on features extracted from 10 subcortical structures. For our data of autistic and typically-developing brains, volume and the radius of the m-rep shape description performed best. Most studies of shape limited to single objects [1,2] Neuroimaging studies interested in group differences [3,4] Mental illness processes likely not isolated to single structures Methodology required to analyze multiple structures jointly Introduction Kevin Gorczowski 1, Martin Styner 1,2, Ja-Yeon Jeong 1, J.S. Marron 3, Joseph Piven 2, Heather Cody Hazlett 2, Stephen M. Pizer 1, Guido Gerig 1,2 1 Department of Computer Science, 2 Department of Psychiatry, 3 Department of Statistics University of North Carolina at Chapel Hill Fig. 1: Left: Sheets of medial atoms describing subcortical structures. Right: Implied surfaces of subcortical structures’ medial descriptions. Fig. 5: Left: Distributions of mean classification scores using volume. Right: Using only radii of m-rep shape description. Fig. 4: Average classification accuracy (percentage of 38 test samples correctly classified) over 100 runs using different training and testing sets. Statistical Discrimination From training set, calculated normal of separating axis using distance weighted discrimination (DWD) [5] Projecting test samples onto DWD normal gives classification score Ran 100 times using training sets of 32 samples (16 autism, 16 control), 38 test samples Classification accuracy: percentage of 38 test samples correctly classified Mean classification score: average projection onto DWD normal for runs in which sample was in testing set Conclusions Subcortical structures: amygdala, caudate, hippocampus, globus pallidus, putamen. Total of 10 structures (left and right). 46 autism and 24 typically-developing samples from a longitudinal, pediatric autism study Scans at age 2 and age 4 (23 autism subjects, 10 control) Corrected for uneven gender distributions (autism samples heavily weighted towards males) by subtracting gender-specific mean Shape: Used the medial representation (m-rep). Deformed template model to fit manual segmentations. Describes objects as a sheet of medial atoms. Atoms have elements position, radius, and normal vectors to surface boundary. Pose: Object sets first aligned jointly (global) to remove image scan differences. Then individual objects aligned (local) to extract relative pose changes. Parameters of local alignments (translation, rotation, scale) used as pose features. Volume: Volumes of individual objects computed from m-rep shape description before any alignment. Feature Dimensionality: Volume - 10, Pose - 70, Shape - 1890 Joint analysis of multiple objects using statistical discrimination Exploration of different features extracted from object sets Volume and local width (m-rep radii) give best performance for our application Feature Extraction Fig. 3: Left: Data samples from two populations with separating axis and normal vector to axis. Right: Data projected onto normal vector of separating axis, used as classification score.
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.