Cortical Surface Analysis and Automatic Parcellation of Human Brain
Wen Li, Ph.D. student
Advisor: Vincent A. Magnotta, Ph.D.
Biomedical Engineering
University of Iowa, Iowa City, IA, USA
Outline
• Background
• Preprocessing
• Surface Generation
• Conformal Flattening
• Spherical Demons Registration
• Conclusion
Outline
• Background
• Preprocessing
• Surface Generation
• Conformal Flattening
• Spherical Demons Registration
• Conclusion
• BRAINS (Brain Research: Analysis of Images, Networks and Systems http://www.nitrc.org/projects/brains)
• Topographic structures vs. functions of human cerebral cortex
• Image-based vs. surface-based human cortical analysis
• Automatic, rapid and reliable parcellation of cerebral cortex (development, aging, disease progression or treatment response)
Outline
• Background
• Preprocessing
• Surface Generation
• Conformal Flattening
• Spherical Demons Registration
• Conclusion
• BRAINS AutoWorkup • AC-PC alignment• T1-T2 image registration• Skull stripping/Brain masking• Tissue classification• left-right hemisphere separation• Removal of cerebellum and brainstem
Image size: 256x256x192 pixel3
Voxel size: 1.0x1.0x1.0 mm3
Intensity range: 8 bits (0~255)10: pure CSF130: pure GM250: pure WM
Outline
• Background
• Preprocessing
• Surface Generation
• Conformal Flattening
• Spherical Demons Registration
• Conclusion
• Topology Correction (Neurolib)
• Marching Cubes
• Surface Decimation (220,000 -> 70,000)
• Surface Smoothing
• Geometry features
Distance from surface point to posterior-commissure [-1.0, 1.0]
Distance from surface point to a convex hull of it [0.0, 1.0]
Mean curvature [-1.0, 1.0]
Outline
• Background
• Preprocessing
• Surface Generation
• Conformal Flattening
• Spherical Demons Registration
• Conclusion
• Linear parameterizations with fixed boundaries (itk::QuadEdgeMeshToSphereFilter)
Split surface into halves Boundary smoothing Resulting sphere
Spheres with geometry features of original surface
Outline
• Background
• Preprocessing
• Surface Generation
• Conformal Flattening
• Spherical Demons Registration
• Conclusion
• Modified diffeomorphic demons registration
• Velocity field is not an arbitrary 3D vector field. It is a tangent vector field on the sphere.
• Spherical Diffeomorphic Demons– Reference: B.T.T. Yeo et al. Spherical
Demons: Fast Diffeomorphic Landmark-Free Surface Registration. IEEE Transactions on Medical Imaging, 29(3):650--668, 2010.
• Muti-resolution spherical demons registration
• 4 resolution levels (Icosahedral regular spheres)
level Resolution Feature Size (cells)
1 IC4DistanceFrom
PC5,120
2 IC5 GeoDepth 20,480
3 IC6 GeoDepth 81,920
4 IC7Mean
Curvature327,680
B.T.T. Yeo IEEE Transactions on Medical Imaging, 29(3):650--668, 2010.
Fixed mesh
Moving mesh
Level 1 Level 2 Level 3 Level 4
Deformation field
Labels on moving mesh
Warped labels
• Two common indexes to calculate similarity (overlapping) of region A and B
• Dice Index:
• Jaccard Index:
• Evaluation by using Dice Index
• The average Dice across subjects and regions is 0.88.
Subject Label 1 Label 2 Label 3 Label 4 Average
1 0.975 0.901 0.886 0.887 0.912
2 0.953 0.870 0.870 0.883 0.894
3 0.957 0.787 0.838 0.853 0.859
4 0.917 0.861 0.762 0.866 0.852
5 0.952 0.859 0.866 0.867 0.886
Outline
• Background
• Preprocessing
• Surface Generation
• Conformal Flattening
• Spherical Demons Registration
• Conclusion
• This pipeline proves to be capable of parcellating the surface of human cortex automatically in reasonable time
• Applying Spherical Demons Registration in pyramid levels helps implement the registration by different geometry features in different resolution levels
• We are currently evaluating the approach for a more refined cortical parcellation. 50 subjects’ data will be applied in the pipeline and further results will be given soon.