Identifying Group-wise Consistent White Matter Landmarks via Novel Fiber Shape Descriptor Hanbo Chen, Tuo Zhang, Tianming Liu the University of Georgia, US. Northwestern Polytechnical University, China.
Motivation How to identify regions of interest (ROIs)?
Challenge 1 - unclear functional or cytoarchitectural boundaries Li et al. 2010
Challenge 2 remarkable individual variability in brains
Challenge 3 highly nonlinear properties of cortical regions T. Liu 2011
Our efforts in identifying ROIs Li et al. NIPS Neuroinformatics Zhang et al. Cerebral Cortex MICCAI Li et al. Human Brain Mapping DICCCOL in Disease Chen et al. MICCAI Chen et al. TMI Ge et al. ISBI Ge et al. IPMI DICCCOL based network 2010 2011 2012 2013 DICCCOL Function Zhu et al. IPMI NeuroImage DICCCOL Zhu et al. Cerebral Cortex Yuan et al. Neuroinformatics Chen et al. MICCAI
Our efforts in identifying ROIs Li et al. Individualized ROI Optimization via Maximization of Group-wise Consistency of Structural and Functional Profiles. 2010 2011 2012 2013
Our efforts in identifying ROIs Zhang et al. Predicting Functional Cortical ROIs via DTI-derived Fiber Shape Models. 2010 2011 2012 2013
Our efforts in identifying ROIs Zhu et al. Discovering Dense and Consistent Landmarks in the Brain. 2010 2011 2012 2013
Our efforts in identifying ROIs Zhu et al. DICCCOL: Dense Individualized and Common Connectivity-based Cortical Landmarks. -minimize distance -random initiate ROIs -358 ROIs defined 2010 2011 2012 2013
Our efforts in identifying ROIs Li et al. NIPS Neuroinformatics Zhang et al. Cerebral Cortex MICCAI Li et al. Human Brain Mapping DICCCOL in Disease Chen et al. MICCAI Chen et al. TMI Ge et al. ISBI Ge et al. IPMI DICCCOL based network Gray Matter 2010 2011 2012 2013 Zhu et al. IPMI NeuroImage DICCCOL Zhu et al. Cerebral Cortex DICCCOL Function Yuan et al. Neuroinformatics White Matter Chen et al. MICCAI
Identifying Group-wise Consistent White Matter Landmarks via Novel Fiber Shape Descriptor
Obtain connection profile of fiber bundle
Obtain connection profile of fiber bundle
Calculate probability distribution to obtain connection map HEALPix framework (Gorski et al. 2005) 12 Regions 48 Regions 192 Regions
Calculate probability distribution to obtain connection map
Fiber bundle shape descriptor Fiber bundle Connection profile Connection map Connection entropy Connection similarity
Two criterions for WM landmarks 1. Network hubs 2. Group consist
Identify & Optimize WM landmarks Linear Alignment Connection Profile Connection Entropy Identify Landmarks Optimize Landmarks
Identify & Optimize WM landmarks Linear Alignment Connection Profile Connection Entropy Identify Landmarks Optimize Landmarks
Identify & Optimize WM landmarks Linear Alignment Connection Profile Connection Entropy Identify Landmarks Optimize Landmarks
Identify & Optimize WM landmarks Linear Alignment Connection Profile Connection Entropy Identify Landmarks Optimize Landmarks
Identify & Optimize WM landmarks Linear Alignment Initial Optimized Connection Profile Connection Entropy Identify Landmarks Optimize Landmarks
Locations of 12 WM landmarks
Fiber bundles of 12 WM landmarks
Optimize & predict to increase connection complexity and consistency
Distances between initial landmarks and final optimized landmarks 15 Distance (mm) 10 5 0 1 2 3 4 5 6 7 8 9 10 11 12
Conclusion Two criterions for WM landmarks Network hubs Consistent across individuals Fiber bundle shape descriptor Connection map Connection entropy Connection similarity Identify, optimize, and predict Identified 12 WM landmarks Reproducible on new individuals
Acknowledgement NIH Career Award EB-006878, NIH R01 DA-033393, NIH R01 AG-042599, NSF Career Award IIS-1149260, NSF BME-1302089, Franklin Foundation Travel Awards.
Now I will answer ANY questions Poster time: Tue 13:30-16:00 O2-03 Two criterions for WM landmarks Network hubs Consistent across individuals Fiber bundle shape descriptor Connection map Connection entropy Connection similarity Identify, optimize, and predict Identified 12 WM landmarks Reproducible on new individuals
Experimental Data & Preprocessing 2 sets of experimental data 18 young healthy subjects 64 healthy subjects from Human Connectome Project Q1 release Preprocessing Eddy current correction Skull removal Streamline fiber tracking
Two properties derived from connection map 1. Entropy (complexity of a fiber bundle) 48 HH(VV) = PP kk (VV)llllll 48 PP kk (VV) kk=1 PP kk (VV) = pprrrrrr. dddddddddddddddddddddddd oooo kk tth ssssssssssss pppppppppp iiii cccccccc. pppppppppppppp VV 2. Similarity between fiber bundles SS PP(VV ii ), PP VV jj = PP(VV ii) PP(VV jj ) PP(VV ii ) PP(VV jj ) PP(VV ii ) = cccccccc. mmmmmm oooo tthee ii tth RRRRRR
Two properties derived from connection map Fiber Bundle: Conn. Map: Entropy: 0.64 0.59 0.81 Similarity: 0.96 0.26
Higher connection entropy indicates increasing connection complexity Fiber Bundle: Conn. Map: Entropy: 0.59 0.64 0.81
High similarity indicates similar shape between fiber bundles Fiber Bundle: Conn. Map: Similarity: 0.96 0.26
Experimental Data & Preprocessing 18 Young Healthy Subjects Matrix: 128 128 60, resolution: 2 2 2mm 3, 30 directions, TR=15s, ASSET=2. Human Connectome Project Q1 Release 64 healthy subjects, FOV=210 180, matrix=168 144, 90 directions, TR=5.5 s, resolution=1.25 1.25 1.25mm 3. Eddy current correction, skull removal, streamline fiber tracking.
Optimized template Predicted subjects
12 WM landmarks i ii iii iv 10 12 14 64 86 Random HCP Q1 Subjects
Two criterions for WM landmarks 1. Network hubs 2. Group consist