Dissecting, Imaging, and Modeling Brain Networks KOCSEA 2008 - PowerPoint PPT Presentation

Dissecting, Imaging, and Modeling Brain Networks KOCSEA 2008 October 26, 2008 Yoonsuck Choe Brain Networks Laboratory Department of Computer Science Texas A&M University Joint work with: Bruce McCormick, Louise Abbott, John Keyser, David Mayerich, Jaerock Kwon, Donghyeop Han, and Pei-San Huang, 1

Introduction Main research questions: 1. How does the brain work? 2. How can we use the knowledge to build intelligent artifacts ? Approach: 1. Computational neuroanatomy Image source: http://www.nervenet.org/papers/Cerebellum2000.html 2

Overview • Connectomics • Knife-Edge Scanning Microscope • Structural reconstruction algorithms 3

Connectomics 4

Connectomics • Connectome : Complete structural description of the connection matrix of the brain (see e.g. Sporns et al. 2005). • Connectomics : Acquisition and mining of the connectome. • The only available connectome: that of the C. elegans (White et al. 1986). Image source: http://www.mouseatlas.org/data/mouse/stages/t47/view http://www.nervenet.org/papers/Cerebellum2000.html 5

Why Connectomics Research? • Structure of the nervous system as a foundation of its function. – Dynamical properties can be estimated from static structure. • Intensive study of single neurons and their molecular properties must be complemented by a system-level, architectural perspective. • Discover modules that make up the brain (motifs, basic circuits). • To understand how the brain works! 6

Goal of the Project Obtain and reconstruct the full mouse connectome at a sub-micrometer resolution. • 77-78d weight 26–30g • 13 mm (A-P) × 9.5 mm (M-L) × 6 mm (D-V) • 75 million neurons (Williams 2000) 7

Knife-Edge Scanning Microscope 8

Knife-Edge Scanning Microscope • Designed by Bruce H. McCormick. • Diamond microtome, LM optics, high-speed linescan camera, precision 3-axis stage [Movie] 9

Operational Principles of the KESM • Aerotech precision stage moves resin-embedded brain tissue across knife (x/y 20 nm, z 25 nm encoder resolution). • Back-illumination through diamond knife. • Nikon CF1 Flour 10X or 40X objectives (NA 0.3/0.8, water imm.). • Dalsa CT-F3 high-speed line scan camera images the tip of the knife at 44KHz. [Movie] 10

KESM Imaging Line−scan Camera M Light source i c r o s c o p e o b j e c t i Diamond knife v e Specimen Brain specimen is embedded in plastic block. 11

KESM Imaging Line−scan Camera M Light source i c r o s c o p e o b j e c t i Diamond knife v e Specimen Plastic block is moved toward the knife. 12

KESM Imaging Line−scan Camera M Light source i c r o s c o p e o b j e c t i Diamond knife v e Specimen Thin tissue slides over knife and gets imaged. 13

KESM Imaging Line−scan Camera M Light source i c r o s c o p e o b j e c t i Diamond knife v e Specimen Successive line scan constructs a long image. 14

KESM Imaging Line−scan Camera M Light source i c r o s c o p e o b j e c t i Diamond knife v e Specimen One sweep results in a ∼ 4 , 000 × 20 , 000 image ( ∼ 80 MB). 15

KESM Imaging Line−scan Camera M Light source i c r o s c o p e o b j e c t i Diamond knife v e Specimen One brain results in ∼ 25 , 000 images. 16

Stair-Step Cutting Kwon et al. (2008) • Width of the knife and the field of view of the objective are not wide enough to cut the entire top facet of the tissue block. 17

Automated Sectioning/Imaging S/W • Automated stage controller and image acquisition system developed in-house. • Fully automated operation without human intervention: 8 hours a day, 5 days a week. 18

KESM Data: Golgi Stain • Mouse cortex (sagittal section). 19

KESM Data: India Ink • Mouse spinal cord vasculature. [Movie] 20

KESM Results: Volume Visualization Nissl (Cortex) India ink (Spinal cord) Golgi (Pyramidal cell) Golgi (Cortex) Golgi (Cerebellum) Golgi (Purkinje cell) [Movie] 21

Structural Reconstruction Algorithms 22

Reconstruction Approaches Raw data or volume visualization is not enough: Structural reconstruction is needed. • Segment-then-connect: the most common approach • 3D convolutional network: Jain et al. (2007) • Template-matching-based vector tracing: Al-Kofahi et al. (2002) 23

Reconstruction: Tracing in 2D 1 c i+2 2 c i+2 c i+1 c i+1 c i c i * intensity step i+1 step i position Choe et al. (2008) • Moving window with cubic tangential trace spline method. • Investigates pixels only on the moving window border and on the interpolated splines for fast processing. 24

Tracing Results Seed Can et al. (1999) 25 Haris et al. (1999) Our method

Robustness Comparison 120 50 100 40 80 30 Error Error 60 20 40 10 20 0 0 20 30 40 50 20 30 40 50 Width Width Open diamonds: Harris et al.; Closed diamonds: Can et al.; Closed boxes: Our approach. • Accuracy tested based on synthetic data (by varying fiber width): Linear (left), curvy (right). • Much more accurate compared to competing approaches such as Can et al. (1999); Haris et al. (1999). 26

Reconstruction: Tracing in 3D Match! t = 3 t = 2 t = 1 Template matching Tangential slices Templates (Mayerich and Keyser 2008; Mayerich et al. 2008) • Use a moving sphere and trace along points on the surface of the sphere. • Use graphics hardware (GPU) for fast matrix operations during template matching. 27

Tracing Results Spinal cord vasculature (KESM) Neuron (Array Tomography, tectum) Vasculature (KESM, cerebellum) 28

Speeding Up Tracing Using GPU 25 Single Core 2.0GHz Single Core 2.0GHz Quad Core 2.0GHz GPU (Sampling Only) 10000 CPU with GPU Sampling 20 Full GPU GeForce 7300 1000 Time (ms) 15 Factor 100 10 10 5 1 0.1 0 1 10 100 1000 10000 1 10 100 1000 10000 Number of Samples Number of Samples Run time Speedup • Performance figures demonstrate the speedup obtained by using GPU computation. • Speedup achieved by using the full capacity of GPUs show an almost 20-fold speedup compared to single-core CPU-based runs. 29

Preliminary Branching Statistics (vasculature) Sample Statistics from Reconstructed KESM Brain Vasculature Data (1 mm 3 volume) Region Segments Length Branches Surface Volume Volume (mm 2 ) (mm 3 ) 5 5 (mm) (% of total) Neocortex 11459.7 758.5 9100.0 10.40 0.0140 1.4% Cerebellum 34911.3 1676.4 19034.4 20.0 0.0252 2.5% Spinal Cord 36791.7 1927.6 26449.1 22.2 0.0236 2.4% • Geometric structures extracted using the automated reconstruction algorithms allow us to conduct quantitative investigation of the structural properties of brain microstructures. 30

Wrap-Up 31

Discussion and Future Work • Main contribution: novel imaging method plus computational algorithms for automated structural analysis. • Future work: – Full-brain reconstruction and validation – Estimating connectivity from sparsely stained data (cf. Kalisman et al. 2003) – Linking structure to function 32

Conclusion • Understanding brain function requires a system-level investigation at a microscopic resolution. • Innovative microscopy technologies are enabling a data-driven investigation linking the microstructure to the system. • The massive data can only be effectively understood through automated computational algorithms. 33

Acknowledgments • People: – Texas A&M: B. McCormick, J. Keyser, L. C. Abbott, D. Mayerich, D. Han, J. Kwon, Y. H. Bai, D. C.-Y. Eng, H.-F. Yang, G. Kazama, K. Manavi, W. Koh, Z. Melek, J. S. Guntupalli, P .-S. Huang, A. Aluri, H. S. Muddana – Stanford: S. J. Smith, K. Micheva, J. Buchanan, B. Busse – UCLA: A. Toga – Others: T. Huffman (Arizona State U), R. Koene (Boston U), Bernard Mesa (Micro Star Technologies) • Funded by: NIH/NINDS (#1R01-NS54252); NSF (MRI #0079874 and ITR #CCR-0220047), Texas Higher Education Coordinating Board (ATP #000512-0146-2001), and the Department of Computer Science, and the Office of the Vice President for Research at Texas A&M University. 34

In Memory of Bruce H. McCormick Bruce H. McCormick (1928–2007) • Designer of the Knife-Edge Scanning Microscope 35