Transcript The big data challenges of connectomics

The big data challenges of
connectomics
JEFF W LICHTMAN, HANSPETER PFISTER NIR SHAVLT
PRESENTED BY YUJIE LI, OCT 21TH,2015
Connectomics
• The study of the structural and
functional connections among brain
cells.
• Product is the “connectome,” a
detailed map of those connections.
• Significant to understanding of the
healthy and diseased brain.
• “I am my connectome” -- Sebastian
Seung
Neuron structures
http://science.kennesaw.edu/~jdirnber/Bio2108/Lecture/LecPhysio/PhysioNervous.html
http://www.ncbi.nlm.nih.gov/books/NBK21535/
How many neurons in a human brain?
100 billion neurons
How many neurons in a Drosophila?
100,000 neurons.
~ 107 synapses
A video to appreciate the challenge faced with connectomics
Brainbow Technique
A Voyage Into the Brain
http://ngm.nationalgeographic.com/2014/02/brain/voyage-video
Acquisition
Analytical problems stand between the
acquired image and having access to the data
in a useful form
• Alignment
• Reconstruction
• Feature detection
• Graph generation
Alignment
sections collected on
a belt may rotate.
Reconstruction
Challenges for automatic segmentation:
• Irregular neuron objects
• Lateral resolution is several-fold finer
than thickness
• Under/over segmentation
Goal : Obtain saturated reconstructions of very
large (1mm3)brain volumes in a fully automatic
way, with minimal errors and reasonably short
time.
Human tracers, cursive handwritings recognition
Feature detection
Subcellular features: mitochondria, synaptic vesicles etc…
Difficult to find cell boundaries
Irregular shape
Reduce error and analysis time
Graph generation
•Data turned in to a
form that represents
the wiring diagram.
•Data reduction step
• How much of original
data to retain?
• How to store the
graph?
• Skip Oct-trees.
Common theme:
Dehumanizing the pipeline
An irony is that humans are especially good at these tasks….
If we know how our brain wires, would be easier to develop tools to
automate these processes.
Big data challenges of connectomics
• Data size
• Data rate
• Computational complexity
• Parallel computing
• Compute system
• A heterogeneous hierarchical approach
• Data management and sharing
Data size
1mm3 rat cortex image = 2 million gigabytes = 2 petabytes
A complete rat cortex 500mm3 = 1,000 petabytes
(Walmart database manages a few petabytes of data)
A complete human cortex ~1000 larger than rodents = 1,000 * 1000 petabytes
= 1 zetabyte
(All information recorded globally today)
Data rate
- Imaging task distributed to different labs
- Complete connectome of a human cortex is the goal!
- Maybe start with substructures.
Data management and sharing
- Assumed we obtained the data, do we store it?
◦ Yes, image and graph.
- How to move from microscope to the computer system? Transfer bandwidth
◦ Placing computer near the microscope.
◦ 500 standard 4-core 3.6 GHz processors would suffice. $1 million.
- Where to store?
◦ Disk or tapes.
- How to share?
◦ Internet Current achievable data rates: 300 megabites/second
◦ Central sharing sites
◦ Reconstructed layout graph is easier to deal with.
Computational complexity
The goal of many big data system is more than to simply
allow storage and access to large amounts of data. Rather, it
is to discover correlations within data.
◦ Sampling
◦ Parallel computing
◦ Image segmentations and feature extraction are embarrassingly
parallel.
A heterogenerous hierarchical approach
Combines bottom-up information from the image data with top-down
information from the assembled layout graph, to dynamically decide on the
appropriate computation level of intensity to be applied to a given sub-volume.
1) Initially apply the lowest cost computations to small volume.
2) The sub-graphs will be tested for consistency.
3) If discrepancies are found, more expensive computation used.
4) The process will continue hierarchically, growing the volume of
merged segments.
Prospects
- The field needs a significant investment to advance.
- Commercial values in connectomics
◦ Treating brain diseases
◦ Appling lessons learnt to making computer smarter
- Challenges beyond the horizon: still big data problem
Comments
No address on the EM technical limitations:
• Samples post-mortem, not in vivo
• Physical damage during section, potential distortion.
• Lack functional information
No comparison with the current popular approaches to the problem
• Two photon, confocal, brightfiled images
• Neuron-labelling approaches (physical dye, genetic approach)
 Big data is not only about handling the super large dataset.
• It is also about finding a smart way to fuse data from different modalities
and different sources to obtain a comprehensive understanding