Skip to content

BigNeuron Imperial College London Hackathon Discussion Notes

Jump to bottom
Xiaoxiao Liu edited this page Jan 22, 2016 · 14 revisions

Qualitative assessment on the distance comparison:

  • break down to different
    • labs

    • model organism

    • cells

    • microscopies

    • staining (any meta data)

    • image quality

Provide detailed analysis sheet and provide feedback to data contributors to improve their own data generation techniques

Assessment of underlying algorithms:

  • get insight into why particular methods rank higher than others (per metric, organism, etc...)
  • can we categorize methods (or eventually modularize them)?
  • can we get ideas/pointers for future method development based on this?

New Metrics :

  • topology
  • L-measure
  • Sholl analysis
  • regional info is important for clustering: Landmarks (pia surface, soma location relative to pia)
  • how fragmented ( # subtrees)
  • dendrogram matching

Simulate images for testing:

  • Synthesis neuron (Lida)
  • simulation => SWCs==> Imaging ( Erik) ==> reconstruction
  • differences between real imaging data
  • how the results vary according to: SNR/cell types/background noise ( extracted from real background)/imperfect staining/PSF/artifacts

Other ways to generate "ground truth" (in addition to manual annotation and computer simulation)

SWC format :

  • document swc possibilities for this project:
  • allow multiple trees/soma/roots
  • sorted ( do sorting as a post-processing step )
  • no loops (detect loops? in sorting already?)

Consensus skeleton:

Data Consumption Discussion: How to make the BigNeuron data available for neuroscience applications

  • In reconstructions it would be good to encode intrinsic tree properties
    ** Most software lacks the context of a neuron (e.g., pial surface, layers, etc) ** Need the staining for the background

Q. In Neurolucida they encode metadata in the output file, but how can this be done in swc? A. Possibly a second, associated file can include this

  • Another challenge is to compare different species?
    ** For clearly-defined cell types, e.g., ChC, PCs, it would be nice to compare statistics of their morphologies and relate it to their evolutionary positionm e.g., How do pyramidal neurons from a rat compare to those from a mouse or a human? This could be a major biology paper focussing on cross species comparison.

Problem. Even two animals of the same type show statistical differences between their neurons,

  • Can we relate cell types from two species that are evolutionarily distant? And to what extent?

** For ill-defined cell types, we have to use rough categories and try to observe differences at single cell or network level. For mammalian data we already have this is hard, but for fruit-fly we probably have the ability to do this.

  • The Data_contribution_summary document lists the contributors of the data.

** NB. U Washington provides zebrafish cells from both larva and adult.

  • Why are insect brains structured so differently from mammalian ones? To answer this we need much more data, and BigNeuron may provide the data needed to address questions what is the role of the structure or anatomy of the brain?

BlastNeuron provides many features for studying and viewing neurons (see Y Wan et al., Neuroinform (2015) 13:487–499). It provides a rapid search of the morphology database and returns neurons whose shapes are similar to the queried neuron. It performs a global comparison followed by a comparison of local features. If the neurites in a neuron are perturbed (e..g, by bending or rotating them), BlastNeuron can still identify them as similar.

Q. Can BlastNeuron analyse swc files that contain two or more cells and locate synapses? A. In principle yes, but the neurons have first to be sorted and two neurons will confuse this. Vaa3d can be used to do the sorting.

Q. How can one tell to which neuron each tree belongs if there are several neurons in the file?

Data distribution

  • 1.6 million reconstructions, 500GB (5 million downloads from neuromorpho.com)
  • estimate traffic to the data website?
  • 50 TB images (will not release images for Phase1)
  • meta data: a)PSF, image modality, lab, resolution... from data contributors; b) analysis, relationship
  • visualization ( compare multiple neurons)
  • run simple tools for analysis
  • representative images should be distributed ( silver datasets)

Demos

  • Ajayrama: neuron distance to gold standard , for ranking neurons : heavily penalize anything that is too far (> 5 voxels) away from gt

  • Przemyslaw: tubular graph ( sampling from tubularity map) ==> minimum spanning tree ==> probability on the edges = >optimally pruning a minimum spanning tree (steiner tree?) github repo: CVLabVaa3DPlugins

  • Julian, Lida, Eleftherios: NeuroM metrics ( section and segment based) on reconstructed SWCs. Postprocesssed to comply with NeuroM SWC standards.

  • Alvar: image profiling with gold standard reconstructions ( vaa3d plugin) on 77 images from the gold training subset. He found a lot of variation in the imaging qualities across labs. Certain labs also produce images with very large variation in those metrics.

  • Julian: reconstruction quality standards for algorithms developers: connected structures are critical

Clone this wiki locally