Synapse Resolution Whole-Brain Atlases
It is well-known that the highest resolution whole brain atlases are currently at BrainMaps.org, which has been compared to a Google Maps for the brain. However, these atlases are 0.46 microns per pixel, and are not sufficient to discern individual synapses, which require nanometer resolution. So in this post, I will consider the problems associated with constructing a synapse resolution (nanometer resolution) whole-brain atlas.
There seem to be two fundamental hurdles to constructing a synapse resolution whole-brain atlas: 1) image acquisition, and 2) digital technologies for working with the images and serving them over a network.
The first hurdle encompasses the time bottleneck and section preparation. If each section is 50 nm thick, then for a 10 mm mouse brain, 200,000 sections are needed, thus requiring some type of automation for section preparation. If we consider the time to scan a single 10mmx10mm section at 10nm pixel size at 1MHz, it comes out to 12 days (additional overheads would come from stage movements). Even with 200,000 TEMs (transmission electron microscopes) in parallel, one for each section, it will take 12 days for the complete scan. An alternative is offered by way of virtual microscopy solutions offered for light microscopy. One way would be to scan over the section, acquiring one column at a time instead of a patchwork of small images for montaging. Another alternative would be to construct a TEM with parallel scanning capabilities (having parallel magnetic lenses and electron beams), so that the entire section could be scanned at once, instead of scanning each little image patch in serial. This solution requires constructing a special type of TEM which implements certain features found in current day virtual microscopy systems for LM (light microscopy), and thus requires a team of hardware and software specialists to specially design, in addition to some physicists who are intimately acquianted with the physics behind TEM.
The second hurdle involves digital technologies, and the observation that even if a whole mouse brain was able to be acquired through TEM, that digital technologies currently would not be able to deal with that much data (200 PB, uncompressed). A single section is 1 x 10^12 pixels, which comes out to 1 TB (uncompressed, 8 bit grayscale), which is still not feasible using today's digital technologies.
In conclusion, for purposes of obtaining information about whole-brain connectivity, a nanometer-resolution whole-brain scan is required, and current-day tracer experiments are suboptimal and will always leave room for ambiguities that can only be resolved by completely mapping every synapse and axon in the brain. However, constructing a synapse resolution (or nanometer resolution) whole-brain atlas for even a mouse brain is so formidable as to be seemingly beyond today's technological capabilities. Maybe in 10-20 years.
Labels: brain atlas, digital technologies, electron microscopy, light microscopy, synapses, virtual microscopy
posted by neubrain at 3:45 PM
9 Comments:
There are a variety of groups that are actively working on the hardware and software to make this work. Winfried Denk (of 2-photon microscopy fame) is leading a project using Serial Block Face scanning EM to rapidly do volumetric EM of the brain. Dimitry Chklovskii is working on automated circuit connectivity algorhitms. There will be an excellent meeting specifically on this challenge at Janelia Farm in September 2007.
One of the major difficulties is even if you can build the total 3D reconstruction, can you discern the function of every synapse just from the EM? Can you accurately classify neurotransmitter class and synaptic strength from morphology? The answer is unclear.
Nonetheless, I bet we will see a total structural database of a mouse brain within 10 years.
-Andrew Hires
Thanks Andrew. I think 10 years is overly optimistic, and that 20 is more likely, given the magnitude of the two technical difficulties discussed. The serial blockface EM you refer to only works for relatively small scanning areas and is thus not suitable in itself for whole-brain EM. Dmitry Chklovskii has published a bit on wiring optimization, but he really hasn't published much at all as it seems he is just starting out in neuroscience, and what he has published doesn't seem that relevant to the problems involved with whole-brain EM. The problem of determining the function of every synapse is a big issue, as is the more boring technical problem of simply getting alignment between massive, multi-terabye EM images. More important than determining synapse function is morphological and structural classification of EM structures, as this remains closer and truer to the raw data than inferences regarding synaptic function. It's the universal problem of inferring function from form, but before we can begin to do that, we must be sure that we precisely understand the forms involved.
thanks for the meeting link. Sounds interesting. I'll try to make it.
I take back my comment about Chklovskii just starting out in neuroscience (which I mistakenly inferred from his relatively low publication output), as I did some digging around, and found out he has done quite a bit of interesting work in the area of neural circuit reconstruction. There is some information about this at http://www.cshl.edu/labs/mitya/research1.html
Also, Winfried Denk has done some interesting work in this regard. It would appear that since EM reconstruction of the c elegans brain has been complete, that the next logical step would be EM reconstruction of either the ant or fruit fry brain. Trying to EM reconstruct a mouse brain is technically impossible at this current time, but something like a fly or ant brain seems well within the realm of possible.
After seeing the talks on this topic at CSHL this week, I agree that the whole mouse may take longer than 10 years, but maybe not... It is hard to determine the rate of increase of the data collection of a 15 year process by looking at the first 2 year timepoints.
Jeff Lichtman noted that this problem is of far greater scope than the human genome project. And far fewer financial resources are being trained on it.
I have brief writeups of the talks by Chklovskii and Denk from the meeting at Brain Windows
On what is possible right now: the Drosophila brain is the next suitable candidate for TEM imaging with current techonologies. Not with Denk's block face -as of today it's not fast enough and lacks proper resolution. But with traditional TEM.
Why Drosophila: the brain is well known at the light microscopy level, with numerous landmarks identified and correlated with gene expression patterns. In addition, one can follow the full sequence of brain development from late embryo throughout the larval and pupal stages. So it is possible to start with a simpler, smaller brain that contains the minimal structure from which the adult brain will develop. As technology gets better, we can grow into larger, more mature fly brains.
And that's what we are doing: last March there was a meeting at Janelia Farms on insect neuroanatomy. My advisor Volker Hartenstein presented some of my work in TEM imaging of the first instar fly brain, available in non-annotated form.
In my experience the main bottleneck is acquisition time, which can be optimized a lot simply by improving the acquisition software. The AMI group at the Scripps institute in San Diego is developing Leginon: a scalable (as in, can work with multiple microscopes at a time) imaging software package which promises to reduce acquisition time from 3 months (for my work above on first instar brain) to 2 weeks. I've tried it out and it's getting to the point where Leginon can stand to its claims.
The second problem is image stitching, editing and browsing. While still incomplete, I've created TrakEM2 for the purpose, which includes a semantic segmentation editor for 3D modeling. Stephan Saalfeld at MPI-CBG has created the web browser modeled after google maps.
I will apply to attend the Neural Circuit Reconstruction conference in Janelia next September as well. What could be better than getting some feedback from people in the field over a couple of beers.
thanks for the great information, Albert. For the EM volumetric data at http://fly.mpi-cbg.de , perhaps a Zoomify interface would be faster. I noticed issues with the montages, that borders between patches were evident and that stitching wasn't perfect, but I must say, those are big montages! 20x20 or 400 patches per montage. Very nice. I'll see you this Sept at Janelia, if you're going.
Definitely, what we need is a fly brain EM reconstruction. Hopefully within 2-3 yrs. Then onto EM mapping the whole mouse brain!
Albert, do you know anyone who is interested in EM mapping the whole fly brain? It doesn't seem your advisor, Volker Hartenstein, is involved with this, though he has done a lot of work on the fly brain, particularly during development.
Hi neubrain (by the way, do you have a name?).
On fly brain montages at http://fly.mpi-cbg.de : yes there are evident stitching problems and others of higher severity such as noise (lead citrate precipitate from the counterstaining). Plus some interspersed sections are missing. The latter problems are addressable with better handling of the material.
The key problem here was in defining the question and evaluating the trade-offs: what do we want? We wanted best resolution and contrast at the nanometer level. For the purpose of following 100 nm dendrites and synapses.
Since no formal model of TEM-induced elastic deformations exists (from heat dilating the plastic sections where they are darker), any attempt to correct such deformations would have just introduced further artifacts. Better have some small errors than artificially "perfect" sections. When 3D modeling sub-volumes, I simply re-register them.
The pixels contrast problem: each image is its own world in the 16-bit range, skewed from hardly controlable on-camera "optimizations". Correcting the contrast for each tile so that the overall montage looks "right" (smooth) ends up destroying local contrast, which is the relevant part for 100 nm structure identification.
In summary: TEM vendors are not aware of the problems arising in high-throughput situations, and still live in a world where single images are relevant. Certainly up to us to awaken them.
As far as I know, Janelia Farms has an ongoing project targeting whole adult fly brain mapping. Chvloskii's group is working on automated segmentation. I have no idea how far they've got, or how good are their results. But they are using Denk's serial block (hearsay).
As for Volker Hartenstein, his (ours) lab approach is a recursive bottom-up: first define landmarks at a coarse level, then dive in and define finer landmarks, and so on. Have a look at recent papers from the lab to see what I mean (neuroblast lineages, main tracts and neuropile compartments are the current coarse level).
Post a Comment
<< Home