UPDATE : DIADEM Final Results

The DIADEM automated neuronal reconstruction contest has finished.  Accurate, fast, and high-resolution automated neuron reconstruction is of vital importance to cracking the mystery of how neural circuits perform. Even with perfect knowledge of the firing patterns of every cell in a circuit, our understanding of how these patterns are produced and how the information is processed would be quite limited.  True understanding requires knowledge of the precise wiring diagram.  This prize is a good first step towards bringing awareness of this tricky problem to the world’s best computer scientists.

$75,000 in prize money was to go to the group that was able to produce high-quality reconstructions of neuronal structures at least 20x faster than by-hand reconstructions.  In the finals, the fastest speed achieved was 10X the by-hand method. Some groups were hindered by slight variances in the source data formatting, which normally isn’t a big deal unless you only have 20 minutes to produce as much reconstruction as possible…

Since no group was able to beat the hard floor, but substantial progress was made, the money was distributed amongst these finalists.

Badrinath Roysam Team, $25,000
“for the better overall generality of their program in producing robust reconstructions by integration of human and machines interactions.”

Armen Stepanyants Team, $25,000
“for the better overall biological results in the spirit of pure automation.”

Eugene Myers Team, $15,000
“for the excellent quality and strength of their algorithm.”

German Gonzalez Team, $10,000
“for their deeper potential, more original approach, and ultimate scalability of their proposed solution.”

Deniz Erdogmus Team
“for elevating themselves above the current state of automated reconstructions…with a deep understanding of the technical and scientific problems.”

Congrats to the placing teams.

Mice aren’t that blind

Just saw a cool informal talk from Andreas Burkhalter about the mouse visual cortex.  He has a fascinating paper, Area Map of Mouse Visual Cortex, in the Journal of Comparative Neurology, in which he identifies not just three or four areas of mouse visual cortex, but twelve! Each area has a complete map of the entire visual field.  He combines triplet injections of Di-I, Di-O and BDA as fiducial markers with a label for callosal connections. He fixes the tissue in a manner that allows the unrolling and flattening of the entire mouse cortex. This allows him to segment and show the orientation of each field in a single cortical layer in the same slice. Different layers give different patterns of projection. Given the richness of the data obtained, I’m surprised that more systems neuroscientists don’t use identical techniques.

Triple labeling of mouse visual cortex

He has also created a wiring diagram of each of these and shown that receptive field size increases with the depth in the visual system hierarchy. He also noted that although Michael Stryker finds 50% of visual cortex neurons are direction selection (when stimulated by drifting gratings), he finds only 10% are direction selective when using random dot patterns. Presumably, drifting gratings provide additional cues beyond direction of motion that confound analysis. For such a ‘blind’ creature, mice sure have a complex pattern of circuitry to process visual information.