The EyeWire structure
E2198
This part of the documentation is meant to describe some of the terms used in EyeWire and the API. If you are looking for more detailed information about how the dataset was obtained, feel free to check out the wiki pages.
The dataset is 4864 x 20992 x 13056 voxels large. A voxel is not a perfect cube, but has a size of 16.5 x 16.5 x 23 nm. All dimensions are based on EyeWires right-handed coordinate system (see image below)
Volume
The datasets consists of volumes. Each EyeWire volume for example is 256x256x256 voxels large and the overlapping region of two adjacent EyeWire volumes is 32 voxel rows large. Volumes are generated as necessary, i.e. a newly traced cell extends to a yet unexplored region.
Also note that the dataset is slightly tilted and some regions therefore don’t have any data/volumes.
Chunk
Each volume consists of smaller chunks (128^3 voxels). They don’t overlap.
Channel & Segmentation Images
There are two kinds of images in those subvolumes, depending on the volume type:
- The EM-images, compressed to 8 bit greyscale .jpg files and
- the 32 bit .png segmentation files
Colors in the segmentation images should be read as an unsigned integer value and refer to the segments ID.
Segment
All connected pixels of the same value are one segment. Note that even though volumes (and therefore segments) overlap, their segment ID is not necessarily the same!
Task
By choosing one or more identified segments in the overlapping regions of two volumes, it is possible to create a new task for players. A volume might contain many different tasks, sometimes even tasks of the very same neuron.
Validation
Finally, each player contributes his or her own solution for the given task. This validation is compared to those of other players. Segments traced by the majority of players form the consensus and are used to spawn new tasks.