Learning Context Cues for Synapse Segmentation


We propose an approach designed to take such contextual cues into account and emulate the human ability to distinguish synapses from regions that merely share a similar texture.We run various filters over the EM stack but compute our features over arbitrarily sized cubes placed at arbitrary locations inside an extended neighborhood of the voxel to be classified. The top row below corresponds to a synapse which can be identified by the presence of a cleft, pre-synaptic vesicles and a post-synaptic density. Bottom row is a subsurface cister which shares the same texture as synaptic clefts but can be excluded due to the surrounding context.


The resulting classifier is thus highly flexible, able to utilize context from a high variety of regions in the neighborhood of the voxel of interest. Our 3D detector is pose-indexed so that the computed features are extracted at consistent locations relative to the orientation of the synaptic cleft.


Qualitative Results

Comparing to State-of-the-art
Our synapse segmentation framework shown slice by slice as compared to the competing method. Detected voxels are shown in red. Both detectors operate in 3D however by focusing on the neighboring context (for ex. presence of absence of vesicles), our method achieves a far lower false alarm rate for the same detection rate.

A segemented synapse
An example synapse segmented using our framework. Note the two punctures which favor connectivity and signaling.

Segmenting a full EM stack and extracting statistics
A fully segmented EM stack along with extracted morphology statistics, which can be used by neuroscientists to characterize brain tissues.



C. Becker, K. Ali, G. Knott and P. Fua
PREPRINT - IEEE Transactions on Medical Image Analysis, 2013
C. Becker, K. Ali, G. Knott and P. Fua
In Medical Image Computing and Computer Assisted Intervention Conference, 2012.