Graph Abstraction for Simplified Proofreading of Slice-based Volume Segmentation
Ronell Sicat1, Markus Hadwiger1, Niloy Mitra1,2
1 King Abdullah University of Science and Technology2 University College London
Motivation
• Extract 3D structures from electron microscopy (EM) data for analysis
• Target application: Connectomics
input segmentation proofreading analysis
Input
• EM scans of mouse cortex (1024 x 1024 x 150 slices )
Segmentation
• Automatic segmentation extracts neural structures (not perfect)
Proofreading
• Search for and correct segmentation errors
Analysis
• Segmented 3D structures are visualized and analyzed
Motivation
• Proofreading – tedious and time consuming• We want abstraction of segmentation data– cheap to compute– provides search and correction support
Graph Abstraction of Segmentation Data
• Node– segmented region– center of mass
• Edge– connected regions
(same object)
Graph Abstraction of Segmentation Data
Inconsistency Weight
node distance
Inconsistency Weight
node distance
Inconsistency Weight
node distance region overlap
Inconsistency Weight
node distance region overlap
Inconsistency Weight
node distance region overlap
Inconsistency Weight
node distance region overlap
Error Visualization using Inconsistency Weights
Directing the User to Error Regions
Automatic Correction for Special Case Errors
• Fixing extensions– average bounding box is
used for clipping– more complex bounding
region can be used
before
Automatic Correction for Special Case Errors
• Fixing extensions– average bounding box is
used for clipping– more complex bounding
region can be used
before
Automatic Correction for Special Case Errors
• Fixing extensions– average bounding box is
used for clipping– more complex bounding
region can be used
after
Automatic Correction for Special Case Errors
• Fixing holes– fill hole if present in both
neighbor regions– more sophisticated
methods can be used
before
Automatic Correction for Special Case Errors
• Fixing holes– fill hole if present in both
neighbor regions– more sophisticated
methods can be used
after
Automatic Correction for Special Case Errors
• Not perfect (reduces manual effort needed)• Automatic correction (with threshold)– all threads– one thread– one node
• Manual correction can be done anytime• Proofreading tool is implemented as Avizo
plugin
Automatic Correction (single node)
Manual Correction (single node)
Automatic Correction (all nodes)
Final Result
Conclusion
• Graph abstraction of segmentation data – very cheap to compute– helps in visualization– directs user to error regions– simple but provides fast method for reducing
special case errors
Thank you!
Inconsistency Weight Equations
Segmentation Details
• Segmentation algorithm - Kaynig, V., Fuchs, T., Buhmann, J. M., Neuron Geometry Extraction by Perceptual Grouping in ssTEM Images, CVPR, 2010.
Tracing Details
• 3D tracing (Euclidean distance of region center, overlap, difference in region size, texture similarity, smooth continuation) - Kaynig, V., Fuchs, T., Buhmann, J. M., Geometrical Consistent 3D Tracing of Neuronal Processes in ssTEM Data , MICCAI, 2010.