Large Scale Image Segmentation with Structured Loss Based Deep Learning for Connectome Reconstruction
Problem
Reconstructing connectomes requires accurately segmenting neurons in large-scale electron microscopy images, which existing methods could not do with sufficient accuracy and scalability.
Why it matters
Connectome maps are foundational to understanding neural circuitry and brain function.
What the AI did
A 3D U-Net predicts affinities between voxels, trained with a structured Malis-based loss, followed by learning-free percentile-based agglomeration to segment neurons.
Results
Significant improvement over state of the art in accuracy and scalability on three diverse EM datasets; relative improvement numbers truncated in abstract.
Impact
Advances accurate, scalable neuron segmentation for connectome reconstruction across three datasets.
Record
| Venue | IEEE Transactions on Pattern Analysis and Machine Intelligence |
|---|---|
| Published | 2018 |
| Work type | article |
| Authors | Jan Funke, Fabian Tschopp, William Grisaitis, Arlo Sheridan, Chandan Singh, Stephan Saalfeld, Srinivas C. Turaga |
| Citations | 220 (OpenAlex, at import time) |
| References | 30 |
| Model or method | 3D U-Net with structured Malis loss |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1109/tpami.2018.2835450 |
| Open access | bronze |
| Keywords | connectomics, electron microscopy, neuron segmentation, 3D U-Net, deep learning |
Importance rating
significant — thousands to millions affected
Connectome mapping is foundational to understanding neural circuitry and brain function; method enables processing of very large EM datasets, but abstract gives no scale of affected population.