Module 05: From Pixels to Proofreading: Image Segmentation and Quality Control
Discover how neural structures are segmented from raw EM images and how humans proofread to ensure accuracy.
🧠 What is Segmentation?
Segmentation is the process of identifying and labeling structures in an image. In connectomics, this means separating each neurite (axon or dendrite) and assigning it a unique label. Most current approaches rely on deep learning to automate this step.
- Flood-filling networks
- U-Net and 3D CNNs
- Segmentation errors: mergers and splits
🛠️ Proofreading 101
No segmentation algorithm is perfect. Humans review and correct the machine-generated output, a process known as proofreading. Tools like Neuroglancer allow users to inspect 3D reconstructions slice-by-slice and validate continuity.
- Common proofreading errors and their consequences
- Using visual cues to spot mergers/splits
- Basic workflow: select, inspect, edit
🔬 Quality Metrics and Feedback Loops
Segmentation quality can be quantified using metrics like Rand score and edge accuracy. Proofread corrections can also be used to retrain models, creating a virtuous cycle of improvement.
- Manual vs. automated QA
- Using corrections to improve models
- Tracking proofreading contributions
🎯 COMPASS Integration
- Knowledge: Understanding EM segmentation outputs
- Skills: Visual discrimination, 3D spatial reasoning, data quality assessment
- Character: Persistence, humility, teamwork
- Meta-Learning: Adapting to evolving tools and methods
📚 References & Resources
- Januszewski et al., 2018. High-precision automated reconstruction of neurons with flood-filling networks. Nature Methods.
- Neuroglancer: github.com/google/neuroglancer
- SNEMI3D Benchmark: cremi.org
- BossDB Cookbook: Uploading Image Stacks
- Notebook: Image and Segmentation Download (source)
✅ Assessment
- Define and explain the purpose of segmentation in connectomics
- Correct a sample proofreading task using Neuroglancer
- Describe how proofreading improves final circuit reconstructions