DeepMulticut: Deep Learning of Multicut Problem for Neuron Segmentation From Electron Microscopy Volume.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Superpixel aggregation is crucial for automated neuron segmentation in electron microscopy (EM) volumes.
• Existing graph partitioning methods involve separate model estimation and solving stages, leading to inherent model errors.

Purpose of the Study:

• To propose an end-to-end deep learning framework, DeepMulticut, for neuron segmentation.
• To integrate model estimation and solving stages to overcome limitations of existing methods.

Main Methods:

• Developed DeepMulticut, a deep learning framework for the minimum cost multicut problem.
• Relaxed the NP-hard multicut problem to a continuous Soft-GAEC algorithm for differentiability.
• Integrated Edge-CNNs as edge cost estimators within a differentiable multicut optimization system.
• Utilized a decision-oriented loss for adaptive discriminative feature learning in Edge-CNNs.

Main Results:

• Demonstrated the effectiveness of the DeepMulticut framework on three public EM datasets.
• Showcased adaptive discriminative feature learning by feeding decision quality back to Edge-CNNs.
• Achieved improved neuron segmentation by directly optimizing partitioning decisions.

Conclusions:

• DeepMulticut offers an effective end-to-end solution for neuron segmentation in EM volumes.
• The framework successfully integrates deep learning with combinatorial optimization for enhanced segmentation.
• The approach addresses inherent model errors in traditional two-stage methods.