Large Scale Image Segmentation with Structured Loss Based Deep Learning for Connectome Reconstruction.

Area of Science:

• Neuroscience
• Computer Vision
• Biotechnology

Background:

• Accurate neuron segmentation in electron microscopy (EM) is crucial for understanding neural circuits.
• Existing methods face challenges in accuracy, scalability, and handling diverse EM data types.

Purpose of the Study:

• To develop a novel method for neuron segmentation that significantly improves accuracy and scalability.
• To enhance the efficiency and robustness of existing segmentation techniques for large EM datasets.

Main Methods:

• A 3D U-Net model predicts voxel affinities, followed by iterative region agglomeration.
• A quasi-linear method for computing loss gradients and a two-pass gradient computation were implemented.
• Structured loss based on Malis was used for topologically correct segmentations.

Main Results:

• The method achieved significant relative improvements of 27%, 15%, and 250% on diverse EM datasets.
• Simple percentile-based agglomeration outperformed complex methods on the improved predictions.
• The approach demonstrated effectiveness on both isotropic and anisotropic EM data.

Conclusions:

• The proposed method offers a scalable and accurate solution for neuron segmentation in electron microscopy.
• It achieves high throughput, processing data at ~2.6 seconds per megavoxel.
• This advancement facilitates the analysis of very large neural datasets.