UNI-EM

UNI-EM automates segmentation of neuronal electron microscopic (EM) images using deep neural networks (DNNs) to enable three-dimensional reconstruction of neuronal networks for micro-connectomics.

Key Features:

• End-to-end segmentation pipeline: Provides ground truth generation, model training, inference, postprocessing, proofreading, and visualization for CNN-based EM image segmentation.
• Convolutional neural network architectures: Implements 2D CNN architectures including U-Net, ResNet, HighwayNet, and DenseNet for image-level segmentation tasks.
• 3D segmentation with flood-filling networks: Incorporates flood-filling networks (FFNs) to perform 3D neuron segmentation from EM image stacks.
• 3D reconstruction from 2D EM stacks: Supports reconstruction of three-dimensional neuronal structures from two-dimensional EM image stacks.
• Example workflows: Includes example workflows for mitochondria segmentation using 2D CNNs and neuron segmentation using FFNs.

Scientific Applications:

• Micro-connectomics: Automated segmentation to reconstruct neuronal networks from EM images for studies of brain connectivity.
• Mitochondria segmentation: Identification and delineation of mitochondria in EM images using 2D CNNs.
• Neuron 3D reconstruction: 3D neuron segmentation and reconstruction from EM stacks using FFNs and CNNs.

Methodology:

Uses deep learning methods including 2D convolutional neural networks (U-Net, ResNet, HighwayNet, DenseNet) and 3D flood-filling networks (FFNs) together with ground truth generation, model training, inference, postprocessing, proofreading, and visualization.