Deep-Learning-Based Automated Neuron Reconstruction From 3D Microscopy Images Using Synthetic Training Images

Area of Science:

• Neuroscience
• Computational Biology
• Image Analysis

Background:

• Quantitative investigation of brain circuits and functions relies on accurate digital reconstruction of neuronal structures from 3D microscopy images.
• Automatic neuron reconstruction methods are crucial for overcoming the challenges associated with manual reconstruction and enabling large-scale analysis.

Purpose of the Study:

• To propose and evaluate SPE-DNR, a novel deep learning-based method for automatic neuron reconstruction from 3D microscopy images.
• To enhance the robustness and accuracy of neuron reconstruction by incorporating spherical-patches extraction and deep neural networks.

Main Methods:

• SPE-DNR combines spherical-patches extraction (SPE) with deep neuron reconstruction (DNR) using 2D Convolutional Neural Networks (CNNs).
• The method utilizes intensity distribution features from SPE to determine tracing directions and classify voxels, enabling automatic tracing of neurite centerlines from seed points.
• A synthetic image generation scheme was developed to create realistic training data, simulating 3D microscopy conditions and structural defects.

Main Results:

• SPE-DNR demonstrated robust performance in tracing neurite centerlines and determining tracing termination points.
• The method achieved competitive results compared to state-of-the-art neuron reconstruction techniques across three diverse datasets.
• Testing on 67 real 3D neuron microscopy images validated the applicability and generalizability of SPE-DNR.

Conclusions:

• SPE-DNR offers an effective and automated solution for neuron reconstruction from 3D microscopy data.
• The developed image synthesizing scheme improves the realism of training data, enhancing the method's performance.
• This approach holds significant potential for advancing quantitative investigations of neural circuits and brain functions.