Using Simulated Training Data of Voxel-Level Generative Models to Improve 3D Neuron Reconstruction

Area of Science:

• Neuroscience
• Computational Biology
• Medical Imaging

Background:

• Neuron morphology reconstruction is vital for neuroscience research.
• Accurate neuron segmentation from fluorescence microscopy images is challenging due to image noise and complexity.
• Deep learning has advanced segmentation, but requires extensive annotated training data, which is laborious to create.

Purpose of the Study:

• To develop a novel strategy for generating synthetic, voxel-level labeled neuron images to overcome the limitations of manual data annotation.
• To train deep learning segmentation models using these synthetic data for improved neuron reconstruction.

Main Methods:

• Utilized two-stage generative models trained on unlabeled neuron images.
• Developed a novel objective function to preserve predefined labels during synthesis.
• Synthesized realistic 3D neuron images with underlying voxel labels.

Main Results:

• Networks trained on synthetic data achieved superior performance compared to those trained on manually labeled data.
• The generated synthetic images effectively served as training data for segmentation networks.
• Segmentation results from synthetic data improved existing neuron reconstruction methods.

Conclusions:

• The proposed generative model strategy offers an efficient solution for creating large-scale, labeled training datasets for neuron segmentation.
• This approach significantly enhances the accuracy and efficiency of neuron morphology reconstruction in neuroscience.
• Synthetic data generation is a viable and powerful alternative to manual annotation for training deep learning models in biological imaging.