Fast light-field 3D microscopy with out-of-distribution detection and adaptation through conditional normalizing...

Area of Science:

• Biomedical imaging
• Neuroscience
• Computational microscopy

Background:

• Real-time 3D fluorescence microscopy is vital for analyzing live biological processes, including neural activity.
• The eXtended field-of-view light field microscope (XLFM) offers rapid 3D data acquisition but suffers from slow traditional reconstruction methods.
• Existing neural network approaches accelerate reconstruction but lack methods to verify the realism of the output.

Purpose of the Study:

• To develop a fast and certifiable 3D reconstruction method for XLFM data.
• To enable real-time spatiotemporal analysis of neural activity in live organisms.
• To address the limitations of speed and realism verification in current XLFM reconstruction techniques.

Main Methods:

• A novel conditional normalizing flow architecture was developed for 3D volume reconstruction.
• The method was trained on a small dataset (50 image-volume pairs) of zebrafish neural activity.
• Reconstruction speed and the ability to detect out-of-distribution samples were evaluated.

Main Results:

• The proposed method achieves 3D reconstructions at 8 Hz for volumes of 512x512x96 voxels.
• Training the model requires less than two hours, demonstrating computational efficiency.
• The normalizing flow enables likelihood computation for sample certification, distinguishing in-distribution from out-of-distribution data.

Conclusions:

• The conditional normalizing flow provides a fast and certifiable solution for 3D XLFM reconstructions.
• This technique significantly enhances the potential for real-time spatiotemporal analysis of neural dynamics.
• The ability to certify reconstructions is crucial for reliable biomedical applications.