Real-time and universal network for volumetric imaging from microscale to macroscale at high resolution

Area of Science:

• Optics and Imaging Science
• Computer Vision and Machine Learning
• Biomedical Engineering

Background:

• Light-field imaging offers broad applications in micro, meso, and macroscale imaging.
• Current deep learning methods for light-field reconstruction primarily focus on microscale applications.
• A universal reconstruction algorithm for multiscale light-field imaging is needed to advance volumetric imaging.

Purpose of the Study:

• To introduce a real-time and universal network (RTU-Net) for high-resolution light-field image reconstruction applicable to all scales.
• To address the limitation of existing methods concentrating on microscale reconstruction.
• To develop a network capable of multiscale light-field image reconstruction for enhanced volumetric imaging.

Main Methods:

• Developed RTU-Net, a novel network architecture for light-field image reconstruction.
• Employed an adaptive loss function based on generative adversarial principles to enhance generalization.
• Validated RTU-Net's performance on diverse multiscale datasets: microscale (tubulin, mitochondrion), mesoscale (synthetic mouse neuro), and macroscale (particle imaging velocimetry).

Main Results:

• RTU-Net achieved real-time, high-resolution reconstruction across a wide range of volumes (300 μm³ to 25 mm³).
• Demonstrated superior resolution compared to existing light-field reconstruction networks.
• Exhibited strong generalization capability due to its adaptive loss function.

Conclusions:

• RTU-Net is the first universal network for multiscale light-field image reconstruction.
• Its high-resolution, robust, efficient, and broadly applicable nature will significantly advance volumetric imaging.
• RTU-Net deepens insights into high-resolution and volumetric imaging across scientific domains.