Deep learning-enhanced light-field imaging with continuous validation.

Area of Science:

• Biophysics
• Microscopy
• Computational Biology

Background:

• High-speed, 3D visualization of dynamic biological processes is crucial.
• Light-field microscopy (LFM) enables fast volumetric imaging but faces challenges in reconstruction speed and accuracy.
• Existing LFM reconstruction is computationally intensive and prone to artifacts, limiting its biological applications.

Purpose of the Study:

• To develop an artificial intelligence-enhanced microscopy framework for faster and more accurate 3D volumetric imaging.
• To integrate deep learning with a hybrid light-field and light-sheet microscope.
• To overcome the computational bottlenecks and artifacts associated with traditional LFM reconstruction.

Main Methods:

• Developed a hybrid light-field light-sheet microscope.
• Implemented a deep learning-based volume reconstruction framework using convolutional neural networks (CNNs).
• Utilized simultaneously acquired 2D light-sheet images for continuous training and validation of the CNN during imaging.

Main Results:

• Achieved high-quality 3D reconstructions at video-rate throughput.
• Demonstrated refinement of reconstructions using high-resolution light-sheet images.
• Successfully imaged dynamic processes in medaka heart and zebrafish neural activity at rates up to 100 Hz.

Conclusions:

• The AI-enhanced microscopy framework significantly improves the speed and quality of volumetric imaging.
• This approach overcomes key limitations of traditional LFM, enabling advanced biological research.
• The system offers a powerful tool for high-throughput, dynamic 3D imaging in life sciences.