Three-dimensional virtual refocusing of fluorescence microscopy images using deep learning.

Area of Science:

• Biophysics
• Neuroscience
• Computational Imaging

Background:

• 3D fluorescence microscopy often faces challenges with limited depth-of-field, sample drift, and aberrations.
• Conventional methods for 3D imaging require axial scanning, which can be slow and may compromise resolution or speed.
• Accurate neuronal activity imaging in 3D is crucial for understanding complex biological processes.

Purpose of the Study:

• To develop a computational method for digitally refocusing 2D fluorescence images into 3D.
• To enhance the depth-of-field in fluorescence microscopy without additional hardware or compromising imaging speed.
• To correct for sample drift and aberrations in post-acquisition processing.

Main Methods:

• A deep neural network, termed Deep-Z, was trained to virtually refocus images.
• The method processes time sequences of 2D fluorescence images acquired at a single focal plane.
• Deep-Z digitally increases the depth-of-field by 20-fold and corrects for aberrations and sample movement.

Main Results:

• Demonstrated 20-fold digital increase in depth-of-field for 3D imaging of neuronal activity in *C. elegans*.
• Achieved aberration correction, including sample drift and tilt, digitally after image acquisition.
• Enabled cross-modal registration, allowing 3D refocusing of wide-field images to match confocal microscopy data.

Conclusions:

• Deep-Z offers a powerful computational approach to overcome limitations in standard 3D fluorescence microscopy.
• The method significantly improves volumetric imaging speed and reduces post-acquisition challenges.
• Deep-Z has broad potential applications in neuroscience and other fields requiring high-resolution 3D imaging.