Recurrent neural network-based volumetric fluorescence microscopy.

Area of Science:

• Microscopy and imaging science
• Biophysics
• Computational biology

Background:

• Volumetric imaging is crucial in physical, medical, and life sciences.
• Standard fluorescence microscopy often requires extensive axial scanning, limiting speed and efficiency.
• Current 3D scanning microscopy tools have inherent limitations in speed and data acquisition.

Purpose of the Study:

• To develop a deep learning framework for efficient volumetric image reconstruction from sparse 2D fluorescence microscopy data.
• To extend the depth-of-field and reduce the number of axial scans needed for 3D imaging.
• To demonstrate the framework's versatility across different imaging conditions and modalities.

Main Methods:

• A recurrent convolutional neural network, termed Recurrent-MZ, was developed for volumetric image inference.
• The framework incorporates sparse 2D fluorescence images captured at arbitrary axial positions.
• Experiments were conducted on C. elegans and nanobead samples using a 63×/1.4NA objective lens.

Main Results:

• Recurrent-MZ significantly increased the depth-of-field and reduced axial scans by 30-fold.
• The network demonstrated resilience to varying imaging conditions, including input image sequences and axial positioning errors.
• Successful wide-field to confocal cross-modality image transformations and 3D reconstructions were achieved.

Conclusions:

• Recurrent-MZ represents the first application of recurrent neural networks for microscopic image reconstruction.
• The framework offers a flexible and rapid solution for volumetric imaging, overcoming limitations of traditional methods.
• This approach enhances the capabilities of fluorescence microscopy for detailed 3D sample analysis.