Retaining spatial resolution multifocal confocal fluorescence microscopy with deep learning.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Microscopy Techniques

Background:

• Confocal microscopy offers high resolution but slow acquisition speeds.
• Multifocal illumination increases speed but reduces spatial resolution.
• A gap exists between imaging speed and resolution in volumetric biological sample imaging.

Purpose of the Study:

• To develop a deep learning method for multifocal confocal microscopy.
• To achieve faster image acquisition without compromising spatial resolution.
• To address the inherent speed-resolution trade-off in multifocal confocal microscopy.

Main Methods:

• Implemented an image-to-image translation model using modified U-Net, ResU-Net, and Attention U-Net architectures.
• Trained and tested the models on paired experimental datasets of biological samples.
• Utilized conventional confocal images as groundtruth and multifocal images as input.

Main Results:

• The modified Attention U-Net significantly improved image quality and structural detail retention.
• Attention U-Net achieved higher peak Signal-to-Noise Ratio (32.83 dB) and Structural Similarity Index Measure (0.935).
• Spatial frequency analysis confirmed superior preservation of low and high-frequency information compared to U-Net.

Conclusions:

• Deep learning, particularly Attention U-Net, effectively matches traditional confocal imaging quality while increasing speed.
• The developed approach successfully addresses the speed-resolution trade-off in multifocal confocal microscopy.
• This deep learning integration shows significant potential for diverse confocal imaging applications.