Descattering and image restoration with a transformer-based neural network in deep tissue imaging

Area of Science:

• Biomedical Imaging
• Deep Learning
• Optical Microscopy

Background:

• Imaging deep within biological tissues is essential for understanding cellular processes.
• Light scattering significantly limits imaging depth and resolution in techniques like two-photon excitation fluorescence (TPEF) microscopy.
• Current methods often require complex optical setups to mitigate scattering.

Purpose of the Study:

• To develop a computational method for enhancing TPEF imaging depth without hardware modifications.
• To create a deep learning model capable of converting scattered TPEF images into high-quality, scattering-free images.
• To restore hidden spatial information lost due to scattering at greater depths.

Main Methods:

• A multiattention deep learning network was designed to directly map degraded TPEF images to restored images.
• The model was trained exclusively on simulated data, eliminating the need for registered real data pairs.
• The network was optimized to simultaneously descatter images and recover lost details.

Main Results:

• Quantitative evaluations on simulated data demonstrated substantial improvements in peak signal-to-noise ratio (23–29 dB) and structural similarity index (23×).
• The framework successfully visualized lipid droplets at depths up to 1,300 μm in ex vivo samples.
• Clear visualization of vascular structures (up to 950 μm) and astrocytes (up to 500 μm) was achieved in live mouse brains.

Conclusions:

• The proposed multiattention network effectively overcomes light scattering limitations in TPEF imaging.
• This computational approach significantly extends the practical imaging depth for TPEF microscopy.
• The method offers a powerful, hardware-independent solution for deep biological tissue visualization.