High-fidelity single-frame computational super-resolution using signal-preserving denoising-enabled deconvolution.

Area of Science:

• Microscopy and Imaging Technologies
• Computational Biology
• Biophysics

Background:

• Computational super-resolution (SR) methods offer nanoscale imaging from standard microscopy but struggle with noise, artifacts, and generalizability.
• Existing statistical SR methods are prone to noise and artifacts, while deep learning approaches often lack broad applicability.

Purpose of the Study:

• To introduce 3Snet-CLID, an advanced computational SR method designed to overcome the limitations of current techniques.
• To enable high-fidelity nanoscale imaging with improved resolution and signal preservation.

Main Methods:

• Integration of a hybrid supervised/self-supervised deep learning network for signal-preserving denoising.
• Application of direct Richardson-Lucy deconvolution for image restoration.
• Implementation of a per-pixel denoising strategy to suppress noise and maintain signal distribution.

Main Results:

• Achieved over 5-fold resolution improvement on conventional microscopes.
• Successfully visualized nanoscale structures like mitochondrial outer membranes, endoplasmic reticulum, and nuclear pores in live and fixed cells.
• Demonstrated enhanced robustness and mitigation of artifacts compared to traditional methods.

Conclusions:

• 3Snet-CLID overcomes key bottlenecks in computational SR, offering superior denoising and resolution enhancement.
• The method provides an accessible platform for high-fidelity nanoscale live-cell imaging without specialized hardware.
• This approach advances the capabilities of conventional microscopy for detailed biological structure visualization.