Understanding the intricate and rapidly evolving dynamics within cells is crucial for advancements in life sciences research. One pivotal aspect is the ability to quickly and accurately capture structural changes within living cells, which is instrumental in unraveling biological processes and investigating pathologies. Structure Illumination Microscopy (SIM) emerges as a powerful tool in dynamically observing living samples due to its ability to surpass the optical diffraction limit, its labeling of samples does not require specific fluorescent dyes. and its minimal phototoxicity.

Moreover, the integration of neural networks adds a new dimension to this field. Neural networks excel in swiftly and comprehensively extracting complex features from images, offering a promising approach to capture inherent structural characteristics. Particularly in challenging low signal-to-noise imaging scenarios, neural networks can mitigate the limitations of traditional SIM reconstruction methods, addressing issues such as poor noise robustness. However, the current neural networks are limited to their low computational speed.

In response to these challenges, this article proposes a novel deep learning based method for super-resolution SIM reconstruction at the video-level, termed VDL-SIM. This method combines the strengths of U-Net and residual networks, optimizes channel numbers through pruning, and utilizes C++ acceleration for efficient computation. By doing so, VDL-SIM achieves remarkable imaging speeds of up to 47 frames per second (fps), 15 to 33 times faster compared to existing deep learning based SIM algorithms.

One of VDL-SIM's notable features is its robust reconstruction capability, even in scenarios with low signal-to-noise (SNR) levels. This attribute is particularly advantageous as it significantly reduces sample damage during imaging. These characteristics position VDL-SIM as a valuable tool for video-level super-resolution imaging, complementing traditional algorithms in challenging imaging conditions.

In the future, further improvements in imaging quality can be achieved by combining mathematical priors with physical imaging models. This approach will enable more flexible and diverse applications in the field of super-resolution imaging, making it a topic worthy of further in-depth research.

Link: Video-level and high-fidelity super-resolution SIM reconstruction enabled by deep learning