Reconstruction of structured illumination microscopy for live imaging in low light with lightweight neural networks

Structured illumination microscopy (SIM) as a type of super-resolution optical microscopy technique has been widely used in the fields of biophysics, neuroscience, and cell biology research. However, this technique often requires high-intensity illumination and multiple image acquisitions to generate a single high-resolution image. This process not only significantly reduces the imaging speed, but also increases the exposure time of samples to intense light, leading to increased phototoxicity and photobleaching issues, especially prominent in live cell imaging. Here, we propose a lightweight Multi-Convolutional UNet (MCU-Net) aiming to maintain efficient super-resolution reconstruction performance by reducing the model parameter quantity. The algorithm integrates multiple convolutional techniques with multi-scale attention mechanisms, enhancing the model's sensitivity to information at different scales and improving its precise recognition ability for image textures and structures, thus enabling high-quality super-resolution reconstruction even under low-light conditions. The overall performance of the model is evaluated in terms of efficiency and accuracy, comparing MCU-Net with deep neural network models (UNet, ScUNet, EDSR, DFCAN) and traditional reconstruction algorithms (Wiener, HiFi, TV) across different cell types, lighting intensities, and various test sets. Experimental results show that compared to other deep learning models, MCU-Net achieves a 12.66% improvement in MS-SSIM and a 50.79% increase in NRMSE index. Its prediction accuracy remains stable even in the presence of low signal-to-noise ratio inputs. Furthermore, it strikes an optimal balance between reconstruction speed and model accuracy, with a 76.10% improvement in inference speed compared to the DFCAN model.