DMF-SIM: Dual-Model Framework for super-resolution reconstruction and denoising in Structured Illumination Microscopy

Super-Resolution Structured Illumination Microscopy (SR-SIM) is usually challenged by the balance between imaging quality and continuous imaging, especially in low-light imaging conditions, where obtaining a high Signal-to-Noise Ratio (SNR) is even tougher due to the intrinsic mechanism of sensor noise and detection limitations. In more recent studies, a promising solution to overcome this limitation is a combination of traditional methods with deep neural networks. However, the incorporation of extrinsic optical information weakens robustness across different imaging conditions due to complex processing requirements. Additionally, the high computational cost and need for extensive training data limit the overall performance from reaching further improvements. Motivated by the complementary strengths of vision transformers and channel attention mechanisms, in this paper, we propose a Dual-Model Framework in Structured Illumination Microscopy (DMF-SIM) to achieve simultaneous super-resolution reconstruction and denoising by integrating a reconstruction model (SIM-Rec) and a denoising model (SIM-Den). DMF-SIM integrates super-resolution and denoising models by effectively combining vision transformer and channel attention mechanisms, and a substantial performance improvement has been achieved over single-task models. The experiments also demonstrate enhanced image quality and noise reduction across various imaging conditions.