12 October 2026 in Biomedical Optics Express
Structured illumination microscopy (SIM) offers advantages such as low phototoxicity and high imaging speed for live-cell super-resolution imaging. However, under low-light imaging conditions, parameters such as the spatial frequency, direction, and phase of the structured light are susceptible to external system interference, leading to estimation errors and amplified reconstruction artifacts. Although existing physics-driven and deep learning-based SIM reconstruction methods can significantly improve imaging quality under low signal-to-noise ratio conditions, deep learning networks—lacking the constraints of physical prior mechanisms—are prone to overfitting when subjected to noise interference. Meanwhile, physics-driven algorithms rely on accurate parameter estimates and exhibit insufficient reconstruction stability when test parameters deviate from training settings. To address this issue, this study proposes DART-SIM, a deep-learning-based SIM reconstruction method that is robust to structured-light parameters. Within a deep learning framework, this method establishes a physically guided mechanism with parameter robustness, reducing the model’s dependence on calibration parameters. Experimental results demonstrate that DART-SIM can effectively suppress reconstruction artifacts caused by parameter perturbations under low-light conditions, restore fine structures, and enhance the robustness and stability of SIM reconstruction.