UT-SIM: A transformer-based feature fusion network for super-resolution reconstruction in structured illumination microscopy

We propose UT-SIM, a novel deep learning framework that integrates transformer-based feature fusion into a U-Net architecture for super-resolution reconstruction in structured illumination microscopy (SIM). To further enhance reconstruction performance, UT-SIM incorporates Channel and Spatial Attention Modules (CSAM), which dynamically reweight feature maps across multiple scales. This attention mechanism enables the model to focus more effectively on critical spatial and channel-level information. Compared with conventional networks such as SRCNN, U-Net, and ML-SIM, UT-SIM achieves significant improvements in image quality, with structural similarity index (SSIM) gains of approximately 0.3–0.6 and peak signal-to-noise ratio (PSNR) improvements of 1–2 dB. Additionally, we propose a synthetic dataset generation method that enables the creation of tailored datasets for various SIM systems. Models trained with this simulated data demonstrate robust transfer learning capabilities, delivering high-quality reconstructions across diverse SIM platforms.