Self-Supervised Structured Illumination Microscopy Image Denoising Based on Multiresolution Analysis Resampling

In live-cell imaging under low-light conditions, acquiring images with low signal-to-noise ratios (SNR) presents a major challenge for structured illumination microscopy (SIM), limiting its effectiveness in super-resolution imaging and subcellular process investigation. The difficulty of obtaining high-SNR images, coupled with the scarcity of sample data in low-photon environments, impedes the application of deep learning-based approaches for SIM image denoising and reconstruction. To address these limitations, we propose MRS-SIM, a self-supervised denoising method based on multiresolution analysis (MRA) resampling, which enables high-fidelity, artifact-free reconstruction without requiring high-SNR references. We systematically evaluate MRS-SIM through both simulated experiments across varying noise levels and real-world microscopy data at different photon counts. Compared with HiFi-SIM, a typical SIM reconstruction baseline, MRS-SIM achieves up to 5.22 dB peak signal-to-noise ratio (PSNR) and 0.32 structural similarity index (SSIM) improvement under extremely low SNR (-12 dB), while maintaining lower variability across trials. Furthermore, ablation studies confirm the critical role of multiresolution analysis resampling in enhancing denoising accuracy, validating the contributions of key components. With its ability to overcome noise-related challenges and facilitate high-quality image reconstruction in low-light conditions, MRS-SIM offers a robust and efficient solution for low-photon live-cell imaging, particularly in applications requiring the preservation of fluorescence integrity for subcellular dynamics studies.