Artifact-suppressed and adaptive self-inspired learning denoising for super-resolution fluorescence microscopy

Self-supervised denoising methods have advanced fluorescence microscopy by improving photon efficiency and eliminating the need for paired training data, enabling fast and long-term live-cell imaging. However, most computational approaches still struggle with background artifacts that compromise biological interpretation. These artifacts—often amplified by normalization strategies in deep learning approaches—manifest as background inconsistencies and fabricated structures that distort relative intensity relationships. Conventional background removal procedures, whether applied as pre- or post-processing steps, prove inadequate when dealing with diverse biological samples that exhibit varying signal distributions and complex background characteristics. Here, we introduce adaptive-Self-inspired Noise2Noise (SN2N), an adaptive-SN2N framework that employs risk-aware adaptive normalization and Gaussian-weighted overlap inference to suppress artifacts. Adaptive-SN2N was validated on both structured illumination and spinning-disk confocal-based structured illumination microscopes, demonstrating improvement in photon efficiency and effective suppression of background artifacts. We further show its potential to improve downstream segmentation accuracy, particularly for challenging biological structures with sparse distributions and dynamic processes. We anticipate this advancement will enable more robust and accurate biological interpretations across a wide range of imaging conditions.