Physics-guided self-supervised learning for high-fidelity reconstruction in live-cell imaging

Long-term live-cell fluorescence imaging requires low excitation intensity or short exposure times to reduce phototoxicity, which leads to insufficient signal-to-noise ratios (SNR) in the acquired data. Under such conditions, existing reconstruction methods struggle to balance noise reduction and fine structure preservation, often resulting in distorted reconstructions or severe artifacts. Here, we present HiFi-DeconvFormer, a physics-guided self-supervised reconstruction framework for super-resolution deconvolution imaging. By combining a spatial-redundancy-based training strategy with a network architecture that incorporates physical imaging models, our approach enables robust self-supervised learning without requiring ground truth data. Leveraging windowed Transformers and multi-modal regularization, the framework effectively captures long-range biological continuity and high-frequency edge features, ensuring accurate recovery of weak signals. Extensive benchmarks show that HiFi-DeconvFormer outperforms state-of-the-art methods under low-SNR conditions, achieving high-fidelity, artifact-free reconstruction of dynamic subcellular processes.