4D self-supervised learning of cross-dimensional representation enables high-performance denoising of volumetric fluorescence imaging

Volumetric time-lapse fluorescence microscopy enables panoramic visualization of biological structures and dynamics in living organisms. However, noise fundamentally limits the fidelity and quantitative analysis of fluorescence recordings. Current self-supervised denoising methods do not fully model the four-dimensional (4D) correlations of volumetric time-lapse data, thereby constraining restoration quality. Here we present SDT-4D, the first 4D self-supervised denoising framework to jointly incorporate lateral, axial, and temporal dependencies without requiring any clean images for training. We designed a unified 4D transformer architecture to capture long-range dependencies across 4D hypervoxels. This design enables SDT-4D to restore biological signals while preserving volumetric continuity and temporal dynamics. To improve reliability, we further developed a physics-informed uncertainty quantification (PhUQ) method and embedded it into SDT-4D, deriving pixel-wise confidence maps of restored images. In volumetric calcium imaging, SDT-4D achieves state-of-the-art denoising performance and reveals neuronal anatomy and calcium transients. In cellular morphodynamic imaging, SDT-4D restores cellular structures obscured by noise and facilitates downstream analysis of immune cell migration and glial morphology evolution. Together, these results demonstrate the effectiveness and reliability of SDT-4D for restoring and interpreting photon-limited biological dynamics.