Physics-aware measurement-supervised deep learning enables point spread function inversion in soft X-ray tomography

Abstract

Soft X-ray tomography (SXT) is an emerging modality for whole-cell 3D imaging in near-native states. However, the effective spatial resolution is limited by optical artifacts characterized by the point spread function (PSF). Standard reconstruction methods force a compromise between structural sharpness and noise, failing to fully resolve these depth-dependent artifacts. By embedding experimentally measured, depth-variant PSFs into a differentiable forward model, we demonstrate a physics-aware computational optimization that bypasses these limitations to recover high-frequency cellular ultrastructure. The structural fidelity was validated using split-tilt Fourier ring correlation (FRC), alongside an experimental bead phantom tomogram, providing supporting evidence that the recovered high-frequency features reflect genuine specimen structure rather than fabricated artifacts. Our method effectively increases FRC spatial resolution and recovers cellular ultrastructure. Furthermore, under sparse-angular subsampling, the framework maintained spatial resolution using half the projection angles, a computational proxy pointing toward the potential for reduced radiation exposure in future acquisitions. This hardware-free, computational approach offers a route toward mitigating the optical and dosimetric constraints that currently limit nanoscale soft X-ray tomography.