Implicit Reconstruct Spatiotemporal Super-Resolution Microscopy in Arbitrary Dimension

ICLR 2026 Conference Submission1601 Authors

Keywords: Biological Imaging, Fluorescence Microscopy, Implicit Neural Representation, Arbitrary Scale Super-Resolution, 4D Medical Imaging

TL;DR: We introduce SpatimeINR, an implicit neural representation framework that achieves arbitrary scale spatiotemporal super-resolution for 4D imaging data while effectively recovering fine spatiotemporal details in complex dynamic scenarios.

Abstract: High-resolution 4D fluorescence microscopy imaging, essential for deciphering dynamic biological processes, is typically challenged by insufficient spatiotemporal resolutions, including restricted t-axis sampling density to prevent photobleaching and issues with anisotropic resolution. To address these challenges, we propose an implicit neural representation-based arbitrary scale super-resolution framework, termed SpatimeINR, which leverages spatiotemporal latent representation in conjunction with a multilayer perceptron for 4D rendering, while incorporating cycle-consistency loss to ensure fidelity with the original data. Extensive experiments on lung cancer cell and C.elegans cell membrane fluorescence datasets demonstrate that our approach can accurately reconstruct the nonlinear dynamic motion of biological samples along the time axis and significantly outperforms state-of-the-art methods in both temporal and spatial (4D) super-resolution tasks.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 1601