High-fidelity, generalizable light-field reconstruction of biological dynamics with physics-informed meta neural representation

Abstract

Fourier light field microscopy (FLFM) has made a name for itself by providing high-speed volumetric imaging of dynamic processes through 2D light field snapshot. However, the inverse reconstruction in FLFM, characterized by a limited-angle tomography 3D reconstruction challenge, suffers the artifacts and axial elongation. Recently, the implicit neural representation (INR) has emerged as a powerful paradigm in 3D reconstruction due to its continuous signal representation, facilitating arbitrary views synthesis with geometry consistency. In this study, we report a novel INR-based FLFM 3D reconstruction method, termed LF-INR, that leverage the inherent continuity from the inductive bias of INR to eliminate the defects of sparse-view reconstruction and enhance the structural fidelity in a self-supervised learning manner. Additionally, an optimized sampling strategy and transfer-learning are adopted to mitigate the internal limitations of conventional INR approach in generalization and optimization speed. We demonstrate its utility on diverse biological simulation data, which shows LF-INR achieves artifacts-free 3D reconstruction and ∼1.4 folds structural fidelity improvement compared with deconvolution approach, as well as ∼10 folds computational acceleration than conventional INR-based model.