LFCD-Net: physics-inspired 3D reconstruction architecture for miniature light-field microscopy

Abstract Miniature light-field microscopy is a vital tool for high-speed volumetric imaging in freely moving animals, yet its spatial resolution is constrained by the insufficient sampling inherent in simultaneous spatial-angular acquisition. To address this, we propose a physics inspired light-field characteristic driven 3D reconstruction network integrating three core innovations: spatial-angular feature blocks for aliasing suppression, multi-scale feature blocks for structural fidelity, and a physics-inspired adaptive weighting loss to ensure high-quality reconstruction of sparse biological signals. These modules tackle the sampling deficiency to effectively restore spatial resolution. Crucially, we developed the first miniature dual-path imaging system to provide high-quality paired training data. Simulations demonstrate that our method achieves a lateral resolution of 5.69   μ m , representing a 22% improvement over state-of-the-art deep learning methods and significantly outperforming traditional physics-based algorithms. Validations on complex biological structures and physical experiments across a large 500   μ m depth-of-field confirm superior axial stability and reconstruction accuracy, providing a reliable technical pathway for high-resolution in vivo 3D imaging.