DLLP: a deep learning-based layer prediction network for three-dimensional fluorescence microscopy

High light-throughput microscopy stands as a potent tool in biological research, spanning applications from cell functional analysis to brain science. The pursuit of achieving elevated light throughput alongside rapid imaging speeds without the need for additional optical hardware has emerged as a pivotal focus in biomedical research. To address this challenge, we introduced a cutting-edge framework known as the deep learning-based layer predictor (DLLP). By integrating a convolutional neural network (CNN) with the inter-layer dynamic and morphological attention mechanism (IDMA) within a transformer architecture, DLLP employed a tomographic prediction technique capable of reducing the number of scanning layers in three-dimensional (3D) microscopy by over 70%, while maintaining light throughput and image fidelity. The DLLP framework significantly improved imaging speed and quality through two complementary strategies: robust unsupervised denoising to effectively mitigate electrical noise in the xy-plane caused by faster imaging and sparse recovery along the Z-axis to ensure high-quality 3D reconstruction without compromising spatial resolution. The DLLP framework has demonstrated consistent performance across various microscopy modalities, including stimulated emission depletion (STED) microscopy, fluorescence modulation optical sectioning tomography (FMOST) microscopy, multi-photon microscopy, and light-sheet microscopy, surpassing traditional methods and existing deep learning approaches in terms of accuracy and image quality.