Recovery of 3D particles distribution from digital hologram using a one-stage detection network

3D micro-particle reconstruction with high accuracy and low latency is an ambitious and essential task, with various applications such as investigating micro-particle dynamics, microorganism swimming and bio-microfluidics. Without using any focusing optics, digital holography (DH) is a high throughput and compact imaging tool that can encode the depth information of the object in the form of a two-dimensional (2D) gray-scale interference pattern. Conventional reconstruction methods perform auto-focusing and reconstruction from slice to slice along the axial direction. Such computationally expensive and time-consuming iterative processes impede their use in real-time monitoring. In this study, a deep learning based one-stage detector is developed for 3D particles distribution recovery from a digital hologram. Specifically, the proposed model is able to detect and localize target particles in a volume in the way of regressing the bounding boxes and the associated depth from a single pass through the neural network, which eschews the traditional pipeline designs. The unified architecture is optimized in an end-to-end training, and the prediction is made at the image pixels level, which challenges the overlapping issue as the particle concentration increases. In order to verify the feasibility of our proposed method, thorough evaluations are presented on different particle concentrations and noise levels. The results demonstrate that our proposed approach can achieve remarkable and robust performance in detecting position accuracy and extraction rate with improved processing speed. This model facilitates the analysis of the dynamic displacements and motions for micro-particles or cells, and can be further extended to various types of computational imaging problems sharing similar traits.