Holographic image reconstruction with phase recovery and autofocusing using recurrent neural networks

Abstract

Holographic imaging plays an essential role in label-free microscopy techniques, and the retrieval of the phase information of a specimen is vital for image reconstruction in holography. Here, we demonstrate recurrent neural network (RNN) based holographic imaging methods that simultaneously perform autofocusing and holographic image reconstruction from multiple holograms captured at different sample-to-sensor distances. The acquired input holograms are individually back propagated to a common axial plane without any phase retrieval, and then fed into a trained RNN which successfully reveals phase-retrieved and auto-focused images of the unknown samples at its output. As an alternative design, we also employed a dilated convolution in our RNN design to demonstrate an end-to-end phase recovery and autofocusing framework without the need for an initial back-propagation step. The efficacy of these RNN-based hologram reconstruction methods was blindly demonstrated using human lung tissue sections and Papanicolaou (Pap) smears. These methods constitute the first demonstration of the use of RNNs for holographic imaging and phase recovery, and would find applications in label-free microscopy and sensing, among other fields.