Artifact Reduction in Phase Imaging by the Transport of Intensity Equation Using a Convolutional Neural Network

In quantitative phase imaging (QPI), various methodologies are employed to generate three-dimensional representations of samples, particularly in the field of biology, facilitating the tracking of cellular parameters like volume. A recent advancement in QPI involves the utilization of partially coherent light to mitigate coherent noise, coupled with the solution of the phase through solution of the (partial differential) transport of intensity equation (TIE), a rapid and easily implemented technique. Despite the numerous advantages associated with phase imaging based on TIE, conventional solution methods often produce significant phase artifacts when samples extend beyond the field of view due to a lack of unknown boundary values. To address this issue, we have trained a convolutional neural network (CNN) to produce solutions to the TIE using a dataset comprising conventional TIE phase images verified through digital holographic microscopy. We subsequently compared CNN's predictions with those of the conventional TIE solver, particularly focusing on cases where the conventional solutions suffered from artifacts. We additionally assessed the performance of the trained model using images obtained in a Raman microscope, a system from which TIE images suffer from artifacts due to poor alignment. Our findings demonstrate that CNN's phase predictions outperform those of the TIE solver, particularly in scenarios involving extended samples or poor optical alignment.