Artifact-free phase reconstruction for differential interference contrast microscopy based on deep learning

Differential interference contrast (DIC) microscopy, as a label-free, high-contrast, and strong capacity for optical section imaging method, is widely used in routine examinations of cells or tissues. However, due to the inherent non-linearity between DIC image intensity and the phase gradient of a specimen, it is difficult to obtain the quantitative phase image accurately. Moreover, although numerical integration has been tried as a means to reconstruct the specimen phase, the unknown integral constant and the sensitivity to gradient noise lead to insufficient phase image results (obscured by severe linear artifacts). Here, we propose a data-driven approach to achieve artifact-free, high-precision and fast reconstruction of the specimen phase. This method initially uses the specimen phase extracted by digital holography and constructs the “specimen phase-differential phase” training database based on the DIC microscopic imaging model. Subsequently, the Pix2Pix GAN network model is employed, where an appropriate loss function and gradient back propagation algorithm are implemented to allow the network to update the weight parameters automatically. This process enables the trained network model to effectively reflect the mapping relationship between the specimen and differential phase. With a trained deep neural network, high-precision artifact-free reconstruction of the specimen phase can be achieved using only a differential phase image along a single shearing direction. We demonstrate the effectiveness and applicability of the proposed method by quantitative phase imaging of polystyrene spherical crown and HeLa cells. The experimental results show that the model can quickly realize the high-fidelity and artifact-free reconstruction of the specimen phase, and also has excellent anti-noise performance. It provides a promising technology for achieving high spatial sensitivity detection of quantitative DIC microscopic imaging technology.