Enhanced phase retrieval via deep concatenation networks for in-line X-ray phase contrast imaging

In-line X-ray phase contrast imaging offers considerable additional information beyond that acquired from conventional absorption contrast X-ray imaging, showing promising potentials in clinical diagnosis, materials characterization and so on. Given the physically intractable factors tangled inside, conventional phase retrieval methods typically suffer from limited feasibility. A deep-learning-augmented reconstruction strategy is proposed to improve the phase retrieval in spatial resolution and noise compression.

The deep network is composed of a phase contrast refinement module and a phase retrieval module to stabilize and generalize the phase retrieval. The two modules are aggregated in a plug-and-play fashion with the final assembly finetuned using limited training data, essentially encouraging a semi-supervised training. Verification experiments were performed on simulated phase contrast images of histopathological images. The results were compared to those from conventional phase-attenuation duality method.

The deep-learning-augmented reconstruction strategy increases structural similarity and peak signal-to-noise ratio of phase retrieval result by more than 8% and 30%, and reduces root mean squared error by 46% compared with conventional phase-attenuation duality method.

The pilot study of deep learning deployment in in-line X-ray phase-contrast imaging exhibit advantages against conventional methods in terms of spatial resolution and noise robustness.