Denoising Grating-Based Phase-Contrast Computed Tomography with Self-Supervised Deep Learning: A Comparative Study

To increase the detectability of breast cancer, recent efforts have been made in the field of grating-based phase-contrast computed tomography, which provides additional information about the refractive index decrement introduced by the sample. Whereas this signal shows an increased soft-tissue contrast, the implementation for polychromatic sources is still tied to higher doses, preventing medical application on patients. To lower the dose, we introduce the self-supervised deep learning network Noise2Inverse into the grating-based phase-contrast computed tomography field and compare its results with other denoising methods, namely the Statistical Iterative Reconstruction, Block Matching 3D, and Patchwise Phase Retrieval. Before the comparison, the behavior of Noise2Inverse parameters on the phase-contrast results are evaluated. We show that Noise2Inverse delivers superior denoising results with respect to the investigated image quality metrics. Its application allows the increase of resolution while maintaining a good contrast-to-noise ratio. This could enable grating-based phase-contrast computed tomography to outperform conventional absorption-based computed tomography in dose-relevant applications.