PMC

Tomographic images of 425–500 m diameter polyethylene spheres (phase contrast acquisition). From left to right: images reconstructed from complete (i.e. dithered; step-and-shoot) sinograms, cycloidal sinograms interpolated with bicubic splines (step-and-shoot), cycloidal sinograms processed with the MS-D CNN (step-and-shoot), cycloidal sinograms interpolated with bicubic splines (flyscan) and cycloidal sinograms processed with the MS-D CNN (flyscan). The top row shows phase retrieved images, while the bottom row shows what these images would look like if phase retrieval were not applied. The PSNR (top left) and MS-SSIM (bottom left) with respect to the dithered images are shown for each low-dose image.

Simulated sinograms for the numerical foam sample, from left to right and top to bottom: complete sinogram, angular subsampling, low exposure time, cycloidal subsampling and bicubic splines interpolation, low exposure time data with projection denoising using a MS-D CNN, cycloidal subsampling and MS-D CNN data recovery.

Accuracy metrics (PSNR, Dice, and MS-SSIM) for reconstructions of the central slice of the foam phantom, using cycloidal subsampling with an MS-D network at various points during the 48 hour training time.

Tomographic images of a chicken bone (attenuation contrast; step-and-shoot acquisition). From left to right: image reconstructed from a complete (i.e. dithered) sinogram, an incomplete rotation-only beamlet sinogram interpolated with bicubic splines, an incomplete cycloidal sinogram interpolated with bicubic splines, an incomplete cycloidal sinogram processed with the MS-D CNN. The PSNR (top left) and MS-SSIM (bottom left) with respect to the dithered image are shown for each low-dose image.