Noise2Ghost: Self-supervised deep convolutional reconstruction for ghost imaging

Abstract

We present a new self-supervised deep-learning-based Ghost Imaging (GI) reconstruction method, which provides unparalleled reconstruction quality for noisy acquisitions among unsupervised methods. We present the supporting mathematical framework and results from theoretical and real data use cases. Self-supervision removes the need for clean reference data while offering strong noise reduction. This provides the necessary tools for addressing signal-to-noise ratio concerns for GI acquisitions in emerging and cutting-edge low-light GI scenarios. Notable examples include micro- and nano-scale x-ray emission imaging, e.g., x-ray fluorescence imaging of dose-sensitive samples. Their applications include in-vivo and in-operando case studies for biological samples and batteries.

Figures (7)

• Figure 1: Schematic representation of diffused emission signal acquisitions (e.g., x-ray fluorescence imaging) using pencil raster beam scanning (\ref{['fig:xrf-img:pb']}) and ghost imaging (\ref{['fig:xrf-img:gi']}). The former uses a point beam to scan every pixel to form an image, while the latter illuminates the sample with a series of structured beams.
• Figure 2: Schematic representation of the proposed method: (\ref{['fig:scheme:partitioning']}) The partitioning of the realizations set, generating the partial reconstructions (sub-recs), (\ref{['fig:scheme:training']}) the training procedure, and (\ref{['fig:scheme:prediction']}) the prediction of the final reconstruction.
• Figure 3: Synthetic data GI reconstructions comparison (chromosomes phantom), for 5$\times$ compression ratio, and maximum emitted photons per pixel per realization in the range [$10^0$, $10^4$]. We compare peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and bandwidths (Nyquist frequency) against the ground truth. Higher values indicate higher reconstruction quality. We also show the corresponding image resolutions in pixels.
• Figure 4: Dose requirements for the chromosomes phantom, with 5$\times$ compression ratio, and varying noise levels: (\ref{['fig:chrom-progr-dose:tot']}) total required dose for each algorithm to obtain equivalent PSNR and SSIM as a PB acquisition; (\ref{['fig:chrom-progr-dose:avg-max']}) same plot as (\ref{['fig:chrom-progr-dose:tot']}) for the average maximum pixel dose per illumination; and (\ref{['fig:chrom-progr-dose:avg-max-n2g']}) for N2G.
• Figure 5: GI reconstructions of the chromosomes phantom, with 10$\times$ compression ratio and moderate Poisson noise: Mean noise fluctuations $\sim 24.5\%$ of the mean clean bucket fluctuations. From (\ref{['fig:chrom-cmprss:ls']}) to (\ref{['fig:chrom-cmprss:n2g']}) we show the reconstructions of (\ref{['fig:chrom-cmprss:ph']}) with LS, TV, GIDC, INR and N2G respectively. In (\ref{['fig:chrom-cmprss:frc']}) and in (\ref{['fig:chrom-cmprss:table']}) we present the Fourier ring correlation against (\ref{['fig:chrom-cmprss:ph']}) and various performance metrics of each reconstruction, respectively.