Deep Learning in Classical X-ray Ghost Imaging for Dose Reduction

TL;DR

This work assesses dose reduction opportunities in x-ray ghost imaging by combining classical GI with deep learning. Using simulations that incorporate Poisson-Gaussian noise and four illumination-pattern families, the authors show that an end-to-end network (DLGI) can reconstruct images from highly undersampled bucket measurements, achieving results comparable to direct imaging at the same dose for simple datasets, and outperforming traditional GI and compressed-sensing approaches in several cases. However, under the same prior knowledge and detector efficiency, direct imaging often remains superior for very low-dose conditions, with DLGI showing robustness to electronic noise and benefiting from PCA/Hartley masks. The study provides guidance on illumination-pattern design and highlights the potential—and current limits—of GI with deep learning for dose-constrained imaging in medical/biological contexts, pointing to future work on complex datasets and real-world validation.

Abstract

Ghost imaging (GI) is an unconventional technique that combines information from two correlated patterned light fields to compute an image of the object of interest. GI can be performed with visible light as well as penetrating radiation such as x-rays, electrons, etc. Penetrating radiation is usually ionizing and damages biological specimens; therefore, minimising the dose of this radiation in a medical or biological imaging context is important. GI has been proposed as a potential way to achieve this. With prior knowledge of the object of interest, such as sparsity in a specific basis (e.g., Fourier basis) or access to a large dataset for neural network training, it is possible to reconstruct an image of the object with a limited number of measurements. However, low sampling does not inherently equate to low dose. Here, we specifically explore the scenario where reduced sampling corresponds to low-dose conditions. In this simulation-based paper, we examine how deep learning (DL) techniques could reduce dose in classical x-ray GI. Since GI is based on illumination patterns, we start by exploring optimal sets of patterns that allow us to reconstruct the image with the fewest measurements, or lowest sampling rate, possible. We then propose a DL neural network that can directly reconstruct images from GI measurements even when the sampling rate is extremely low. We demonstrate that our deep learning-based GI (DLGI) approach has potential in image reconstruction, with results comparable to direct imaging (DI) at the same dose. However, given the same prior knowledge and detector quantum efficiency, it is very challenging for DLGI to outperform DI under low-dose conditions. We discuss how it may be achievable due to the higher sensitivity of bucket detectors over pixel detectors.

Figures (10)

• Figure 1: (A) Ghost imaging experimental set-up using entangled photon pairs, (B) Computational ghost imaging.
• Figure 2: (A) Structure of a simple neural network with two hidden layers (classifiers), (B) Structure of an autoencoder.
• Figure 3: (A) The first ten random binary masks, (B) The first ten lowest-frequency masks of the Hadamard set, (C) The first ten lowest-frequency masks of the Hartley set, (D) The first ten PCA masks, constructed using the first ten most-important components of the dataset.
• Figure 4: (A) Original images, (B) Classical GI reconstructed images using 294 random binary masks, (C) using 98 Hadamard masks, (D) using 49 Hartley masks, (E) using 20 PCA masks.
• Figure 5: Our end-to-end neural network structure, which can reconstruct object images directly from the measured bucket signals.