Scatter Correction in X-ray CT by Physics-Inspired Deep Learning

TL;DR

This work addresses scatter-induced artifacts in X-ray CT by introducing two physics-inspired deep-learning methods, PhILSCAT and OV-PhILSCAT, which jointly leverage scatter-corrupted projections and an initial reconstruction to estimate and subtract the scatter in the projection domain. A backbone loss is formulated in projection space, avoiding backpropagation through the filtered backprojection, and a two-tap high-pass filter enables efficient training while preserving perceptual image quality via a $\mathcal{Q}$-weighted norm. The methods extend to 3D and CBCT via slice-wise processing and small-cone-angle approximations, and are validated on Monte Carlo–simulated parallel-beam and cone-beam phantoms, showing consistent improvements over a recent data-driven projection-based baseline (DSE) in PSNR, SSIM, and MAE, with OV-PhILSCAT offering faster inference. The results indicate a promising software-only route to reduce scatter artifacts in CT without hardware modifications, with potential impact on clinical and industrial imaging, and suggest directions for real-data validation and further theoretical analysis.

Abstract

Scatter due to interaction of photons with the imaged object is a fundamental problem in X-ray Computed Tomography (CT). It manifests as various artifacts in the reconstruction, making its abatement or correction critical for image quality. Despite success in specific settings, hardware-based methods require modification in the hardware, or increase in the scan time or dose. This accounts for the great interest in software-based methods, including Monte-Carlo based scatter estimation, analytical-numerical, and kernel-based methods, with data-driven learning-based approaches demonstrated recently. In this work, two novel physics-inspired deep-learning-based methods, PhILSCAT and OV-PhILSCAT, are proposed. The methods estimate and correct for the scatter in the acquired projection measurements. Different from previous works, they incorporate both an initial reconstruction of the object of interest and the scatter-corrupted measurements related to it, and use a deep neural network architecture and cost function, both specifically tailored to the problem. Numerical experiments with data generated by Monte-Carlo simulations of the imaging of phantoms reveal consistent improvement over a recent purely projection-domain deep neural network scatter correction method.

Figures (10)

• Figure 1: CT geometries and X-ray scatter. (a) 2D parallel-beam geometry. A scattered x-ray reaches the detector at $(t_2,\theta)$ instead of $(0, \theta)$, making the total measurement $\tau(t_2, \theta)$ differ from the primary $p(t_2, \theta)$. (b) 3D imaging geometries for cone and parallel beam CT reconstruction experiments shown together. The central axial 2D slice on the x-y plane indicated by the gray rectangle at $z=0$.
• Figure 2: Block diagram of PhILSCAT.
• Figure 3: Cumulative fraction of the total energy contained in the frequency components $[0,q]$.
• Figure 4: Network architecture for PhILSCAT for 3D reconstruction (illustrated for $d=64$). Each red solid arrow represents two consecutive convolutional layers with ReLU nonlinearity modules. The number of channels in each intermediate output $s_\mathrm{ij} \in \mathbb{R}^{d \times d \times d/2^j}$ is provided next to its label in the diagram.
• Figure 5: Monochromatic parallel-beam reconstructions: (a), (b) axial and sagittal slices of $p_\theta$-reconstructions; (c), (d) error magnitude using total measurements $\tau_\theta$; vs. (e), (f) using primary measurements $p_\theta^*$ estimated by DSE; (g), (h) estimated by PhILSCAT; and (i), (j) estimated by OV-PhILSCAT. Display windows in HU (different for the reconstructions and for the error maps) are indicated by the colorbars.