Scatter correction for multi-source linear static CT via a geometry-aware parameter regression network with differentiable forward modeling.

OBJECTIVE

Multi-source linear static CT (SLCT) offers high temporal resolution and a compact form factor by eliminating mechanical rotation. However, lateral offsets of peripheral X-ray beams break the rotational symmetry assumed by conventional scatter correction methods, resulting in location-dependent and asymmetric scatter point spread functions (PSFs).

APPROACH: To address this challenge, we propose a scatter correction framework that integrates a geometry-aware analytical kernel with a parameter regression network (GAPR-Net). An explicit correction term is incorporated into the conventional cone-beam scatter kernel to model asymmetric PSF distortion induced by lateral source offsets. GAPR-Net learns scatter kernel hyperparameters associated with source geometry and object thickness from Monte Carlo pencil-beam data, while a differentiable forward model analytically estimates scatter and enables stable optimization under a geometric consistency loss.

MAIN

RESULTS

Extensive simulation experiments were performed using water, electron-density, and ICRP phantoms, together with a DICOM-derived head phantom reconstructed using the derived three-dimensional SLCT reconstruction algorithm. Compared with SKS, DSE, ResNet, and a static-CT-specific 1D U-Net, GAPR-Net achieved consistent improvements. For the electron-density and ICRP phantoms, GAPR-Net achieved scatter-estimation mean absolute percentage errors of 4.43% and 5.31%, respectively. In the reconstructed images, it achieved a peak signal-to-noise ratio of 29.51 dB for the electron-density phantom and reduced the mean absolute percentage error of the ICRP thorax phantom from 14.69% to 6.07%. Head-phantom experiments at 80 kVp, 100 kVp, and 120 kVp further demonstrated stable performance across different X-ray spectra. A physical testbed experiment was conducted to evaluate the performance of the proposed GAPR-Net under real conditions.

SIGNIFICANCE: GAPR-Net provides a physically interpretable and computationally efficient approach for correcting non-stationary asymmetric scatter artifacts in SLCT.&#xD.