Shared-Structure 4D Spectral Gaussian Representation for Sparse-View Spectral CT Reconstruction
2026-08-17 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors developed a new method called 4D-SG to improve reconstructing detailed 3D images from a small number of X-ray scans taken at different energy levels. Their approach separates the shared shape information from how the material absorbs different energies, allowing better use of sparse data and enabling predictions at energy levels not directly measured. Tests on multiple datasets showed that their method outperforms existing Gaussian-based approaches by producing clearer and more accurate images. This helps in making spectral CT scans more efficient and informative even with fewer views.
spectral computed tomographysparse-view reconstructionGaussian representationspectral couplingPSNRSSIMLPIPSattenuation volumeprojection viewsspectral density
Authors
Jiancheng Fang, Shaoyu Wang, Wenjun Xia, Yang Chen, Qiegen Liu
Abstract
Sparse-view spectral computed tomography (CT) reconstructs energy-resolved attenuation volumes from limited projection views, requiring simultaneous handling of angular undersampling and spectral coupling. We propose a SharedStructure 4D Spectral Gaussian Representation (4D-SG) that learns shared Gaussian geometry from full spectrum structural projections and uses a Gaussian-wise Spectral Density Curve Network (GSC-Net) to predict Gaussian raw density transformations. This factorization separates shared spatial structure from spectral attenuation variation, avoids independent channel geometry optimization, and establishes a continuous 4D-SG representation from discrete spectral measurements for unobserved spectral channel queries. Experiments on six synthesized, simulated projection, and real projection datasets with 50 views demonstrate the best average performance. Compared with the strongest Gaussian baseline, 4D-SG improves PSNR from 35.56 dB to 36.61 dB, increases SSIM from 0.909 to 0.914, and reduces LPIPS from 0.208 to 0.194, demonstrating its effectiveness for sparse-view spectral CT reconstruction.