Residual Gaussian Splatting for Ultra Sparse-View CBCT Reconstruction

2026-04-30

AI Summary

In ultra-sparse-view cone-beam CT reconstruction, photometric optimization often induces spectral bias, leading to loss of high-frequency anatomical details and excessive smoothing. This work proposes a residual Gaussian splatting approach that integrates wavelet multi-resolution analysis with 3D Gaussian splatting. By employing a hierarchical representation separating geometric bases from residual details, the method reformulates high-frequency recovery as a physically consistent implicit residual compensation task. A novel spectral–spatial co-optimization strategy is introduced to jointly refine geometric localization and textural fidelity. Crucially, the method incorporates a spectrally decoupled Gaussian representation that resolves the mathematical mismatch between the non-negativity of X-ray attenuation and the bipolar nature of wavelet high-frequency coefficients, thereby suppressing spectral crosstalk. Evaluated on clinical data, the approach significantly enhances reconstruction fidelity of intricate structures such as trabecular bone and vasculature, outperforming existing neural rendering techniques by achieving a superior balance between artifact suppression and detail preservation.

Abstract

While 3D Gaussian splatting (3DGS) offers explicit and efficient scene representations for cone-beam computed tomography reconstruction, conventional photometric optimization inherently suffers from spectral bias under ultra sparse-view conditions, leading to over-smoothing and a loss of high-frequency anatomical details. Since wavelet transforms provide rich high-frequency information and have been widely utilized to enhance sparse reconstruction, this work integrates wavelet multi-resolution analysis with 3DGS. To circumvent the mathematical mismatch between the strict non-negativity of physical X-ray attenuation and the bipolar nature of high-frequency wavelet coefficients, we propose Residual Gaussian Splatting (RGS). Methodologically, we introduce a spectrally-decoupled Gaussian representation that stratifies the volumetric field into a geometric base component and a residual detail component. This decomposition systematically transforms explicit high-frequency fitting into a physically consistent, implicit residual compensation task. Furthermore, we devise a spectral-spatial collaborative optimization strategy to coordinate the interplay between geometric anchoring and texture refinement, effectively preventing spectral crosstalk. Extensive experiments on clinical datasets demonstrate that RGS enables the reconstructed images to capture highly refined geometric textures. It successfully resolves the trade-off between artifact suppression and detail preservation, yielding superior visual fidelity in complex trabecular and vascular structures compared to existing neural rendering baselines.

Problem

• sparse-view CBCT reconstruction
• spectral bias
• high-frequency detail loss
• 3D Gaussian splatting
• over-smoothing

Innovation

• Residual Gaussian Splatting
• Wavelet Multi-resolution Analysis
• Spectral Decoupling
• Sparse-view CBCT Reconstruction
• Spectral-spatial Optimization