Multimedia Systems Computer Networks and Communications

SiP-DIF: Siddon sampling-enabled deep intensity field with physical consistency supervision for sparse-view CBCT reconstruction

Li, Hangyu, Guo, Qin, Wang, Wei, Liu, Moquan, Wang, Nan, Zhu, Yu
Springer Nature · 2026-11-14 · DOI 10.1007/s00530-026-02716-w

Abstract

Cone-beam computed tomography (CBCT) is an important imaging technique that is widely used clinically. However, CBCT scans generate radiation, which may pose health risks to patients. Reducing the number of projection views is one of the mainstream strategies for lowering the radiation dose. Implicit neural representations enable reasonable CBCT reconstruction under sparse views, but existing methods rely on supervision signals in a single dimension and therefore cannot establish an effective connection between the projection domain and the image domain. To address this problem, we propose the Siddon Sampling-Enabled Deep Intensity Field (SiP-DIF). By exploiting the attenuation law of X-rays and introducing physical consistency constraints, SiP-DIF establishes a physical mapping between projection data and three-dimensional voxels, which ensures the physical rationality of the reconstruction results. In addition, we improve the feature extraction network and the feature fusion network to obtain more accurate and detailed CBCT reconstruction. Extensive experiments on datasets with diverse anatomical structures show that SiP-DIF achieves better performance than previous state-of-the-art algorithms. Compared with the baseline model DIF-Net, SiP-DIF improves the average PSNR by 1.8 dB and the average SSIM by 7.98 percentage points. These results indicate that SiP-DIF has strong feature representation and physical modeling capability, and it shows potential for accelerating CBCT synthesis and reducing patient radiation exposure.

Read at publisher →