Multi-Feature Fusion Based Neural Attenuation Field for Sparse-View CBCT Reconstruction

In the field of oral imaging, digital panoramic X-ray (PX) is widely used for its convenience, low cost, and low radiation dose. However, it remains limited in representing the depth and complexity of three-dimensional (3D) oral structures. In contrast, Cone-Beam Computed Tomography (CBCT) provides high-precision 3D visualization but entails higher radiation exposure and greater economic cost. To address this gap, we propose MFF-NAF, a multi-feature fusion–based neural attenuation field method for 3D reconstruction from 2D panoramic images. MFF-NAF integrates a multi-feature fusion module that extracts and concatenates 3D point positional, pixel-aligned, and voxel-aligned features and learn the scene’s density distribution and geometry structure. A density estimation module employs feature-wise linear modulation to dynamically adjust the weights and biases of the activation functions, enabling the capture of anatomical consistency across patients. Furthermore, an optimization module refines structural details to enhance reconstruction accuracy. Experiments on two in-house datasets comprising 160 cases and six comparative methods demonstrate that, under an acceptable parameter budget, MFF-NAF achieves quantitative improvements of 0.73–3.98 dB in PSNR, 0.53–20.66 percentage points in SSIM, and 1.41 – 14.28 percentage points in DICE, outperforming existing methods in 3D oral structure reconstruction.