NeAS: 3D Reconstruction from X-ray Images using Neural Attenuation Surface

TL;DR

This work tackles reconstructing accurate 3D surfaces from sparse X-ray images, addressing the limitations of surface fidelity in prior implicit neural representations. The authors propose NeAS, an implicit framework that jointly learns a neural attenuation field and a surface via a signed distance function, augmented by a surface boundary function to confine attenuation within object boundaries. They extend to multiple materials with 2M-NeAS and introduce a pose refinement strategy to cope with imperfect extrinsics, enabling improved surface reconstruction and novel-view synthesis. Experiments on simulated and real X-ray data demonstrate superior surface accuracy and competitive rendering quality, highlighting the practical potential for low-dose 3D imaging and bone-density estimation. The approach advances X-ray INR methods by integrating explicit geometric priors (SDF) with attenuation modeling and robust optimization strategies.

Abstract

Reconstructing three-dimensional (3D) structures from two-dimensional (2D) X-ray images is a valuable and efficient technique in medical applications that requires less radiation exposure than computed tomography scans. Recent approaches that use implicit neural representations have enabled the synthesis of novel views from sparse X-ray images. However, although image synthesis has improved the accuracy, the accuracy of surface shape estimation remains insufficient. Therefore, we propose a novel approach for reconstructing 3D scenes using a Neural Attenuation Surface (NeAS) that simultaneously captures the surface geometry and attenuation coefficient fields. NeAS incorporates a signed distance function (SDF), which defines the attenuation field and aids in extracting the 3D surface within the scene. We conducted experiments using simulated and authentic X-ray images, and the results demonstrated that NeAS could accurately extract 3D surfaces within a scene using only 2D X-ray images.

Figures (10)

• Figure 1: Overview of our framework. To render a pixel of an image, first sample 3D points along the corresponding ray, and then query the SDF and attenuation coefficient of these points using MLP $\Theta_{sdf}$ and $\Theta_{att}$. A surface-bound function was calculated from the SDF and used to constrain attenuation, retaining the attenuation only inside the surface. Then, attenuations are accumulated for volume rendering. The final pixel intensity is compared with the ground truth by MSE loss, which is backpropagated to optimize MLPs.
• Figure 2: Surface boundary function (SBF) $\Omega(d,s)$ with SDF values $d$ on a certain ray and different $s$ parameters. As the $s$ value increases, the more accurate the surface position is determined.
• Figure 3: Pipeline to query attenuation coefficient for two materials.
• Figure 4: Brief illustration of the attenuation coefficient in a bone and muscle area.
• Figure 5: Phantoms used for X-ray image capturing.