X-Field: A Physically Grounded Representation for 3D X-ray Reconstruction

Feiran Wang, Jiachen Tao, Junyi Wu, Haoxuan Wang, Bin Duan, Zongxin Yang, Yan Yan

Abstract: X-ray is essential in medical diagnostics for visualizing internal body structures, yet its use is strictly regulated due to the health risks posed by radiation exposure. To mitigate these risks, recent research has explored generating novel views from sparse input to minimize radiation doses, employing methods like NeRF and 3D Gaussian Splatting. However, these approaches primarily adhere to principles for visible light imaging, failing to account for the distinct characteristics of X-ray imaging. In this paper, we propose a novel 3D representation, named X-Field, specifically designed to align with the intrinsic characteristics of X-ray imaging. In lieu of the continuous, view-dependent representations used in visible light, X-Field models X-ray representation as discrete and view-independent, rooted in the physical property of energy absorption rate. To capture such property, we employ ellipsoids with uniform energy absorption rates, effectively representing complex material distributions in internal structures. Our method further empowers a hybrid progressive initialization strategy, leveraging structural priors from CT imaging, and optimizes via a material-based approach that dynamically adjusts to local variations in material composition. Experimental results demonstrate that X-Field achieves state-of-the-art visual fidelity in reconstructing human organs and other objects, highlighting its potential to transform medical imaging by enhancing safety and diagnostic precision.