Joint Spectrum Estimation and Material Decomposition for One-Kvp-Switching Dual-Energy CBCT Using Implicit Neural Representations (JSEMD-INR)
Abstract
Purpose
Standard Cone-Beam CT (CBCT) enables cost-effective Dual-Energy CT (DECT) via a one-kVp-switching protocol(switching energy once in the middle of a full scan), but inherently suffers from geometry-inconsistent data acquisition and unknown spectral characteristics. This work proposes a physics-informed machine learning framework (JSEMD-INR) to jointly estimate energy spectra and basis material decomposition directly from the ill-posed, limited-angle projections, enabling quantitative imaging on standard hardware without complex pre-calibration.
Methods
We formulate the unknown energy spectra as a weighted sum of known spectral basis components. The spatial distribution of basis materials is parameterized using an Implicit Neural Representation (INR) with multi-resolution hash encoding. A nonlinear polychromatic forward model connects the INR to the measured projections. To address the inherent non-convexity, we employ a robust three-stage optimization pipeline: (1) Pretraining, where the INR is initialized using Feldkamp-Davis-Kress (FDK) reconstructions from combined projections to mitigate limited-angle artifacts; (2) Spectral Initialization, where spectrum parameters are estimated using the fixed spatial structure; and (3) Joint Optimization, where spectral weights and INR parameters are updated via an alternating minimization strategy to minimize the discrepancy between predicted and measured projections. JSEMD-INR was validated on 2D and 3D phantoms using a simulated one-kVp-switching, geometry-inconsistent protocol.
Results
JSEMD-INR successfully suppressed limited-angle artifacts and achieved high-fidelity basis material decomposition. Compared to the Image-Domain-Decomposition (IDD) baseline, our method demonstrated a substantial PSNR improvement of over 30 dB for both water and bone basis images. Furthermore, the ground truth energy spectra were accurately recovered with relative errors of 1.28% and 0.77% for the low- and high-energy settings, respectively, verifying the robustness of the JSEMD-INR.
Conclusion
JSEMD-INR effectively overcomes the challenges of geometry-inconsistent acquisition, achieving superior basis material decomposition and accurate spectral characterization compared to conventional methods. This approach demonstrates the feasibility of performing quantitative DECT on standard CBCT systems without specialized hardware.