Enabling Quantitative Cone-Beam CT with a Deep Learning Framework: Monte Carlo-Informed Material Decomposition for Adaptive Radiotherapy
Abstract
Purpose
Conventional cone-beam CT (CBCT) on linear accelerators suffers from limited soft-tissue contrast and quantitative inaccuracy, hindering its use for precision radiotherapy tasks. This study aims to develop and clinically validate a deep learning-based material decomposition framework for dual-energy CBCT (DE-CBCT) to generate high-quality virtual monoenergetic images (VMIs) and accurate relative electron density (RED) maps.
Methods
A fully-connected neural network (NN) was trained on a high-fidelity Monte Carlo-simulated database to directly decompose preprocessed 80 kV and 140 kV projections into line integrals of aluminum and PMMA basis materials. The framework was implemented on a Varian Edge linac. Its performance was evaluated against a conventional polynomial fitting (POLY) method using an electron density phantom and an anthropomorphic head phantom. Key metrics included RED accuracy, contrast-to-noise ratio (CNR) and signal-to-noise ratio (SNR) of VMIs (40-150 keV), and the reduction of metal and streak artifacts.
Results
The NN method significantly outperformed the POLY method. It reduced RED error from 26.2% to 4.5% for high-density bone and from 45.6% to 19.9% for lung-equivalent material. NN-derived VMIs exhibited markedly improved CNR (up to 88.2% improvement for high-Z materials) and an average SNR increase of 78.9%. Substantial artifact suppression was achieved, with metal and streak artifact indices reduced by 34.3% and 85.1%, respectively. In the head phantom, the NN method improved CNR by 12.5-14.0% over single-energy CBCT while reducing image noise by 22.3%.
Conclusion
The proposed Monte Carlo-validated neural network provides a robust and accurate solution for DE-CBCT material decomposition. It enables the generation of VMIs with superior contrast, reduced noise, and suppressed artifacts, effectively addressing major limitations of current CBCT systems. This work demonstrates a significant step toward enabling reliable, quantitative CBCT for image-guided and adaptive radiotherapy.