Deep Learning-Based 4D Dual-Energy CBCT Generation for Material Decomposition
Abstract
Purpose
To propose a deep-learning approach for predicting high from low-energy 4D-CBCT.
Methods
Paired 4D dual-energy (DE)-CBCT datasets were generated using XCAT and the TIGRE reconstruction toolbox. Five male and five female adults were simulated by applying bounded multiplicative Gaussian perturbations (mean=1, SD=0.12) to major anatomical scaling parameters (e.g.: phantom/torso scale). Three breathing-curves were simulated per subject (temporal-scale deviation ±0.1/±0.3/±0.5) with respiratory rates randomly sampled within 10–20 breaths/min and amplitude negatively correlated with rate. 4D CBCT volumes at 80 kVp and 150kVp were reconstructed into 10 respiratory phase bins, yielding 30 paired 4D sets (300 volumes per energy; cropped to 76×512×512 voxels). A 3D U-Net was trained to predict 150kVp from 80kVp images using a mean squared error loss with subject-level splitting (8/1/1 for training/validation/internal testing). Four XCAT paediatric phantoms (1, 10, 15 years) were used for external testing. A two-channel U-Net (80kVp + sparse 150kVp) was compared against the baseline network (single-channel 80kVp) to ascertain the benefit of including high-energy attenuation information in the learning process. Similarity to the ground-truth was calculated on the synthesized 150kVp CBCT and derived relative electron density and stopping power (RED/RSP) via basis-material decomposition.
Results
On the adult internal test set, the baseline model synthesized images with NMAE=0.005279, NCC=0.9994, and SSIM=0.9902. The two-channel configuration resulted in modest improvement (NMAE→0.004790 and NCC→0.9996). Comparable performance was achieved in the paediatric testing dataset (NMAE=0.004124, NCC=0.9997, SSIM=0.9906). The basis-material decomposition using the synthesized 4D DE-CBCT yielded comparable RED/RSP maps for dose calculation, with the ground-truth based on the paired images (RED: NMAE=0.0006512; RSP: NMAE=0.006761, both NCC≈0.99999, SSIM≈0.9998–0.9999).
Conclusion
This study demonstrates the feasibility of phase-aligned 4D DE-CBCT synthetization without scanner upgrade, displaying satisfactory generalizability from adult to paediatric data.