Author: Xinyi Fu, Dan Ruan, Ke Sheng 👨🔬
Affiliation: Department of Radiation Oncology, University of California, Los Angeles, University of California, San Francisco, Department of Radiation Oncology, University of California, San Francisco 🌍
Purpose:
Digitally reconstructed radiographs (DRRs) are easy to generate and widely used to establish research protocols in pulmonary diagnosis and image-guided radiotherapy tasks. A question remains as to whether DRRs accurately reflect realistic X-ray digital radiographs (XDRs) clinically. We compare lung tissue conspicuity on 3DCT and tumor tracking performance on 4DCT between these paired imaging modalities, addressing the clinical translatability of DRR-based models to real-world X-ray applications.
Methods:
80 thoracic 3DCT scans and 20 thoracic 4DCT scans with physician-contoured planning CTs were included. DRRs were generated using ray tracing from CTs, while XDRs were simulated using the physics-driven GEANT4 Monte Carlo model, both resampled to 1 mm isotropic resolution. XDR simulation employed cone beam geometry tracking 10^9 particles (2 hours/image on 128-core CPU). Analysis on 3DCT-derived images included pixel-wise comparison of normalized DRR-XDR pairs and nodule visualization assessment. For 4DCTs, tumor tracking accuracy was evaluated on end-inhale and end-exhale phases using template matching, with templates derived from DRR/XDR projections of planning CTs.
Results:
DRRs, accounting for only primary photons, demonstrated marginally superior lung tissue conspicuity compared to XDRs with secondary particle effects. Analysis of 3DCT-derived paired images achieved MSE of 0.0102±0.0029, SSIM of 0.8293±0.0161, and PSNR of 19.9167±1.2542, with DRRs showing slightly enhanced nodule visibility. For the 4DCTs, including challenging cases with organ overlap, high tumor mobility, and small tumors, DRRs exhibited lower tracking errors (11.71±12.77mm end-inhale, 2.05±2.28mm end-exhale) versus XDRs (14.65±11.25mm end-inhale, 3.82±3.35mm end-exhale). Paired t-tests confirmed a significant correlation between paired DRR-XDR tracking errors (p=0.0115 end-inhale, p=0.0003 end-exhale).
Conclusion:
This study quantifies the differences introduced by using DRR as a surrogate of XDRs for nodule visualization and lung tumor tracking. Caution is exercised due to significant degradation in tumor tracking when scatter contamination, beam-hardening, and heavy noise are present, potentially necessitating domain adaptation strategies to calibrate these applications.