Author: Ross I. Berbeco, Vera Birrer, Raphael Bruegger, Pablo Corral Arroyo, Roshanak Etemadpour, Dianne M. Ferguson, Rony Fueglistaller, Thomas C. Harris, Yue-Houng Hu, Matthew W. Jacobson, Mathias Lehmann, Nicholas Lowther, Daniel Morf, Marios Myronakis 👨🔬
Affiliation: Brigham and Women's Hospital, Harvard Medial School, Dana-Farber Cancer Institute, Department of Radiation Oncology, Dana Farber/Brigham and Women's Cancer Center, Department of Radiation Oncology, Brigham and Women’s Hospital, Dana-Farber Cancer Institute, Harvard Medical School, Brigham and Womens Hospital, Dana Farber Cancer Institute, Harvard Medical School, Brigham and Women's Hospital, Varian Imaging Laboratory, Dana-Farber Cancer Institute 🌍
Purpose: A challenge for dual energy CBCT is that noise and residual errors in material decomposition steps can become amplified when forming low energy, high contrast virtual mono-energetic images (VMIs). This applies especially when spectral separation is narrow, as with dual-layer detector data. To mitigate this, we propose a deep learning method for directly obtaining a VMI, skipping explicit material decomposition.
Methods: CBCT projections of a Catphan 604 (125 kVp, 1340mAs) and CIRS pelvis phantom (140 kVp, 1680mAs) were acquired with a dual-layer kV-imager (DLI) prototype, mounted on a Varian Truebeam linac. A deep U-Net was trained to infer 40 keV and 100 keV mono-energetic projections from each pair of DLI projections. Training data was obtained from a digitization of the phantoms in several unconventionally tilted positions. The simulations produced ideal, poly-energetic projections, except for Gaussian noise added for augmentation purposes. Scatter was not simulated. The trained U-Net was then applied to the real, non-tilted phantom scans to obtain mono-energetic projections. For comparison, 40 keV mono-energetic projections of the Catphan were also generated using a basic table look-up method, operating pixel-by-pixel on the projection frames. The mono-energetic data sets were reconstructed into VMIs with conventional FDK and compared to single-energy reconstructions of the original projections.
Results: The proposed method generated VMIs with substantially improved contrast and with minimal additional noise. In Catphan’s 20% bone insert, contrast-to-noise (CNR) improved by 60% over single-energy CBCT. Conversely, the basic table-lookup method suffered a CNR drop of 40%, and showed significant non-uniformity artifacts. The 100 keV VMI of the pelvis phantom demonstrated notably reduced streaks through the hip.
Conclusion: This work demonstrates the potential for generating CNR-enhancing VMIs with a novel dual-layer imager for use in radiotherapy clinics. Moreover, the ability to mono-energize projections can simplify subsequent material identification tasks, when desired.