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: Applications of combined kV-MV CBCT include metal artifact correction and material identification. Difficulties arise, however, when the imagers have misaligned geometric perspectives of the scan subject (e.g., from patient motion). This will often occur since kV and MV scans are acquired with different source-detector hardware. A method to align projections using dual-layer kV-CBCT and a deep neural network is proposed. Notably, the method operates entirely with projections, and does not rely on registration with prior 3D images.
Methods: CBCT projections of several phantoms were acquired with a prototype dual-layer kV-imager (DLI) and with different MV-CBCT imagers on other linacs. Additional misalignment was added by randomly disordering the MV-CBCT projections and altering pixel size. A Catphan 604 was acquired at 125 kVp, 600 mAs and at 2.5 MV, 7 MU. A pelvis phantom was scanned at 125 kVp, 750 mAs and at 6 MV, 7MU. A U-Net with three input channels was trained with digitized phantoms. Two channels were kV projection pairs from the two DLI layers. The third was an MV-CBCT projection geometrically misaligned with the DLI. The U-Net output inferred the MV-CBCT projection as when properly aligned with the kV projections. Our evaluation compared 3D reconstructions of the kV-CBCT, the U-Net aligned MV-CBCT, and a plain MV-CBCT reference.
Results: Tissue boundaries in the U-Net aligned image agreed strongly with kV-CBCT, while Hounsfields agreed well with the MV-CBCT reference (and with 40% lower noise for Catphan). Quantitative agreement to within 3% of reference ROI means were observed, except in adipose and dense bone (5%).
Conclusion: A proposed remedy for misalignment between MV and kV frames shows promise in real phantom scans. The method was sometimes observed to reduce MV-CBCT noise, mitigating traditional shortcomings of MV detectors. This stands to unlock many applications for kV-MV-CBCT in radiotherapy image guidance.