Main Session
Sep
28
PQA 01 - Radiation and Cancer Physics, Sarcoma and Cutaneous Tumors
2121 - Evaluation of Methods for Enhancing Energy-Optimized Cone-Beam CT Virtual Monoenergetic Images
Presenter(s)
Andrew Keeler, PhD - Loyola University Medical Center, Maywood, IL
A. Keeler1, J. Luce1, A. Yunker2, M. Nikolova3, M. Georgeson4, H. Nguyen1, J. C. Roeske1, and H. Kang1; 1Department of Radiation Oncology, Stritch School of Medicine, Cardinal Bernardin Cancer Center, Loyola University Chicago, Maywood, IL, 2Data Science and Learning Division, Argonne National Laboratory, Lemont, IL, 3Loyola University of Chicago, Chicago, IL, 4Creighton University, Omaha, NE
Purpose/Objective(s): Energy-optimized dual energy cone beam CT (DE-CBCT) has shown promise in producing virtual monoenergetic images (VMIs) with enhanced soft-tissue contrast over polychromatic CBCT images. These enhanced, energy-selective images may prove beneficial in image-guided radiotherapy and emerging adaptive radiotherapy applications. Past research has shown that optimal VMI energies correlate strongly with those that minimize the noise levels of the VMI. The purpose of this study is to investigate various reconstruction and postprocessing approaches for reducing noise to further enhance CBCT-VMI. Materials/Methods: Two CBCT full-fan, full trajectory scans of a Catphan 604 phantom were obtained at 80 kV, 792 mAs and 140 kV, 132 mAs on a commercial linac. The scans were decomposed into equivalent thicknesses of aluminum (Al) and PMMA using an open-source toolkit, and optimal VMI energies were determined using an in-house analysis program developed for fast synthesis and evaluation of VMI. VMI projections were then synthesized at this optimal energy from the decomposed Al and PMMA projections and reconstructed into VMI using four noise-reduction methods: FDK plus median filtering using a 3x3 kernel, FDK plus enhancement using the Noise2Inverse (N2I) deep learning network, and two common iterative reconstruction algorithms, a maximum likelihood method (MLEM) and a total variation minimization method (OS-AwASD-POCS). VMI produced using these methods were compared to VMI from the default method, FDK reconstruction, and were assessed for CNR enhancement relative to the FDK image (rCNR) averaged over all material inserts, HU accuracy, and spatial resolution at the 5% level of the modulation transfer function. Results: As seen in the table below, all enhancement methods reduced image noise, with rCNR ranging from 1.18 to 3.58. However, FDK+median, N2I, and MLEM suffered from decreased image resolution (blurring), with N2I and MLEM additionally suffering from reduced HU accuracy. Conversely, OS-AwASD-POCS not only showed the highest rCNR, but also increased the HU accuracy and spatial resolution. Conclusion: To further enhance the image quality of CBCT-VMI, we explored several advanced image reconstruction methods. Our results demonstrated that both iterative reconstruction algorithms and postprocessing methods prove effective in reducing the noise in VMI. OS-AwASD-POCS produced the highest-quality images with improvement in all image quality metrics over baseline FDK, while the other methods improved rCNR but showed losses in HU fidelity and resolution. Future work will include fine-tuning OS-AwASD-POCS to reduce noise in patient data and better quantify the gains of using enhanced CBCT-VMI in clinical settings. Abstract 2121 - Table 1
| FDK | FDK+median | FDK+N2I | MLEM | OS-AwASD-POCS | |
| Relative CNR (rCNR) | 1 | 1.18 | 2.35 | 1.54 | 3.58 |
| HU Accuracy (RMSE) | 13.7 | 13.7 | 23.2 | 14.9 | 12.7 |
| Resolution @ 5% Modulation Transfer (lp/mm) | 7.3 | 5.9 | 5.9 | 6.1 | 7.5 |