Patient-Specific Synthetic MRI Generation From CBCT for Image Guidance in Liver SBRT
Z Zhang1*, Z Jiang2, K Lu3, F Yin4, L Ren5, (1) Duke University, Durham, NC, (2) Duke University, Durham, NC, (3) Duke University, Chapel Hill, NC, (4) Duke University Medical Center, Chapel Hill, NC, (5) University of Maryland, Baltimore, MD
Presentations
SU-F-BRC-6 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Ballroom C
Purpose: Cone-beam CT (CBCT) is the most common onboard imaging modality for tumor localization of liver SBRT. However, CBCT has poor soft-tissue contrast, making it extremely challenging to localize tumors in the liver. Consequently, indirect localization methods based on liver boundary are often used with limited accuracy. MR-Linac provides MRI images with high soft-tissue contrast for target localization. However, such machines are expensive and not accessible in most clinics. In this study, we developed a patient-specific deep learning model to generate synthetic MRI from CBCT and planning MRI to improve the precision of CBCT-based liver SBRT.
Methods: The model was implemented based on the U-Net architecture with L1 loss function. A key innovation is that the model was trained using the patient-specific CBCT-MRI image pairs to generate synthetic MRI from CBCT. Data augmentation was used to generate enough patient-specific training data. Specifically, our dataset contains seven MRI and one planning CT image from the same patient for training and evaluation purpose. The planning CT was deformably registered to each MRI, and then used to simulate CBCT on each day using DRR and FDK reconstruction. The model was trained to generate synthetic MRI from CBCT in the liver with the real MRI as ground-truth, and tested in a different CBCT dataset. The synthetic MRI was quantitatively evaluated against ground-truth MRI using structural-similarity-index(SSIM), peak-signal-to-noise-ratio(PSNR), mean-square-error(MSE), and mutual-information(MI).
Results: The synthetic MRI demonstrated superb soft-tissue contrast with clear tumor visualization. On average, the synthetic MRI achieved 28.01, 0.025, 0.929, and 0.406 for PSNR, MSE, SSIM, and MI, respectively, outperforming CBCT images. The model performance was consistent across all three patients tested.
Conclusion: Our study demonstrated the feasibility of a patient-specific model to generate synthetic MRI from CBCT for liver tumor localization, opening up a potential to democratize MRI guidance in regular linacs.
Funding Support, Disclosures, and Conflict of Interest: This study is supported by NIH grant R01-EB028324, R01-EB001838, R01-CA226899.
Keywords
Taxonomy
IM- Cone Beam CT: Development (New Technology and Techniques)
Contact Email