(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Abdominal tumors do not enhance in CBCT images during treatment; therefore radiation oncologists must rely on bony landmarks and fiducial markers to identify the targeted structure. As such, our objective was to implement a deep learning methodology to convert non-contrast CBCT images into synthesized contrast-enhanced CT images for the purpose of on-line adaptive radiation delivery.
Methods: We have implemented a Cycle Consistent Generative Adversarial Networks architecture to synthesize a contrast enhanced CT image from the daily acquired CBCT image. Daily CBCT images and planning CT images (non-enhanced, arterial and venous phase) from 15 liver/prostate SBRT patients were used for training while data from 10 liver/prostate SBRT cases were used for testing. A deformable registered CT (CT-to-CBCT) was created as the reference to assess HU accuracy of the synthetic CBCT. Additionally, an intravenous contrast enhanced CBCT will be acquired on the first treatment fraction of the testing patients to assess enhancement accuracy of the synthetically enhanced CBCT.
Results: results on prostate SBRT datasets showed HU correspondence within 3% between the synthetic CT and planning CT in addition to increased synthetic SNR compared to CBCT. Synthetic images were successfully imported into the clinical treatment planning system and dose calculation was performed. A passing rate of 100% was achieved with a 3%/3mm, 10% threshold criteria with the 3D gamma test. Results on liver non-contrast to contrast training showed accurate synthetic contrast enhancement for the arterial phase.
Conclusion: on these results, we anticipate that CBCT images can be synthetically enhanced with minimal error using this novel deep learning algorithm. Synthetic contrast allows direct visualization of hepatic tumors and surrounding vasculature without the side effects of daily intravenous contrast agents. This visualization should translate into better image alignment with the treatment plan, increasing treatment delivery accuracy, and robust applicability within an adaptive delivery framework.
Keywords
Image Processing, Cone-beam CT, Contrast
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email
