Daily Coverage Assessment Using Deep Learning Generated Synthetic CT for Lung SBRT Patients
R LJ Qiu1, Y Lei1, J Shelton1,2, K Higgins1, J Bradley1, W Curran1, T Liu1, A Kesarwala1, X Yang1, (1) Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, GA (2) Department of Radiation Oncology, Grady Memorial Hospital, Atlanta, GA
Presentations
PO-GePV-M-163 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: The daily dose in stereotactic body radiation therapy (SBRT) is much higher than conventional RT. Therefore, suboptimal RT delivery could be substantial and lead to inferior treatment outcomes. Lung SBRT patients are more susceptible to suboptimal RT delivery due to breathing motion and interfraction ventilation changes. However, a reliable method to evaluate daily target coverage has not been developed. This study proposes to calculate daily dose distribution and assess coverage on synthetic CTs (sCT) generated by deep learning algorithms.
Methods: A deep-learning (DL) model that integrates histogram matching (HM) and perceptual supervision into a cycle-consistent adversarial network (Cycle-GAN) framework, was trained to learn mapping between thoracic CBCTs and paired planning CTs. The proposed algorithm was evaluated using data from 20 patients with early-stage lung cancer who were treated with SBRT. To further access the accuracy of dose calculation on sCTs generated by our model, 3 patients that got replanning CT (rCT) simulation were chosen, to minimize anatomical change between the rCT and prior day CBCT. sCTs were generated by feeding the prior day CBCT into the trained DL model. Dose were calculated on both sCT and rCT (ground truth) using the replanned clinical plan. Target contours on rCT was transferred to sCT via deformable image registration and reviewed by physician.
Results: The dose-volume histogram (DVH) shows that the % of the planning target volume (PTV) receiving prescription dose was 89.1%/91.8%/95.0% (pt1/pt2/pt3) on the rCT, 92.5%/91.8%/98.3% on the sCT, and 94.5%/90.8%/99.5% on the CBCT. Similarly, the DVH also shows that both the gross target volume (GTV) and PTV dose curves from the sCT were closer to the rCT than the CBCT.
Conclusion: Dose calculation on sCT generated by DL model is much more accurate than that on CBCT. It can be a useful tool to access daily RT target coverage.
Keywords
Cone-beam CT, Computer Vision, Image-guided Therapy
Taxonomy
IM/TH- Cone Beam CT: Machine learning, computer vision
Contact Email