H Lee*, A Lalonde, B Winey, G Sharp, H Paganetti, Massachusetts General Hospital / Harvard Medical School, Boston, MA

• 
  H Lee

PO-GePV-I-12 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: In recent efforts to personalize radiotherapy and to improve the clinical outcome, cone-beam CT (CBCT) is utilized to calculate delivered doses and adapt treatment plans based on the patients’ anatomy on the day of treatment. However, scatter artifacts degrade image quality and compromise the usability of CBCT beyond patient positioning. In this study, we developed a deep learning-based scatter correction method considering system-specific geometrical information.Materials &

Methods: The proposed method estimates scattered X-ray signals using a kernel-based method which models the scatter-to-primary ratio using point-spread functions (scatter kernels) for the CBCT geometry, X-ray spectrum, and patients’ thickness. The estimations need to be refined to remove the effect of scattered photons in scatter estimation. In the proposed method, we refined the estimated signals using a deep neural network (DNN). We modeled Elekta’s XVI CBCT system using Monte Carlo (MC) simulations and obtained the scatter kernels. We collected planning CT images of 37 H&N patients and obtained scatter-free and scatter-contaminated CBCT projection data using MC simulation to train the DNN. Scatter signals were estimated using the kernels from the projection data. The U-Net was trained to find final scattered signals from the scatter-contaminated projection data and estimated scatter signals.

Results: We used the proposed method to remove scattered signals for the 5 patients’ data not included in the training phase of the DNN. The projection data were reconstructed using the FDK algorithm. Scatter-corrected images and scatter-contaminated images were compared to scatter-free images. The average root-mean-square error (ARMSE) of the scatter corrected images was 48.35 HU, while the ARMSE of scatter-contaminated images was 146.84 HU.

Conclusion: We developed a hybrid scatter correction algorithm to incorporate geometrical information with the DNN. The developed methods successfully estimated and removed scatter artifacts from the projection data.

Keywords

Cone-beam CT, Scatter

Taxonomy

IM- Cone Beam CT: Machine learning, computer vision

Contact Email