Y Chan¹*, M Li¹, K Parodi², C Belka^1,3, G Landry¹, C Kurz¹, (1) Department of Radiation Oncology, University Hospital, LMU Munich, Munich, Germany, (2) Department of Medical Physics, Ludwig-Maximilians-Universitaet Muenchen (LMU Munich), Garching (Munich), Germany, (3) German Cancer Consortium (DKTK), Partner Site Munich, Munich, Germany

• 
  Y Chan

TH-F-BRC-2 (Thursday, 7/14/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: Daily cone-beam CT (CBCT) imaging during the course of fractionated radiotherapy treatment can expose patients to a considerable amount of radiation dose. This work investigates the feasibility of low dose CBCT imaging capable of enabling accurate prostate radiotherapy dose calculation with only 25% projections by overcoming under-sampling artefacts and correcting CT numbers by employing a cycle-consistent GAN.

Methods: Uncorrected CBCT imaging data of 41 prostate cancer patients, acquired with about 350 projections at a medical linear accelerator, were retrospectively under-sampled to 25% dose images (CBCT(LD)) with only about 90 projections. CBCTs were reconstructed using FDK on an isotropic 1 mm³ grid, onto which corresponding planning CTs (pCT) were also resampled. We adapted a cycleGAN in 2D to translate CBCT(LD) into pCT equivalent images (CBCT(LD_gan)) including shape loss. Unpaired 4-fold cross-validation (33 patients) was performed to allow using the median of the 4 models as output. A previously validated non-AI CBCT scatter-correction technique (CBCT(cor)) was used for HU accuracy evaluation on 8 additional test patients. Volumetric modulated arc therapy (VMAT) plans were optimized on CBCT(cor) and recalculated on CBCT(LD_gan) to determine dose calculation accuracy.

Results: The mean absolute HU error with respect to CBCT(cor) decreased from 137 HU for CBCT(LD) to 110 HU for CBCT(LD_gan). The median differences of D₉₈, D₅₀ and D₂ comparing CBCT(LD_gan) to CBCT(cor) were 0.3%, 0.4% and 0.4% for the PTV. Dose accuracy was high with 2% and 1% dose difference pass rates of 99.5% and 98.0% (median, 10% dose threshold). The computational time for generating CBCT(LD_gan) was approximately 10s per patient.

Conclusion: This study demonstrated the feasibility of adapting cycleGAN for simultaneously removing under-sampling artefacts and correcting image intensity of 25% dose CBCT images. High dose calculation and HU accuracy was achieved, paving the way towards low dose CBCT imaging for treatment adaptation.

Funding Support, Disclosures, and Conflict of Interest: Acknowledgements: DFG: 399148265, GRK2274

Keywords

Cone-beam CT, Image-guided Therapy

Taxonomy

IM/TH- Cone Beam CT: Machine learning, computer vision

Contact Email