Click here to


Are you sure ?

Yes, do it No, cancel

Deep-Learning Based CBCT Image Correction for CBCT-Guided Adaptive Radiation Therapy

J Harms1*, Y Lei1 , T Wang1 , R Zhang1 , J Zhou1 , X Dong1 , P Patel1 , K Higgins1 , X Tang2 , W Curran1 , T Liu1 , X Yang1 , (1) Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, GA 30322, (2) Department of Radiology and Imaging Sciences and Winship Cancer Institute, Emory University, Atlanta, GA 30322


(Thursday, 7/18/2019) 7:30 AM - 9:30 AM

Room: Stars at Night Ballroom 2-3

Purpose: Adaptive radiation therapy (ART), where dose delivery is changed daily based on patient setup and anatomy, has been a goal in most clinics since the introduction of cone-beam CT (CBCT) since its introduction to the radiation therapy workflow. However, the large scatter-to-primary ratio typical of CBCT leads to degraded image quality and the loss of quantitative information in CBCT images. In this work, we propose a deep-learning method to correct CBCT artifacts and restore HU levels to those typical of planning CT images.

Methods: The proposed method learns a mapping from a CBCT HU distribution to a planning CT HU distribution. The powerful cycle-consistent generative adversarial network (cycle-GAN) framework is used. During training, a generator is continually optimized to produce corrected CBCT (CCBCT) images, while a discriminator is optimized to identify the differences between a CCBCT image and a planning CT image. As these two are pitted against each other, convergence of the overall network optimization is improved. Compared with a GAN, a cycle-GAN includes an inverse transformation from CBCT to CT images, which constrains the model by forcing calculation of both a CCBCT and a synthetic CBCT. The proposed algorithm was evaluated using 24 brain patient datasets and 20 pelvis patient datasets.

Results: Overall, mean absolute error, peak signal-to-noise ratio, normalized cross-correlation and spatial non-uniformity were 18 HU, 37.18 dB, 0.99 and 0.05 for the proposed method, improvements of 45%, 12%, 1%, and 65%, respectively, over the CBCT image. The proposed method showed superior image quality as compared to a conventional scatter correction method, reducing noise and artifact severity.

Conclusion: The authors have developed a novel deep learning-based method to generate high-quality corrected CBCT images. With further evaluation and clinical implementation, this method could lead to quantitative adaptive radiation therapy.

Funding Support, Disclosures, and Conflict of Interest: This research is supported in part by the National Cancer Institute of the National Institutes of Health under Award Number R01CA215718 (XY).


Not Applicable / None Entered.


IM- Cone Beam CT: Machine learning, computer vision

Contact Email