MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

CBCT Image Quality Augmentation Using Deep Learning Models: A Comparison Study

Y Zhao1*, Z Jiang2 , X Teng1 , L Ren2 , (1) Duke Kunshan University, Kunshan,(2) Duke Univeristy, Durham, NC

Presentations

(Sunday, 7/14/2019) 2:00 PM - 3:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: To investigate the feasibility of using deep learning networks to augment the CBCT image quality either in the image domain by learning a direct mapping from CBCT to CT or in the projection domain by predicting and correcting scatter signals.

Methods: For image domain augmentation, two deep learning networks, i.e. a symmetric residual convolutional neural network (SR-CNN) and a U-net network, were trained to augment the quality of CBCT to match with CT quality. The training data included head patient CT and CBCT images after rigid registration and lung patient CT and Monte Carlo (MC) simulated CBCT images. After model training, a new set of CBCT was fed into the trained networks for augmentation. The networks’ performances were evaluated and compared both qualitatively and quantitatively, using the peak-signal-to-noise ratio (PSNR) and the structural similarity index measure (SSIM).For projection domain augmentation, a U-net was trained to predict the scatter signals from the CBCT projections. The training data included MC simulated projection and scatter signals. The trained network was used to predict scatter signal in projections from the real liver patient dataset for scatter correction before CBCT reconstruction.

Results: Regarding image domain augmentation, the quality of CBCT was augmented by both SR-CNN and U-net substantially with higher PSNR and SSIM than the original CBCT. U-net achieved better augmentation and faster speed than SR-CNN.Regarding projection domain augmentation, U-net was able to predict the scatter component accurately in the projections. The scatter-corrected CBCT demonstrated substantial improvement of the image contrast and anatomical details compared to the original CBCT.

Conclusion: The proposed deep learning models can effectively augment CBCT quality in both image and projection domains. Given their relatively fast computational speeds, they can be effectively implemented in clinics to improve our precision for target localization and adaptive radiotherapy using CBCT.

Keywords

Cone-beam CT

Taxonomy

IM- Cone Beam CT: Machine learning, computer vision

Contact Email