MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Patient-Independent CT Intensity Correction Method Using Generative Adversarial Networks (GAN) for Single X-Ray Based Tumor Localization

R Wei1*, F Zhou1 , B Liu1 , X Bai1 ,Q Wu2 , (1) Image Processing Center, Beihang University, Beijing, ,(2) Duke University Medical Center, Durham, NC

Presentations

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

Room: 221AB

Purpose: Recently we developed a tumor localization technique using single X-ray projection. The intensity inconsistency between DRR and CBCT projection image has a significant impact on the performance, and the intensity correction was made utilizing CT and CBCT from the same patient. In this study, we developed and validated a patient-independent intensity correction method based on Generative Adversarial Networks (GAN).

Methods: In our proposed method, a GAN was trained to simulate 3D-CBCT from planning CT which was subsequently used to generate DRR that resemble the true X-ray projection. The GAN was trained with 3D patches from CT and 3D-CBCT of multiple patients. The architecture of our GAN was a variant of Pix2Pix GAN, where 3D convolution was used instead of 2D convolution. To apply to the new patient, the simulated 3D-CBCT can be derived with planning CT fed into the trained GAN. The DRR of this synthetic 3D-CBCT can be applied in the formulism of the single X-ray projection based tumor localization.

Results: Our method was evaluated using images from three patients. Two patients’ data was utilized to train the GAN, and the third one was used to test the network. The CBCT projections at angle ranging from 0° to 90° were sorted into different phases and utilized as the ground truth. The corresponding CT was selected from the 4D-CT and used as the input to the network. For this experiment, the relative error of proposed method was below 9.8%, with average at 7.6%, while the relative error of previous method was above 10% with a mean value of 12.3%.

Conclusion: We proposed a GAN based CT intensity correction method for tumor localization. This method allows for fast implementation without the need of additional CBCT scan before treatment. Moreover it achieves better performance than previous method using same patient images.

Keywords

CT, Cone-beam CT, Image Processing

Taxonomy

IM- Cone Beam CT: Machine learning, computer vision

Contact Email