Purpose: On-treatment kV X-ray images have been used in tracking patient motion during spine SBRT. One major challenge is the reduced contrast due to the passing of the X-ray through a large portion of the patient's body from the lateral direction. Besides, the overlap of the spine with the moving soft tissues could lead to significant errors in auto-registration. Our goal in this work is to automatically extract the spine component from the conventional 2D X-ray images to facilitate more robust and accurate motion management.
Methods: A ResNet generative adversarial network (ResNetGAN) consisting of one generator and one discriminator was developed to learn the conditional mapping between the 2D kV X-ray images and the spine-only DRR. The trained model took a kV X-ray image as input and learned to generate the spine component of the X-ray image. The training dataset included 699 x-ray images from 11 patients and the corresponding matched spine-only DRRs. Another 185 x-ray images from additional three patients were used for validation. The resulted spine-only X-ray images and the original X-ray images were registered to the DRRs, to compare the spine tracking accuracy.
Results: The decomposed spine-only X-ray images matched with the calculated DRRs with submillimeter accuracy (165 out of 185), indicating that the model retains accurate information of the spine structure from the original X-ray images. For cases with low contrast in the original X-ray image, the decomposed spine-only X-ray significantly reduced the mean matching errors (in mm) to 0.62 and 0.16 from 0.69 and 0.21, and the maximum errors to 3.6 and 0.4 from 7.2 and 1.4, in x- and y- directions, respectively.
Conclusion: We have developed a deep learning-based approach to suppress soft tissues in a kV projection image, leading to more accurate markerless spine tracking in paraspinal SBRT.
Funding Support, Disclosures, and Conflict of Interest: The presented study is partially supported by Varian Medical Systems Research Collaborations.
Not Applicable / None Entered.