Click here to

Session: Functional, Biologic and Anatomic IGRT [Return to Session]

X-Ray Image Decomposition Using Personalized Prior Knowledge for Improved Motion Tracking in Paraspinal SBRT

X He*, W Cai, F Li, Q Fan, L Cervino, J Moran, P Zhang, X Li, T Li, Memorial Sloan Kettering Cancer Center, New York, NY


SU-F-BRC-3 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: X-ray image quality is critical for accurate motion tracking to assure safe delivery of radiation therapy. This study aims to develop a deep learning algorithm to improve kV image contrast by decomposing the image into bony and soft-tissue components. In particular, we want to incorporate patient-specific prior knowledge into the algorithm for optimal decomposition. We demonstrated its use in tracking motion in paraspinal SBRT with online kV imaging.

Methods: By utilizing a Multi-Head Cross Attention (MHCA) mechanism, a three-branch DecNet was designed to extract common features from the online kV projection image and patient-specific prior images, which were subsequently used to derive the spine component of the online kV image. The training dataset included 1275 x-ray images from 19 patients. The patient-specific prior images were randomly transformed (translation + rotation) DRRs, calculated based on patients’ planning CTs. The approach was validated using 370 x-ray images from six additional patients. The resulting spine-only x-ray images and the original x-ray images were registered to the DRRs to compare the spine tracking accuracy.

Results: The trained DecNet could effectively retain and enhance the spine structure information while suppressing the soft tissues from the kV projection images. The decomposed spine-only x-ray images had submillimeter matching accuracy with the DRRs at all beam angles. The decomposed spine-only x-ray significantly reduced the maximum errors to 0.84 mm and 0.61 mm from 1.5 mm and 1.3 mm, respectively, in x- and y- directions. The mean matching errors were reduced to 0.16 mm and 0.12 mm from 0.25 mm and 0.22 mm with p-values <<0.001.

Conclusion: The incorporation of the patient-specific prior knowledge into the deep learning algorithm significantly improved x-ray image contrast through the kV projection image decomposition, leading to submillimeter motion tracking accuracy in paraspinal SBRT.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email