Click here to

Session: Deep Learning for Treatment Verification and Assessment [Return to Session]

Accurate Correction of Abdominopelvic CBCT with Internal Air Pockets Using a Hybrid Method of Deep Learning and Deformable Registration for Online Verification of Proton Range

J Uh*, C Wang, C Hua, St. Jude Children's Research Hospital, Memphis, TN


TU-D930-IePD-F2-3 (Tuesday, 7/12/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: To develop an accurate method of cone-beam CT (CBCT) correction for verifying the ranges of proton beams that potentially intersect with internal air, wherein a clear and precise definition of air-tissue boundary is often challenging even with recent deep learning methods.

Methods: Our previous method for correcting CBCT, which hybridizes low-frequency components of deformed prior CT (pCT) with CycleGAN-processed high-frequency components of CBCT, has been further advanced focusing on improvements in internal air regions. The incorrect air pocket position in deformably registered pCT is mitigated by iterating the hybrid process, i.e., the initial correction is used to improve the pCT for an updated correction. In addition, the patient-specific HU in the original pCT is more weighted in regions of smaller deformation fields where the registration is presumed to be more accurate. The corrected CBCT by the new hybrid method was evaluated in HU and proton range accuracy against the same-day CT that had been manually edited for the displacement of internal air between scans. CT/CBCT images from 81 patients were used for training (n=60), hyperparameter-tuning (n=6), and evaluation (n=15).

Results: The mean absolute error in HU of the corrected CBCT (29±7) was significantly smaller than those of deformed pCT (39±20), CycleGAN-only correction (54±5), and the previous hybrid method (39±7) (P<0.01). The Dice coefficient of internal air (<-300 HU) was 0.74±0.11, 0.20±0.12, 0.66±0.11, and 0.66±0.13, respectively, which was higher with the new method (P<0.001). Proton range (R80) analysis focusing on the cases exhibiting air pockets in V80 (n=7) showed an error of 1.8±0.6 mm for 90% beamlets in comparison to 5.7±6.6 mm by deformed pCT and 4.1±0.9 mm by CycleGAN-only correction (P<0.01).

Conclusion: The new hybrid method significantly improved the accuracy of the corrected CBCT with air pockets and showed promises in range verification of abdominal and pelvic proton therapy.


Cone-beam CT, Image Artifacts, Image-guided Therapy


IM/TH- Cone Beam CT: Scatter correction/ reduction

Contact Email