Room: Track 4
Purpose: To develop a machine learning-based method of correcting CBCT images for online detection of changes in proton range and plan adaptation to daily anatomy.
Methods: Planning CT and pre-treatment CBCT datasets from 35 patients aged 0.7-23 years who received proton therapy for abdominal or pelvic tumors were analyzed. The conventional Cycle Generative Adversarial Network (CycleGAN) was refined with a residual network in the generator to convert CBCT to CT-like images suitable for proton dose calculation. Seven datasets were selected to test the neural network trained using the other datasets (n=28), wherein the selection criteria included target volume within the field-of-view of CBCT and availability of repeat CT on the same day of CBCT. To handle the large variation in body size, the bodies in the training CBCT-CT pairs were linearly scaled to a reference size. Testing CBCT images were scaled accordingly prior to correction, followed by rescaling the corrected images back to the original size. The corrected CBCT was compared to the same-day CT in HU accuracy, dose distribution, and proton ranges.
Results: Correcting a typical CBCT image took 20 seconds on one GPU node after 3 days of training. The corrected CBCT showed high levels of similarity to the same-day CT. The structural similarity index was 0.84±0.05. Mean absolute error in HU were 151±33 and 41±7 for bone and soft tissues, respectively. The gamma passing rate (2%/2mm criteria) of dose map was 98.3±1.8%. CTV V95 only differed by -0.45±0.94%. The deviation in proton range (R80) from the reference values was 0.14±1.23mm.
Conclusion: In spite of the variation in body size, patient position, and CBCT coverage, the CycleGAN shows great promise in enabling online verification of proton range and adaptive planning in pediatric patients based on CBCT of the day.
Cone-beam CT, Image-guided Therapy, Protons