Purpose: Previously we developed a PCTV method to enhance the edge sharpness for low-dose CBCT reconstruction. However, deformable registration, which used to deform the edge from planning CT to on-board volumes, is time-consuming and user dependent. This study aims to compare a supervised and an unsupervised learning model for predicting daily on-board edge deformation to bypass deformable registration to improve the PCTV reconstruction efficiency.
Methods: The new method used supervised/unsupervised CNN deformation prediction and PCTV reconstruction. For the supervised model, deformation vector field (DVF) registered from CT to full-sampled CBCTs was used as the ground-truth to train the model to register CT and retrospectively under-sampled low-dose CBCT on the first day. The model was then updated with the following daysâ€™ data. For the unsupervised model, only CT and CBCT image pairs were used as input to train the model with no ground-truth DVF needed. The model was initially trained on a large amount of MRI data, and then was fine-tuned on our lung patient data. Both supervised and unsupervised learning model were evaluated using lung SBRT patient data. The first n-1 dayâ€™s CBCTs were used for model training to predict náµ—Ê° day edge information (n=2, 3, 4, 5). 45 half-fan projections covering 360Ëš from nth day CBCT were used for reconstruction. Results from Edge-preserving total variation (EPTV), PCTV and PCTV-CNN were compared.
Results: The cross correlation between the predicted edge map and the reference is around 0.74 and 0.87, for supervised and unsupervised model respectively. PCTV-CNN with the unsupervised model was robust to the daily changes.
Conclusion: Preliminary results demonstrated the feasibility to use the unsupervised CNN to predict daily deformation of on-board edge information for PCTV based low-dose CBCT reconstruction. The accuracy of the supervised CNN might be affected by the limited data size and inaccurate ground truth DVF.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH grant R01 CA-184173.