Synthetic CT Generation From Contrast Enhanced Planning CT and Low-Quality Treatment CBCT Using a Novel Dual Convolutional Neural Network
Y Cao*, A Vassantachart, S X Bian, D C Ling, Z Fan, E Chang, W Yang, University of Southern California, Los Angeles, CA
Presentations
MO-IePD-TRACK 3-6 (Monday, 7/26/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: The study aims to produce quantitative cone-beam computed tomography (CBCT) by learning HU features from planning CT while maintaining the structural integrity of on-treatment CBCT using a dual convolutional neural network (CNN) architecture.
Methods: The study included paired CT and CBCT images from 49 patients (33 training, 8 validation, and 8 testing). A primal-dual 3D CNN architecture was built based on a loop, which contains a primal network and a dual network. For the primal network, CBCT was treated as initial synthetic CBCT (S-CBCT) to obtain the synthetic CT (S-CT), aiming to capture the HU information from planning CT. In the dual network, the S-CT were treated as input data to generate S-CBCT using another CNN, aiming to capture the geometry information from CBCT. The energy function was computed by a combination of the pixel intensity loss of S-CT with CT, and the pixel+texture loss of S-CBCT with CBCT. Thus, the full objective was: E_T=aE_(SCT-CT)+bE'_(SCBCT-CBCT), where a and b are weightings to control the contribution of each network. Mean absolute error (MAE), peak signal-to-noise ratio (PSNR), normalized cross-correlation (NCC), and structural similarity index (SSIM) were used to evaluate the model performance.
Results: The dual CNN model performed better when the total loss was close to the loss between S-CT and CT using unequal weightings (a=0.9 and b=0.1). The visual inspection of S-CT showed improved image quality. All parameters tested were improved from CBCT to S-CT (equal weightings) and S-CT (unequal weightings). The results for the 4 abovementioned images are: MAE (78.88, 65.45, and 59.97); NCC (0.89, 0.92 and 0.95); PSNR (24.28, 26.30 and 28.65); SSIM (0.58, 0.67, and 0.75), respectively.
Conclusion: In this study, a primal-dual CNN was built to generate synthetic CT with more accurate HU, reduced noise, and organ structure integrity for the treatment day.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email