MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Deep-Learning-Based Synthetic-CT Generation Method for CBCT-Guided Proton Therapy

Y Liu1*, Y Lei1 , L Lin2 , A Dhabaan2 , T Liu1 , W Curran1 , J Zhou2 , X Yang1 , (1) Emory University, Atlanta, GA, (2) Emory Proton Therapy Center, Atlanta, GA,

Presentations

(Wednesday, 7/17/2019) 1:45 PM - 3:45 PM

Room: Stars at Night Ballroom 2-3

Purpose: The widely applied CBCT has allowed for improved daily 3D image-guided radiotherapy. However, the severe artifacts found in CBCT limit its clinic use. In this study, we explored the feasibility of generating CBCT-based synthetic-CT on which proton dose can be calculated with comparable accuracy to the CT-based plan.

Methods: Registered CT-CBCT pairs are usually used to train the deep-learning-based networks which can predict a mapping from CBCT to synthetic-CT. Due to the patient motion during image acquisition and anatomic change throughout the treatment course, perfectly-registered pairs are difficult to obtain. To minimize the mismatch-induced prediction error, a novel residual-cycle-GAN was employed by supervising both forward and backward projections to achieve one-to-one mapping. Four generators with residual blocks and two discriminators were used in the training networks to enhance the similarity between CT and synthetic-CT. Through the residual connection in each residual block, an input bypasses the hidden layers from which specific differences between CBCT and CT are learned. A cohort of 7 head and neck patients, consisting 13 CBCT scan and 5 pelvic patients, consisting 9 CBCT scan was used to evaluate the algorithm by leave-one-out cross-validation. Initial CT-based plans were transferred to synthetic-CT for proton dosimetric comparison.

Results: Mean absolute error (MAE) and peak signal-to-noise ratio (PSNR) were 18.11±4.45 HU and 34.45±2.67 dB for the head and neck patients and 16.36±4.74 HU and 35.47±2.26 dB for the pelvic patients. The differences in PTV V100, D95 and D1 between CT- and synthetic-CT-based plans are 0.06%±2.00%, 0.00%±0.80%, and 0.07%±0.70%.

Conclusion: This work generated synthetic-CT from routine CBCT images based on our proposed residual-cycle-GAN algorithm. The image similarity and dosimetric agreement between the CT and synthetic-CT warrant further development of CBCT-based adaptive radiotherapy to update structural contours and dose calculations.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email