(Wednesday, 7/17/2019) 1:45 PM - 3:45 PM
Room: Stars at Night Ballroom 2-3
Purpose: To develop a novel deep-learning-based method for generating synthetic CT (SCT) from routine MRI for proton treatment planning for base of skull tumors.
Methods: We propose a novel 3D Cycle-GAN framework for proficient nonlinear mapping between MR and planning CT (pCT) image pairs. A training cohort of 25 patients with co-registered pCT and MR images were used. Another 25 patients were used as the study cohort to evaluate the efficacy of the method in the following aspects: 1) Image quality of SCT was evaluated using mean absolute error (MAE) and normalized cross-correlation (NCC) metrics; 2) Dosimetric accuracy was investigated by comparing several dose volume histogram (DVH) endpoints obtained from plans calculated using pCT and SCT, respectively. A 3D gamma-index test was conducted to quantify dosimetric accuracy; 3) Proton distal range was calculated along each beam.
Results: Mean MAE and NCC of the study cohort is 54.55 ± 6.81 HU and 0.96 ± 0.01, respectively. Dmean, Dmax, D10%, D50% of PTV show no statistically significant differences. D95% shows an average of 0.259 Gy difference which is within 0.47% relative dose difference. Average passing rate of 90.8%, 98.42%, and 99.96% were achieved for our gamma-index test with 1%-1 mm, 2%-2 mm, 3%-3 mm criteria, respectively. Proton distal range differences for 50 beams fall within tolerance adopted by three institutions with a maximum value of 4 mm.
Conclusion: The proposed deep-leaning-based method is capable of generating high quality SCTs from routine MRIs for base of skull tumors. High image similarity, good dosimetric and proton distal range agreement between sCT and pCT warrant further development of a MRI-only workflow for proton radiotherapy of base of skull tumors.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
