Clinical implementation of deep learning-based synthetic CT for MRI-only volumetric modulated arc therapy in head and neck and pelvic cancer patients

Abstract

Rationale and objective: This study investigates the implementation of an MRI-only workflow in radiotherapy, focusing on synthetic CT (sCT) images generated through a deep learning-based commercial software to facilitate dose calculation and treatment delivery for volumetric modulated arc therapy (VMAT) in head and neck (H&N) and pelvic cancer patients. Methods: This retrospective analysis included 33 patients (10 H&N, 9 prostate, 9 rectum, and 5 cervical cancer). All patients underwent CT and MRI for radiotherapy planning. sCT images were generated using deep learning-based MRI Planner™ software. Clinical treatment plans were initially optimized on CT and then recalculated on sCT with identical parameters. Image quality was evaluated using the dice similarity coefficient (DSC), mean error (ME), and mean absolute error (MAE). Dosimetric accuracy was assessed by comparing dose-volume histogram (DVH) differences and performing global gamma analysis (3%/2 mm to 1%/1 mm) between CT- and sCT-based plans. Treatment plan quality assurance (QA) conducted with an electronic portal imaging device (EPID) was compared between sCT and CT using global gamma analysis (3%/2 mm to 2%/2 mm). Patient set-up verification was evaluated by comparing CT-cone beam CT (CBCT) and sCT-CBCT. Results: sCT images demonstrated high accuracy, with average body DSCs of 0.99 (SD = 0.00), ME of -8.94 HU (SD = 9.50), and MAE of 67.30 HU (SD = 6.94) for H&N, and DSCs of 1.00 (SD = 0.00), ME of -8.74 HU (SD = 6.26), and MAE of 33.64 HU (SD = 3.91) for the pelvis. Lower DSCs were observed in bone, with 0.87 (SD = 0.03) in H&N and 0.88 (SD = 0.02) in pelvis. Dose differences were within 2%, with average gamma pass rates of all criteria 92% for H&N and 96% for pelvis, across all criteria. QA plan evaluation revealed the gamma pass rate between CT and sCT of 0.05% for H&N and 0.12% for pelvis. Mean positioning differences between CT-CBCT and sCT-CBCT were 0.13 mm and 0.06 for the H&N and 0.17 mm and 0.13 for the pelvis in all directions. Conclusion: Deep learning-based software successfully generated accurate sCT images for both H&N and pelvic cancer patients, supporting reliable dose calculation, treatment plan QA, and patient set-up verification. This enables the potential implementation of MRI-only workflows for VMAT in H&N and pelvic cancer treatment.