Abstract
AIMS: This study evaluates the feasibility of deep learning-generated synthetic computed tomography (sCT) for proton dose calculation. MATERIALS AND METHODS: The sCT images were generated from T2-weighted magnetic resonance imaging (MRI) of 10 retrospectively collected prostate cancer patients using MRI Planner (Spectronic Medical, Sweden). The sCT images were compared with CT images to assess image quality, proton range, and dosimetric evaluation. Image quality was evaluated using mean Hounsfield unit (HU) difference, mean absolute error (MAE), and mean error. Geometric agreement between CT and sCT images was measured using the dice similarity coefficient (DSC). Ground truth CT images were employed to generate single-field uniform dose (SFUD) and intensity-modulated proton therapy plans for each patient. The proton range shift (RS) between CT and sCT images was calculated at the central 80% distal dose falloff in each SFUD plan. Dosimetric evaluations were performed using dose-volume histogram comparison and gamma index analysis. RESULTS: The mean MAE values for the body, bone, and soft tissue were 43.42 ± 3.96, 118.40 ± 9.70, and 33.90 ± 4.04 HU, respectively. DSC values revealed good geometric agreement between sCT and CT images. All RS values fell within clinically acceptable criteria, with a relative RS ranging from 0.02%-0.67%. The relative dose differences for all clinical target volume metrics were less than ±1.0%, and gamma pass rates exceeded 90% at 3%/3 mm, 3%/2 mm, and 2%/2 mm criteria. CONCLUSION: sCT images generated by MRI Planner were feasibly used for proton dose calculation in prostate cancer and could potentially be used for implementing MRI-only proton therapy.