Purpose: To identify the model training parameters that result in the synthetic CTs with the best fiducial visualization and normal tissue accuracy for MR-only stereotactic radiation therapy to the pelvis.
Methods: We previously developed a machine learning model to create synthetic CTs from the paired MRs. The model is applied for 11 CT/MR (conventional T2) image sets used for previous Cyberknife (Accuray, Inc) radiosurgery treatments. The MR images were pre-processed to mimic the appearance of fiducial-enhancing images. Model performance was evaluated on three factors, fiducial contrast (“realistic” vs manually enhanced), patch sizes and numbers of image pairs in the training. Four observers identified the fiducial centers in the synthetic and real CTs and associated DRRs to evaluate the fiducial reconstruction accuracy. The CT and DRR images were also qualitatively compared using mean squared error (MSE) and peak-signal-to-noise-ratio (PSNR). The same treatment plan was calculated on the real and synthetic CT.
Results: We found that the manually enhanced images, largest patch size (128x128 pixels), and 9 image pairs in the training set gave the best clinically applicable synthetic CTs and the highest contrast for the reconstructed fiducials. The fiducial appearance in the synthetic CTs, as well as the DRRs generated from them had similar contrast and visualization to the real CT. There was no significant difference found for the fiducial localization for the CTs and DRRs (p>0.05). The mean MSE between the true and synthetic CTs was 2.77%±1.31% and the mean PSNR was 32.45±4.61. The MSE and PSNR for DRRs were 1.2±1.5% and 22.8±5.7, respectively. 3D gamma analysis on the dose distributions calculated on the real and synthetic CTs both pass 100% (2%/2mm)
Conclusion: Our machine learning based method generates synthetic CTs for accurate treatment planning and DRRs with sufficient image contrast for target tracking in prostate stereotactic radiotherapy.