Purpose: To evaluate the accuracy of synthetic CT from MRI using a conditional generative adversarial network (cGAN) for stereotactic radiosurgery (SRS).
Methods: A method to generate synthetic CT was developed using an image-to-image translation cGAN with a loss function including a least-square adversarial term, mean-absolute-error (MAE) regularization and mutual information. Training datasets included 25 pairs of CT-sim images and diagnostic T1-weighted MRI with gadolinium contrast acquired for SRS planning. Before training, non-uniformity corrections and image normalization using z-score were applied to MR images. Synthetic CT images were compared with true CT images for 5 cases (brain metastases, arteriovenous malformation and vestibular schwannoma) in terms of image quality and dosimetric accuracy. To evaluate the dosimetric accuracy, clinical plans were recalculated on synthetic CT images with the fixed MUs in an Eclipse treatment planning system. Dose volume histograms (DVHs) and 3D gamma analyses (1%/1mm) were compared between synthetic CT (with and without density corrections) and true CT datasets.
Results: For the test datasets, CT number MAEs for air, soft tissues and bone were 456, 53 and 320 respectively. The peak signal-to-noise ratios (PSNR) were in the range of 23.2-25.2 and the structural index similarity measures (SSIM) were in the range of 0.905-0.92. The 3D gamma analyses (γ<1) for true CT data against synthetic CT data with and without density corrections were 98%-100% and 87%-95%, respectively. Compared to DVHs for the true CT data, D99.5% for synthetic CT data with and without density corrections was -0.4%±0.7% and 1.2%±1.9%, respectively, while D2% for the data with and without corrections was -0.2%±0.6% and 2.2%±0.5%.
Conclusion: Our study demonstrated the potential of using synthetic CT from MRI for SRS treatment planning. Our future work will validate our method on more cases and also investigate the accuracy of synthetic CT for SRS position verification.
Not Applicable / None Entered.