Room: AAPM ePoster Library
Purpose: Synthesizing CT from MR is needed for dose calculation in MR-only radiotherapy planning. A lot of work has been done to use deep learning for this purpose, including CycleGAN. However, all of these works tested the model on data acquired with the same protocols, so it is currently unknown how the model performs for different MR scanning protocols. The purpose of his work is to address the model generalizability by using transfer learning to adapt the model from T2 MR to T1 MR.
Methods: Three different types of MR images were collected, including T2 MR, T1-FLAIR MR and T1-POST MR. We took T2 MR and corresponding CT images as the source dataset, T1-FLAIR MR and T1-POST MR as the target datasets. Firstly, the source model was trained with CycleGAN using the source dataset. Then, the pre-trained source model was adapted to one of the target datasets. Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) were used to evaluate the quality of the synthetic CT.
Results: For adapted model, we achieved lowest MAE and RMSE of 75.70 and 195.86 on FLAIR dataset, 75.45 and 195.49 on POST dataset and highest PSNR and SSIM of 28.17 and 0.84 on FLAIR dataset, 27.72 and 0.82 on POST dataset. Quantitative results indicate that adapted model outperforms source model and target model (re-trained on target dataset only) on FLAIR and POST dataset. Qualitative visual evaluation also shows that adapted model could generate more accurate details than the source model and less error than source model and target model on both datasets.
Conclusion: This work indicates that the pre-trained CycleGAN model for MR to CT conversion can be transferred to the datasets acquired through different scanning protocols.