Evaluation of U-Net Based Architectures for Synthetic CT Generation from Dual-Contrast MRI

Deep learning, especially U-Net-based models, has become the dominant approach for synthetic CT (sCT) generation for MRIonly radiotherapy and PET/MR attenuation correction. However, the field lacks direct comparisons of different U-Net variants under identical conditions. In this study, we evaluated five U-Net-based architectures, U-Net, ResU-Net, Attention U-Net, U-Net++, and attention deep residual U-Net (ADR-U-Net) for generating sCT from paired T1-weighted and FLAIR brain MRI. All models were trained on the same dataset using the same input format, preprocessing pipeline, architecture depth, optimizer, loss function, and training setup. A combined loss of mean absolute error (MAE) and Structural Similarity Index Measure (SSIM) was used. Performance was assessed at both voxel and patient levels using Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), MAE, and SSIM. All models successfully produced pseudo-CT images. Among them, ADR-U-Net achieved the best performance, with the lowest MAE (35.3 HU) and RMSE (87.2 HU), and the highest SSIM (0.923) and PSNR (36.4 dB). Qualitative results showed that all networks preserved brain structure, while ADR-U-Net was more accurate near bone-air interfaces. Patient-wise analysis also confirmed its superior consistency and robustness. This study demonstrates the value of combining residual and attention modules in the same architecture, along with dual-contrast MRI input, for improved sCT accuracy and stability. These findings provide guidance for selecting and developing deep learning models in MRI-only radiotherapy and PET/MR workflows. Future work may include clinical validation, incorporation of pathological cases, and extension to 3D or dose-aware models.