Evaluation of U-Net based architectures for synthetic CT generation from dual-contrast MRI

Abstract

Deep learning–based MRI-to-CT synthesis supports MR-only radiotherapy and PET/MR attenuation correction, but comparisons of U-Net variants are confounded by differences in datasets, preprocessing, training protocols, and evaluation metrics. This study performed a controlled, tissue-aware, and complexity-aware comparison of five U-Net-based architectures, including U-Net, ResU-Net, Attention U-Net, U-Net++, and attention deep residual U-Net (ADR-U-Net), for brain sCT generation from T1-weighted and FLAIR MRI in 37 subjects from the CERMEP-IDB-MRXFDG database. Subject-level five-fold cross-validation reduced split-dependent bias and prevented slice-level information leakage. Evaluation included mean absolute error (MAE), root mean square error (RMSE), peak signal-to-noise ratio, structural similarity index measure, tissue-specific errors in Hounsfield units (HU), paired Wilcoxon signed-rank tests with Bonferroni correction, and computational complexity analysis. ADR-U-Net achieved the lowest whole-image MAE of 37.38 ± 6.01 HU and RMSE of 109.72 ± 17.35 HU. Its MAE improvement was 1.60 HU relative to ResU-Net, indicating a modest gain over the closest comparator. A focused ADR-U-Net modality-ablation analysis showed lower whole-image MAE with dual-contrast input (37.38 ± 6.01 HU) than with T1-only (40.79 ± 5.00 HU) or FLAIR-only input (38.41 ± 6.30 HU). Bone remained the dominant source of HU error, and U-Net++ had the highest number of floating-point operations without achieving the best accuracy. These findings suggest that model selection should consider tissue-specific HU errors, subject-level robustness, and computational cost. Because no external validation or downstream clinical endpoint was included, the results should be interpreted as controlled within-dataset technical evidence rather than evidence of clinical readiness.