Skull-Base Synthetic CT from Brain MRI: Development, Independent-Center Validation, and Radiologist Evaluation

We developed and evaluated a deep learning pipeline for skull-base synthetic computed tomography (sCT) generation from routine non-contrast T1-weighted and T2-weighted brain magnetic resonance imaging (MRI), focusing on clinically relevant bony anatomy for neurosurgical and radiotherapy workflows. We retrospectively collected 184 manually registered T1/T2/CT triplets for training and internal validation (166/18 split) and 43 cases from a second center for independent-center validation. A three-dimensional Swin-UNETR generator was trained with bone-weighted multiscale reconstruction, perceptual, and adversarial losses. Two board-certified radiologists independently scored 16 skull-base structures, artifacts, and diagnostic acceptability on raw sCT in Hounsfield units (HU), without histogram matching; this blinded review was the primary endpoint. Acceptability was 76.7% for Rater 1 and 67.4% for Rater 2 (pooled 72.1%, cluster-robust 95% CI 58.5-82.6%), with mean structure scores from 2.67 to 4.85 out of 5. Raw sCT showed limited absolute HU calibration (global MAE $364\pm 50$  HU, bone Dice $0.38\pm 0.09$ ), with a compressed dynamic range in which air and soft tissue were overestimated and dense bone underestimated; histogram-matched outputs, reported as a secondary analysis, reached bone Dice $0.88\pm 0.03$ and bone MAE $26.31\pm 5.03$  HU. Halo artifacts were near-universal, and false-canal artifacts occurred in 21-30% of cases. The method shows promise as a radiation-free adjunct for skull-base visualization, but the present results do not support replacing dedicated CT for diagnostic, dosimetric, or physics-based applications.