MedSAM-prior-guided coarse-to-fine CBCT synthesis from ZTE MRI

Cone-beam computed tomography (CBCT) is valuable for assessing fine bony structures in dental and craniofacial imaging but entails ionizing radiation. Zero echo time (ZTE) MRI captures bone-related signals without radiation, yet it falls short of reproducing CBCT-level boundaries and textures. We propose a MedSAM-prior–guided coarse-to-fine framework to synthesize CBCT (sCBCT) from ZTE MRI while preserving low radiation benefits. A frozen MedSAM image encoder provides stable coarse anatomy; targeted low-rank adaptation (LoRA) confined to upper encoder blocks and a skipfusion decoder restore details. Using a single-institution dataset of 26 paired ZTE MRI–CBCT scans (21 training, 5 test), the proposed method reduced RMSE and MAE and increased PSNR and SSIM over a baseline, with consistent gains across all five test subjects. In clinical image assessment, real CBCT scored highest, and among synthetic images our method outperformed the baseline and a broader upper-layer adaptation model. Artifacts and noise were comparable or improved, whereas resolution remained lower than real CBCT. These results support combining a pretrained medical prior for coarse structure with data-aligned restoration of fine detail for sCBCT generation from ZTE MRI.