Prior-refined Segment Anything Model enables quality-robust liver segmentation on cone-beam CT for adaptive radiotherapy

Online adaptive radiotherapy (ART) requires accurate organ contours on daily cone-beam computed tomography (CBCT). However, imaging artifacts and low soft-tissue contrast often render segmentation that relies solely on daily-image appearance unreliable, hindering the automation of treatment adaptation. Clinicians therefore propagate planning contours onto the daily image with deformable image registration (DIR), but registration errors and anatomical changes can turn these transferred contours into misleading guidance. Using a synthetic, paired CBCT/CT dataset of 131 liver cases spanning five image quality levels (32 to 490 projections), we developed Prior-Refined Segment Anything Model (PR-SAM), a parameter-efficient framework that conditions a frozen Segment Anything Model on planning contours and corrects domain shift and prior misalignment through lightweight trainable modules and dual-branch decoding. PR-SAM achieved a mean Dice similarity coefficient of 0.9611 and a 95th-percentile Hausdorff distance (HD95) of 2.46 mm, with negligible variation across image quality levels, significantly outperforming seven convolutional-neural-network baselines and MedSAM (paired Wilcoxon signed-rank test: p < 0.001 at slice level; p = 0.0002 at patient level). Against DIR-based contour propagation, the clinical standard, PR-SAM was significantly more accurate under simulated inter-fraction misalignment (Dice p = 7.6 × 10 −5 ; HD95 p = 2.2 × 10 −4 ); DIR was numerically higher only under exact geometric alignment, an oracle condition intrinsic to synthetic data and unattainable clinically. Inference required approximately 7.3 s per volume in deployment configuration, comfortably within on-couch time constraints. These results suggest that prior-refined foundation models can support daily liver contouring and may reduce manual correction workload in abdominal ART.