Conditional flow matching framework for metal artifact reduction in head-and-neck radiotherapy planning CT

Abstract We present a conditional flow-matching framework for image-domain metal artifact reduction (CFM-MAR) in head-and-neck CT, compared with linear-interpolation MAR (LI-MAR) and normalized MAR (NMAR). Trained on the public AAPM CT-MAR head subset (1,626 slices), CFM-MAR reduced patient-region root-mean-square error (RMSE) on the slice-level test set from 228.1 to 44.5 ± 23.0 Hounsfield units (HU), against 53.0 ± 72.5 HU for NMAR and 75.1 HU for LI-MAR. Per slice, CFM-MAR and NMAR were not significantly different ( p = 0.14); the lower mean was driven by a shorter error tail: no slice was worse than no correction and the worst slice reached 172 HU, against four worsened slices and a maximum RMSE of 854 HU for NMAR. Under leave-one-examination-out training, mean RMSE increased to 64.1 HU, exceeding NMAR (56.2 HU), while the tail advantage persisted. On 50 clinical planning CTs CFM-MAR suppressed streaks but partially filled internal air cavities. Blinded readers preferred its output to that of a matched direct-regression network with lower benchmark distortion, including for contouring. These findings are based solely on imaging measurements; dosimetric analysis, contouring validation, workflow validation and external HU validation remain to be performed.