We have previously developed a retrospective 4DâMRI technique using body area as the respiratory surrogate, but generally, the reconstructed 4D MR images suffer from severe or mild artifacts mainly caused by irregular motion during image acquisition. Those image artifacts may potentially affect the accuracy of tumor target delineation or the shape representation of surrounding nontarget tissues and organs. So the purpose of this study is to propose an approach employing principal component analysis (PCA), combined with a linear polynomial fitting model, to remodel the displacement vector fields (DVFs) obtained from deformable image registration (DIR), with the main goal of reducing the motion artifacts in 4D MR images. Seven patients with hepatocellular carcinoma (2/7) or liver metastases (5/7) in the liver, as well as a patient with nonâsmall cell lung cancer (NSCLC), were enrolled in an IRBâapproved prospective study. Both CT and MR simulations were performed for each patient for treatment planning. Multipleâslice, multipleâphase, cineâMRI images were acquired in the axial plane for 4DâMRI reconstruction. Singleâslice 2D cineâMR images were acquired across the center of the tumor in axial, coronal, and sagittal planes. For a 4D MR image dataset, the DVFs in three orthogonal direction (inferiorâsuperior (SI), anteriorâposterior (AP), and medialâlateral (ML)) relative to a specific reference phase were calculated using an inâhouse DIR algorithm. The DVFs were preprocessed in three temporal and spatial dimensions using a polynomial fitting model, with the goal of correcting the potential registration errors introduced by threeâdimensional DIR. Then PCA was used to decompose each fitted DVF into a linear combination of three principal motion bases whose spanned subspaces combined with their projections had been validated to be sufficient to represent the regular respiratory motion. By wrapping the reference MR image using the remodeled DVFs, âsyntheticâ MR images with reduced motion artifacts were generated at selected phase. Tumor motion trajectories derived from cineâMRI, 4D CT, original 4D MRI, and âsyntheticâ 4D MRI were analyzed in the SI, AP, and ML directions, respectively. Their correlation coefficient (CC) and difference (D) in motion amplitude were calculated for comparison. Of all the patients, the means and standard deviations (SDs) of CC comparing âsyntheticâ 4D MRI and cineâMRI were 0.98±0.01,0.98±0,01, and 0.99±0.01 in SI, AP, and ML directions, respectively. The mean±SD Ds were 0.59±0.09âmm,0.29±0.10âmm, and 0.15±0.05âmm in SI, AP and ML directions, respectively. The means and SDs of CC comparing âsyntheticâ 4D MRI and 4D CT were 0.96±0.01,0.95±0.01, and 0.95±0.01 in SI, AP, and ML directions, respectively. The mean±SD Ds were 0.76±0.20âmm,0.33±0.14âmm, and 0.19±0.07âmm in SI, AP, and ML directions, respectively. The means and SDs of CC comparing âsyntheticâ 4D MRI and original 4D MRI were 0.98±0.01,0.98±0.01, and 0.97±0.01 in SI, AP, and ML directions, respectively. The mean±SD Ds were 0.58±0.10âmm...