Contour-regulated registration framework for Liver CT perfusion images

Liver computed tomography perfusion (CTp) imaging enables quantitative assessment of vascular dynamics but is highly susceptible to motion-induced artifacts arising from breath-hold variability and involuntary patient motion. These effects cause different portions of the liver to fall within the imaging field of view across time frames and introduce in-plane deformation, creating substantial challenges for image registration. Misregistration between distinct regions of interest disrupts temporal signal consistency and perfusion estimation, underscoring the need for robust registration strategies tailored for Liver CTp.

To develop and evaluate a contour-regulated automated registration framework for correcting motion artifacts and aligning time-resolved images in Liver CTp series.

Forty-six CTp datasets from 23 patients were analyzed; each dataset consisted of a pre-contrast static scan, a post-contrast high temporal resolution cine scan (59-79 time frames), and a post-contrast low temporal resolution helical scan (8-10 time frames). A 2D liver auto-segmentation nnU-Net model was trained on 23 whole liver contours that had been manually delineated on pre-contrast static images and applied to all 46 cine and helical series. The auto-segmented contours were used to regulate a two-stage registration pipeline consisting of affine registration followed by TransMorph-based deformable registration. The models with different parameters were trained by using intensity-corrected images on 36 randomly selected cases and tested on the remaining 10 cases. Performance of multiple registration model variants were evaluated against the unregistered series by using mean surface distance (MSD) and 95th percentile Hausdorff distance (HD95) on simplified 3D surfaces reconstructed from manually placed anchor points.

Liver auto-segmentation achieved mean Dice similarity coefficients of 0.92 for helical scans and 0.89 for cine scans. Structure alignment was improved significantly compared with raw data after registration. The best-performing model had an auto-determined contour regularization weighting factor ( ${\lambda }_S)$ of 0 or 0.8, reduced the MSD from 5.1 mm in the unregistered data to 1.0 mm and HD95 from 9.7 mm to 3.0 mm in cine scans. In helical scans, MSD decreased from 3.2 mm to 1.0 mm and HD95 from 6.6 mm to 2.9 mm.

The proposed contour-regulated, two-stage registration framework improved temporal alignment in dynamic Liver CTp images, providing a robust foundation for subsequent analysis and clinical assessment.