Purpose: Original PCA lung model formulated by Zhang (2007) was validated using centroid position of a target volume. We (1) extend the analysis to the lung region using fiducial registration-error (FRE) analysis, which is more sensitive to accuracy of the deformation algorithm used to construct the model, and a more granulated evaluation of the deformation vector field (DVF). We (2) carry out preliminary analysis of the model to estimate DVF for lung ROI as well as target ROI for subsequent CT session, which can be used for margin assessment from respiration changes.
Methods: We used public DIR-LAB 4DCT datasets and perform B-spline based deformable image registration (DIR) to acquire the displacement vector fields. Then perform FRE analysis between the actual and predicted landmark positions. Next, using two 4DCT images of the same patient acquired 30 minutes apart, the PCA model constructed from initial 4DCT was used to estimate subsequent 4DCT. Residual analysis was performed over each phase of breathing cycle to quantify differences between estimated and actual motions.
Results: Mean FRE value for the DIR-LAB 4DCT cases was 1.6mm, consistent with gold-standard image-registration results, validating high fidelity of the PCA algorithm. When predicting a second 4DCT session, the mean residual over voxels in the lung region was 2.4mm. The directional components in eigenmode one and two indicate the first component is dominated by superior-inferior motion near the diaphragm, the second component is less predictable and has comparable contribution in all directions.
Conclusion: The FRE evaluation is consistent with the previous centroid position evaluation. The PCA model can be used with surrogate information acquired during treatment to reconstruct target and organ motion for the purpose of dose accumulation. Our patient case study may indicate that the first principal component explains common diaphragm-dominated motion and the second component describes cycle-specific information.