Evaluation of a Patient-Specific Respiratory Motion Model in Thoracic and Abdominal Phantom and Patient CT Images
F Liu*, Q Zhang, Y Hu, K Goodman, G Mageras, Memorial Sloan Kettering Cancer Center, New York, NYTU-E-BRA-9 Tuesday 2:00:00 PM - 3:50:00 PM Room: Ballroom A
Purpose: We have previously described a model of patient-specific respiratory motion to predict organ deformations without assuming repeatable breath cycles. The model is derived from deformable image registration (DIR) between respiration-correlated images (RCCT), followed by a principal component analysis (PCA) which relates the first two principal components of 3D deformations to the position and direction of motion of the diaphragm or implanted fiducials. This study examines model accuracy in phantom and patient images.
Methods: We compare model and DIR accuracy using 3 types of image sets, each exhibiting different deformation patterns: (1) synthetic images in lung and abdomen from the 4D NURBS-based cardiac torso (NCAT) phantom with known deformations; (2) CT scans of physical deformable phantom with implanted markers in liver; and (3) liver structures in patient RCCT images using rigid registration in a small VOI as approximate ground truth. The model is calibrated by applying fast free-form DIR between a reference image set at end expiration and each of the other images at different motion states, defined by diaphragm or, in some patient cases, implanted fiducials as surrogate signals. Following PCA, the first two principal components are selected to yield a model-predicted displacement field for the given surrogate signal.
Results: Discrepancy between model prediction and ground truth (mean ± stand deviation) in 3D displacements is 3.3±2.0 mm in lung and 3.7±1.9 mm in abdomen in NCAT phantom, 3.8±2.7 mm in physical deformable phantom and 2.8±2.9 mm in patient data (N=7). Corresponding DIR discrepancies are 3.8±2.0 mm (NCAT lung), 3.7±1.8 mm (NCAT abdomen), 3.6±2.8 mm (physical phantom), and 2.0±2.2 mm (patient data).
Conclusions: Motion model accuracy is found to be comparable to fast free-form in all three types of images, indicating that the assumption of two principal components is sufficient to describe the fast free-form DIR-derived deformations.