Abstract 3417: Four-Dimensional Active Appearance Model Segmentation of Cardiac Magnetic Resonance Images
Analysis of cardiac magnetic resonance (CMR) images is often performed using a limited number of cardiac phases and/or spatial slices and fails to correctly represent the 4D nature of cardiac motion. We developed an active appearance model (AAM) method for 4-dimensional (3D+time) segmentation of the complete left and right ventricles (LV endocardial and epicardial surfaces, RV endocardial surface). Ten patients with repaired tetralogy of Fallot (TOF) and resultant pulmonary regurgitation (3 M, 7 F, 30±9 years) and 25 normal subjects (14 M, 11 F, 25±5 years) underwent steady-state free precession 2D cine CMR imaging of the heart. To avoid incomplete LV/RV coverage in either short or long axis data when taken separately, the long and short axis images were registered and fused to form 16-cardiac-phase 4D data sets with cubic voxels (1.5–2.0 mm3) in which LV/RV surfaces were traced to form an independent standard. The expert knowledge embedded in the tracings was learned by a 4D AAM as variations of ventricular shape, gray level, and motion. The AAM was used for automated LV/RV segmentation. The 4D segmentation results were compared to the independent standard using a hold-out strategy where the AAM was repeatedly created and trained on a subset of 20 normal and 8 TOF subjects and then had its performance tested in the remaining 7 subjects. The hold-out training and testing was repeated 5 times. The automated AAM segmentation achieved accurate LV/RV surface detection in 16-phase CMR data in normal and TOF subjects judged by border positioning errors (see Table⇓). Overall, larger segmentation variability was observed in TOF subjects due to the limited TOF sample size and the corresponding lack of representative shapes and motion in some of the hold-out training sets. Our 4D AAM method successfully segmented normal and dilated LV/RV when a representative training set was used. The 4D character of the reported segmentation is a major advance in automated ventricular analysis techniques.