Registration based filtering: an acceptable tool for noise reduction in left ventricular dynamic rotational angiography images?

VT ablations could benefit from Dynamic 3D (4D) left ventricle (LV) visualization as road-map for anatomy-guided procedures. We developed a registration-based method that combines information of several cardiac phases to filter out noise and artifacts in low-dose 3D Rotational Angiography (3DRA) images. This also enables generation of accurate multi-phase surface models by semi-automatic segmentation (SAS). The method uses B-spline non-rigid inter-phase registration (IPR) and subsequent averaging of the registered 3DRA images of 4 cardiac phases, acquired with a slow atrial pacing protocol, and was validated on data from 5 porcine experiments. IPR parameter settings were optimized against manual delineations of the LVs using a composed similarity score (Q), dependent on DICE-coefficient, RMSDistance, Hausdorff (HD) and the percentage of inter-surface distances <= 3mm and <= 4mm. The latter are clinically acceptable error cut-off values. Validation was performed after SAS for varying voxel intensity thresholds (ISO), by comparison between models with and without prior use of IPR. Distances to the manual delineations at optimal ISO were reduced to <= 3mm for 95.6 +/- 2.7% and to <= 4mm for 97.1 +/- 2.0% of model surfaces. Improved quality was proven by significant mean Q-increase irrespective of ISO (7.6% at optimal ISO (95% CI 4.6-10.5, p < 0.0001)). Quality improvement was more important at suboptimal ISO values. Significant (p < 0.0001) differences were also noted in HD (-20.5%; 95% CI -12.1%-29.0%), RMSD (-28.3%; 95% CI -21.7%--35.0%) and DICE (1.7%; 95% CI 0.9%-2.6%). Generating 4D LV models proved feasible, with sufficient accuracy for clinical applications, opening the perspective of more accurate overlay and guidance during ablation in locations with high degrees of movement.