3D segmentation of the left atrial appendage in computed tomography for planning of transcatheter occlusion

Abstract

Anatomical evaluation of the Left Atrial Appendage (LAA) is paramount in patients undergoing percutaneous LAA occlusion as an alternative to oral anticoagulation. Three-dimension assessment and visualization of the LAA is valuable to understand the anatomy of this very variable anatomic structure aiding device-selection andsizing. In this context computed tomography (CT) has been successfully used, however, the clinical practice is mainly based on manual analysis, making it tedious and observer dependent.In this work, we propose a methodology to perform a fast 3D segmentation of the LAA in CT based on our curvilinear B-spline Explicit Active Surface framework. Starting from a manually defined centerline, a blind-ended model is initialized and refined through: (i) a fast contour growing using a global formulation and (ii) a contour refinement based on local energies. Its feasibility in terms of segmentation error and clinical metrics (the diameter of the landing zone [LZ] and the LAA depth) is compared with manual references.The proposed methodology was tested in a database with 15 patients. It registered a point-to-surface error of 0.80±0.24mm, a Dice value of 88.08±3.57%, and a Hausdorff distance (95th percentile) of 2.55±1.19mm. It was found a bias error of -0.36mm (p>0.05) and -0.55mm (p<0.05) for the LZ and LAA depth, with narrow limits-of-agreement for both cases (LZ: [-1.98,1.27]mm; LAA depth: [-2.20,1.10]mm).The proposed methodology showed its high feasibility and performance in this preliminary study, reinforcing its potential for the planning of LAA occlusion.

Deep learning-based automatic left atrial appendage filling defects assessment on cardiac computed tomography for clinical and subclinical atrial fibrillation patients

“…For automatic LAA segmentation, Morais et al achieved a Dice score of 0.88 for their semi-automatic 3D LAA segmentation framework [ 20 ]; however, it relied on manual setting of the LAA centerlines. Takayashiki et al utilized a U-Net based model for LAA segmentation on heart CT images [ 21 ], but the reported accuracies were poor.…”