Left Atrial Appendage Segmentation Based on Ranking 2-D Segmentation Proposals

摘要

The left atrial appendage (LAA) is the main source of thrombus in patients with atrial fibrillation (AF). Automated segmentation of the LAA can greatly help doctors diagnose thrombosis and plan LAA closure surgery. Considering large anatomical variations of the LAA, we present a non-model based semi-automated approach for LAA segmentation on CTA data. The method requires only manual selection of four fiducial points to obtain the bounding box for the LAA. Subsequently we generate a pool of segmentation proposals using parametric max-flow for each 2-D slice. Then a random forest regressor is trained to pick out the best 2-D proposal for each slice. Finally all selected 2-D proposals are merged into a 3-D model using spatial continuity. Experimental results on 60 CTA data showed that our approach was robust when dealing with large anatomical variations. Compared to manual annotation, we obtained an average dice overlap of 95.12%.