Isolation of the left atrial surface from cardiac multi-detector CT images based on marker controlled watershed segmentation

Abstract 

The delineation of left atrium (LA) and pulmonary veins (PVs) anatomy from high resolution images holds importance for atrial fibrillation (AF) investigation and treatment. In this study, a semiautomatic segmentation procedure for LA and PVs inner surface from contrast enhanced CT data was developed. The procedure consists of a three dimensional marker controlled watershed segmentation applied to the external morphological gradient, followed by variable threshold surface extraction from the original intensity image. A preliminary anisotropic non-linear filtering was implemented to improve the S/N ratio of CT images. The performance of segmentation was evaluated on cardiac CT scans of 12 AF patients both qualitatively and quantitatively. The qualitative evaluation by expert radiologist assessed the segmentation as overall successful in all patients and capable of extracting both the LA body and the connected vascular trees. The quantitative validation, by computing discrepancy measures with respect to a manually segmented gold standard, indicated an average of about 90% of voxels correctly classified and an average border mismatch lower than 1.5 voxels (1.2mm). The accurate extraction of the inner LA-PVs walls provided by this method, along with the minimal required human intervention, should facilitate the use of anatomical atrial models for the non-pharmacological treatment of AF.

Keywords: Image segmentation, Watershed, Anisotropic filtering, Mathematical morphology, Computed tomography, Left atrium