Feasibility of a novel algorithm for automated reconstruction of the left atrial anatomy based on intracardiac echocardiography

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Background Intracardiac echocardiography (ICE) is frequently used to guide electrophysiology procedures. The novel pre-commercial automated algorithm is a model-based algorithm, developed using machine learning methodology, which reconstructs a 3D anatomy of the left atrium (LA) and its structures based on a set of 2D ultrasound (US) frames, without the need to manually annotate US contours. Potential advantages are shortening of mapping time and better anatomy resolution when compared to conventional anatomical mapping techniques. Purpose Determine the early feasibility of the automated LA anatomy algorithm in routine clinical setting. Methods We included 16 patients (64±9 years, 88% males, body mass index 27±3, parasternal long axis LA diameter 40±3mm) undergoing LA mapping/catheter ablation (conscious sedation in 56% and general anaesthesia in the remaining cases). 2D US frames were acquired from three ICE-positions: superior vena cava/high right atrium (RA) junction, RA and right ventricular outflow tract. The automated LA anatomy map was validated in two steps: 1) identification of anatomical structures (pulmonary veins [PV] and left atrial appendage) by alignment of the ablation catheter to the automated map; and 2) assessment of PV anatomical accuracy by analysing the relationship with the automated lesion tags (3mm lesion radius, 3mm for 8sec stability criteria and 5-20g average contact force) and the PV antrum of the automated map in 9 patients with paroxysmal atrial fibrillation undergoing first time point-by-point radiofrequency PV isolation (PVI). In step 2, PV pairs were divided into 6 segments (total of 12 segments per patient) and were classified as accurate, insufficient ("floating" lesion tags), or excessive anatomy (invisible lesion tags). Results Mean 2D US frames per patient was 29±6 and acquisition time was 16±4 min. In the step 1 validation, all anatomical structures were correctly identified in the 16 patients by the automated algorithm. The step 2 validation showed anatomical accuracy in 76% of all 108 PV segments analysed. There was a range of 0-2 and 0-5 of the total 12 segments per patient with insufficient and excessive anatomy, respectively. In all cases with PV segment anatomical discrepancy, this was corrected with standard anatomy collection with the ablation catheter in ≤2min. In the whole patient group, catheter ablation was successfully performed on the automated map without complications. Conclusions The automated ICE-based 3D LA map algorithm performed overall well and correctly identified the LA anatomical structures in all patients. Anatomical PV antrum accuracy was high, and the majority of PV segments needed no manual correction before proceeding to PVI.