Advancing 3D ICE: Challenges and Deep Learning Strategies in Atrium Segmentation

Intracardiac echocardiography (ICE) is increasingly used worldwide during electrophysiology and structural heart procedures. Novel advancements in ICE technology have enabled live 3D visualisation of the heart, however, volumetric images captured with 3D ICE can be difficult to interpret. To address this, we investigated the automatic segmentation of the atria from 3D ICE using a modified nnU-Net framework. We evaluated the performance of 2D and 3D models in combination with and without dilated convolutions based on dice scores. The training dataset comprised 133 3D orbital ICE images acquired during in-vivo porcine studies. For 3D model training, ICE images were pre-processed using either delay-and-sum (DAS) or iterative maximum-a-posteriori (IMAP) beamforming, with or without additional filtering via a speckle-to-speckle deep-learning model. The 3D DAS model achieved the highest dice score of 0.89 without dilated convolutions or speckle filtering.