Right atrial remodelling in atrial fibrillation evaluated with automatic segmentation using deep learning analysis

Evaluation of right atrial cavities and right atrial (RA) remodeling in atrial fibrillation (AF) has been poorly studied.

The aim of this study was to evaluate in AF, the interest of automatic segmentation using transthoracic echocardiography (TTE) to improve the evaluation of the RA.

A complete 2D TTE and 3D TTE with biplane analysis of the RA was performed in each patient hospitalized for AF and included in the ongoing FASTRHAC study. We developed a complete segmentation pipeline based on the U-Net architecture for RA delimitation in echocardiographic images. Our network's encoder, a ResNet34, was pre-trained on ImageNet (transfer learning) to optimize relevant feature extraction. The model was trained on a dataset of 1218 annotated RA images (ground truth), complemented by data augmentation techniques to improve its generalization and robustness.

Our model demonstrated high and precise performance for RA segmentation, achieving a Dice score (measure of the spatial similarity between the prediction and the ground truth) of 0.9618 and an IoU (Intersection over Union, also known as Jaccard Index) of 0.9264 on the validation set. The visualization of predictions, particularly in the form of probability maps, confirms our pipeline's ability to precisely identify the RA, even on complex contours. This automatic and reliable segmentation is fundamental for our pipeline to calculate the RA area and volume (Figure 1). The next step will be the development of automatic segmentation of the biplane RA analysis (Figure 2) to improve the RA volume analysis and then evaluate the potential association with clinical outcome.

RA automatic segmentation is a promising tool for evaluating the anatomical and functional RA remodeling in AF.