Multi-Depth Boundary-Aware Left Atrial Scar Segmentation Network

TL;DR

This paper introduces a boundary-aware neural network for automatic segmentation of left atrial scars from cardiac MRI images, leveraging spatial relationships between LA and scars to improve accuracy.

Contribution

It proposes a novel two-branch network with a Sobel fusion module that propagates boundary information, enhancing scar segmentation performance.

Findings

Achieved an average Dice score of 0.608 for LA scar segmentation.

Utilized 40 images for training and 20 for evaluation.

Demonstrated the effectiveness of boundary-aware segmentation in cardiac MRI.

Abstract

Automatic segmentation of left atrial (LA) scars from late gadolinium enhanced CMR images is a crucial step for atrial fibrillation (AF) recurrence analysis. However, delineating LA scars is tedious and error-prone due to the variation of scar shapes. In this work, we propose a boundary-aware LA scar segmentation network, which is composed of two branches to segment LA and LA scars, respectively. We explore the inherent spatial relationship between LA and LA scars. By introducing a Sobel fusion module between the two segmentation branches, the spatial information of LA boundaries can be propagated from the LA branch to the scar branch. Thus, LA scar segmentation can be performed condition on the LA boundaries regions. In our experiments, 40 labeled images were used to train the proposed network, and the remaining 20 labeled images were used for evaluation. The network achieved an average Dice score of 0.608 for LA scar segmentation.