Segmenting Bi-Atrial Structures Using ResNext Based Framework

TL;DR

This work addresses the need for robust bi-atrial segmentation from 3D LGE-MRI to aid AF ablation planning. It introduces TASSNet, a two-stage framework that localizes atrial ROIs with a coarse 3D U-Net and then performs fine bi-atrial segmentation using a 2D-3D ensemble of U-Nets with ResNext encoders and a cyclical learning rate. The approach is validated on the Utah LGE-MRI dataset and an out-of-distribution dataset, showing superior Dice scores for atrial cavities and walls and strong generalization. The framework facilitates comprehensive atrial characterization, enabling fibrosis quantification and personalized therapy planning, while highlighting challenges in exactly delineating thin atrial walls and the potential benefits of multi-modal data fusion.

Abstract

Atrial Fibrillation (AF), the most common sustained cardiac arrhythmia worldwide, increasingly requires accurate bi-atrial structural assessment to guide ablation strategies, particularly in persistent AF. Late gadolinium-enhanced magnetic resonance imaging (LGE-MRI) enables visualisation of atrial fibrosis, but precise manual segmentation remains time-consuming, operator-dependent, and prone to variability. We propose TASSNet, a novel two-stage deep learning framework for fully automated segmentation of both left atrium (LA) and right atrium (RA), including atrial walls and cavities, from 3D LGE-MRI. TASSNet introduces two main innovations: (i) a ResNeXt-based encoder to enhance feature extraction from limited medical datasets, and (ii) a cyclical learning rate schedule to address convergence instability in highly imbalanced, small-batch 3D segmentation tasks. We evaluated our method on two datasets, one of which was completely out-of-distribution, without any additional training. In both cases, TASSNet successfully segmented atrial structures with high accuracy. These results highlight TASSNet's potential for robust and reproducible bi-atrial segmentation, enabling advanced fibrosis quantification and personalised ablation planning in clinical AF management.

Figures (3)

• Figure 1: Overview of the TASSNet framework. In Stage 1, a 3D U-Net is employed to extract regions of interest (ROIs) from the input LGE-MRI volume, concentrating on the atrial structures. The output of stage I is a localised ROI, reducing spatial complexity for more precise segmentation in stage 2. In Stage 2, both 2D and 3D U-Net architectures are used to segment the atrial structures within the extracted ROIs finely. The 2D U-Net processes data slice-by-slice, generating 2D probability maps, while the 3D U-Net uses volumetric information to produce 3D probability maps. The predicted 3D model, delineating the left and right atrial walls and cavities, is the ensemble output of both networks.
• Figure 2: Performance visualisation of the ensemble model in 2D (the first row) and 3D views (the second row) for four different hearts in the Utah dataset. The left atrial (LA) cavity is highlighted in yellow, the right atrial (RA) cavity in blue, the LA wall in green, and the RA wall in red.
• Figure 3: Performance visualisation of the ensemble model in 2D (the first row) and 3D views (the second row) for four different hearts in the OOD dataset. The left atrial (LA) cavity is highlighted in yellow, the right atrial (RA) cavity in blue, the LA wall in green, and the RA wall in red.