Evaluating Convolution, Attention, and Mamba Based U-Net Models for Multi-class Bi-Atrial Segmentation from LGE-MRI

Claas Thesing, Alfonso Bueno-Orovio, Abhirup Banerjee

Abstract: Atrial Fibrillation (AF) is the most prevalent cardiac arrhythmia, associated with a significantly increased risk of stroke and heart failure. Thus, it is critical to analyse the progression of AF, which requires accurate segmentation of the atria in medical images. This study addresses the challenge of automating this segmentation process for Late Gadolinium-Enhanced Magnetic Resonance Imaging (LGE-MRI) scans through a multi-class segmentation approach targeting the left and right atria and atrial walls. Instead of introducing a novel architecture, we evaluate existing U-Net-based models employing different backbone architectures, including convolutional layers, attention-based layers, and Mamba blocks, and assess their performance over the Multi-class Bi-Atrial Segmentation (MBAS) Challenge dataset. Additionally, we compare the impact of different preprocessing techniques on the segmentation performance. Achieving the third rank in the MBAS Challenge with a convolutional nnU-Net model, we demonstrate its superior performance compared to other models. Our findings aim to inform the development of more efficient and accurate segmentation methods in the clinical evaluation of AF.

External IDs:doi:10.1007/978-3-031-87756-8_22