Biatrial Segmentation from Cine MRI Using Convolutional and Attention Networks

Accurate segmentation of the atria, including their appendages, is fundamental for quantitative assessment of cardiac function and as a first step in the development of digital twins. While recent transformer-based models such as MedFormer have introduced novel mechanisms to capture global anatomical context, their actual benefit over more established convolutional architectures in non-contrast cine MRI remains underexplored. In this work, we evaluate the performance of MedFormer in comparison to the widely adopted U-Net architecture for atrial segmentation from long-axis cine cardiac MRI, using a private dataset comprising 3-chamber (3CH) and 4-chamber (4CH) views. Segmentation performance in terms of Dice Similarity Coefficient (DSC), Average Surface Distance (ASD), and Hausdorff Distance (HD) were computed. In the 3CH view, U-Net achieved a DSC of 0.980±0.001, ASD of 0.023±0.004 mm, and HD of 0.221±0.050 mm, with MedFormer reporting comparable results (DSC: 0.979±0.001, ASD: 0.027±0.053 mm, HD: 0.225±0.048 mm). In the 4CH view, both models demonstrated comparable performance (U-Net: DSC 0.977±0.012; MedFormer: DSC 0.983±0.009). The segmentation performance with both networks matches that reported in recent state-of-the-art methods for atrial segmentation, highlighting the suitability of both architectures for this task.