Adaptive multi-scale attention-gated 3D U-Net for accurate cardiac MRI segmentation

Accurate segmentation of cardiac substructures from three-dimensional cardiac magnetic resonance imaging (3D cardiac MRI) is essential for quantitative functional assessment and computer-aided diagnosis. However, existing U-Net–based architectures often struggle with multi-scale anatomical variability, low inter-tissue contrast, and limited ability to model spatial–channel feature dependencies, motivating the need for more advanced feature selection and context aggregation strategies. This study aims to develop a robust 3D segmentation framework that addresses these limitations. This paper proposes YASAM-Net, an enhanced 3D U-Net that integrates Residual Squeeze-and-Excitation blocks and a dual-path attention mechanism (YASAM) capable of jointly modeling spatial and channel-wise dependencies while adaptively fusing multi-scale contextual information. The model achieves a mean Dice score of 0.90 on the ACDC dataset, with class-wise Dice values of 0.92 (left ventricle), 0.90 (right ventricle), and 0.89 (myocardium). It further demonstrates high structural similarity (SSIM = 0.95) and low boundary error (HD95 = 7.71 mm; ASSD = 1.83 mm). Additional cross-dataset validation on the M&Ms and LVSC datasets confirms the method’s robustness under domain shift. Overall, YASAM-Net provides a precise, stable, and computationally efficient solution for automated cardiac MRI segmentation, with strong potential for integration into real-world multi-center clinical workflows.