Boundary Aware Multi Scale Attention U Net for Automated Cardiac MRI Segmentation

Abstract

Accurate segmentation of cardiac structures in cine cardiac magnetic resonance (CMR) images is essential for quantitative assessment of cardiac function and disease. However, automatic delineation of the right ventricle, myocardium, and left ventricle remains challenging because of anatomical scale variations, weak tissue boundaries, irregular right-ventricular geometry, and substantial pixel-level class imbalance. This study proposes a Boundary-Aware Multi-Scale Attention U-Net (BMAA-U-Net) for automated segmentation of cardiac structures from CMR images. The proposed framework combines parallel multi-scale convolutional operations to capture spatial information at different resolutions with attention-gated skip connections to focus on relevant anatomical features. A dedicated boundary prediction branch is incorporated to enhance the localization of cardiac structure boundaries, while residual feature fusion is used to strengthen feature representation across the network. To address class imbalance, a class-weighted Dice–cross-entropy loss is employed during training. The proposed approach is evaluated using a reconstructed Automated Cardiac Diagnosis Challenge dataset containing 150 patients, 300 annotated cardiac phases, and 2,978 two-dimensional slices. The study investigates the effectiveness of integrating multi-scale feature extraction, attention mechanisms, and explicit boundary supervision for reliable automated cardiac MRI segmentation.