Enhanced PVTFormer with Boundary-Aware Attention for Robust Cardiac MRI Segmentation Under Limited Data Conditions

Abstract

Cardiac magnetic resonance imaging (CMR) is widely used for quantitative assessment of cardiac morphology and function because it enables reproducible measurement of ventricular volumes, myocardial mass, and ejection fraction. Accurate segmentation of the left ventricle (LV), right ventricle (RV), and myocardium (MYO) is therefore a critical step in computer-aided cardiac analysis. However, manual delineation remains labor-intensive and observer-dependent, while automated models may degrade when boundaries are weak, noisy, or poorly contrasted. Convolutional neural networks (CNNs), especially U-Net-based architectures, have improved medical image segmentation by combining encoder–decoder learning with skip connections, but their local convolutional operations can limit long-range anatomical modeling. Transformer-based methods address this limitation through self-attention; however, they can introduce high computational cost and may still require explicit local refinement to recover fine anatomical boundaries. This paper proposes an Enhanced Pyramid Vision Transformer framework that integrates a lightweight Boundary-Aware Attention Module (BAAM) for cardiac MRI segmentation. The method combines hierarchical transformer encoding with attention-guided boundary refinement to improve the segmentation of LV, RV, and MYO structures, with specific focus on low-contrast myocardial interfaces. Experiments were conducted on the ACDC 2017 dataset and compared with U-Net, Attention U-Net, TransUNet, nnU-Net, and a baseline PVTFormer. The proposed method achieved Dice Similarity Coefficients of 0.942 (LV), 0.914 (RV), and 0.901 (MYO), with a 95th-percentile Hausdorff Distance (HD95) of 5.83 mm. It outperformed the baseline PVTFormer by 1.1, 1.5, and 2.0 percentage points on LV, RV, and MYO Dice respectively, and reducing HD95 by 18.0% relative to the baseline. The largest gain was observed for myocardium segmentation, consistent with the design goal of BAAM. These results support the use of boundary-aware hierarchical attention for improving cardiac MRI segmentation under limited annotated data conditions.