Deep learning based U-net variants for cardiac MRI segmentation

Abstract

Accurate segmentation of ventricular structures and the myocardium from Cardiac Magnetic Resonance (CMR) images is essential for the diagnosis and management of cardiovascular diseases. This study presents a comprehensive approach to cardiac MRI segmentation by developing and evaluating six U-Net variants: Original U-Net, Residual U-Net, Attention U-Net, Feature Pyramid U-Net, Feedback Residual U-Net, and Transformer-Based U-Net, each incorporating architectural enhancements tailored to address specific challenges in segmenting complex cardiac anatomy. These architectures incorporate advanced enhancements such as deeper encoder levels, attention mechanisms, residual connections, multi-scale feature fusion, transformer modules, and feedback mechanisms. To improve segmentation robustness, a novel hybrid loss function, combining Dice Loss and Cross-Entropy Loss, was proposed to effectively manage class imbalance and improve segmentation precision. Among the evaluated models, the Feature Pyramid U-Net achieved the highest performance, with Dice coefficients of 0.9388 (Left Ventricle), 0.8759 (Right Ventricle), and 0.8426 (Myocardium), demonstrating its superior ability to capture multi-scale contextual information. To bridge the gap between research and clinical application, an interactive web application was developed and deployed, enabling real-time inference, visual inspection of annotated segmentations, and region-specific descriptions through a user-friendly interface. This work not only advances the design of deep learning architectures for medical image segmentation, but also demonstrates a practical pathway for integrating these models into clinical workflows.