Advancing Cardiac MRI segmentation techniques using a hybrid U-Net architecture with ViT, and Residual Blocks

Abstract

"Semantic segmentation annotates each class in an image pixel by pixel. Segmenting cardiac magnetic resonance imaging (MRI) is crucial for diagnosing various cardiac pathologies such as coronary diseases. This process typically requires significant manual effort from multiple radiologists and cardiac imaging specialists. Although automated approaches exist, there is still much room for improvement in accuracy, making this a continuously researched area. This study introduces a novel artificial intelligence-based method for segmenting CMRIs, focusing particularly on the left ventricle (LV), right ventricle (RV), and myocardium (MYO). Precise segmentation of these components is essential for evaluating heart health and managing disorders. The method uses a hybrid U-Net architecture enhanced with residual blocks and Vision Transformers (ViT) to improve accuracy, efficiency, and flexibility of cardiac segmentation. Utilizing the M&Ms-2 dataset, this study also explores the benefits of using PReLU instead of the traditional ReLU activation function in convolutional neural networks. Additionally, it examines optimal weight distribution among chosen loss functions. With the proposed hybrid U-Net models, the Dice Similarity Coefficient (DSC) scores surpassed the 6th rank among the published top results in the MnMs-2 challenge with 0.91 for short-axis (3D) and 0.89 for long-axis (2D), which is considered a benchmark."