Hybrid 3D U-Net and Attention Mechanisms for Whole Heart Segmentation from CT Images

Accurate delineation of heart structures from multimodal images is crucial for the treatment and investigation of different cardiovascular diseases. Automated whole-heart segmentation remains a challenging task due to its complex structure and imbalances in sample data. Convolutional Neural Networks (CNNs) are popular due to their efficiency in segmenting medical images. However, they often struggle to capture long-range dependencies and lack the precision needed for complex anatomic structures such as the heart. To overcome these limitations, this study presents a hybrid 3D U-Net framework that utilizes residual connections with attention mechanisms to improve feature learning and localization of cardiac structures. Residual connections stabilize training in deeper networks and attention blocks focus on relevant regions, refining segmentation quality. This network focuses on relevant regions and uses attention blocks to enhance quality. The proposed architecture was evaluated on 40 volumetric CT images of the Multi-Modality Whole Heart Segmentation (MM-WHS) dataset, achieving an average dice score of 85%. These results demonstrate the effectiveness and high accuracy of the proposed method for delineating cardiac substructures, offering potential clinical utility for automated cardiac analysis.