The Hierarchical Feature Learning for Accurate Pixel-Wise Segmentation of Cardiac Structures in CT Images: A Comprehensive Evaluation of U-Net Variants

Abstract

Precise segmentation of cardiac structures from computed tomography angiography (CTA) images is essential for the early analysis of cardiovascular conditions. This study presents a robust deep learning model for precise segmentation of cardiac structures from CTA images, with particular emphasis on pixel-level accuracy and hierarchical convolutional feature learning to delineate thin and low-contrast vessels. Four different variants of the U-Net architecture such as Vanilla U-Net, U-Net4, Attention U-Net, and U-Net++ were evaluated on a publicly available CTA image dataset.  The Vanilla U-Net had the best overall accuracy of 99% and precision of 98%, indicating excellent feature learning with very few false positives. However, its recall of 91% was slightly lower, indicating some under-segmentation of the distal vessels. Attention U-Net and U-Net4 had the same accuracy of 98% and F1-score of 95%. Attention U-Net improved precision to 98% with attention-driven feature refinement, while U-Net4 improved recall to 93% with multi-scale feature aggregation. U-Net++ had a slightly lower accuracy of 97% but the best recall of 94%, indicating its dense skip connections that enable the refinement of fine details of the vessels. For all models, the Dice score varied from 92-94% and IoU from 88-89%, while AUC scores were all above 0.99, indicating excellent segmentation performance. Without performing any statistical test or confidence interval calculation, the results show a consistent pattern across all evaluation metrics considered in the study, which suggests that the comparisons between models are reliable. These results demonstrate that by leveraging deep learning-based feature learning accurate and continuous coronary artery segmentations can be obtained. The proposed method has promising potential in improving the non-invasive diagnosis of CAD and reducing the dependence on invasive procedures.