The proposed U-Net architecture with a customized loss function achieved an F1 Score of 61.47% for coronary artery stenosis segmentation, outperforming previous research which achieved 53.4%.
Does a U-Net architecture with dense blocks and a customised loss function improve automated coronary artery stenosis segmentation in X-ray angiography images compared to standard architectures?
A customized U-Net deep learning architecture with dense blocks significantly improves automated coronary artery stenosis segmentation in X-ray angiography images.
Absolute Event Rate: 61.47% vs 53.4%
The disorders that affect our heart and blood vessels are cardiovascular disorders, and they are the leading cause of death worldwide. A significant disorder among these is coronary artery stenosis. Stenosis detection is a time-consuming process that requires an expert cardiologist and is also prone to human error. A fully automated system can handle all these challenges. Therefore, we presented a deep learning-based framework for binary stenosis segmentation in the coronary arteries using an X-ray angiography dataset. In this research work, three architectures 1st standard U-Net, then the U-Net enhanced with squeeze-and-excitation blocks, and last the U-Net incorporating dense blocks, where the final configuration achieved the best performance in the stenosis segmentation task. A custom loss function is employed to enhance model performance, utilising the publicly available ARCADE dataset. This dataset comprises X-ray angiography images from 1,500 patients, with 1,000 for training, 300 for validation, and 200 for testing. The training dataset was augmented to address limited data availability and enhance model generalizability. The model achieved a precision of 0.5985, a recall of 0.6319, and an F1 Score of 61.47%, whereas previous research in this challenge has achieved only a 53.4% F1 Score. The experimental results show that our method can achieve promising performance, which successfully segments the stenotic regions under the effect of small vessel size and low contrast.
Iman et al. (2026) studied Coronary artery stenosis (n=1,500). U-Net architecture with customised loss function (Weighted Categorical Entropy and Dice loss) vs. Previous research / Standard U-Net was evaluated on F1 Score for binary stenosis segmentation. The proposed U-Net architecture with a customized loss function achieved an F1 Score of 61.47% for coronary artery stenosis segmentation, outperforming previous research which achieved 53.4%.