The accumulation of atherosclerotic atheroma in coronary arteries leads to significant cardiac risks. Accurate segmentation of coronary arteries in X-ray coronary angiographic (XCA) images is essential for diagnosing atherosclerotic disease and guiding interventional procedures. However, low image quality and complex vascular morphology pose significant challenges. To address these issues, we proposed Angio-Fusion Net, a deep learning framework designed to enhance vessel delineation in challenging angiographic conditions. Angio-Fusion Net employs a single Attention-VGG16-U-Net model for segmentation, while applying two different preprocessing methods to improve image quality. The first combines Top-Hat Morphology and Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance fine vascular structures, while the second integrates Gamma Correction with CLAHE to improve visibility in low-contrast conditions. The segmentation model leverages VGG16's hierarchical feature extraction and U-Net's spatial precision, enhanced by attention mechanisms that highlight salient vascular regions and reduce background interference. Skip connections further preserve the integrity of complex coronary morphology. Experimental results on low-quality XCA images demonstrate that Angio-Fusion Net outperforms existing state-of-the-art methods. The Gamma–CLAHE + Attention-VGG16-U-Net version achieved a Dice score of 96.15% ± 0.47%, Jaccard index of 92.61% ± 0.65%, and Accuracy of 98.02%, while the Top-Hat–CLAHE version yielded a Dice score of 93.21% ± 0.45%, Jaccard index of 87.39% ± 0.67%, and Accuracy of 93.27%. These results show that Angio-Fusion Net works well in different imaging conditions and can help cardiologists detect blocked arteries early, thereby improving treatment decisions.
Sunilkumar et al. (Wed,) studied this question.