Currently, coronary artery disease still the principal etiology of mortality worldwide. X-ray Coronary Angiography (XRCA) remains the gold standard to the diagnosis and treatment of this disease. However, the complex vessel structure, image noise, uneven illumination, poor contrast stenosis and many vessel overlaps. Recently, deep learning methods have been proposed in a wide range of medical applications by learning features directly from images in a fully supervised manner and have out performed traditional methods with better performance. In fact, the U-Net architecture is the baseline for most of the state-of-the-art of segmentation methods.

Comparing three CNN models of coronary artery segmentation based on artificial intelligence.

We presents the implementation process of CNN models: U-Net, Res-Net, and DenseNet networks are widely used in medical imaging and have achieved very high performance in segmentation, 20, 46, 3, 28, 49NThe first step is the pre-processing of coronary images. The second step is data augmentation using different methods in order to improve performance. The last step is the segmentation of blood vessels from enhanced images using the previous CNN architectures.

Twenty-five obtained images are selected for coronary artery segmentation in 2D X-ray angiograms acquired during routine cardiac catheter examinations using a single plane Siemens Artis Zee angiography system at the Cardiology Department (A) of The University Hospital Fattouma Bourguiba, Monastir, Tunisia. After data augmentation, we obtained 102 images. To have a comparative study, we test the three presented methods: U-Net, Res-Net and Dens-Net on our dataset. It provides the higher precision (0.98) yielding a higher Dice metric but not far to U-net and Res-net precision values (0.97 for both). Also, it provides the higher values of sensitivity and F-score metrics. However, U-Net provides the lowest values of deferent metrics: 0.97 for accuracy, 0.66 for F-score and 0.57 for sensitivity, but accuracy was best for 0.79.

For better refining the coronary artery segmentation task and assist clinicians when acute coronary syndrome is diagnosed, we have to work on the impact of the patch size on the detection of different types of coronary lesions and to predict in the future the evolution of the coronary atherosclerotic plaque.