X-ray angiography is the most used tool by clinician to diagnose the majority of cardiovascular disease and deformations in coronary arteries like stenosis. In most applications involving angiograms interpretation, accurate segmentation is essential to extract the coronary artery tree and thus speed up the medical intervention.

In this paper, we propose a multiscale algorithm based on Graph cuts for vessel extraction. The proposed method introduces the direction information into an adapted energy functional combining the vesselness measure, the geodesic path and the edgeness measure. The direction information allows to guide the segmentation along arteries structures and promote the extraction of relevant vessels. In the multiscale analysis, we study two scales adaptation (local and global). In the local approach, the image is divided into regions and scales are selected within a range including the smallest and largest vessel diameters in each region, while the global approach computes these diameters considering the whole image. Experiments are conducted on three datasets DS1, DS2 and DS3, having different characteristics and the proposed method is compared with four other methods namely fuzzy c-means clustering (FC), hysteresis thresholding (HT), region growing (RG) and accurate quantitative coronary artery segmentation (AQCA).

Comparing the two proposed scale adaptation, results show that they give similar precision values on DS1 and DS2 and the local adaptation improve the precision on DS3. Standard quantitative measures were used for algorithms evaluation including Dice Similarity measure (DSM), sensitivity and precision. The proposed method outperforms the four considered methods in terms of DSM and sensitivity. The precision values of the proposed method are slightly lower than the AQCA but it remains higher than the three other methods.

The proposed method in this paper allows to automatically segment coronary arteries in angiography images. A multiscale approach is adopted to introduce the direction information in a graph cuts based method in order to guide this method to better detect curvilinear structures. Quantitative evaluation of the method shows promising segmentation results compared to some segmentation methods from the state-of-the-art.