Int J CARS DOI 10.1007/s11548-014-1144-3 ORIGINAL ARTICLE Automatic detection of coronary artery anastomoses in epicardial ultrasound images Alex Skovsbo Jørgensen · Samuel Emil Schmidt · Niels-Henrik Staalsen · Lasse Riis Østergaard Received: 30 September 2014 / Accepted: 18 December 2014 © CARS 2015 Abstract Purpose Epicardial ultrasound (EUS) can be used to assess the quality of coronary artery bypass graft surgery (CABG) anastomoses by determining stenotic rates. Currently, no objective quantitative methods are available for the analy- sis of EUS images. Therefore, surgeons have to be trained in interpreting EUS images, which may limit the use of EUS in clinical practice. Automatic detection of vessel structures can enable the objective and quantitative quality assessment of anastomoses without user interaction to facilitate the revi- sion of anastomoses during the primary surgery. Methods An automatic vessel detection algorithm extracted and detected image regions that uniquely intersected with the vessel lumen of anastomotic structures. First, an initial pixel- based segmentation was performed from regional minimums using a watershed segmentation and an adaptive thresholding approach. A region-based merging step was then performed to merge oversegmented vessel structures using a Bayesian classification of different region combinations constructed from the pixel-based segmentations. Finally, a vessel classi- fication step was performed on the extracted regions after the region-based merging to determine the probabilities that the regions contained vessel structures. A. S. Jørgensen (B ) · S. E. Schmidt · L. R. Østergaard Department of Health Science and Technology, Aalborg University, Fredrik Bajersvej 7D2, 9220 Aalborg, Denmark e-mail: asj@hst.aau.dk N.-H. Staalsen Department of Cardiothoracic Surgery, Center for Cardiovascular research, Aalborg University Hospital, Hobrovej 18-22, 9100 Aalborg, Denmark N.-H. Staalsen Institute of Clinical Research, Skejby Sygehus, Aarhus University Hospital, Norrebrogade 44, 8000 Aarhus, Denmark Results The performance of the vessel classifier was tested using m-fold cross-validation of 320 EUS images containing anastomotic vessel structures from 16 anastomoses made on healthy porcine vessels. An area under the curve of 0.966 (95 % CI 0.951–0.984) and 0.989 (95 % CI 0.985–0.993, p < 0.001) of a precision–recall and receiver operator character- istic curve, respectively, was obtained when detecting vessel regions extracted from the EUS images. Conclusions The vessel detection algorithm can detect ves- sel regions in EUS images at a high accuracy. It can be used to enable the automatic analysis of EUS images for the qual- ity assessment of CABG anastomoses. Keywords Ultrasound · Vessel · Pattern recognition · Object detection · Coronary artery bypass graft surgery Introduction Coronary artery bypass graft surgery (CABG) is used to treat severe cases of coronary heart disease. However, up to 9% of CABG anastomoses contain stenoses above 50 % post- operation [1]. This can lead to early graft failure causing perioperative myocardial infarction or fatal heart failure [2]. Some stenoses can occur due to technical errors made by the surgeon. This can be detected using intraoperative anastomo- sis quality assessment tools. The detection of technical errors can enable the revision of the anastomosis during the primary surgery [2]. The current gold standard for intraoperative qual- ity assessment is coronary angiography, which visualizes the morphology of anastomoses in X-ray images using a contrast agent. However, because coronary angiography is not avail- able in many operating rooms, ultrasound-based methods, such as transit time flow measurement and epicardial ultra- 123