Introduction: Intraoperative quality assessment of cardiac surgery is an essential tool to ensure the quality of the performed surgery and prevent potential revision.
Coronary artery bypass graft (CABG) surgery has been shown to be related to a series of operative technical complications but remains one of the only major vascular surgeries that are not routinely intraoperatively assessed. Additionally, methods widely used for evaluating the anastomoses quality of the CABG-surgery are impractical to use or provide false insight into the quality of the anastomosis, to which epicardial ultrasonography (EUS) has been shown as a promising alternative as it can visualize structural information of the anastomosis. However, evaluation of anastomoses remains subjective to the physician, leading to cases of lacking or unnecessary revision to which the purpose of this project was to develop an objective quality assessment method.
Methods: 367 longitudinal EUS-frames recordings of anastomoses were available for this project to which 96 frames were used to develop the proposed methods. The methods comprised of a vessel detection algorithm, vessel lumen segmentation, and patency estimation. Two different segmentations were tested and compared: local-phase based snake and Chan-Vese to investigate which method produced the best anastomosis estimate.
Results: The methods were tested on the remaining 271 EUS-frames to which it was able to detect full or partial vessel structures from 89.67 \% of the test frames, achieving an average Dice coefficient from sufficient detections of 0.8134 and 0.8187 for the local-phase based snake and Chan-Vese, respectively. This lead to a patency estimation of a maximum of 135 EUS-frames.
Validation, when applying the patency estimation to corresponding manual annotations, resulted in the highest agreement of 88.15 \% when compared with the Chan-Vese segmentation.
Conclusion: This project showed that the proposed method was able to detect and estimate the patency of anastomosis vessels from EUS-recordings, however, challenges persist when estimating the edges defined by manual expert annotations due to artifacts in the EUS-frames.