Segmentation of Media and Lumen in Intravascular Ultrasound Image Using Guided Multiscale Normalized Cut

Segmentation of vessel membranes (media and lumen) in intravascular ultrasound (IVUS) frames is helpful for the diagnosis of the atherosclerosis. However, the manual delineation of media and lumen is laborious. Furthermore, many existing fully automatic segmentation methods for media is underperforming, since they cannot reduce the interference of artifact shadow and speckle noise effectively. To overcome these problems, we design a special mask based on the geometric property of media and an assumption about the intensity distribution. For media segmentation, the guided multiscale normalized cut (GMNC) extracts the region of interest (ROI) by this mask so that the interference can be reduced and the following processing can focus on the ROI. For lumen segmentation, the media region becomes the new ROI naturally. Then, a targeted binarization processing can extract the rough lumen region in this ROI. Next, the multiscale normalized cut is introduced to detect the rough edge and then the final result is obtained by smoothing. To investigate the effectiveness of the GMNC, it was evaluated on a public dataset including 435 IVUS images and the Hausdorff distance (HD), Jaccard measure (JM) and percentage of area difference (PAD) were adopted as evaluation metrics. The result showed that the average HD, PAD for media segmentation decreased by 32% and 26%, compared with the best one of three existing methods (the sequential algorithm based, the region detection based and the statistical model based). The mean of JM also increased by 9%. Thus, it can be concluded that the GMNC has advantage in media segmentation. Additionally, the average HD, JM and PAD for lumen segmentation were 0.37 mm, 0.82 and 0.19 respectively, these evaluation parameters were still advantageous over those counterparts of two compared methods.