4. 1 Segmentation of lumen region and assessment of lumen diameter
Lumen is the area between the lumen-intima (LI) boundary of the near-wall (LI-near) and far-wall (LI-far). The distance between LI-near and LI-far is defined as lumen diameter (LD) whereas the gap between intima and media is labelled as IMT. Figure 3. shows a longitudinal segment of the common carotid artery (CCA) and the delineation of LI and MA boundaries. Characterization of the lumen is primary for the assessment of LD and IMT. However, lumen recognition is challenging because of variability in dataset, plaque composition-and-morphology, arterial structure, presence of stenosis, jugular vein and importantly the imaging standards [96,97]. Algorithms applied several theories to overcome these challenges. The most
prominent assumptions are that (a) the brightest boundary corresponds to media-adventitia (MA) edge, and (b) the arterial blood flow is laminar [88]. These theories aid in identifying the edges and ROI. In previous studies of Sifakis et al. [98], statistical estimates were used for automatic recognition of lumen, both in multi-frame and single frame. The success rate was 99% and the algorithm was effective even in the presence of plaques and moderate amounts of “mimicking arteries” [98]. Surprisingly, the algorithm failed in images with poor far-wall representation and sharp arterial curvature [98]. Over the years, an enormous amount of work was done on automated lumen characterization. Araki et al. [88] used automated ROI detection, spectral analysis and K-means classifier for lumen characterization. The algorithm performed well (Dice similarity ~1) but was susceptible to noise, motion artefacts and presence of plaque. Molinari et al. [99] suggested a novel approach for seed point selection followed by curve fitting and classification. The method was successful in 92% cases except in the presence of backscattering and plaque [99]. Most algorithms in the past preferred straight vessels in the frame for better segmentation. Kumar et al. [100] applied spatial transformation for pre-processing (straightening curved vessels) and scale-space for boundary segmentation. Dice similarity and Jaccard index were 94.2, 89.1and 93.9, 88.6 respectively on two different sets of GT. Carvalho et al. [95] used contrast-enhanced ultrasound along with the B-mode ultrasound to overcome motion artefacts and noise. Further, joint-histogram and graph-based methods were employed for segmentation. The authors informed segmentation error as the average root mean square error (RMSE) of 112±73 µm compared with the expert recording on two datasets [95]. Rocha et al. [101] used a linear-Bayes classifier along with dynamic programming (DP) for segmentation on 199 images. The authors reported a success rate of 99.5% and robustness to shadowing, scanner variability and plaque irregularity [101]. Further, reviews on a comparison between different segmentation approaches are presented elsewhere [102]. Table I. summarizes benchmarking studies on lumen segmentation and LD assessment.
Table I. Benchmarking studies on lumen segmentation and LD assessment.