4. 2 Segmentation LI-far wall and IMT assessment.
Although IMT can be assessed at both walls, far-wall is more reliable than the near-wall. Further, far-wall IMT is associated with coronary heart disease and mean IMT, especially those measured 1 cm away from the bulb, is correlated with risk factors [103,104]. Previous research has largely overlooked computer-aided methods in IMT measurements [105]. Liguori et al. [106] used pattern recognition and edge detection techniques for LI-MA delineation. This semi-automated method reported high-correlation (0.97) with manual methods. Dynamic programming (DP) is a highly optimised search-based method useful for edge detection [95,101,107]. However, the method showed limited success in curved vessels. Zhou et al. [108] applied a modified DP for intima-media segmentation. The method was efficient and invariant to the rotation but refrained from IMT estimation. Important studies showed that integrated methods provide a better approximation. Faita et al. [109] applied gradient-based edge detection fused with a robust edge detector (“FOAM”) followed by a heuristic search. The algorithm performed well when validated with two expert readings on 150 scans. Further, the authors displayed real-time processing using video of the cardiac cycle and the Bland-Altman plots showed zero inter-observer variability [109].
Another important study by Rossi et al. [110] presented segmentation based on two methods: adventitial delineation using sustain attack filter and intimal delineation using multiscale anisotropic barycentre (MAB). The authors reported a variation of 1.3% for diameter and 3% for IMT in 36 recordings. However, accuracy of the technique varied considerably based on probe positioning, patient orientation, artefacts and image quality [110]. Later a more robust method, the Hough transform (HT) often used for the detection of lines and circles, was used for carotid artery segmentation [111]. Golemati et al. [112] applied HT and observed that accuracy and specificity were >0.96 for both transversal and longitudinal images of non-atherosclerotic subjects. However, the results depend on plaque, shadow and scanner variation [112]. Some scholars have researched snake-algorithm. Loizou et al. [113], Petroudi et al. [90] and Santhiyakumari et al. [114] used snake-algorithm (active contour) for intima-media segmentation. Despite good results, the method is limited by lengthy processing time and initialization. Destrempes et al. [115] argued that distribution of echogenicity in a vertical strip of pixels is a mixture of three Nakagami distributions related to intima, media and adventitia. The authors then applied maximum-a-posteriori (MAP) estimation and expectation-maximization for segmentation [115].
Molinari et al. and the team proposed several techniques for segmentation and IMT measurement [116,117]. The team performed LI segmentation based on local statistics and signal analysis and reported improvement in both LI detection (+32.3%±6.7%) and IMT measurement (+43.5%±2.4%) [118]. In another research, the team applied, feature extraction, curve fitting and classification, more suitable for MA segmentation [117]. As stated earlier, the integrated approach improved IMT estimation (+3.6±1.4%) and segmentation efficiency [117]. Both IMT and LD show variations in a cardiac cycle. Tracking these variations in IMT is a serious challenge. Ilea et al. [119] adopted segmentation on a video of the cardiac cycle using a Canny edge detector and adaptive normalisation [119]. Table II review various findings on IMT measurement.
Table II. Benchmarking studies on IMT measurement