Table 1.
List of inclusion and exclusion criteria.
Fig 1.
Local inter-expert variability and contour uncertainty.
Example of 3 expert external elastic membrane (EEM) and lumen contours and corresponding lumen local inter-expert variability (local IEV) on a given scanline, together with the algorithmic contour with uncertain contour points displayed in red. On the left panel, the 3 experts’ contours (in yellow) and the corresponding algorithmic contours (in green) are displayed for both the EEM and lumen. The points of the algorithmic contour that have high uncertainty are displayed in red. A scanline is shown in cyan, together with the 3 points (circles in yellow) of experts’ contours intersecting the scanline (2 of these points almost coincide in this example). The right panel is a zoomed portion of the left image that lies within the cyan rectangle. The LIEV is the maximal pairwise distance between the 3 points. This measure can be computed for each scanline (corresponding to an angular position within the two dimensional IVUS image).
Fig 2.
Example of IVUS images with gold standard and segmentation contours.
Each panel corresponds to a different patient. Panels a and b display IVUS images acquired with a 40 MHz probe; the image in panel a was acquired pre-intervention, whereas panel b shows a post-intervention acquisition. Panels c and d display IVUS images acquired with a 60 MHz probe; the image in panel c was acquired pre-intervention whereas panel d shows a post-intervention acquisition. The external elastic membrane (EEM) and lumen contours computed with the fast-marching method (FMM) algorithm are superimposed on the images (green full curves with uncertainties in red) together with the EEM and lumen contours manually drawn by an expert (yellow full curves). The plaque burden computed from the segmentation contours and from the expert’s curves is displayed in each panel. The plaque in panel c contains a lipid pool visible from 1 to 4 o’clock. Both post-intervention images (panels b and d) were acquired in a stented vessel segment.
Table 2.
Comparison of local inter-expert variability with pointwise contour uncertainty at 40 MHz.
Table 3.
Comparison of local inter-expert variability with pointwise contour uncertainty at 60 MHz.
Table 4.
Comparison of inter-expert variability between 40 and 60 MHz data.
Fig 3.
Inter-expert variability at 40 and 60 MHz.
Linear regressions between inter-expert variability in external elastic membrane (EEM), lumen, and plaque cross-sectional areas (CSA), and plaque burden, on one hand, and average EEM CSA, on the other hand, at 40 and 60 MHz. Inter-expert variability in a measurement is defined as its maximal pairwise difference. The EEM CSA is defined as the average over the 3 experts’ measurement in EEM CSA obtained from their contours.
Table 5.
Comparison of algorithmic to expert discrepancy with inter-expert variability at 40 MHz.
Table 6.
Comparison of algorithmic to expert discrepancy with inter-expert variability at 60 MHz.
Table 7.
Comparison of algorithmic to expert discrepancy with inter-expert variability at 40 MHz in cross-sectional areas and plaque burden.
Table 8.
Comparison of algorithmic to expert discrepancy with inter-expert variability at 60 MHz in cross-sectional areas and plaque burden.
Table 9.
Comparison of algorithmic to expert discrepancy with inter-expert variability at 40 MHz in Balocco et al. [15] measurements (EEM).
Table 10.
Comparison of algorithmic to expert discrepancy with inter-expert variability at 40 MHz in Balocco et al. [15] measurements (lumen).
Table 11.
Comparison of algorithmic to expert discrepancy with inter-expert variability at 60 MHz in Balocco et al. [15] measurements (EEM).
Table 12.
Comparison of algorithmic to expert discrepancy with inter-expert variability at 60 MHz in Balocco et al. [15] measurements (lumen).
Fig 4.
Gamma correction function implemented as a lookup table (LUT).
Illustration of the LUT function of Eq (A-4) used in Eqs (A-3) and (A-5).
Fig 5.
Illustration of the 23 × 23 pixels Gabor filter of Eq (A-6).
Fig 6.
Example of an IVUS image and corresponding gradient images.
Panel a) displays an IVUS image. Panels b) and d) display gradient images obtained by processing the image with a gray level gradient filter (panel b)) and with a probability density function (PDF) textural gradient filter (panel d)), respectively. The former gradient filter reveals contour-based information, whereas the latter filter detects region-based information. Both types of information are combined into the speed function of the fast-marching method (FMM) algorithm used in the external elastic membrane (EEM) contour segmentation (c.f. Eq (A-7)). The resulting EEM contour is superimposed on both gradient images (green full curve). Panels c) and e) display gradient images obtained by processing the image with a Gabor filter (panel c)) and with a PDF label gradient filter (panel e)), respectively. As for the EEM, the former gradient filter reveals contour-based information, whereas the latter detects region-based information. Both types of information are combined into the speed function of the FMM algorithm used in the lumen contour segmentation (c.f. Eq (A-8)).
Fig 7.
Semi-automatic initialization.
Left: example of initial semi-automatically traced contours (red curves) on a longitudinal cut of a pullback. Right: example of a cross-section from the same pullback; the yellow curves represent the contours obtained by a B-spline interpolation of the points obtained from the manually traced boundaries on the L-views. The initial boundaries used by the algorithmic segmentation are obtained from these yellow contours.