3-D Stent Detection in Intravascular OCT Using a Bayesian Network and Graph Search

Abstract

Worldwide, many hundreds of thousands of stents are implanted each year to revascularize occlusions in coronary arteries. Intravascular optical coherence tomography is an important emerging imaging technique, which has the resolution and contrast necessary to quantitatively analyze stent deployment and tissue coverage following stent implantation. Automation is needed, as current, it takes up to 16 h to manually analyze hundreds of images and thousands of stent struts from a single pullback. For automated strut detection, we used image formation physics and machine learning via a Bayesian network, and 3-D knowledge of stent structure via graph search. Graph search was done on en face projections using minimum spanning tree algorithms. Depths of all struts in a pullback were simultaneously determined using graph cut. To assess the method, we employed the largest validation data set used so far, involving more than 8000 clinical images from 103 pullbacks from 72 patients. Automated strut detection achieved a 0.91±0.04 recall, and 0.84±0.08 precision. Performance was robust in images of varying quality. This method can improve the workflow for analysis of stent clinical trial data, and can potentially be used in the clinic to facilitate real-time stent analysis and visualization, aiding stent implantation.

Fig. 1

Top: Some popular stent designs showing different, but regularized structures. Bottom: Two adjacent frames from an OCT pullback showing stent struts with bright reflections followed by dark shadows. In frame i, the oval arrows at 4–5 o’clock point to ambiguous struts. It is clear that the ambiguous structures in frame i are leading edges of clearly identified struts in frame i+1 marked with arrows, demonstrating the value of using 3-D information to identify stent struts.

Fig. 2

Overview of the automated stent detection method.

Fig. 3

The Bayesian network for inference of strut presence. (a) Original OCT image in polar coordinates. The green contour indicates the detected lumen boundary of the vessel. (b) By calculating the mean intensity of the A-line within a fixed depth from the lumen boundary, the 2-D image is projected into a 1-D curve. Struts generate local minima and have large shadow contrasts, SC. (c) The probability of strut presence of each A-line generated by the Bayesian network in (d). (d) The Bayesian network representation based on principles of OCT image formation. Known variables are marked in blue.

Fig. 4

Top: Individual frames in polar coordinates from the stented portion of a pullback. Bottom: Synthesized en face projection image showing the structure of the stent.

Fig. 5

Transforming the all-strut depth determination into a graph search problem. Representative image frames from an OCT pullback showing a stent implanted in a coronary artery with new tissue growth, displayed in cartesian (left) and polar (right) coordinates. Left: To maintain the tubular structure, the distances from adjacent struts (blue dots) to the lumen centroid are kept within a certain constraint. Right: With the deformation constraint, the optimal depths (blue line) for all the struts form the globally optimal surface in the graph constructed using only the pixels in the strut lines.

Fig. 6

Generation of stent contours (black) based on interpolation for quantification of clinically relevant metrics. Lumen contours are shown in green. (a) Stents with neointima. (b) Malapposition.

Fig. 7

Human analyzed data used as the gold standard for validation. Manually marked struts are indicated by blue dots in the image. Only stent struts with bright bloom were analyzed by human analysts in order to minimize inter-observer variability in strut-level analysis. For example, the yellow circled struts were not analyzed by human analysts. Inset: Human analysts marked the front edge, instead of the center of bloom for analysis.

Fig. 8

The effect of the size of the training data set on the performance of Bayesian classification. The testing data are a subset of randomly selected 10 pullbacks from the entire validation data set.

Fig. 9

En face stent segmentation results for two representative types of stents. Top: Xience V stent. Bottom: Nobori stent. Dice’s coefficient values for the two stents are 0.86±0.02 and 0.92±0.06, respectively.

Fig. 10

Performance of the stent strut detection in 8332 clinical images from 103 pullbacks. As the gold standard is “biased” that not every strut was analyzed manually (Section III B), the actual precision of the method is expected to be significantly underestimated. (a) and (b): Recall and precision for struts with different thickness of neointima coverage, respectively. These metrics are derived on a frame-by-frame basis. Tissue coverage is determined by the average thickness of all struts in a frame. The numbers under the bars indicate the number of images in each category. (c) Overall performance using all the images. Here the metrics are derived on a pullback-by-pullback basis.

Fig. 11

Examples of automated stent strut detection in cases with different amounts of neointima coverage, images of varying quality and in the presence of various artifacts. Automatically segmented luminal boundaries were marked in green. Stent struts were marked in blue.

Fig. 12

Examples of cases where the algorithm failed. (a) False positives could be generated when artifacts cast shadows (yellow arrows). (b) Struts with very thick coverage and almost no shadows may not be detected by the automated algorithm (yellow arrows). (c): Near a stent strut branching (yellow arrow), the algorithm detected only one strut whereas there were actually two. A false positive was also detected by the algorithm (red arrow).

Fig. 13

Left: Correlation of the stent area measurements based on stent strut detection by the automated algorithm and by human analysts. Right: Bland-Altman plot. Automatically derived stent areas correlate well with areas determined by analysts but show a bias because we used the front edge of the bloom in manual analysis without correction.

Fig. 14

3-D reconstructions of an implanted stent from an intracoronary OCT pullback. The vessel was volume-rendered in orange, and the segmented stent was rendered in silver white. The voxels inside the lumen boundary were not rendered. (a) Stent rendering using only manually-marked struts in 2-D cross-sectional frames. All possible struts were segmented by an analyst, and confirmed by a second observer. This “perfect” manual segmentation is too sparse to make a complete stent mesh in 3D. For better visualization, only half the vessel is shown. (b) Using en face projection processing, and mapping back to 2-D frames, the 3-D stent is very well visualized. (c) Fly-though view shows malapposed struts (green arrows). The dark band along the vessel is the region blocked by the guide wire.