Int J Cardiovasc Imaging (2014) 30:839–848 DOI 10.1007/s10554-014-0406-z

ORIGINAL PAPER

Automated detection and quantification of clusters of malapposed and uncovered intracoronary stent struts assessed with optical coherence tomography Tom Adriaenssens • Giovanni J. Ughi • Christophe Dubois • Kevin Onsea • Dries De Cock • Johan Bennett • Stefanus Wiyono • Maarten Vanhaverbeke Peter Sinnaeve • Ann Belmans • Jan D’hooge • Walter Desmet



Received: 2 September 2013 / Accepted: 18 March 2014 / Published online: 26 March 2014 Ó Springer Science+Business Media Dordrecht 2014

Abstract To date, accurate quantification and localization of malapposed and uncovered struts needs manual and time consuming analysis of large datasets. To develop an algorithm for automated detection and quantification of clusters of malapposed and uncovered struts in optical coherence tomography (OCT) pullbacks, including comprehensive information about their three-dimensional spatial distribution. 64 lesions in 64 patients treated with drugeluting stent underwent assessment with OCT immediately after implantation and at 9-month follow-up (55 patients). An automated algorithm was used to detect and quantify stent strut malapposition at baseline and coverage at follow-up on an individual strut level. We subsequently applied an algorithm for the automated clustering of malapposed and uncovered struts and for the quantification of clusters’ properties. In the 64 baseline examinations, a total of 24,013 struts were analyzed, of which 1,519 (6 %) were malapposed. Most malapposed struts (78 %) occurred in clusters and more than half of patients had malapposition clusters. The mean number of struts per cluster was

Tom Adriaenssens and Giovanni J. Ughi have contributed equally to this work. T. Adriaenssens (&)  G. J. Ughi  C. Dubois  K. Onsea  D. De Cock  J. Bennett  S. Wiyono  M. Vanhaverbeke  P. Sinnaeve  J. D’hooge  W. Desmet Department of Cardiovascular Medicine, University Hospitals Leuven, Louvain, Belgium e-mail: [email protected] T. Adriaenssens  G. J. Ughi  C. Dubois  P. Sinnaeve  A. Belmans  J. D’hooge  W. Desmet Department of Cardiovascular Sciences, KU Leuven, Louvain, Belgium

19.7 ± 11.8 with a mean malapposition distance of 213 ± 66 lm. In the 55 follow-up pullbacks, a total of 20,484 struts were analyzed, of which 1,320 (6 %) were uncovered. Again, most uncovered struts (85 %) occurred in clusters. The mean number of struts per cluster was 21.1 ± 14.7. We developed an automated algorithm for studying clustering of malapposed or uncovered struts. This algorithm might facilitate future investigations of the prognostic impact of clusters of malapposed or uncovered struts. Keywords Optical coherence tomography  Coronary stent  Drug-eluting stents  Strut apposition  Cluster of malapposition  Cluster of uncovered struts

Introduction The use of intracoronary optical coherence tomography (OCT), with a resolution of 10–15 lm (ten times better than intravascular ultrasound—IVUS), has provided new insights into the incidence and clinical implications of stent strut malapposition, both in its acute (detected immediately after stent implantation) and late (persistent or acquired) forms as well as to the incidence and distribution of uncovered struts in a stented segment. However, the introduction of OCT in clinical practice has confronted researchers and practitioners with methodological and organizational challenges, as it provides enormous amounts of data that are cumbersome and time-consuming to analyze, interpret and report. The observation in earlier IVUS studies that only more pronounced forms of malapposition correlate with adverse clinical outcome emphasizes the need for an integrated analysis method, preferably fast and automated with specific attention to location and the degree

123

840

of malapposition in a specific stented segment [1]. Due to the lack of an automated process, OCT studies so far often only reported the amount and percentage of malapposed and uncovered struts, lacking detailed information on their spatial distribution (i.e. whether they accumulate in groups or are more homogeneously distributed along the entire stented segment), even when this could prove to be valuable information for better understanding stent behavior characteristics. Our group has previously developed and validated a robust algorithm that allows an automated and comprehensive analysis of stent strut apposition and coverage in coronary arteries [2]. Building on this work, we set out to develop an algorithm to automatically classify malapposed/ uncovered stent struts as ‘isolated’ or ‘clustered’, i.e. belonging to a ‘cluster of malapposed/uncovered struts’. The secondary aim of this study was to describe clusters of malapposition in pullbacks acquired immediately after DES implantation and clusters of uncovered struts in pullbacks acquired at follow-up, in order to quantify the degree and severity of these events.

Methods Patient population Between June 2009 and November 2010, 64 patients were examined with OCT after stent implantation at our institution. The study population consisted of patients included in the STACCATO (Stent sTrut Apposition and Coverage in Coronary ArTeries: an Optical Coherence Tomography Study, ClinicalTrials.gov NCT01065519) randomized controlled trial, which is a comparative study between the Xience everolimus-eluting (Abbott Vascular, Santa Clara, USA) and the Biomatrix (Biosensors, Switzerland) biolimus A9-eluting stent in patients with ST segment elevation myocardial infarction (STEMI), non-ST segment elevation myocardial infarction (NSTEMI), unstable and stable angina, studying healing characteristics of these stents with OCT. The study was conducted in compliance with the University Hospitals Leuven Bioethics Committee guidelines and received its approval. Written informed consent was obtained for every patient. Patients were excluded if they had significant left main disease, congestive heart failure, cardiogenic shock, renal insufficiency with a baseline serum creatinine level of [2.0 mg/dl or a coexistent condition associated with a limited life expectancy (i.e., advanced cancer). In addition, patients with extremely tortuous vessels or heavy calcification and chronic total occlusion were excluded because of anticipated difficulty in advancing the OCT catheter.

123

Int J Cardiovasc Imaging (2014) 30:839–848

OCT image acquisition and analysis Optical coherence tomography images were acquired with the C7-XRTM FD-OCT system and DragonFlyTM imaging catheters (St Jude Medical, Minnesota, USA). This system is equipped with a tunable laser light source with sweep range of 1,250–1,370 nm. The catheter is delivered over a 0.014 in. guidewire through a 6 Fr or larger guiding catheter, after administration of intracoronary nitrates. For an effective clearing of blood from the imaging field, angiographic contrast medium is injected through the guiding catheter adopting an automated modality. Injection of 12–18 ml of contrast at a rate \4 ml/s is sufficient to achieve an imaging period of 2–3 s consistently in all of the major coronary branches. At a pullback rate of 20 mm/ s, an imaging period of 2 s is long enough to scan a 4 cm vessel segment. FD-OCT images were calibrated adjusting the Z-offset [3, 4]. Images from pullbacks acquired immediately after stent implantation were used for the assessment of acute apposition, while coverage of stent struts was assessed in pullbacks acquired during a 9-month control angiogram. The image data were digitally stored for offline analysis. All cross-sectional images were initially screened for quality assessment and excluded if images were not suitable for analysis according to standardized methods [4]. Quantitative strut level analysis and morphometric analysis were performed every three frames (i.e. 0.6 mm intervals) along the entire target segment. Automated algorithm for stent detection and malapposition The main principle of the proposed algorithm is the clustering of malapposed/uncovered struts according to their spatial three-dimensional (3D) location. As previously described, a fully automated algorithm for lumen contour and stent strut detection and segmentation was used for OCT data analysis [2, 5, 6]. In summary, after lumen borders are completely automatically traced and stent struts automatically located, measurements representing individual stent strut apposition (Fig. 1) and coverage (Fig. 2) are automatically obtained. Discrimination between apposed/malapposed struts is obtained applying the following threshold: T = real strut thickness ? polymer thickness (in case of drug-eluting stent) ? blooming artefact (estimated at 18 lm for current generation OCT systems) [4]. In case a strut is floating over a side-branch (Fig. 3), it is labelled accordingly and excluded from further analysis. If the measured strut/lumen distance is inferior to T, the relative strut is labelled as apposed; otherwise it is labelled as malapposed. Concerning stent strut coverage, the central point of every strut is

Int J Cardiovasc Imaging (2014) 30:839–848

841

Fig. 1 Examples of individual stent strut segmentation and quantification of malapposition distances. Red and green dots indicate malapposed and apposed struts, respectively. The inlets illustrate a magnification (yellow boxes) for apposed struts (left image) and

malapposed struts (right image). The apposition distance is indicated by white arrows. The white asterisk indicates the guide-wire shadowing artifact

Fig. 2 Examples of individual stent strut segmentation and quantification of coverage. Green dots indicate covered stent struts and red dots indicate uncovered ones. The inlets show a magnification (yellow

boxes) for covered struts (left image) and uncovered struts (right image). The white asterisk indicates the guide-wire shadowing artifact

identified and from that point, the distance perpendicular to the lumen contour is computed. To take into account the blooming artefact, a threshold of 20 lm was applied. All stents with strut to lumen border below this threshold were considered as malapposed. The accuracy of this automated software algorithm has been validated by comparison with manual assessment and by comparison with histopathology in a rabbit model [4, 7]. Automatic results are visually inspected and approved by a trained user, and in case of segmentation errors (e.g. caused by severe image artefacts), the user can manually adjust stent strut quantification results. Once an entire pullback is analysed, this method results in a 3D position of every strut and discrimination

between apposed and malapposed (or covered and uncovered) struts. Automated algorithm for cluster analysis Generally speaking, clustering is a well-known technique which aims to define groups of variables with homogeneous properties within the group they belong to and inhomogeneous properties between other groups [8]. In this particular case, we propose to cluster malapposed and uncovered struts into different groups according to their 3D position in space. To do so, a non-supervised clustering technique based on distance measurements (i.e. Euclidean

123

842

Int J Cardiovasc Imaging (2014) 30:839–848

the stent length was decisive of ‘proximal’ or ‘distal’ allocation. Location of the point of gravity at the middle (±15 %) of the stented segment pointed to a ‘middle’ allocation, with the remaining positions being ‘proximalmiddle’ and ‘distal middle’. With respect to the determination of the number of quadrants a cluster is spreading; 0°–90° was categorized as one quadrant, 90°–180° as two, 180°–270° as three and 270°–360° as four quadrants. Three-dimensional visualization of strut clustering

Fig. 3 Examples of stent struts (red circles) overlying a side-branch. The white asterisk indicates the guide-wire shadowing artifact

Because OCT generates a high number of cross-sectional images, an accurate 3D visualization of the vessel can be obtained. To highlight and visually report final results of the clustering procedure, a three-dimensional map of malapposed (uncovered) struts is reported for each dataset analyzed. Figures 4 and 5 report representative examples of clustering of malapposed (and uncovered) struts, respectively. Implementation details

distance) is applied. Clusters are defined through a hierarchical agglomerative clustering algorithm based on nearest neighbor’s technique (also known as ‘‘single linkage’’). The proposed algorithm automatically creates groups of malapposed (uncovered) struts as follows: starting from the two most similar elements (i.e. closer struts, in our case), malapposed (uncovered) struts are subsequently merged together in clusters becoming larger step-by-step (agglomerative method). The procedure ends when the distance between ‘‘two groups’’ is larger than a predefined threshold, or alternatively, as done in the current analysis, it can be stopped manually by selecting the final number of different clusters based on the 3D visualization. We decided to adopt this strategy because defining a threshold based on minimum distance between two clusters may be difficult as stated in the limitation section. The current proceeding allows for a very intuitive and robust clustering of struts, taking into account their exact spatial distribution with respect to the longitudinal axis (i.e. pullback acquisition direction) and angular distribution. The minimal number of struts needed to have a cluster was arbitrarily set at five in the current study. Finally, for every cluster, properties like long-axis position (i.e. distal, middle or proximal), cluster length (according to the long-axis direction), angle distribution, number of struts belonging to the group and maximum and mean strut malapposition distance values (obviously not applicable for clusters of uncovered struts) were quantified. Based on the point of gravity of the cluster, the location of its middle point in the 15 % proximal or distal segment of

123

Both segmentation and cluster algorithms were implemented through the use Matlab R2011a and related toolboxes (MathWorkÒ, Natick, MA, USA): image processingTM and statisticsTM toolboxes. Run time for the clustering algorithm was quantified to be in the order of 10-2 s while the entire clustering procedures for an entire datasets (including 3D visualization) was \1 s. Computation time was quantified on a standard office pc (i.e. Dual-core 1.86 GHz Intel Xeon E5502 processor, 3, 5 Gb of RAM running Windows XP sp3 32-bit version) without any specific optimization. Statistical analysis To characterize malapposed clusters, the number of struts in the cluster, the mean and maximum malapposition distance within each cluster, the length of the cluster and the number of quadrants across which the cluster was positioned, were calculated. To characterize clusters of uncovered struts, the number of struts in the cluster, the length of the cluster and the number of quadrants across which the cluster was positioned, were calculated. For patients with multiple clusters, the average value across the clusters was calculated for the above mentioned parameters and used in the data summaries and statistical analyses. Summary statistics are given for the whole patient group and for patients presenting with and without clusters of malapposed/uncovered struts. For continuous measurements, mean and standard deviation or median and

Int J Cardiovasc Imaging (2014) 30:839–848

843

Fig. 4 Schematic representation of the number of isolated and clustered malapposed and uncovered struts and their spatial distribution along the long axis of the stent

Fig. 5 Two examples of stent strut clustering for malapposition. Malapposed stent struts are depicted as colored circles in the 3D space and the blue line represents the stent length. a An example of two clusters of malapposed struts at both stent edges. Also an isolated malapposed strut (blue circle) is present. b Presents an example of a single cluster of malapposed struts located on the proximal stent edge. From the image it is possible to immediately understand about cluster position, angular distribution and cluster dimension. 2D cross-sectional images (B1, B2, and B3) give examples of IVOCT images generating the cluster. Red arrows indicate malapposed struts. The white asterisk indicate the guide-wire shadowing artifact

interquartile range, as appropriate, are presented. For categorical variables, the observed frequencies and percentages are reported. Comparisons between groups were made using an independent t test or Wilcoxon rank-sum test, as

appropriate, for continuous variables. For categorical variables, a Chi squared test was used. All tests are 2-sided and assessed at a significance level of 5 %. Due to the exploratory nature of the study, no

123

844

Int J Cardiovasc Imaging (2014) 30:839–848

adjustments were made to the significance level to account for multiple testing.

Results Patient and lesion characteristics are described in Tables 1 and 2. There is an equal distribution between patients with stable or unstable angina, NSTEMI and STEMI. As severe calcification was a contraindication for participation in the study, most patients presented without or with mild calcification. Total stent length was 23.4 ± 7.7 mm. Most patients were treated with one stent, only a minority (17 %) was treated with two overlapping stents. Postdilation with non-compliant balloons was performed in over a third of the lesions. Strut malapposition clustering Table 3 describes the results of the malapposition and cluster data. Representative examples of clusters of

Table 2 Angiographic and procedural characteristics N = 64 Lesion length (mm) (mean ± SD)

14.1 ± 6.2

Max. stent diameter per lesion (mm) (mean ± SD)

3.1 ± 0.4

Total stent length per lesion (mm) (mean ± SD)

23.4 ± 7.7

DES implantation

64 (100 %)

Direct stenting

23 (36 %)

Number of stents per lesion 1 2

53 (83 %) 11 (17 %)

Postdilation performed

23 (36 %)

Final OCT performed

64 (100 %)

Preprocedural RVD (mm) (mean ± SD) Preprocedural MLD (mm) (mean ± SD)

3.1 ± 0.5 0.6 ± 0.5

Postprocedural MLD (mm) (mean ± SD) In-stent

2.8 ± 0.5

In-segment

2.4 ± 0.6

RVD reference vessel diameter, MLD minimal lumen diameter, DES drug-eluting stent, OCT optical coherence tomography

Table 3 Malapposition cluster data for the total population Table 1 Patient and lesion characteristics

N (%) N = 64

Total number of struts analyzed

24,013

Age (years) (mean ± SD)

60.7 ± 14.9

Total number of malapposed struts

1,519 (6 %)

Diabetes mellitus

8 (12.5 %)

Number of isolated malapposed struts

332 (22 %)

Arterial hypertension

40 (62.5 %)

Number of malapposed struts within clusters

1,187 (78 %)

Hypercholesterolemia

53 (82.8 %)

Total number of clusters

61

Clinical presentation

Number of patients with clusters

37

Stable/unstable angina

22 (34 %)

Number of struts per cluster (mean ± SD)

19.7 ± 11.8

NSTEMI

21 (33 %)

Malapposition distance per cluster (lm) (mean ± SD)

213 ± 66

Maximum malapposition distance per cluster (lm) (mean ± SD)

321 ± 186

Length of a cluster (mm) (mean ± SD)

6.7 ± 5.3

STEMI Target lesion coronary artery

21 (33 %)

LAD

28 (44 %)

LCx

15 (23 %)

RCA De novo lesion

21 (33 %) 64 (100 %)

Preprocedural TIMI flow 0

14 (22 %)

1

2 (3 %)

2

7 (11 %)

3

41 (64 %)

Severity of calcification None

48 (75 %)

Mild

13 (20 %)

Moderate

3 (5 %)

Severe

0 (0 %)

NSTEMI non ST segment elevation myocardial infarction, STEMI ST segment elevation myocardial infarction, LAD left anterior descending, LCx left circumflex, RCA right coronary artery, TIMI thrombolysis in myocardial infarction

123

Length of a cluster (mm) [median (IQR)]

4.8 (6.4)

Number of quadrants per cluster (mean ± SD)

2.4 ± 0.9

Number of quadrants per cluster [median (IQR)]

2 (2)

Position of cluster Proximal

27/61 (44 %)

Proximal-middle

5/61 (8 %)

Middle

14/61 (23 %)

Distal-middle

3/61 (5 %)

Distal

11/61 (18 %)

All stent

1/61 (2 %)

malapposed struts are shown in Fig. 5. A total of 24,013 struts were analyzed, of which 135 struts were labeled as ‘struts overlying a side-branch’ and excluded from further analysis. A total of 1,519 struts were assessed to be

Int J Cardiovasc Imaging (2014) 30:839–848

845

malapposed. The number of isolated malapposed struts was 332 (22 %) and the number of malapposed struts within clusters was 1,187 (78 %). In 37 of 64 patients, at least one cluster could be seen. A total of 61 clusters of malapposed struts were counted. The mean number of struts per cluster was 19.7 ± 11.8 and the mean malapposition distance per cluster was 213 ± 66 lm. Clusters had a mean length of 6.7 ± 5.3 mm and spread over 2.4 ± 0.9 quadrants. Malapposition clusters often tended to be localized in the proximal part of the stent (44 %), and less frequently in the proximal-middle (8 %), middle (23 %), distal-middle (5 %), distal (18 %) and over the whole treated segment (2 %) part of the stent. A graphical illustration of this distribution is represented in Fig. 4. In Table 4, the frequencies of malapposed struts are shown from the patient’s perspective, categorized by the presence or absence of clusters. Significantly more struts (mean of 412 ± 124 vs. 325 ± 110 in the ‘cluster’ vs. ‘no cluster’ group, p = 0.006) and a higher percentage of malapposed struts (9.5 ± 7.5 % in the ‘cluster’ vs. 1.8 ± 1.5 % in the ‘no cluster’ group, p \ 0.001) were seen. In parallel, there was a longer total stent length (26.5 ± 7.8 vs. 19.2 ± 5.4 mm) and a higher incidence of overlapping stents (27 vs. 4 %) in patients with versus without clusters of malapposed struts. Taken together, clusters of malapposed struts occur more frequently in longer stents and with overlapping stents, independently of post-dilation.

Table 4 Distribution of malapposed struts by clustering No clusters (n = 27)

Clusters (n = 37)

All patients (n = 64)

p value

325 ± 110

412 ± 124

375 ± 125

0.006

1.8 ± 1.5

9.5 ± 6.3

6.2 ± 6.2

\0.001

[5 % malapposed struts

2 (7 %)

27 (73 %)

29 (45 %)

\0.001

[10 % malapposed struts

0 (0 %)

13 (35 %)

13 (20 %)

\0.001

Total stent length (mm)

19.2 ± 5.4

26.5 ± 7.8

23.4 ± 7.7

\0.001

Postdilation performed (%)

9 (33 %)

14 (38 %)

23 (36 %)

0.711

Overlapping stents

1 (4 %)

10 (27 %)

11 (17 %)

0.015

Total number of struts (mean) % of malapposed struts (mean)

Table 5 Uncovered struts data for the total population N (%) Total number of struts analyzed

20,484

Total number of uncovered struts

1,320 (6 %)

Number of isolated uncovered struts Number of uncovered struts within clusters Total number of clusters

203 (15 %) 1,117 (85 %) 55

Number of patients with clusters

36

Number of struts per cluster (mean ± SD)

21.1 ± 14.7

Length of a cluster (mm) (mean ± SD)

7.1 ± 4.7

Length of a cluster (mm) [median (IQR)]

5.7 (4.8)

Number of quadrants per cluster (mean ± SD)

2.3 ± 0.9

Number of quadrants per cluster [median (IQR)] Position of cluster

2 (2)

Proximal

15/54 (28 %)

Proximal-middle

6/54 (11 %)

Middle

12/54 (22 %)

Distal-middle

2/54 (4 %)

Distal

19/54 (35 %)

All stent

0/54

Strut coverage clustering In the 55 follow-up pullbacks, a total of 20,484 struts were analyzed for coverage, of which 1,320 were uncovered. Table 5 summarizes the results of coverage and cluster data. Figure 6 shows a representative example of clustering of uncovered struts. The number of isolated uncovered struts was 203 (15 %) and the number of uncovered struts within clusters was 1,117 (85 %). In 36 of 55 patients, at least one cluster could be seen with a mean of 1.5 clusters in these patients. A total of 54 clusters of uncovered struts were counted. The mean number of struts per cluster was 21.1 ± 14.7. Clusters had a mean length of 7.1 ± 4.7 mm and spread over 2.3 ± 0.9 quadrants. In contrast to malapposition clusters, uncovered strut clusters tended to be more evenly distributed along the stent axis: proximal (28 %), proximal-middle (11 %), middle (22 %), distalmiddle (4 %) and distal (35 %). A graphical distribution is represented in Fig. 4. In Table 6, results of numbers of uncovered struts are compared from the perspective of patients with versus without clusters of uncovered struts. On aggregate, isolated uncovered struts (i.e. not occurring in a cluster) appear to be infrequent, while most uncovered struts occur in clusters. Figure 7 shows a graphical distribution of clusters of malapposed (in the baseline OCT acquisition) and uncovered struts (in the follow-up OCT acquisition). Twenty-one (21) of the 54 clusters of uncovered struts appear to have

123

846

Int J Cardiovasc Imaging (2014) 30:839–848

Fig. 6 Example of stent strut clustering for lack of coverage. In analogy to Fig. 3, uncovered stent struts are depicted as colored circles in the 3D space and the blue line represents the stent length. An example of a single cluster of uncovered struts located on the distal stent edge is given. The image also contains two isolated uncovered struts. From the image it is possible to immediately understand about cluster position, angular distribution and cluster dimension. 2D cross-sectional images illustrate some of the IVOCT images generating the cluster. Red arrows indicate the uncovered struts. The inlet shows a magnification of the IVOCT image in the center (yellow box). The white asterisk indicate the guide-wire shadowing artifact

Table 6 Distribution of uncovered struts by clustering No clusters (n = 19)

Clusters (n = 36)

All patients (n = 55)

p value

Total number of struts (mean)

342 ± 302

389 ± 356

372 ± 123

0.119

% of uncovered struts (mean)

1.0 ± 1.0

9.6 ± 6.8

6.6 ± 6.9

\0.01

[5 % uncovered struts

0 (0 %)

28 (78 %)

28 (51 %)

\0.001

[10 % uncovered struts

0 (0 %)

13 (36 %)

13 (24 %)

\0.001

cluster of malapposed struts in the same region in the baseline acquisition, indicating some, but no strong relation between both observations.

Discussion We developed a fully automatic algorithm for clustering of malapposed stent struts in OCT pullbacks as an addition to a previously described method for automated detection and quantification of stent strut apposition and coverage [5]. The automatic algorithm facilitates a time efficient quantification of the number of malapposed or uncovered struts within a stent, as well as their spatial distribution in three

123

dimensions. The clinical application of this algorithm in a population of patients undergoing OCT immediately after PCI and at 9-month follow-up, illustrates its potential use in cardiovascular research as well as in daily clinical practice. In this DES patient population, apparent similarities were noticed between the presence of malapposed strut clusters at base line and uncovered strut clusters at followup. At both instances, the majority of malapposed/uncovered struts appeared in clustered groups, while only a minority was isolated. Clusters were observed in almost 2/3 of patients and typically consisted of close to 20 struts, with a length of close to 7 mm and spreading over 2–3 quadrants. Clusters of malapposed struts occurred predominantly in the proximal segment of the stent (consistent with a higher incidence of malapposition proximally in tapering vessels), while clusters of uncovered struts were more evenly distributed over the whole stent length. The occurrence of malapposed and uncovered struts has been linked to an increased risk of acute and late thrombotic events [1, 7]. This would imply that a first possible use of our new algorithm could be the optimization of PCI with OCT. Several factors contributing to short and longterm prognosis of implanted stents, such as stent expansion and apposition, the presence of edge dissections or incomplete lesion coverage can be assessed with OCT. Before OCT will be implemented on a wide scale in every

Int J Cardiovasc Imaging (2014) 30:839–848

847

Fig. 7 Graphical representation of the occurrence of clusters of acute malapposed (red color, upper row) and uncovered struts (green color, lower row), in baseline and follow-up OCT acquisitions, respectively.

For each stented segment, the proximal and distal 3 mm are depicted in pale color. In case only one bar is depicted, this means no followup OCT acquisition was available for this patient

day interventional practice, however, it will be crucial to (1) provide robust scientific evidence which of these observations and measurements impact on clinical outcome and are amenable to additional intervention and (2) develop automated tools for fast and reliable analysis that give the operator an ad-hoc graphical ‘guide’ for optimal intervention. The here described algorithm for clustering of malapposed and uncovered struts, as an extension of our previously described software for automated strut detection, can be a valuable aid to achieve such goal. The method for automatic clustering offers the possibility for

advanced assessment of spatial distribution of malapposed or uncovered stent struts as well. We designed the present algorithm with the intention to be able to have an impact on several tunable parameters during an analysis procedure. The most important choice for the operator is to decide on the minimal number of struts needed to have a cluster. In the currently described illustration of the algorithm, this was arbitrarily set at 5. Further studies of in vivo human OCT data, as well as studies of clusters in human pathology specimens of patients with stent thrombosis, will be needed to more

123

848

Int J Cardiovasc Imaging (2014) 30:839–848

accurately assess the minimal dimensions of a cluster of malapposed or uncovered struts to be clinically relevant. A second important variable is the threshold for how far malapposed or uncovered struts can be located from their most neighboring strut but still be considered belonging to the same group. Again, such a parameter can be standardized, or modified accordingly, after validation in larger datasets. We deliberately refrained from the classical reporting of malapposition in areas and volumes, as these calculations may suffer from bias.

uncovered struts. A conveniently automated assessment and characterization of accumulations of malapposed and uncovered struts might facilitate further studies in this field. However, further investigations to assess the clinical importance of such accumulations for long-term clinical outcome are needed.

Limitations

Conflict of interest The authors have no conflict of interest to declare. Tom Adriaenssens received consultancy and speaker’s fees for St Jude Medical.

A first limitation of the clustering algorithm is related to the accuracy of the automated stent strut identification procedure, which was extensively discussed in our previous publication [2]. The accuracy of the algorithm for strut detection and quantification was consistently validated versus manual assessment and versus histopathology in an animal model [2, 6]. When applied to large datasets of OCT images by our research group, only very minor corrections usually need to be made after visual inspection of the results for final discrimination of malapposed and uncovered struts. Furthermore, a few arbitrary choices were made, including the minimum number of struts considered as cluster and the maximal distance of malapposed or uncovered struts to their most neighboring malapposed or uncovered strut for still being considered to belong to the same cluster. Further studies are required to define appropriate thresholds and to assess the relevance of accumulations of malapposed/uncovered struts with respect to safety outcomes (e.g. late stent thrombosis). Finally, we did not assess the repeatability across readings and between readers.

Conclusion The proposed algorithm facilitates the quantification of strut apposition and coverage in combination with comprehensive information about spatial strut distribution in an automated way. In a series of 64 and 55 patients analyzed at baseline and at follow-up, respectively, we observed a significant number of clusters of malapposed and

123

Acknowledgments The research program is supported by Grant G.0690.09N of the Research Foundation-Flanders (FWO) and IOF (Industrial research Foundation) grant of the Catholic University Leuven ZKC2992. Tom Adriaenssens is supported partially by a Clinical Doctoral Grant of the Research Foundation Flanders.

References 1. Cook S, Wenaweser P, Togni M, Billinger M, Morger C, Seiler C, Vogel R, Hess O, Meier B, Windecker S (2007) Incomplete stent apposition and very late stent thrombosis after drug-eluting stent implantation. Circulation 115:2426–2434. doi:10.1161/circulatio naha.106.658237 2. Ughi GJ, Adriaenssens T, Onsea K, Kayaert P, Dubois C, Sinnaeve P, Coosemans M, Desmet W, D’Hooge J (2012) Automatic segmentation of in vivo intra-coronary optical coherence tomography images to assess stent strut apposition and coverage. Int J Cardiovasc Imaging 28:229–241. doi:10.1007/s10554-011-9824-3 3. Imola F, Mallus MT, Ramazzotti V, Manzoli A, Pappalardo A, Di Giorgio A, Albertucci M, Prati F (2010) Safety and feasibility of frequency domain optical coherence tomography to guide decision making in percutaneous coronary intervention. EuroIntervention 6:575–581. doi:10.4244/eijv6i5a97 4. Bezerra HG, Costa MA, Guagliumi G, Rollins AM, Simon DI (2009) Intracoronary optical coherence tomography: a comprehensive review clinical and research applications. JACC Cardiovasc Interv 2:1035–1046. doi:10.1016/j.jcin.2009.06.019 5. Ughi GJ, Adriaenssens T, Desmet W, D’Hooge J (2012) Fully automatic three-dimensional visualization of intravascular optical coherence tomography images: methods and feasibility in vivo. Biomed Opt Express 3:3291–3303. doi:10.1364/boe.3.003291 6. Ughi G, Vandyck C, Adriaenssens T, Hoymans V, Sinnaeve P, Timmermans J, Desmet W, Vrints C, D’Hooge J (2013) Automatic assessment of stent neointimal coverage by intravascular optical coherence tomography. Eur Heart J CardiovascImaging. doi:10. 1093/ehjci/jet134 7. Finn AV, Joner M, Nakazawa G, Kolodgie F, Newell J, John MC, Gold HK, Virmani R (2007) Pathological correlates of late drugeluting stent thrombosis: strut coverage as a marker of endothelialization. Circulation 115:2435–2441. doi:10.1161/CIRCULA TIONAHA.107.693739 8. Tan PSMKV (2005) Introduction to data mining. Addison-Wesley publishing company, Boston, USA, pp 487–568

Automated detection and quantification of clusters of malapposed and uncovered intracoronary stent struts assessed with optical coherence tomography.

To date, accurate quantification and localization of malapposed and uncovered struts needs manual and time consuming analysis of large datasets. To de...
3MB Sizes 0 Downloads 3 Views