Automated knowledge-based detection of nonobstructive and obstructive arterial lesions from coronary CT angiography

Automated knowledge-based detection of nonobstructive and obstructive arterial lesions from coronary CT angiography

Republished with permission from Medical Physics

Dongwoo Kang1, Piotr J. Slomka2, Ryo Nakazato2, Reza Arsanjani2, Victor Y. Cheng2, James K. Min2, Debiao Li3, Daniel S. Berman2, C.-C. Jay Kuo1, and Damini Dey3,a
1 Department of Electrical Engineering, University of Southern California, Los Angeles, California 90089
2 Departments of Imaging and Medicine and Cedars-Sinai Heart Institute, Cedars-Sinai Medical Center, Los Angeles, California 90048
3 Biomedical Imaging Research Institute, Department of Biomedical Sciences, Cedars-Sinai Medical Center, Los Angeles, California 90048

Received 12 September 2012; accepted 20 February 2013; published online 22 March 2013


Coronary artery disease (CAD) is the leading cause of death worldwide for both men and women.1 Three-dimensional (3D) coronary computed tomography angiography (CCTA) with the use of multidetector CT scanners is increasingly employed for noninvasive evaluation of CAD, having shown high accuracy and negative predictive value for detection of coronary artery stenosis in comparison with invasive coronary angiography.2,3,4,5,6 Beyond stenosis, CCTA also permits noninvasive assessment of atherosclerotic plaque and coronary artery remodeling.7,8,9

Although computer-aided extraction of the coronary arteries is often employed to aid visual analysis,10,11,12,13,14 clinical assessment of CCTA and lesion detection is based on visual analysis, which is time consuming and subject to observer variability.15 It was reported that acquiring expertise in CCTA may take more than 1 yr.15 Automatic software that detects and identifies the coronary artery lesions would reduce such observer variability as well as the time for assessment.

Many computer-aided algorithms for detecting and diagnosing various abnormalities have been developed for medical imaging, such as detection and quantification of chronic obstructive pulmonary disease in lung,16,17,18,19,20 colon cancer,21,22,23,24 and lesions in mammograms.25,26,27,28 Detection and quantification of coronary artery lesions are particularly challenging due to limited spatial resolution and coronary artery motion, even smaller plaque size than arteries, and complex and variable coronary artery anatomies. Automated lesion detection requires accurate extraction of coronary artery centerlines, classification of normal and abnormal lumen cross sections, quantification of luminal stenosis, and classification of lesions with the different degree of stenosis.

Obtaining a reliable coronary centerline from CCTA is a key step which serves as a starting point for lumen segmentation and stenosis grade calculation for lesion detection. Many approaches have been proposed on centerline extraction: thinning based techniques,23,29,30,31 tracking methods,32,33,34 minimal path techniques,10,13,35 and distance transform methods36,37,38,39 have been presented.

To date, only a few studies have attempted automatic detection of lesions;40,41,42,43,44 only obstructive lesions (with stenosis ≥50%) were detected. To our knowledge, there has been no attempt of automatic detection of nonobstructive lesions (25%–49%). However, nonobstructive lesions (25%–49%) have been shown to be a clinically significant predictor of future coronary events.45,46 We have previously described a preliminary algorithm for automated detection of both obstructive and nonobstructive lesions, validated on 19 patients.47 In this work, we describe an improved automated algorithm, which was validated on detection of lesions in the left anterior descending (LAD), left circumflex (LCX) and right coronary artery (RCA) of 42 consecutive patients (126 arteries, 252 proximal and mid segments). The aim in our study was to develop a robust automated algorithm to detect both obstructive and nonobstructive lesions from CCTA and validate it in comparison with three experienced expert readers.


Our algorithm can be divided into following four main steps, as described below: (1) centerline extraction and vessel classification, (2) vessel linearization, (3) lumen segmentation, and (4) lesion location detection. The algorithm requires two user-defined points as input, at the ostium of RCA and the left main (LM) coronary artery, and second, the placing of a standard circular region of interest in the aorta at the level of the LM ostium, to obtain scan-specific attenuation range for luminal contrast as previously described.48,49 Note that algorithms for the automatic segmentation of aorta and detection of the origin of coronary arteries have been previously proposed;50 thus, they were not the focus of this work. The algorithm uses a 3D CCTA image dataset (typically the best phase) as input and a binary mask of coronary tree that is initially computed by the commercially available coronary artery segmentation software (Circulation, Syngo MMWP Version VE31A, Siemens Medical Solutions, Forchheim, Germany). The binary mask was used as a first step for centerline extraction. Following this step, our approach uses scan-specific threshold levels with recursive region growing for accurate lumen segmentation. The remaining process of lesion detection is automated. The luminal attenuation in CCTA differs with acquisition protocols and between patients scanned with the same protocol. Plaque attenuation thresholds have been shown to vary significantly with intracoronary lumen attenuation51 and reconstruction parameters choice52 and are patient and scan-specific. Therefore, scan-specific attenuation threshold levels for lumen and plaque48,49 are important for accurate lumen segmentation and were utilized in our algorithm. Our knowledge-based algorithm allows normal tapering of the coronary arteries and identifies and adjusts for arterial branch points. The flowchart of the proposed automated lesion detection algorithm is given in Fig. 1.

Figure 1- Flowchart of the proposed method.

Our study consisted of 42 consecutive patients, who underwent CCTA for clinical reasons at the Cedars-Sinai Medical Center. All CCTA datasets were acquired on the dual-source 64-slice CT scanner (Definition Siemens Medical Solution, Forchheim, Germany) with a gantry-rotation time of 330 ms and standard collimation of 0.6 mm, and had good-excellent image quality. CCTA scan parameters were 512 × 512 matrix, voxel size 0.38 × 0.38 × 0.3 mm3, and typically consisted of 400–500 slices per dataset. Of the 42 patients, 10 patients subsequently underwent invasive coronary angiography (ICA) within 1 month of the CCTA scan.

Centerline extraction and classification of three main arteries

In this section, we present an automatic centerline extraction algorithm as our first step, which also classifies the three main coronary arteries: LAD, LCX, and RCA.

Initially, 3D thinning53 is applied to the input mask (CCTA dataset bounded by the initial lumen mask), resulting in a collection of N center points xi in the arteries including aorta. Subsequently, these N points are connected using graph theory. An undirected graph G = ⟨V, E⟩ is defined as a set of nodes (V = {p1, p2, ...,pN}) and a set of edges (E) that connect adjacent nodes. A non-negative weight we is assigned on each edge e ∈ E, where we is defined as the Euclidean distance between two nodes that are connected by an edge e. We deleted the thinning points inside the aorta by detecting the center of aorta pa which is found by using Dijkstra's shortest path method54 applied between the ostium point of RCA pR and the ostium point of the LM pL, so that the points in the left and right coronary arteries are not connected to each other.

A temporal end point ptemp1 ∈ V of RCA, the farthest point from the ostium pR, is found using Dijkstra's shortest path method. Among the points pi ∈ {V − {point s on the path from pR to ptemp1}}, another temporal end point ptemp2 ∈ V of RCA is found by searching the maximum cost that is the sum of distances from pR and ptemp1 ∈ V using Dijkstra's shortest path. The end points ptemp1 ∈ V and ptemp2 ∈ V are further classified into the end points of posterior descending artery (PDA)-RCA and posterior lateral branch (PLB)-RCA utilizing anatomical knowledge of relative artery positions.

Similarly, the end point of the first left coronary artery (either LAD or LCX) is also found using Dijkstra's shortest path method. The LAD and the LCX are further classified using knowledge-based rules of relative artery positions. The end point of the Ramus artery is also identified if it exists. In each of the three main coronary arteries, the end points of small branches, such as the first diagonal branch (D1) on LAD and the first obtuse marginal branch (OM1) on LCX, are also detected using Dijkstra's shortest path method and anatomical knowledge of relative artery position. The first row in Fig. 2 shows the results of an extracted centerline and a vessel classified as LAD.

Figure 2 - The first row shows an example of extracted centerlines in mid LAD and D1. Each column shows an image at different angles with axes perpendicular to each other. The second row shows vessel linearization (LAD) of the first row in three orthogonal directions. Both rows show the same location of a lesion (25%–49% stenosis by expert visual grading). The third row also shows linearized vessel (LAD) of a normal CCTA dataset. The outline (third row) shows segmented lumen using our method. Detected lesion locations are marked by points around the detected locations.

Vessel linearization

The vessels are subsequently converted to a linearized representation for further image processing. At each point pi ∈ V of the extracted centerlines, two basis vectors, which span the cross-sectional plane perpendicular to the centerline, are calculated. The basis vectors define the cross-sectional planes perpendicular to the centerline, which are used to map the arteries to linearized image coordinates. These cross-sectional planes (size 8 × 8 mm2, corresponding to 21 × 21 matrix with a voxel-size of 0.38 mm, the smallest voxel dimension of the coronary CTA volume) are stacked up along the centerlines, resulting in 3D linear volume of coronary arteries. The second row in Fig. 2 shows an example of the resulting linearized vessels.

Lumen segmentation and calcium volume measurement

Plaque attenuation has been shown to vary significantly with intracoronary lumen attenuation55,56 and reconstruction kernels.51,57 Thus, application of scan-specific or even plaque-specific attenuation thresholds is essential.48,49 Since attenuation ranges for lumen and plaque can depend on the patient and the acquisition protocol, we compute these attenuation ranges automatically from the scan using a validated method previously described by our group.48,49 Attenuation thresholds for the lumen, noncalcified, and calcified plaque are found from the image histogram of the “normal blood pool” region-of-interest placed in the aortic root, and are adjusted for proximal-to-distal decrease in contrast.48,49 Scan-specific attenuation range for the lumen is defined by the upper attenuation threshold for noncalcified plaque and the lower attenuation threshold for calcified plaque.48,49 The lumen is then identified by recursive region-growing in the linearized volume, using the computed attenuation range, which allows exclusion of calcified and noncalcified plaque. An example of the lumen segmentation is shown in the third row in Fig. 2. Additionally, calcium volumes are measured using the attenuation threshold for calcified plaque from the linearized volume.

Detection of lesions with stenosis

From the previous steps, presence and location of lesions are identified by a knowledge-based algorithm, using the lumen segmentation performed with scan-specific lumen attenuation range. Lumen diameters for each 2D cross section are first obtained from the segmented lumen areas. Expected or “normal” luminal diameter is derived from the scan by automated piecewise least squares line fitting (Fig. 3) over the proximal and mid segments (67%) of the coronary artery between branch points detected; this computation allows us to take into account any “normal tapering” present in the dataset. The lumen diameters at all positions are first cropped, considering expected dimensions of the coronary arteries,58 and then the lumen diameters after branch points are cropped again by using the lumen diameters before the branch points. Lesion detection is then performed in multiple passes through the coronary artery, using the following steps:

In the first pass, all possible lesions are found by considering the difference, d, from the piecewise fitted line, as below:

where s is the distance in mm from the ostium, ls is the luminal diameter at s, and a and b are obtained from least-squares line fitting. In this first pass, branch-points are excluded by cropping above the fitted line, as shown in Fig. 3.

For each possible lesion, the algorithm searches locally for proximal and distal normal references by considering d in Eq. (1).

In a second pass, we compute a stenosis estimate Se(%) for each possible lesion, as below:

If Se is greater than or equal to 25%, the lesion is considered for the next step.
For lesions with Se ≥ 25%, stenosis is computed for each cross section corresponding to s between the detected proximal and distal limits, considering both reference limits, as previously published:59

where ls, lp, ld are the luminal diameters for cross section corresponding to s, proximal, and distal references; sp and sd are the linear distances between the proximal reference and the distal reference, and the cross section corresponding to s, respectively.

Figure 3 - Example of lumen segmentation and lesion detection in LAD. Range of proximal LAD lesion (stenosis 25%–49%) marked by expert is shown as a small box at around x = 27 mm – 48 mm. Lumen diameters computed from the segmented lumen are shown and their cropped lumen diameters by anatomical knowledge are also shown. Expected “normal” luminal diameter is derived from the scan by automated piecewise line fitting between branch points, and takes into account “normal tapering” present in the dataset. The locations of the lesions with ≥25% stenosis detected by the algorithm, concordant with the expert observer, is marked with vertical arrows.

Lesions with maximum stenosis St ≥ 25% are marked (Fig. 4) in the region of maximum stenosis as well as the proximal and distal reference segments.

Figure 4 - An example of 3D volume rendering (a), detected nonobstructive lesion (mixed plaque) in a CCTA image (b), and according ICA image (c). Arrows in (a)–(c) indicate the location of the same lesion (25%–49% stenosis by expert visual grading from CCTA and 34.0% stenosis by quantitative analysis from ICA).

At branch point locations, lesions can be missed due to the wider lumen diameter at the branch. Therefore, in addition to the above lesions with stenosis detection algorithm, a search for calcified lesions was performed at each branch point by searching above automated measuring volume using scan-specific attenuation threshold levels for lumen and plaque,48,49 and a calcified lesion was added at that location if detected. The user has the option in our software to manually accept or reject any identified lesions as needed (calcified or noncalcified). Once rejected, the lesion is marked as corrected and the annotated limits of the lesions are deleted.

Visual assessment and reference standard

All datasets were first visually assessed in a standard and systematic way, by three experienced expert readers, using consensus reading to minimize the interobserver variability; these readers were three experienced imaging cardiologists. Segmental analysis was based on the standard 15-segment American Heart Association.60 Each segment of the coronary artery tree was graded for the presence and type of plaque or stenosis, as recommended by published guidelines of the Society of Cardiovascular CT,61 and all coronary lesions with stenosis ≥25% were identified. This was used as the reference standard for algorithm performance in this study. For comparison, a second blinded reader (imaging cardiologist with Level III CT certification, with 1 yr of experience with cardiac CT) also independently identified all coronary lesions with stenosis ≥25%. The observer agreement for this reader with the reference standard was 94.8% (kappa 0.84, 95% confidence interval 0.75 to 0.92, p < 0.0001).

Statistical analysis

Our algorithm requires three user-interactions–-placing circular region-of-interest at the aortic root, and marking the origin of the left main and the right coronary artery. The performance of our algorithm was examined where two different readers ran the program independently with the three interactions on all the cases. The results of the reproducibility evaluation are shown in the result. The agreement of the lesion detection results derived from the proposed automated lesion detection software was measured using the kappa statistics, as well as sensitivity, specificity, and receiver-operator characteristic (ROC) analysis, with Analyse-it software.62 A p-value < 0.05 was considered statistically significant.


We tested the algorithm on 42 consecutive patients [26 male]. The mean age was 60 ± 12 yr and the mean body weight was 83 ± 10 kg. About 21 patients had coronary lesions with stenosis greater than or equal to 25%. In these patients, 45 lesions with stenosis ≥25% were identified. Eight out of the remaining 21 patients had lesions with stenosis <25% and 13 patients did not have any lesions (no luminal stenosis or plaque). In the 45 lesions with stenosis ≥25%, 20 lesions were obstructive (≥50%).

The proposed algorithm ran successfully (Figs. 5,6,7) on all the proximal and mid coronary artery segments in all patients, with an execution time of around 50 s for centerline extraction for all coronary arteries and <2 s for all subsequent steps on a standard 2.5 GHz personal computer running windows XP.

Figure 5 - The first row shows an example of extracted centerlines in mid LAD and D1. Each column shows an image at different angles with axes perpendicular to each other. The second row shows vessel linearization (LAD) of the first row in three orthogonal directions. Both rows show the same location of a lesion (25%–49% stenosis by expert visual grading). The outline (third row) shows segmented lumen using our method.


Figure 6 - Example of lumen segmentation and lesion detection in LAD. Range of proximal LAD lesion (stenosis 25%–49%) marked by expert is shown as boxes at around x = 0 mm – 38 mm and 53 mm – 68 mm. Lumen diameters computed from the segmented lumen are shown and their cropped lumen diameters by anatomical knowledge are also shown. Expected “normal” luminal diameter is derived from the scan by automated piecewise line fitting between branch points, and takes into account “normal tapering” present in the dataset. The locations of the lesions with ≥25% stenosis detected by the algorithm, concordant with the expert observer, is marked with vertical arrows. The second vertical line, which is at around x = 60 mm, is the lesion detected additionally by calcium volume measurement.

Figure 7 - An example of 3D volume rendering (a), detected nonobstructive lesion (mixed plaque) by stenosis calculation (CCTA image) (b), and detected nonobstructive lesion (mixed plaque) by calcium volume measurement at a branch point (CCTA image) (c). Arrows in (a)–(c) indicate the locations of the detected lesions (25%–49% stenosis). This patient did not undergo ICA.


The algorithm was validated by standard 10-fold cross-validation, as described in Ref. 63 by dividing the 42 patient datasets into 10 subsamples, and training and testing 10 times. In the 45 lesions with stenosis ≥25%, the proposed automated algorithm correctly identified 21/24 lesions in the LAD, 10/10 in the LCX, and 11/11 in the RCA. Figures 8,9 show two patient examples from our study.

Figure 8 - Detection of lesion with stenosis. Arrows indicate the location of lesions. Detected lesions with stenosis by primarily noncalcified plaque in the mid segment (the first row, 59% stenosis by quantitative analysis and 50%–69% stenosis by expert visual grading). The second row shows according ICA images of the first row (36.2% stenosis by quantitative analysis).

Figure 9 - Detection of lesion with stenosis. Arrows indicate the location of lesions. Detected lesions with stenosis by mixed plaque in the proximal segment (the first row, 70% stenosis by quantitative analysis and 90%–99% stenosis by expert visual grading). The second row shows according ICA images of the first row (62.4% stenosis by quantitative analysis).

In total, the proposed automated algorithm correctly identified true lesions yielding a sensitivity of 93% (42/45) on a per-segment basis (Table 1). Three lesions (25%–49% stenosis) were missed by our approach, when there is no proximal or distal normal references around the location of stenosis. Also, on a per-segment basis, the proposed algorithm showed 81% specificity, 83% accuracy, 98% negative predictive value, and 52% positive predictive value. For obstructive lesions, the sensitivity was 100%, the specificity was 84%, and the accuracy was 86%.

There were 39 false positive detections (Table 2) in the 252 coronary artery segments of 42 patients by the proposed algorithm resulting in an average of 0.15 per segment; the reasons for these false positives are described in Table 2. Twenty three false positives out of 39 (59%) were lesions with <25% stenosis as assessed by the expert observer (Fig. 10). When excluding the 23 detections that were lesions having stenosis <25% and considering only the detections that were not related with stenotic atherosclerosis as false positives, specificity was increased to 92% and the average false positive ratio was 0.06 per segment. Out of the remaining 16 false positive detection marks, five cases were associated with normal segment with narrowing without plaque, nine cases were associated with undetected small branches, and two cases were associated with unclear image contrast and blurring.

Figure 10 - An example of “false positive” lesion. The algorithm detected this location as a lesion with stenosis ≥25%, but expert readers graded it <25% stenosis.

Additionally, 10 out of our 42 patients underwent ICA, using the Inova digital x-ray system from GE Healthcare with multiple views of the left and right coronary artery to identify the projection in which the segment appeared most stenotic. Standard cardiac catheterization technique was employed. Acquired images were transferred to an AGFA Heartlab workstation for quantitative coronary catheter angiography (QCA) analysis. Ten cases were interpreted for stenosis by consensus of two experienced readers, and a separated investigator independently performed QCA for all 10 cases. Reference luminal proximal and distal positions to the stenosis were defined by readers, and then QCA software on the workstation detected luminal edges, located site of maximal stenosis, and quantified the maximal stenosis64 (Figs. 8,9). Based on ICA-based QCA, 10 out of 10 patients had stenosis ≥25%, in agreement with the results of our algorithm on CCTA (Figs. 8,9).

Program reproducibility

There was very good agreement between two independent readers for running the lesion detection software, with an observed agreement of 94.8% (kappa 0.89, 95% confidence interval 0.83 to 0.95, p < 0.0001). For comparison, this program agreement was comparable to agreement of the second blinded reader with the reference standard (agreement 94.8%, kappa 0.84, 95% confidence interval 0.75 to 0.92, p < 0.0001).
Table 3 shows the sensitivity, specificity, and ROC-area-under-the-curve (ROC-AUC) for the two readers running the program, compared to the reference standard. The ROC-AUC was 0.87 for both readers running the program.


Our results demonstrate high agreement with three expert readers in consensus in a sizable patient population (42 patients), and the algorithm was validated by standard 10-fold cross-validation. The algorithm showed a sensitivity of 93% and a negative predictive value of 98% on a per-segment basis in the main three coronary arteries, which is the most desirable. Our specificity was relatively low (81%) due to 39 additional detections on a per-segments basis, which could be manually rejected, as in clinically used software for coronary calcium scoring.65,66 From a clinical standpoint, it is essential for the computer-aided system to have a high sensitivity since the false positive results can be discarded by one-button-click. By identifying all potential lesions, the system would aid the physician by quickly identifying all lesions. Furthermore, plaque burden could be measured automatically over all identified lesions.48,49

Previous studies

Our proposed study is in line with the few previous studies attempting automated lesion detection from CCTA. Halpern42 and Arnoldi40 published validation papers using commercial software with expert human interpretation, where they detected obstructive lesions only (with ≥50% stenosis). Dinesh41 proposed a method, which utilized manual centerlines and artery classification, and did not provide specific stenosis calculation, and were evaluated with the small number of patients (eight patients). Recently, Kelm43 and Goldberg44 also published automated detection of obstructive (≥50% stenosis) coronary artery lesions from CCTA.

One of the main advances as compared to previously published work is that our method can accurately detect both obstructive and nonobstructive lesions (25%–49% stenosis), whereas previous studies detected obstructive lesions only (≥50% stenosis). This is of particular clinical value since lesions with nonobstructive stenosis have been shown to contribute to cardiovascular events.45,46 Nonobstructive lesions (25%–49%) are more challenging to detect due to the subtle narrowing of the lumen. Additionally, compared to other previous methods that detected only obstructive lesions, our proposed algorithm also identified arterial branch points to improve automated detection of lesions.

Our automated software showed high reproducibility when run by two independent observers (94.8% agreement). The small interobserver variability was due to the three manual user-interactions; setting two ostial points and setting a ROI in aorta. Those manual interactions produced slightly different lumen segmentation and estimated normal luminal reference diameters, resulting in different stenosis calculations. Automatic algorithms have been previously proposed for these steps50 and these could be combined with our lesion detection technique. The remaining processes were all automatic.

There were a few limitations in our study. Invasive coronary angiography was not performed for all patients. In our study, the reference standard was clinically utilized visual detection and grading of lesions, by three expert readers in consensus. However, stenosis calculation by expert readers was not available. The algorithm was not fully automated, but required three mouse clicks to define ostial locations and aortic region-of-interest. A binary mask was used as a first step for centerline extraction. However, the algorithm can be combined with existing well-established commercial software performing lumen segmentation, to provide clinical tools for lesion detection. Reduced specificity due to the additional lesions found by the software poses a challenge. Currently, only the major proximal and mid branches are classified. Automated or semiautomated labeling of all coronary branches would potentially reduce the number of false positives. For robust performance, even in CCTA datasets with fair image quality, our algorithm used only the lumen for automated identification of lesions; however, assessment of the vessel wall in combination with lumen may improve our lesion detection results, and this needs to be further evaluated. Lesions with stenosis ≥25% only were considered in our study; we have, however, also evaluated the results for the detection of lesions with stenosis ≥50%.


In conclusion, we developed a novel automated algorithm for detection and localization of obstructive and nonobstructive arterial lesions from CCTA, which performed with high sensitivity compared to three experienced expert readers.


The authors wish to thank Mr. Amit Ramesh for his help with the software and Mr. David Choi for help with data anonymization. This study was supported by AHA Grant No. 09GRNT2330000 (PI: Damini Dey).


Table I. Performance characteristics of the proposed algorithm of lesion (≥25% stenosis) detection in N = 42 patients (13 complete normal). In a total of 45 lesions with ≥25% stenosis, 6 were of severe stenosis (≥70%), and 14 were obstructive stenosis (≥50%). Sensitivity was 93%, specificity was 81%, and accuracy was 83% per segment.


  1. S. Yusuf, S. Reddy, S. Ounpuu, and S. Anand, “Global burden of cardiovascular diseases: Part I: General considerations, the epidemiologic transition, risk factors, and impact of urbanization,” Circulation 104, 2746–2753 (2001).
  2. S. Achenbach, U. Ropers, A. Kuettner, K. Anders, T. Pflederer, S. Komatsu, W. Bautz, W. G. Daniel, and D. Ropers, “Randomized comparison of 64-slice single- and dual-source computed tomography coronary angiography for the detection of coronary artery disease,” JACC Cardiovasc. Imaging 1, 177–186 (2008).
  3. M. J. Budoff, D. Dowe, J. G. Jollis, M. Gitter, J. Sutherland, E. Halamert, M. Scherer, R. Bellinger, A. Martin, R. Benton, A. Delago, and J. K. Min, “Diagnostic performance of 64-multidetector row coronary computed tomographic angiography for evaluation of coronary artery stenosis in individuals without known coronary artery disease: Results from the prospective multicenter ACCURACY (Assessment by Coronary Computed Tomographic Angiography of Individuals Undergoing Invasive Coronary Angiography) trial,” J. Am. Coll. Cardiol. 52, 1724–1732 (2008).
  4. J. Hausleiter, T. Meyer, M. Hadamitzky, M. Zankl, P. Gerein, K. Dorrler, A. Kastrati, S. Martinoff, and A. Schomig, “Non-invasive coronary computed tomographic angiography for patients with suspected coronary artery disease: The Coronary Angiography by Computed Tomography with the Use of a Submillimeter resolution (CACTUS) trial,” Eur. Heart J. 28, 3034–3041 (2007).
  5. W. B. Meijboom, C. A. van Mieghem, N. R. Mollet, F. Pugliese, A. C. Weustink, N. van Pelt, F. Cademartiri, K. Nieman, E. Boersma, P. de Jaegere, G. P. Krestin, and P. J. de Feyter, “64-slice computed tomography coronary angiography in patients with high, intermediate, or low pretest probability of significant coronary artery disease,” J. Am. Coll. Cardiol. 50, 1469–1475 (2007).
  6. J. M. Miller, C. E. Rochitte, M. Dewey, A. Arbab-Zadeh, H. Niinuma, I. Gottlieb, N. Paul, M. E. Clouse, E. P. Shapiro, J. Hoe, A. C. Lardo, D. E. Bush, A. de Roos, C. Cox, J. Brinker, and J. A. Lima, “Diagnostic performance of coronary angiography by 64-row CT,” N. Engl. J. Med. 359, 2324–2336 (2008). [MEDLINE]
  7. S. Achenbach, F. Moselewski, D. Ropers, M. Ferencik, U. Hoffmann, B. MacNeill, K. Pohle, U. Baum, K. Anders, I. K. Jang, W. G. Daniel, and T. J. Brady, “Detection of calcified and noncalcified coronary atherosclerotic plaque by contrast-enhanced, submillimeter multidetector spiral computed tomography: A segment-based comparison with intravascular ultrasound,” Circulation 109, 14–17 (2004). [MEDLINE]
  8. A. W. Leber, A. Becker, A. Knez, F. von Ziegler, M. Sirol, K. Nikolaou, B. Ohnesorge, Z. A. Fayad, C. R. Becker, M. Reiser, G. Steinbeck, and P. Boekstegers, “Accuracy of 64-slice computed tomography to classify and quantify plaque volumes in the proximal coronary system: A comparative study using intravascular ultrasound,” J. Am. Coll. Cardiol. 47, 672–677 (2006). [MEDLINE]
  9. M. Petranovic, A. Soni, H. Bezzera, R. Loureiro, A. Sarwar, C. Raffel, E. Pomerantsev, I. K. Jang, T. J. Brady, S. Achenbach, and R. C. Cury, “Assessment of nonstenotic coronary lesions by 64-slice multidetector computed tomography in comparison to intravascular ultrasound: Evaluation of nonculprit coronary lesions,” J. Cardiovasc. Comput. Tomogr. 3, 24–31 (2009).
  10. T. Deschamps and L. D. Cohen, “Fast extraction of minimal paths in 3D images and applications to virtual endoscopy,” Med. Image Anal. 5, 281–299 (2001). [ISI] [MEDLINE]
  11. L. M. Lorigo, O. D. Faugeras, W. E. Grimson, R. Keriven, R. Kikinis, A. Nabavi, and C. F. Westin, “CURVES: Curve evolution for vessel segmentation,” Med. Image Anal. 5, 195–206 (2001). [ISI] [MEDLINE]
  12. R. Manniesing, M. Schaap, S. Rozie, R. Hameeteman, D. Vukadinovic, A. van der Lugt, and W. Niessen, “Robust CTA lumen segmentation of the atherosclerotic carotid artery bifurcation in a large patient population,” Med. Image Anal. 14, 759–769 (2010).
  13. C. T. Metz, M. Schaap, A. C. Weustink, N. R. Mollet, T. van Walsum, and W. J. Niessen, “Coronary centerline extraction from CT coronary angiography images using a minimum cost path approach,” Med. Phys. 36, 5568–5579 (2009).
  14. M. Schaap, L. Neefjes, C. Metz, A. van der Giessen, A. Weustink, N. Mollet, J. Wentzel, T. van Walsum, and W. Niessen, “Coronary lumen segmentation using graph cuts and robust kernel regression,” in Proc. Information Processing in Medical Imaging, Vol. 5636 (Springer, 2009), pp. 528–539.
  15. F. Pugliese, M. G. Hunink, K. Gruszczynska, F. Alberghina, R. Malago, N. van Pelt, N. R. Mollet, F. Cademartiri, A. C. Weustink, W. B. Meijboom, C. L. Witteman, P. J. de Feyter, and G. P. Krestin, “Learning curve for coronary CT angiography: What constitutes sufficient training?,” Radiology 251, 359–368 (2009).
  16. I. C. Sluimer, P. F. van Waes, M. A. Viergever, and B. van Ginneken, “Computer-aided diagnosis in high resolution CT of the lungs,” Med. Phys. 30, 3081–3090 (2003). [MEDLINE]
  17. L. Sorensen, S. B. Shaker, and M. de Bruijne, “Quantitative analysis of pulmonary emphysema using local binary patterns,” IEEE Trans. Med. Imaging 29, 559–569 (2010). | First Citation first citation in article
  18. T. Stavngaard, S. B. Shaker, K. S. Bach, B. C. Stoel, and A. Dirksen, “Quantitative assessment of regional emphysema distribution in patients with chronic obstructive pulmonary disease (COPD),” Acta Radiol. 47, 914–921 (2006).
  19. R. Uppaluri, E. A. Hoffman, M. Sonka, P. G. Hartley, G. W. Hunninghake, and G. McLennan, “Computer recognition of regional lung disease patterns,” Am. J. Respir. Crit. Care Med. 160(2), 648–654 (1999). [ISI] [MEDLINE]
  20. D. Wormanns, M. Fiebich, M. Saidi, S. Diederich, and W. Heindel, “Automatic detection of pulmonary nodules at spiral CT: Clinical application of a computer-aided diagnosis system,” Eur. Radiol. 12, 1052–1057 (2002). [MEDLINE]
  21. S. B. Gokturk, C. Tomasi, B. Acar, C. F. Beaulieu, D. S. Paik, R. B. Jeffrey Jr., J. Yee, and S. Napel, “A statistical 3D pattern processing method for computer-aided detection of polyps in CT colonography,” IEEE Trans. Med. Imaging 20, 1251–1260 (2001). [Inspec] [ISI] [MEDLINE]
  22. G. Kiss, J. Van Cleynenbreugel, M. Thomeer, P. Suetens, and G. Marchal, “Computer-aided diagnosis in virtual colonography via combination of surface normal and sphere fitting methods,” Eur. Radiol. 12, 77–81 (2002). [MEDLINE]
  23. R. J. T. Sadleir and P. F. Whelan, “Colon centreline calculation for CT colonography using optimised 3D topological thinning,” in Proc. First International Symposium on 3D Data Processing Visualization and Transmission, 800–803 (2002).
  24. P. Sundaram, A. Zomorodian, C. Beaulieu, and S. Napel, “Colon polyp detection using smoothed shape operators: Preliminary results,” Med. Image Anal. 12, 99–119 (2008). [Inspec] [MEDLINE]
  25. T. W. Freer and M. J. Ulissey, “Screening mammography with computer-aided detection: Prospective study of 12,860 patients in a community breast center,” Radiology 220, 781–786 (2001). [ISI] [MEDLINE] | First Citation first citation in article
  26. S. Nawano, K. Murakami, N. Moriyama, H. Kobatake, H. Takeo, and K. Shimura, “Computer-aided diagnosis in full digital mammography,” Invest. Radiol. 34, 310–316 (1999). [ISI] [MEDLINE] | First Citation first citation in article
  27. R. M. Nishikawa, “Current status and future directions of computer-aided diagnosis in mammography,” Comput. Med. Imaging Graph. 31, 224–235 (2007). [Inspec] [MEDLINE]
  28. T. Tanaka, N. Nitta, S. Ohta, T. Kobayashi, A. Kano, K. Tsuchiya, Y. Murakami, S. Kitahara, M. Wakamiya, A. Furukawa, M. Takahashi, and K. Murata, “Evaluation of computer-aided detection of lesions in mammograms obtained with a digital phase-contrast mammography system,” Eur. Radiol. 19, 2886–2895 (2009).
  29. Y. Ge, D. R. Stelts, J. Wang, and D. J. Vining, “Computing the centerline of a colon: A robust and efficient method based on 3D skeletons,” J. Comput. Assist. Tomogr. 23, 786–794 (1999). [ISI] | First Citation first citation in article
  30. C. M. Ma and M. Sonka, “A fully parallel 3D thinning algorithm and its applications,” Comput. Vis. Image Underst. 64, 420–433 (1996). [ISI] | First Citation first citation in article
  31. C. M. Ma, S. Y. Wan, and J. D. Lee, “Three-dimensional topology preserving reduction on the 4-subfields,” IEEE Trans. Pattern Anal. Mach. Intell. 24, 1594–1605 (2002).
  32. S. R. Aylward and E. Bullitt, “Initialization, noise, singularities, and scale in height ridge traversal for tubular object centerline extraction,” IEEE Trans. Med. Imaging 21, 61–75 (2002). [Inspec] [ISI] [MEDLINE]
  33. O. Wink, W. J. Niessen, and M. A. Viergever, “Fast delineation and visualization of vessels in 3D angiographic images,” IEEE Trans. Med. Imaging 19, 337–346 (2000). [Inspec] [ISI] [MEDLINE]
  34. O. Wink, W. J. Niessen, and M. A. Viergever, “Multiscale vessel tracking,” IEEE Trans. Med. Imaging 23, 130–133 (2004). [Inspec] [MEDLINE]
  35. L. D. Cohen and R. Kimmel, “Global minimum for active contour models: A minimal path approach,” Int. J. Comput. Vision 24, 57–78 (1997). | First Citation first citation in article
  36. I. Bitter, A. E. Kaufman, and M. Sato, “Penalized-distance volumetric skeleton algorithm,” IEEE Trans. Vis. Comput. Graph. 7, 195–206 (2001). [Inspec] [ISI]
  37. T. Boskamp, D. Rinck, F. Link, B. Kummerlen, G. Stamm, and P. Mildenberger, “New vessel analysis tool for morphometric quantification and visualization of vessels in CT and MR imaging datasets,” Radiographics 24, 287–297 (2004). [MEDLINE]
  38. D. Q. Chen, B. Li, Z. R. Liang, M. Wan, A. Kaufman, and M. Wax, “A tree-branch searching, multiresolution approach to skeletonization for virtual endoscopy,” Med. Imaging Image Process 1(Parts 1 and 2), 726–734 (2000).
  39. Y. Zhou and A. W. Toga, “Efficient skeletonization of volumetric objects,” IEEE Trans. Vis. Comput. Graph. 5, 196–209 (1999). [Inspec] [ISI]
  40. E. Arnoldi, M. Gebregziabher, U. J. Schoepf, R. Goldenberg, L. Ramos-Duran, P. L. Zwerner, K. Nikolaou, M. F. Reiser, P. Costello, and C. Thilo, “Automated computer-aided stenosis detection at coronary CT angiography: Initial experience,” Eur. Radiol. 20, 1160–1167 (2010).
  41. M. S. Dinesh, P. Devarakota, and J. Kumar, “Automatic detection of plaques with severe stenosis in coronary vessels of CT angiography,” in Proc. SPIE Medical Imaging, Vol. 7624, Medical Imaging 2010: Computer-Aided Diagnosis, pp. 76242Q.
  42. E. J. Halpern and D. J. Halpern, “Diagnosis of coronary stenosis with CT angiography comparison of automated computer diagnosis with expert readings,” Acad. Radiol. 18, 324–333 (2011).
  43. B. M. Kelm, S. Mittal, Y. Zheng, A. Tsymbal, D. Bernhardt, F. Vega-Higuera, S. K. Zhou, P. Meer, and D. Comaniciu, “Detection, grading and classification of coronary stenoses in computed tomography angiography,” Med. Image. Comput. Comput. Assist. Interv. 14, 25–32 (2011).
  44. R. Goldenberg, D. Eilot, G. Begelman, E. Walach, E. Ben-Ishai, and N. Peled, “Computer-aided simple triage (CAST) for coronary CT angiography (CCTA),” Int. J. Comput. Assist. Radiol. Surg. 7(6), 819–827 (2012).
  45. T. S. Kristensen, K. F. Kofoed, J. T. Kuhl, W. B. Nielsen, M. B. Nielsen, and H. Kelbaek, “Prognostic implications of nonobstructive coronary plaques in patients with non-ST-segment elevation myocardial infarction: A multidetector computed tomography study,” J. Am. Coll. Cardiol. 58, 502–509 (2011).
  46. G. W. Stone, A. Maehara, A. J. Lansky, B. de Bruyne, E. Cristea, G. S. Mintz, R. Mehran, J. McPherson, N. Farhat, S. P. Marso, H. Parise, B. Templin, R. White, Z. Zhang, and P. W. Serruys, “A prospective natural-history study of coronary atherosclerosis,” N. Engl. J. Med. 364, 226–235 (2011).
  47. D. Kang, P. J. Slomka, R. Nakazato, V. Y. Cheng, J. K. Min, D. Li, D. S. Berman, C.-C. J. Kuo, and D. Dey, “Automatic detection of significant and subtle arterial lesions from Coronary CT Angiography,” in Proc. SPIE Medical Imaging, Vol. 8314, Medical Imaging 2012: Image Processing, pp. 831435.
  48. D. Dey, V. Y. Cheng, P. J. Slomka, R. Nakazato, A. Ramesh, S. Gurudevan, G. Germano, and D. S. Berman, “Automated 3-dimensional quantification of noncalcified and calcified coronary plaque from coronary CT angiography,” J. Cardiovasc. Comput. Tomogr. 3, 372–382 (2009).
  49. D. Dey, T. Schepis, M. Marwan, P. J. Slomka, D. S. Berman, and S. Achenbach, “Automated three-dimensional quantification of noncalcified coronary plaque from coronary CT angiography: Comparison with intravascular US,” Radiology 257, 516–522 (2010).
  50. G. Yang, P. Kitslaar, M. Frenay, A. Broersen, M. J. Boogers, J. J. Bax, J. H. Reiber, and J. Dijkstra, “Automatic centerline extraction of coronary arteries in coronary computed tomographic angiography,” in International Journal of Cardiovascular Imaging (Springer, Netherlands, 2011), pp. 1–13.
  51. F. Cademartiri, L. La Grutta, G. Runza, A. Palumbo, E. Maffei, N. R. Mollet, T. V. Bartolotta, P. Somers, M. Knaapen, S. Verheye, M. Midiri, R. Hamers, and N. Bruining, “Influence of convolution filtering on coronary plaque attenuation values: Observations in an ex vivo model of multislice computed tomography coronary angiography,” Eur. Radiol. 17, 1842–1849 (2007).
  52. S. Achenbach, “Cardiac CT: State of the art for the detection of coronary arterial stenosis,” J. Cardiovasc. Comput. Tomogr. 1, 3–20 (2007). [MEDLINE]
  53. T.-C. Lee, R. L. Kashyap, and C.-N. Chu, “Building skeleton models via 3-D medial surface/axis thinning algorithms,” Graph. Models Image Process. 56, 462–478 (1994). [Inspec] [ISI]
  54. E. W. Dijkstra, “A note on two problems in connexion with graphs,” Numer. Math. 1, 269–271 (1959).
  55. F. Cademartiri, N. R. Mollet, G. Runza, N. Bruining, R. Hamers, P. Somers, M. Knaapen, S. Verheye, M. Midiri, G. P. Krestin, and P. J. de Feyter, “Influence of intracoronary attenuation on coronary plaque measurements using multislice computed tomography: Observations in an ex vivo model of coronary computed tomography angiography,” Eur. Radiol. 15, 1426–1431 (2005).
  56. S. S. Halliburton, P. Schoenhagen, A. Nair, A. Stillman, M. Lieber, E. M. Tuzcu, D. G. Vince, and R. D. White, “Contrast enhancement of coronary atherosclerotic plaque: A high-resolution, multidetector-row computed tomography study of pressure-perfused, human ex-vivo coronary arteries,” Coron Artery Dis. 17, 553–560 (2006).
  57. S. Achenbach, K. Boehmer, T. Pflederer, D. Ropers, M. Seltmann, M. Lell, K. Anders, A. Kuettner, M. Uder, W. G. Daniel, and M. Marwan, “Influence of slice thickness and reconstruction kernel on the computed tomographic attenuation of coronary atherosclerotic plaque,” J. Cardiovasc. Comput. Tomogr. 4, 110–115 (2010).
  58. S. Achenbach, “Computed tomography coronary angiography,” J. Am. Coll. Cardiol. 48, 1919–1928 (2006). [MEDLINE]
  59. V. Cheng, A. Gutstein, A. Wolak, Y. Suzuki, D. Dey, H. Gransar, L. E. Thomson, S. W. Hayes, J. D. Friedman, and D. S. Berman, “Moving beyond binary grading of coronary arterial stenoses on coronary computed tomographic angiography: Insights for the imager and referring clinician,” JACC: Cardiovasc. Imaging 1, 460–471 (2008).
  60. J. T. Dodge Jr., B. G. Brown, E. L. Bolson, and H. T. Dodge, “Intrathoracic spatial location of specified coronary segments on the normal human heart. Applications in quantitative arteriography, assessment of regional risk and contraction, and anatomic display,” Circulation 78, 1167–1180 (1988). [MEDLINE]
  61. G. L. Raff, A. Abidov, S. Achenbach, D. S. Berman, L. M. Boxt, M. J. Budoff, V. Cheng, T. DeFrance, J. C. Hellinger, and R. P. Karlsberg, “SCCT guidelines for the interpretation and reporting of coronary computed tomographic angiography,” J. Cardiovasc. Comput. Tomogr. 3, 122–136 (2009).
  62. “Analyse-it for Microsoft Excel (version 2.20),” Analyse-it Software, Ltd., see (2009).
  63. P. A. Devijver and J. Kittler, Pattern Recognition: A Statistical Approach (Prentice-Hall, 1982).
  64. J. H. Reiber, P. W. Serruys, C. J. Kooijman, W. Wijns, C. J. Slager, J. J. Gerbrands, J. C. Schuurbiers, A. den Boer, and P. G. Hugenholtz, “Assessment of short-, medium-, and long-term variations in arterial dimensions from computer-assisted quantitation of coronary cineangiograms,” Circulation 71, 280–288 (1985). [ISI] [MEDLINE]
  65. A. S. Agatston, W. R. Janowitz, F. J. Hildner, N. R. Zusmer, M. Viamonte Jr., and R. Detrano, “Quantification of coronary artery calcium using ultrafast computed tomography,” J. Am. Coll. Cardiol. 15, 827–832 (1990). [ISI] [MEDLINE]
  66. T. Q. Callister, B. Cooil, S. P. Raya, N. J. Lippolis, D. J. Russo, and P. Raggi, “Coronary artery disease: Improved reproducibility of calcium scoring with an electron-beam CT volumetric method,” Radiology 208(3), 807–814 (1998). [MEDLINE]