The detection of stenotic plaques strongly depends on the quality of the coronary arterial tree imaged with coronary CT angiography (cCTA). However, it is time consuming for the radiologist to select the best-quality vessels from the multiple-phase cCTA for interpretation in clinical practice. We are developing an automated method for selection of the best-quality vessels from coronary arterial trees in multiple-phase cCTA to facilitate radiologist’s reading or computerized analysis. Our automated method consists of vessel segmentation, vessel registration, corresponding vessel branch matching, vessel quality measure (VQM) estimation, and automatic selection of best branches based on VQM. For every branch, the VQM was calculated as the average radial gradient. An observer preference study was conducted to visually compare the quality of the selected vessels. 167 corresponding branch pairs were evaluated by two radiologists. The agreement between the first radiologist and the automated selection was 76% with kappa of 0.49. The agreement between the second radiologist and the automated selection was also 76% with kappa of 0.45. The agreement between the two radiologists was 81% with kappa of 0.57. The observer preference study demonstrated the feasibility of the proposed automated method for the selection of the best-quality vessels from multiple cCTA phases.
Coronary CT angiography (cCTA) is a useful noninvasive modality for imaging of the heart and evaluation of the extent of plaques. However, due to the coordinated motion of the heart chambers, different arterial segments may be blurred at different phases of the cardiac cycle [
A number of studies reported methods of automatic selection of the best-quality phase of the entire coronary arterial tree or motion correction during reconstruction of cCTA or 3D angiograms. Rasche et al. [
The above studies focused on selection of the best-quality cCTA phase, which inevitably have to make compromise among the individual arteries because there may not be a single phase in which all arteries attain their best quality. Lessick et al. [
The multistage framework for automated best-quality vessel selection in cCTA is shown in Figure
Multistage framework for selection of the best-quality phase for individual arterial branches in multiphase cCTA.
A cCTA examination acquired with ECG-gating and reconstructed at 6 cardiac phases (e.g., 80%, 75%, 70%, 50%, 45%, and 40%) is shown in Figure
cCTA scans acquired with ECG-gating and reconstructed at multiple cardiac phases: (a) 3D rendered cCTA volumes for 6 acquired cardiac phases (80%, 75%, 70%, 50%, 45%, and 40%). Left (LCA) and right (RCA) coronary arterial tree in phase 50% are marked by white arrows. (b) Segmented LCA and RCA tree in each phase (80%, 75%, 70%, 50%, 45%, and 40%) using multiscale enhancement and dynamic balloon tracking method [
LCA trees from Figure
To identify the coronary tree branches, the first stage is the determination of the corresponding branching points. The branching points of the coronary trees in every phase are automatically detected (Figure
Automatically detected vessel branching points (in red), which are propagated to all phases.
For a branching point in a given phase, the above process will generate multiple potential branching point candidates in its proximity, due to inaccuracies in the detection of branching point locations in the different phases, as well as the inaccuracies in the coronary tree registration. Therefore, we developed a method to automatically identify the most likely branching point in a given phase and then identify the corresponding branching points in all phases (Figure
LCA tree: automatically identified corresponding vessel branches (shown by matching colors) in six phases. Note that some colors are repeatedly used for different branches because of the limited number of colors available. The different branches of the same color can be distinguished by locations.
In the next step, the corresponding branches of the same vessel appearing in different phases are identified. The correspondence between the branches is established by using the following criteria:
(
(
The process is repeated for all phase pairs to determine the correspondence of all branches (Figure
Vessel straightening. Curved planar reformation was used to straighten each of the branches.
In this preliminary study, we defined a simple vessel quality measure (VQM) to automatically estimate the vessel quality. First, the gradients
A straightened branch from the LCA tree in Figure
The distance between the point inside and the point outside the vessel in the denominator of the gradient calculation is not explicitly included in (
We conducted an observer preference study to visually compare the relative quality of the vessels and compared with the automatic ranking with VQM. Because of the large number of possible vessel pairs that can be formed by exhaustive pairing of the corresponding vessel branches from multiple phases, to limit the radiologists’ effort required for reading, we used a single pair of each vessel branch in our data set for the observer study. First, a pair of the best- and worst-quality branches among the available phases for each branch was automatically identified using the VQM (Figure
A pair of best- and worst-quality vessel branches selected automatically from the corresponding arterial branches in Figure
With Institutional Review Board (IRB) approval, cCTA examinations for seven patients (2 men and 5 women; age range, 31–65 years; mean age, 49.4 years) were collected retrospectively from the patient files at the University of Michigan Health System and used in this preliminary study. The cCTA cases were acquired with a clinical protocol in which an isoosmolar nonionic contrast medium (Visipaque; GE Healthcare) was administered using an 18-gauge cannula in an upper extremity vein. A test bolus of 15 to 20 mL at the rate of 4 to 5 mL/s was administered with sequential scanning every 2 seconds at the level of the left main coronary artery, with a region of interest placed in the aortic root, to determine the optimum scan delay for each patient. For the coronary CT angiograms 80 mL of contrast medium was injected (60 mL at 5 mL/s and 20 mL at 3.5 mL/s) followed by a saline chase bolus of 50 mL at 5 mL/s. The mean and standard deviation heart rate of the patients were 62.7 ± 5.2 bpm.
The cCTA images were acquired by helical retrospective gating at 120 kVp and 440–800 mA with GE multidetector CT scanners (GE Healthcare Lightspeed VCT (6 patients) and Discovery CT750 HD (1 patient)). The cCTA image slices were reconstructed at 0.625 mm slice interval and 0.488 mm in-plane pixel size. The clinical protocol of cCTA in our department reconstructed 6 phases in a range from 40% to 87%. While more phases can be reconstructed from the data, 6 phases are used in the clinical protocol at our health system because they are clinically sufficient to cover a broad range of heart rates. This strategy provides optimal yield and a balance between diagnostic accuracy and efficiency. Routine reading of more than 6 phases is clinically impractical and overly burdens radiologists’ workload. In our study, we tried to emulate clinical practice, where only select phases are generated for assessment but our method is applicable to more or fewer phases.
The data set therefore contained a total of 42 cCTA volumes (7 patients with 6 phase cCTA scans each) with 84 coronary arterial trees (42 LCA trees and 42 RCA trees). After automatic registration and identification, 167 groups of corresponding branches were established (102 LCA and 65 RCA). The VQM of 833 branches in 6 phases were calculated (531 LCA branches and 302 RCA branches). Note that not all vessel branches had a complete set of 6 phases because some vessels might be lost at segmentation and tracking due to poor image quality. For each group of corresponding branches, two branches, the branch with the highest VQM and the branch with the lowest VQM, were selected based on the VQM, as described in the previous section. This resulted in 167 branch image pairs (102 pairs from LCAs and 65 pairs from RCAs). Detailed information for the number of established branch image pairs for each case is given in Table
Corresponding branch image pairs for the 7 cases.
Case # | Total | LCA | RCA |
---|---|---|---|
1 | 31 | 17 | 14 |
2 | 14 | 11 | 3 |
3 | 25 | 12 | 13 |
4 | 38 | 23 | 15 |
5 | 18 | 13 | 5 |
6 | 26 | 16 | 10 |
7 | 15 | 10 | 5 |
| |||
Total | 167 | 102 | 65 |
The performance of the automatic selection using VQM was evaluated by the following two methods: Estimation of the percentage of the total number of vessel pairs for which the automatic selection agreed with the radiologist’s selection of the higher quality branch in the pair. Cohen’s kappa statistics to estimate the agreement between the automatic selection and the radiologist’s selection of the higher quality branch in the pair.
For comparison, the agreement between the two radiologists was also evaluated with the two methods.
The overall agreement between radiologist 1 and the automated selection of the best-quality branches was 76% for the
Agreement between radiologist 1 and the automated selection using VQM of the best-quality branch in the corresponding vessel pairs, between radiologist 2 and the automated selection, and between the two radiologists.
Number of branches | % agreement | |||
---|---|---|---|---|
Rad 1-computer | Rad 2-computer | Rad 1-Rad 2 | ||
Case 1 | 31 | 90% | 90% | 94% |
Case 2 | 14 | 93% | 93% | 86% |
Case 3 | 25 | 84% | 76% | 84% |
Case 4 | 38 | 79% | 74% | 82% |
Case 5 | 18 | 61% | 67% | 83% |
Case 6 | 26 | 62% | 77% | 62% |
Case 7 | 15 | 53% | 47% | 80% |
| ||||
Overall | 167 | 76% | 76% | 81% |
The average kappa for the agreement between radiologist 1 and the automated selection was 0.49 (range: 0.04 to 0.87), which corresponds to a moderate agreement, based on the commonly used scale [
Cohen’s kappa statistics estimation of the agreement between radiologist 1 and the automated selection using VQM of the best-quality branch in the corresponding vessel pairs, between radiologist 2 and the automated selection, and between the two radiologists.
Number of branches | Kappa | |||
---|---|---|---|---|
Rad 1-computer | Rad 2-computer | Rad 1-Rad 2 | ||
Case 1 | 31 | 0.87 | 0.80 | 0.87 |
Case 2 | 14 | 0.81 | 0.81 | 0.65 |
Case 3 | 25 | 0.68 | 0.51 | 0.68 |
Case 4 | 38 | 0.56 | 0.50 | 0.61 |
Case 5 | 18 | 0.24 | 0.33 | 0.67 |
Case 6 | 26 | 0.21 | 0.32 | 0.18 |
Case 7 | 15 | 0.04 | −0.13 | 0.33 |
| ||||
Overall | 167 | 0.49 | 0.45 | 0.57 |
Agreement between radiologist 1 and the automated selection using VQM of the best-quality branch in the corresponding vessel pairs, between radiologist 2 and the automated selection, and between the two radiologists estimated in terms of the commonly used categories based on Cohen’s kappa statistics [
Kappa | |||
| |||
Rad 1-computer | Rad 2-computer | Rad 1-Rad 2 | |
| |||
Case 1 | Almost perfect | Substantial | Almost perfect |
Case 2 | Almost perfect | Almost perfect | Substantial |
Case 3 | Substantial | Moderate | Substantial |
Case 4 | Moderate | Moderate | Substantial |
Case 5 | Fair | Fair | Substantial |
Case 6 | Fair | Fair | Slight |
Case 7 | Slight | Less than chance | Fair |
| |||
Overall | Moderate | Moderate | Moderate |
Agreement categories based on Cohen’s kappa statistics [
Kappa < 0: less than chance agreement.
Kappa 0.01–0.20: slight agreement.
Kappa 0.21–0.40: fair agreement.
Kappa 0.41–0.60: moderate agreement.
Kappa 0.61–0.80: substantial agreement.
Kappa 0.81–0.99: almost perfect agreement.
Figures
Automatically selected pair of best- and worst-quality vessel branches from the available corresponding arterial branches in six phases using VQM and radiologists’ preferences. The automated selection matched the radiologists’ preferences (branch (a)).
Automatically selected pair of best- and worst-quality vessel branches from the available corresponding arterial branches in six phases using VQM and radiologists’ preferences. The automated selection matched the radiologists’ preferences (branch (a)).
Automatically selected pair of best- and worst-quality vessel branches from the available corresponding arterial branches in six phases using VQM and radiologists’ preferences. The automated selection (branch (a)) did not match the radiologists’ preferences (branch (b)).
Automatically selected pair of best- and worst-quality vessel branches from the available corresponding arterial branches in six phases using VQM and radiologists’ preferences. The automated selection (branch (a)) matched the preference of radiologist 1 but did not match the preference of radiologist 2 (branch (b)).
A pair of best- and worst-quality vessel branches selected automatically from the corresponding arterial branches in six phases using VQM and the radiologists’ preferences. The automated selection (branch (a)) and the radiologists’ preferences (branch (b)) did not match in this case.
In this study, we used cCTA cases with 6 phases for evaluation of the best-quality vessel selection method and the cases were acquired with retrospective gating techniques. Prospective gating, if applicable, is the current state of the art that provides sufficient quality scans with less radiation. However, prospective gating techniques can be effective only in appropriately selected cases when the heart rate is stable with low beat-to-beat variations and is below about 65 bpm. This allows acquisition in a selected phase of the cardiac cycle. If these conditions are not met, retrospective gating is still used clinically to generate cCTA of multiple cardiac phases. In addition, even with prospective gating, more than one phase may be reconstructed, depending on the amount of padding used. Our proposed method can be applied to cCTA examinations with more than one phase and should be independent of whether the multiple phases are obtained by retrospective or prospective gating techniques.
The image pairs of three branches of different quality are shown in Figures
Automatically selected pair of best- and worst-quality vessel branches from the available corresponding arterial branches in six phases using VQM and radiologists’ preferences. The automated selection matched the radiologists’ preferences (branch (a)).
Automatically selected pair of best- and worst-quality vessel branches from the available corresponding arterial branches in six phases using VQM and radiologists’ preferences. The automated selection matched the radiologists’ preferences (branch (a)).
The agreement between the radiologists for Cases
The low kappa for the agreement between the radiologists for a case, such as Case
The average agreements between the automated selection and radiologist 1 and between the automated selection and radiologist 2 were very close in terms of both percent agreement and kappa. The average agreement between the two radiologists was slightly higher than the average agreement between the automated selection and either radiologist.
In this preliminary study, we used only the average radial gradient along the vessel branch as the vessel quality measure. The average radial gradient estimates the sharpness of the vessel wall. It can be expected that other measures such as the contrast and smoothness of the vessels may also be useful as descriptors for the quality of the vessels. We will further develop the VQM to improve the accuracy of the automated ranking of the corresponding branches from multiple phases.
One limitation of the study is the small number of cases. Although this pilot study did demonstrate the feasibility of our approach, a larger data set and more observers have to be used in future studies to further develop and validate the robustness of the methods. A second limitation is that, in clinical practice, cCTA interpretation is not based on one view and all display formats (straightened multiplanar reformation (MPR)s, curved MPRs, and original axial data) are available for diagnosis. However, in our observer study for assessment of the performance of the automatic vessel selection method, the observers were provided with the straightened MPR display to visually judge the vessel quality based on information similar to that used by the computer. This experimental design reduced the reading time to a more practical level for the pilot study and also focused the comparison on the vessel selection stage based on the VQM rather than the accuracy of the entire process including vessel segmentation and straightening. It may also be noted that should the proposed method become practical for clinical use, it would only be used for initial selection of the best phase for a given vessel segment; all the possible (or the preferred) formats of the vessel segment at the selected phase could be displayed automatically so that the radiologist could make use of all diagnostic information as desired. A third limitation is that we did not include all possible branch pairs from all phases in the observer study because the total number of pairings to be evaluated would be over 2000, which would impose excessive demand on the radiologists’ effort. Since the comparison of the best- and worst-quality phases was an easier task, the observer study could not reveal the performance of the automatic ranking among all phases. Nevertheless, the observer study did show a correlation of the magnitude of the difference in VQM with the visual similarity in the vessel quality and that small differences in VQM would indicate very similar quality so that the choice of one or the other might not be as critical. We will further improve the methods and perform more extensive validation studies in the future.
In this study we proposed a method for automatic selection of the best-quality vessels from multiple cCTA phases. The method utilizes a number of image analysis techniques specifically designed for cCTA, including coronary arterial tree segmentation and registration, identification of coronary tree branches and their correspondence among the multiple phases, and assessment of vessel quality by a quantitative measure that guides the selection of the best-quality phase for each vessel from multiple cCTA phases. An observer study with two cardiothoracic radiologists as observers was conducted to evaluate the proposed method. The results demonstrate that the automatic method agreed well with the radiologists’ selections and thus the feasibility of the approach. The best-quality arterial segments constitute the building blocks for a virtual best-quality composite coronary arterial tree. The automated selection of the best-quality arterial segments from all available phases of the cCTA is expected to improve the efficiency and facilitate the detection of plaques by either the radiologist or a CAD system.
The authors declare that they have no competing interests.
This work is supported by National Institutes of Health Grant no. R01 HL106545.