Coronary Artery Stent Evaluation by CTA: Impact of Deep Learning Reconstruction and Subtraction Technique
Abstract
To listen to the podcast associated with this article, please select one of the following: iTunes, Google Play, or direct download.
BACKGROUND. Coronary CTA with hybrid iterative reconstruction (HIR) is prone to false-positive results for in-stent restenosis due to stent-related blooming artifact.
OBJECTIVE. The purpose of this study is to assess the impact of deep learning reconstruction (DLR), subtraction images, and the combination of DLR and subtraction images on the diagnostic performance of coronary CTA for the detection of in-stent restenosis.
METHODS. This prospective study included patients with coronary stents who underwent coronary CTA between March 2020 and August 2021. CTA used a technique with two breath-holds (noncontrast and contrast-enhanced acquisitions). Conventional and subtraction images were reconstructed for HIR and DLR. The maximum visible instent lumen diameter was measured. Two readers independently evaluated images for in-stent restenosis (≥ 50% stenosis). A simulated assessment of combined conventional and subtraction images was generated, reflecting assessment of conventional and subtraction images in the presence or absence of severe misregistration artifact, respectively. Invasive angiography served as reference standard.
RESULTS. The study enrolled 30 patients (22 men and eight women; mean age, 63.6 ± 7.4 [SD] years) with a total of 59 stents; severe misregistration artifact was present for 32 stents. Maximum visible in-stent lumen diameter was higher for DLR than for HIR (2.3 ± 0.5 vs 2.1 ± 0.5 mm, p < .001), and among stents without severe misregistration artifact, it was higher for subtraction than conventional DLR (3.0 ± 0.5 vs 2.4 ± 0.5, p < .001). Among conventional CTA with HIR, conventional CTA with DLR, combination (conventional and subtraction) approach with HIR, and combination (conventional and subtraction) approach with DLR, the highest patient-level diagnostic performance measures were as follows: for reader 1, sensitivity was identical (62.5%), specificity was highest for combination with DLR (90.1%), PPV was highest for combination with DLR (71.4%), NPV was highest for combination with DLR (87.0%), and accuracy was highest for combination with DLR (83.3%); for reader 2, sensitivity was identical (50.0%), specificity was highest for combination with HIR or DLR (both 95.5%), PPV was highest for combination with HIR or DLR (both 80.0%), NPV was highest for combination with HIR or DLR (84.0%), and accuracy was highest for combination with HIR or DLR (both 83.3%).
CONCLUSION. The combined DLR and subtraction technique yielded optimal diagnostic performance for detecting in-stent restenosis by coronary CTA.
CLINICAL IMPACT. The described technique could guide patient selection for invasive coronary stent evaluation.
HIGHLIGHTS
Key Finding
▪ Coronary CTA using DLR and a two-breath-hold subtraction technique, with a combined (conventional and subtraction images) interpretation, yielded PPV, NPV, and accuracy for in-stent restenosis of 73.3%, 93.2%, and 88.1% for reader 1 and 87.5%, 86.3%, and 86.4% for reader 2, respectively.
Importance
▪ The combination of DLR and subtraction technique may help overcome challenges related to stent-related blooming artifact when evaluating patients with coronary stents by CTA.
In-stent restenosis is an important potential complication after coronary artery stent implantation, associated with recurrent angina and possibly acute coronary syndrome [1]. Coronary CTA is commonly used for noninvasive evaluation for restenosis [2, 3]. The use of iterative reconstruction (IR) has substantially improved the diagnostic performance of coronary CTA for detecting restenosis while lowering radiation exposure [4–6]. Nonetheless, the diagnostic performance of coronary CTA for detection of restenosis remains suboptimal, in part because of false-positive results for restenosis due to stent-related blooming artifact [6–8].
Deep learning reconstruction (DLR) based on a convolutional neural network has the potential to overcome the limitations of CT image reconstruction by use of traditional IR. Studies have shown that DLR that uses low-dose hybrid IR (HIR) images as training input and high-dose model-based IR images as a training target for learning can produce high-quality images that are free of noise contamination [9, 10]. For coronary CTA, DLR has been shown to improve edge delineation and lower image noise [9, 11]. Subtraction images obtained from coronary CTA examinations have also shown the potential to reduce blooming artifact in patients with severe coronary calcification or coronary artery stents [12–15], and they potentially could be obtained in combination with DLR to further improve the evaluation for restenosis.
To our knowledge, the diagnostic performance of DLR for the evaluation of the patency of coronary artery stents has not been investigated. Therefore, the purpose of this study was to assess the impact of DLR, subtraction images, and the combination of DLR and subtraction images on the image quality and diagnostic performance of coronary CTA for the detection of in-stent restenosis.
Methods
Patient Population and Study Protocol
The protocol for this prospective study was reviewed and approved by the ethics committee at Peking Union Medical College Hospital, Chinese Academy of Medical Sciences, and Peking Union Medical College. Written informed consent was obtained from all patients.
Patients with a coronary artery stent who were referred by a cardiologist to undergo coronary CTA between March 2020 and August 2021 were eligible for study participation. Study participants underwent coronary CTA using a two-breath-hold subtraction protocol rather than a standard protocol. All patients who were eligible during the study period were approached for potential study enrollment. Enrolled patients were subsequently excluded if they were unable to successfully sustain a breath-hold at the time that coronary CTA was performed using the subtraction protocol or if they did not undergo standard-of-care invasive coronary angiography within 4 weeks after CTA to serve as a reference standard. For patients included in the final analysis, recorded stent properties included stent location within the coronary artery (proximal, mid, or distal), stent diameter (2.25, 2.5, 2.75, 3.0, or 3.5 mm or unknown), and stent type (bare metal, drug eluting, or unknown).
Coronary CTA Acquisition
All coronary CTA acquisitions were obtained using a 320-MDCT scanner (Aquilion ONE Genesis Edition, Canon Medical Systems). Patients with a heart rate of 70 beats/min or higher were administered β-blockers (metoprolol tartrate in 25- to 50-mg tablets [Betaloc, AstraZeneca]) before the examination. The two-breath-hold subtraction protocol [16] included a noncontrast acquisition followed by a contrast-enhanced acquisition, obtained during separate breath-holds; both acquisitions were performed with prospective ECG triggering. Sublingual nitroglycerin (0.5–1.0 mg) was administered 1–2 minutes before the start of the examination. Iodinated contrast media (iopamidol, 370 mg I/mL [Isovue, Bracco Sine Pharmaceutical]) was administered in the antecubital vein through a 20-gauge trocar by use of a dual-syringe power injector (Dual Shot GX, Nemoto Kyorindo). A weight-based volume of contrast agent (body weight [expressed as kilograms] × 0.053 mL/s) was injected over a fixed duration of 10 seconds and was followed by 30 mL of saline administered at the same rate. Images were acquired from the carina to the level of the diaphragm, thus including the entire heart. A bolus-tracking technique was used with a trigger threshold of 280 HU in the ascending aorta. Other acquisition parameters were as follows: tube voltage, 100 kVp; collimation, 320 × 0.5 mm; gantry rotation time, 275 ms; and z coverage, 120–160 mm. The tube current was adjusted automatically with a noise index (SD) of 33.
CT Image Reconstruction
The noncontrast and contrast-enhanced coronary CTA image acquisitions were reconstructed using HIR (Adaptive Iterative Dose Reduction 3D [AIDR 3D], FC09 cardiac kernel, Canon Medical Systems) and DLR (Advanced Intelligent Clear-IQ Engine [AiCE], cardiac kernel, Canon Medical Systems) with a slice thickness of 0.5 mm and an interval of 0.25 mm. Corresponding subtraction images were generated using postprocessing software (SURESubtraction, Canon Medical Systems). The CTDIvol and dose-length product (DLP) were recorded. The effective dose was calculated by multiplying the DLP by a cardiac-specific conversion factor (κ = 0.026) [17]. All images were transferred to a dedicated workstation (syngo.via VB10, Siemens Healthcare) for analysis.
Image Evaluation
Two cardiothoracic radiologists (C.X. and Y.Y., with 5 and 8 years of posttraining experience), blinded to the reconstruction approach (HIR or DLR) and the results of invasive coronary angiography, reviewed images to perform subjective assessments of image quality, diagnostic confidence, and stent patency. The examinations were reviewed in two sessions, which were separated by 1 month. In each session, for each patient, the radiologists were presented for review either the conventional HIR or conventional DLR images and either the subtraction HIR or subtraction DLR images, both of which were selected at random and were reviewed in random order within the session. Within a session, the radiologists were not informed as to which conventional image set and which subtraction image set corresponded with the same patient. Two weeks before the start of the first session, the two radiologists reviewed all image sets for all patients in consensus, solely to assess for the presence of severe misregistration artifact on subtraction images. These examinations were included in the analysis of conventional HIR and DLR images but were excluded from the analysis of subtraction HIR and DLR images.
The readers assigned each conventional image set an image quality score using a 4-point Likert scale [18] (with 1 indicating nondiagnostic, poor image quality due to severe artifacts; 2, adequate, reduced image quality due to moderate artifacts; 3, good, minor artifacts; and 4, excellent, no artifacts). The subtraction image sets were assigned a score as follows: 1 indicated that the images were nondiagnostic due to severe misregistration artifact precluding evaluation of the stent lumen; 2, adequate, stent moderately subtracted, sufficient image quality to exclude restenosis; 3, good, stent subtracted, diagnostic image quality; and 4, excellent, stent totally subtracted). The readers also recorded diagnostic confidence using a 4-point scale, which ranged from 1 (indicating lowest confidence) to 4 (denoting highest confidence). Finally, the readers classified the stent as showing no stenosis, less than 50% stenosis (i.e., intimal hyperplasia), or stenosis of 50% or more (i.e., in-stent restenosis).
One of the previously noted radiologists (C.X.) measured objective parameters (maximum visible in-stent lumen diameter, stent lumen attenuation increase ratio [SAIR], SNR, and CNR) using a dedicated workstation (syngo.via VB10, Siemens Healthcare). To measure the maximum visible in-stent lumen diameter, the radiologist manually traced the margin of the vessel lumen on a single cross-sectional slice through the stent, and the software then automatically determined the maximum diameter. The mean attenuation and SD of attenuation (representing image noise) were measured inside the stent and within the adjacent adipose tissue. The mean attenuation of the coronary arterial lumen was measured proximal and distal to the stent, and the mean value of the proximal and distal measurements was used to indicate coronary lumen. ROIs were drawn as large as possible, with the stent or lumen edges avoided. Based on these values, SAIR, SNR, and CNR [19] were calculated using the following formulas:
and
For SAIR, lower values indicated better quality; for remaining objective measures, higher values indicated better quality.
Invasive Coronary Angiography
Invasive coronary angiography was performed by one of six cardiologists (all nonauthors, with 6–30 years of posttraining experience) on an Allura Xper UNIQ FD10 system (Philips Health-care). During angiography, images were acquired in multiple projections, including at least two orthogonal projections of the target vessels. For the purposes of this investigation, two of the cardiologists, who were blinded to the coronary CTA results, evaluated each stent on the angiographic images for presence of in-stent restenosis, defined as luminal stenosis of 50% or more.
Statistical Analysis
Quantitative parameters were expressed as mean ± SD, and qualitative parameters were expressed as frequency and percentage. Interobserver agreement on subjective image quality, diagnostic confidence, and stent patency (which were all handled in an ordinal fashion for the purposes of assessing interob-server agreement) was evaluated by kappa coefficients (with κ ≤ 0.40 denoting poor agreement; 0.41–0.60, moderate; 0.61–0.80, good; > 0.80, excellent). Measures were statistically compared between conventional HIR and DLR as well as between conventional DLR and subtraction DLR. In the subset of patients without severe misregistration artifact, measures were summarized for all combinations of acquisition and reconstruction techniques, although not all pairwise combinations were statically compared. The Shapiro-Wilk test was used to assess the normality of quantitative data. Normally distributed variables were compared using a paired t test, and nonnormally distributed variables were compared using a Wilcoxon test. The sensitivity, specificity, NPV, PPV, and accuracy of the reader interpretations were compared between methods by bootstrapping, with findings from invasive angiography serving as the reference standard. To assess diagnostic performance, reader assessments were classified in a binary fashion as positive (in-stent restenosis) or negative (patent or intimal hyperplasia). A simulated assessment of the combination of conventional coronary CTA and subtraction images was generated, reflecting the interpretation of subtraction CTA in the absence of severe misregistration artifact and reflecting the interpretation of conventional CTA in the presence of severe misregistration artifacts. The diagnostic performance of this simulated combined interpretation was compared between HIR and DLR in all stents. Measures were also computed when stratified by stent diameter (< 3 mm vs ≥ 3 mm) and by stent location (proximal, middle, or distal coronary artery). Finally, a patient-level analysis of diagnostic performance was performed in which patients were classified as having true-positive findings as long as at least one stent with restenosis based on the reference standard was correctly classified by the reader as showing restenosis on CTA.
A p value less than .05 was considered statistically significant. Statistical analyses were performed using R software (version 3.6.1).
Results
Patient Sample
A total of 185 patients with a coronary stent were referred for coronary CTA during the study period. A total of 49 patients declined participation, and 136 patients were enrolled. Of the patients who were enrolled, four were excluded due to inability to sustain a breath-hold for the purposes of undergoing a two-breath-hold coronary CTA, and 102 were excluded due to not having undergone invasive coronary angiography within 4 weeks after CTA. These exclusions resulted in a final study sample of 30 patients (22 men and eight women; mean age, 63.6 ± 7.4 years) with 59 coronary stents. The two-breath-hold coronary CTA examination had a mean CTDIvol of 39.2 ± 7.9 mGy, mean DLP of 163.4 ± 85.1 mGy × cm, and mean effective radiation dose of 4.2 ± 2.2 mSv. Figure 1 presents a flowchart showing patient inclusion. Patient and stent characteristics are summarized in Table 1.
| Characteristic | Value |
|---|---|
| Patient | |
| Total no. of patients | 30 |
| Age (y), mean ± SD | 63.6 ± 7.4 |
| Sex | |
| Male | 22 |
| Female | 8 |
| BMI, mean ± SD | 27.0 ± 3.6 |
| Diabetes | 16 |
| Hypertension | 23 |
| Hypercholesterolemia | 16 |
| Smoking | 21 |
| Family history of CAD | 9 |
| Heart rate (beats/min), mean ± SD | 60.1 ± 5.2 |
| Stent | |
| Total no. of stents | 59 |
| Locationa | |
| LM | 2 |
| LAD (proximal, middle, distal) | 30 (14, 15, 1) |
| D1 | 3 |
| LCX (proximal, middle, distal) | 10 (5, 3, 2) |
| OM | 1 |
| RCA (proximal, middle, distal) | 13 (2, 6, 5) |
| Diameter (mm) | |
| 2.25 | 6 |
| 2.5 | 6 |
| 2.75 | 8 |
| 3.0 | 15 |
| 3.5 | 11 |
| Unknown | 13 |
| Type | |
| Bare metal | 5 |
| Drug eluting | 41 |
| Unknown | 13 |
Note—Except where otherwise indicated, data are number of patients or stents. CAD = coronary artery disease, LM = left main artery, LAD = left anterior descending, D1 = diagonal 1, LCX = left circumflex, OM = obtuse marginal, RCA = right coronary artery.
a
All stents in LM, D1, and OM were classified as proximal. LAD, LCX, and RCA divided into three equal parts with respect to vessel length.
Interobserver Agreement
Interobserver agreement for subjective image quality was moderate to good for conventional CTA with DLR (κ = 0.75), conventional CTA with HIR (κ = 0.65), subtraction CTA with DLR (κ = 0.58), and subtraction CTA with HIR (κ = 0.69). Interobserver agreement for diagnostic confidence was moderate to excellent for conventional CTA with DLR (κ = 0.82), conventional CTA with HIR (κ = 0.48), subtraction CTA with DLR (κ = 0.56), and subtraction CTA with HIR (κ = 0.48). Interobserver agreement for stent patency was moderate to good for conventional CTA with DLR (κ = 0.79), conventional CTA with HIR (κ = 0.60), subtraction CTA with DLR (κ = 0.57), and subtraction CTA with HIR (κ = 0.73).
Image Quality
Tables 2 and 3 present subjective and objective measures of image quality. For conventional CTA, subjective image quality was significantly higher for DLR than for HIR for reader 1 (3.4 ± 0.7 vs 3.0 ± 0.6, p < .001) and reader 2 (3.2 ± 0.7 vs 2.7 ± 0.7, p < .001). Diagnostic confidence was significantly higher for DLR than for HIR for reader 1 (3.0 ± 0.7 vs 2.7 ± 0.6, p < .001) and reader 2 (2.9 ± 0.8 vs 2.3 ± 0.8, p < .001). The maximum visible in-stent lumen diameter was significantly higher for DLR than for HIR (2.3 ± 0.5 vs 2.1 ± 0.5 mm, p < .001). SAIR was not significantly different between DLR and HIR (0.77 ± 0.42 vs 0.81 ± 0.54, p = .73). SNR was not significantly different between DLR and HIR (29.0 ± 18.1 vs 29.8 ± 16.4, p = .58). CNR was significantly higher for DLR than for HIR (51.6 ± 24.3 vs 37.7 ± 18.8, p < .001).
| Measure | All Stents (n = 59) | Stents Without Severe Misregistration Artifact (n = 27) | ||||||
|---|---|---|---|---|---|---|---|---|
| Conv HIR | Conv DLR | p | Conv HIR | Conv DLR | Sub HIR | Sub DLR | p (Conv vs Sub DLR) | |
| Subjective image quality | ||||||||
| Reader 1 | 3.0 ± 0.6 | 3.4 ± 0.7 | < .001 | 3.2 ± 0.5 | 3.6 ± 0.5 | 2.9 ± 0.7 | 3.0 ± 0.7 | .001 |
| Reader 2 | 2.7 ± 0.7 | 3.2 ± 0.7 | < .001 | 2.9 ± 0.6 | 3.5 ± 0.6 | 2.7 ± 0.7 | 3.0 ± 0.7 | .007 |
| Diagnostic confidence | ||||||||
| Reader 1 | 2.7 ± 0.6 | 3.0 ± 0.7 | < .001 | 2.9 ± 0.6 | 3.2 ± 0.7 | 3.1 ± 0.6 | 3.0 ± 0.7 | .43 |
| Reader 2 | 2.3 ± 0.8 | 2.9 ± 0.8 | < .001 | 2.6 ± 0.7 | 3.1 ± 0.7 | 2.8 ± 0.8 | 3.2 ± 0.8 | .80 |
Note—Except where otherwise indicated, data are mean ± SD. HIR = hybrid iterative reconstruction, DLR = deep learning–based image reconstruction, conv = conventional, sub = subtraction.
| Measure | All Stents (n = 59) | Stents Without Severe Misregistration Artifact (n = 27) | ||||||
|---|---|---|---|---|---|---|---|---|
| Conv HIR | Conv DLR | p | Conv HIR | Conv DLR | Sub HIR | Sub DLR | p (Conv vs Sub DLR) | |
| Maximum visualized in-stent lumen diameter | 2.1 ± 0.5 | 2.3 ± 0.5 | < .001 | 2.2 ± 0.4 | 2.4 ± 0.5 | 2.9 ± 0.5 | 3.0 ± 0.5 | < .001 |
| SAIR | 0.8 ± 0.5 | 0.8 ± 0.4 | .73 | 0.7 ± 0.4 | 0.7 ± 0.3 | 0.0 ± 0.2 | 0.1 ± 0.2 | < .001 |
| SNR | 29.8 ± 16.4 | 29.0 ± 18.1 | .58 | 31.5 ± 16.4 | 30.4 ± 20.5 | 13.3 ± 7.1 | 18.0 ± 9.6 | < .001 |
| CNR | 37.7 ± 18.8 | 51.6 ± 24.3 | < .001 | 38.7 ± 17.9 | 49.1 ± 19.4 | 22.6 ± 15.5 | 33.6 ± 16.0 | < .001 |
Note—Except where otherwise indicated, data are mean ± SD. Conv = conventional, HIR = hybrid iterative reconstruction, DLR = deep learning–based image reconstruction, SAIR = stent-lumen attenuation increase ratio, sub = subtraction.
Severe misregistration artifact was observed for 32 of 59 (54.2%) stents. The remaining 27 of 59 (45.8%) stents were included in the analysis of subtraction images. For DLR, subjective image quality was significantly lower for subtraction than for conventional images for reader 1 (3.0 ± 0.7 vs 3.6 ± 0.5, p < .001) and reader 2 (3.0 ± 0.7 vs 3.5 ± 0.6, p < .001). For DLR, no significant differences of diagnostic confidence were found between subtraction and conventional images for reader 1 (3.0 ± 0.7 vs 3.2 ± 0.7, p = .43) and reader 2 (3.2 ± 0.8 vs 3.1± 0.7, p = .80). For DLR, maximum visible in-stent lumen diameter was significantly higher for subtraction than for conventional images (3.0 ± 0.5 vs 2.4 ± 0.5, p < .001). For DLR, SAIR was significantly lower for subtraction than for conventional images (0.05 ± 0.25 vs 0.65 ± 0.34, p < .001). For DLR, SNR was significantly lower for subtraction than for conventional images (0.1 ± 0.2 vs 0.7 ± 0.3, p < .001). For DLR, CNR was significantly lower for subtraction than for conventional images (33.6 ± 16.0 vs 49.1 ± 19.4, p < .001).
Tables S1–S4 (available in the online supplement) show comparisons of the subjective and objective measures of image quality between stents with a diameter less than 3 mm versus 3 mm or greater and among stents with a proximal, middle, or distal location.
Diagnostic Performance
A total of 14 of 59 (23.7%) stents showed in-stent restenosis at invasive angiography. Tables 4 and 5 summarize the diagnostic performance of the two readers for detecting in-stent restenosis by CTA performed at the stent level. For conventional CTA for reader 1, DLR compared with HIR showed no significant difference in sensitivity (78.6% vs 78.6%, p = .40) but revealed significantly higher specificity (82.2% vs 68.9%, p < .001), PPV (57.9% vs 44.0%, p < .001), NPV (92.5% vs 91.2%, p < .001), and accuracy (81.4% vs 71.2%, p < .001). For conventional CTA for reader 2, DLR compared with HIR showed no significant difference in sensitivity (50.0% vs 50.0%, p = .04) but showed significantly higher specificity (91.1% vs 86.7%, p < .001), PPV (63.6% vs 58.3%, p < .001), NPV (85.4% vs 85.1%, p < .001), and accuracy (81.3% vs 78.0%, p < .001).
| Measure | Sensitivity | Specificity | PPV | NPV | Accuracy |
|---|---|---|---|---|---|
| All stents (n = 59) | |||||
| Conv HIR | 78.6 (53.6–100.0) [11/14] | 68.9 (55.8–81.6) [31/45] | 44.0 (25.0–61.3) [11/25] | 91.2 (80.8–100.0) [31/34] | 71.2 (59.3–82.2) [42/59] |
| Conv DLR | 78.6 (55.1–100.0) [11/14] | 82.2 (69.7–93.3) [37/45] | 57.9 (33.3–80.0) [11/19] | 92.5 (84.0–100.0) [37/40] | 81.4 (71.2–91.5) [48/59] |
| Combined HIR | 78.6 (53.6–100.0) [11/14] | 80.0 (67.4–91.2) [36/45] | 55.0 (31.7–76.5) [11/20] | 92.3 (82.9–100.0) [36/39] | 79.7 (67.8–89.8) [47/59] |
| Combined DLR | 78.6 (55.1–100.0) [11/14] | 91.1 (81.8–99.0) [41/45] | 73.3 (48.6–96.9) [11/15] | 93.2 (84.9–100.0) [41/44] | 88.1 (79.7–96.6) [52/59] |
| p | |||||
| Conv HIR vs conv DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Conv HIR vs combined HIR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Conv DLR vs combined DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Combined HIR vs combined DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Stents without severe misregistration artifact (n = 27) | |||||
| Conv HIR | 83.3 (46.7–100.0) [5/6] | 71.4 (57.2–90.0) [15/21] | 45.5 (16.1–75.0) [5/11] | 93.8 (80.0–100.0) [15/16] | 74.1 (57.5–88.9) [20/27] |
| Conv DLR | 83.3 (46.7–100.0) [5/6] | 81.0 (63.6–95.5) [17/21] | 55.6 (20.0–87.5) [5/9] | 94.4 (81.8–100.0) [17/18] | 81.5 (64.9–92.6) [22/27] |
| Sub HIR | 83.3 (46.7–100.0) [5/6] | 95.2 (83.8–100.0) [20/21] | 83.3 (40.0–100.0) [5/6] | 95.2 (83.3–100.0) [20/21] | 92.6 (81.5–100.0) [25/27] |
| Sub DLR | 83.3 (46.7–100.0) [5/6] | 100.0 (100.0–100.0) [21/21] | 100.0 (100.0–100.0) [5/5] | 95.5 (85.0–100.0) [21/22] | 96.3 (88.9–100.0) [26/27] |
| p | |||||
| Conv HIR vs conv DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Conv HIR vs sub HIR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Conv DLR vs sub DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Sub HIR vs sub DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
Note—Except where otherwise indicated, data are percentage with 95% CI in parentheses and numerator and denominator in brackets. HIR = hybrid iterative reconstruction, DLR = deep learning–based image restoration, conv = conventional, sub = subtraction.
| Measure | Sensitivity | Specificity | PPV | NPV | Accuracy |
|---|---|---|---|---|---|
| All stents (n = 59) | |||||
| Conv HIR | 50.0 (18.8–78.6) [7/14] | 86.7 (75.3–95.6) [39/45] | 58.3 (25.1–81.3) [7/13] | 85.1 (74.5–95.5) [39/46] | 78.0 (67.8–89.8) [46/59] |
| Conv DLR | 50.0 (22.7-78.6) [7/14] | 91.1 (80.5-97.9) [41/45] | 63.6 (29.3-90.9) [7/11] | 85.4 (74.8-95.5) [41/48] | 81.3 (70.4-91.5) [48/59] |
| Combined HIR | 50.0 (18.7–78.6) [7/14] | 95.6 (88.0–100.0) [43/45] | 77.8 (42.9–100.0) [7/9] | 86.0 (76.5–95.9) [43/50] | 84.7 (74.6–93.2) [50/59] |
| Combined DLR | 50.0 (22.7–78.6) [7/14] | 97.8 (92.0–100.0) [44/45] | 87.5 (50.0–100.0) [7/8] | 86.3 (76.4–95.8) [44/51] | 86.4 (78.0–94.9) [51/59] |
| p | |||||
| Conv HIR vs conv DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Conv HIR vs combined HIR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Conv DLR vs combined DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Combined HIR vs combined DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Stents without severe misregistration artifact (n = 27) | |||||
| Conv HIR | 66.7 (20.0–100.0) [4/6] | 76.2 (56.9–94.7) [16/21] | 44.4 (12.5–80.0) [4/9] | 88.9 (71.9–100.0) [16/18] | 74.1 (55.6–92.6) [20/27] |
| Conv DLR | 66.7 (20.0–100.0) [4/6] | 85.7 (69.8–97.9) [18/21] | 57.1 (16.7–95.2) [4/7] | 90.0 (75.0–100.0) [18/20] | 81.5 (66.7–96.3) [22/27] |
| Sub HIR | 66.7 (20.0–100.0) [4/6] | 95.2 (83.8–100.0) [20/21] | 80.0 (33.3–100.0) [4/5] | 90.9 (77.3–100.0) [20/22] | 88.9 (77.8–100.0) [24/27] |
| Sub DLR | 66.7 (20.0–100.0) [4/6] | 100.0 (100.0–100.0) [21/21] | 100.0 (100.0–100.0) [4/4] | 91.3 (78.6–100.0) [21/23] | 92.6 (81.5–100.0) [25/27] |
| p | |||||
| Conv HIR vs conv DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Conv HIR vs sub HIR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Conv DLR vs sub DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
| Sub HIR vs sub DLR | > .99 | < .001 | < .001 | < .001 | < .001 |
Note—Except where otherwise indicated, data are percentage with 95% CI in parentheses and numerator and denominator in brackets. HIR = hybrid iterative reconstruction, DLR = deep learning–based image restoration, conv = conventional, sub = subtraction.
Of the stents included in the subtraction analysis, 7/27 (25.9%) showed in-stent restenosis based on invasive angiography. For DLR for reader 1, subtraction compared with conventional images showed no significant difference in sensitivity (83.3% vs 83.3%, p > .99) but significantly greater specificity (100.0% vs 81.0%, p < .001), PPV (100.0% vs 55.6%, p < .001), NPV (95.5% vs 94.4%, p < .001), and accuracy (96.3% vs 81.5%, p < .001). For DLR for reader 2, subtraction compared with conventional images showed no significant difference in sensitivity (66.7% vs 66.7%, p > .99) but significantly higher specificity (100.0% vs 85.7%, p < .001), PPV (100.0% vs 57.1%, p < .001), NPV (91.3% vs 90.0%, p < .001), and accuracy (92.6% vs 81.5%, p < .001).
The simulated interpretation of the combination of conventional and subtraction CTA images used the reader assessments of subtraction images in the 27 stents without severe misregistration artifact and the reader assessments of conventional images in the 32 stents with severe misregistration artifact. Among conventional CTA with HIR, conventional CTA with DLR, combination approach with HIR, and combination approach with DLR, the highest lesion-level diagnostic performance measures were as follows: for reader 1, sensitivity was the same for all approaches (78.6%), specificity was highest for combination with DLR (91.1%), PPV was highest for combination with DLR (73.3%), NPV was highest for combination with DLR (93.2%), and accuracy was highest for combination with DLR (88.1%); for reader 2, sensitivity was the same for all approaches (50.0%), specificity was highest for combination with DLR (97.8%), PPV was highest for combination with DLR (87.5%), NPV was highest for combination with DLR (86.3%), and accuracy was highest for combination with DLR (86.4%).
Tables S5 and S6 (available in the online supplement) show comparisons of diagnostic performance between stents with a diameter less than 3 mm versus 3 mm or greater and among stents with a proximal, middle, or distal location.
Tables S7 and S8 (available in the online supplement) show patient-level assessments of diagnostic performance. A total of 8 of 30 (23.3%) patients were positive for in-stent restenosis. Among patients with severe misregistration artifact, 5 of 17 (29.4%) were positive for in-stent restenosis. Among conventional CTA with HIR, conventional CTA with DLR, combination approach with HIR, and combination approach with DLR, the highest patient-level diagnostic performance measures were as follows: for reader 1, sensitivity was the same for all approaches (62.5%), specificity was highest for combination with DLR (90.1%), PPV was highest for combination with DLR (71.4%), NPV was highest for combination with DLR (87.0%), and accuracy was highest for combination with DLR (83.3%); for reader 2, sensitivity was the same for all approaches (50.0%), specificity was highest for combination with HIR or DLR (both 95.5%), PPV was highest for combination with HIR or DLR (both 80.0%), NPV was highest for combination with HIR or DLR (84.0%), and accuracy was highest for combination with HIR or DLR (both 83.3%).
Figures 2 and 3 show representative cases of the utility of the DLR and subtraction techniques for detection of in-stent restenosis.
Fig. 2A —57-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
A, Conventional hybrid iterative reconstruction (HIR) image (A) and conventional deep learning reconstruction (DLR) image (B) show blurred appearance of proximal aspect of stent (arrow).
Fig. 2B —57-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
B, Conventional hybrid iterative reconstruction (HIR) image (A) and conventional deep learning reconstruction (DLR) image (B) show blurred appearance of proximal aspect of stent (arrow).
Fig. 2C —57-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
C, Subtraction HIR image (C) and subtraction DLR image (D) show discrete region of hypoattenuation at proximal aspect of stent (arrow). Both readers provided diagnosis of in-stent restenosis only for subtraction HIR and subtraction DLR. Diagnostic confidence score for four methods for reader 1 was 2, 2, 3, and 4, respectively, and for reader 2 it was 1, 2, 3, and 3, respectively. Patient subsequently underwent invasive catheter angiography.
Fig. 2D —57-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
D, Subtraction HIR image (C) and subtraction DLR image (D) show discrete region of hypoattenuation at proximal aspect of stent (arrow). Both readers provided diagnosis of in-stent restenosis only for subtraction HIR and subtraction DLR. Diagnostic confidence score for four methods for reader 1 was 2, 2, 3, and 4, respectively, and for reader 2 it was 1, 2, 3, and 3, respectively. Patient subsequently underwent invasive catheter angiography.
Fig. 2E —57-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
E, Fluoroscopic images obtained before (E) and after (F) contrast media injection show in-stent restenosis of proximal aspect of stent (arrow).
Fig. 2F —57-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
F, Fluoroscopic images obtained before (E) and after (F) contrast media injection show in-stent restenosis of proximal aspect of stent (arrow).
Fig. 3A —53-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
A, Conventional hybrid iterative reconstruction (HIR) image (A), conventional deep learning reconstruction (DLR) image (B), and subtraction HIR image (C) show area of hypoattenuation at proximal aspect of stent (arrow). On subtraction DLR image (D), this region is less apparent. Both readers provided diagnosis of in-stent restenosis for conventional HIR, conventional DLR, and subtraction HIR but not for subtraction DLR. Diagnostic confidence score for four methods for reader 1 was 2, 2, 2, and 4, respectively, and for reader 2, it was 2, 2, 2, and 3, respectively. Patient subsequently underwent invasive catheter angiography.
Fig. 3B —53-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
B, Conventional hybrid iterative reconstruction (HIR) image (A), conventional deep learning reconstruction (DLR) image (B), and subtraction HIR image (C) show area of hypoattenuation at proximal aspect of stent (arrow). On subtraction DLR image (D), this region is less apparent. Both readers provided diagnosis of in-stent restenosis for conventional HIR, conventional DLR, and subtraction HIR but not for subtraction DLR. Diagnostic confidence score for four methods for reader 1 was 2, 2, 2, and 4, respectively, and for reader 2, it was 2, 2, 2, and 3, respectively. Patient subsequently underwent invasive catheter angiography.
Fig. 3C —53-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
C, Conventional hybrid iterative reconstruction (HIR) image (A), conventional deep learning reconstruction (DLR) image (B), and subtraction HIR image (C) show area of hypoattenuation at proximal aspect of stent (arrow). On subtraction DLR image (D), this region is less apparent. Both readers provided diagnosis of in-stent restenosis for conventional HIR, conventional DLR, and subtraction HIR but not for subtraction DLR. Diagnostic confidence score for four methods for reader 1 was 2, 2, 2, and 4, respectively, and for reader 2, it was 2, 2, 2, and 3, respectively. Patient subsequently underwent invasive catheter angiography.
Fig. 3D —53-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
D, Conventional hybrid iterative reconstruction (HIR) image (A), conventional deep learning reconstruction (DLR) image (B), and subtraction HIR image (C) show area of hypoattenuation at proximal aspect of stent (arrow). On subtraction DLR image (D), this region is less apparent. Both readers provided diagnosis of in-stent restenosis for conventional HIR, conventional DLR, and subtraction HIR but not for subtraction DLR. Diagnostic confidence score for four methods for reader 1 was 2, 2, 2, and 4, respectively, and for reader 2, it was 2, 2, 2, and 3, respectively. Patient subsequently underwent invasive catheter angiography.
Fig. 3E —53-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
E, Fluoroscopic images obtained before (E) and after (F) contrast media injection show less than 50% stenosis of proximal aspect of stent (arrow), indicating absence of in-stent restenosis.
Fig. 3F —53-year-old man who underwent coronary CTA to assess patency of coronary artery stent. CTA was performed using two-breath-hold subtraction technique.
F, Fluoroscopic images obtained before (E) and after (F) contrast media injection show less than 50% stenosis of proximal aspect of stent (arrow), indicating absence of in-stent restenosis.
Discussion
The present study compared approaches for performing coronary CTA in patients with coronary arterial stents. In comparison with HIR, DLR showed improvements in various measures of diagnostic performance for the detection of in-stent restenosis, with further improvements achieved by performing DLR in combination with a two-breath-hold subtraction technique. The technique could help guide the selection of patients with coronary stents to undergo further invasive evaluation.
In the present study, HIR yielded poor specificity and PPV for restenosis, consistent with the study by Amanuma et al. [13]. The main cause of false-positive results on coronary CTA is stent-related blooming artifact, which appears as blurred stent edges with hypoattenuation within the stent. Tatsugami et al. [9] reported a significantly shorter mean edge rise distance for DLR than for HIR, resulting in sharper edges on DLR, but they did not compare methods in terms of diagnostic performance. In the current study, the visible in-stent lumen diameter was greater for DLR than for HIR, allowing improved evaluation of stent patency. The diagnostic utility of DLR relative to HIR is evidenced by DLR showing significantly higher specificity, PPV, and accuracy for both readers and significantly higher diagnostic confidence for both readers.
In the analysis of subtraction CTA for DLR, the maximum visible in-stent lumen diameter was significantly higher for subtraction than for conventional images. Further, subtraction CTA with DLR achieved an accuracy of 92.6–96.3% for the two readers as well as specificity and PPV of 100.0% for both readers. These values are higher than those reported in prior studies using a range of approaches for stent evaluation by coronary CTA [6, 7, 13, 20, 21]. The overall optimal performance was achieved through a combination approach that incorporated both conventional DLR and subtraction DLR, selecting between the two methods on the basis of the presence versus absence of severe misregistration artifact. The combination approach achieved an accuracy of 86.4% and 88.1% for the two readers, with high PPV and NPV for both readers.
The subtraction method in the current study used a two-breath-hold technique. This approach was used because of the relatively longer breath-hold that would have been required for a one-breath-hold subtraction technique, which would be more likely to exceed patients' breath-holding capacity. The two-breath-hold technique may have contributed to the observed rate of severe misregistration artifact of 54.2%, which is somewhat higher than reported in prior studies [15, 22]. Also, the two-breath-hold subtraction technique requires an additional non-contrast acquisition and thereby inherently has a higher radiation dose [14]. DLR allows a reduction in radiation exposure, while improving image quality [9]. Thus, the use of DLR may facilitate the adoption of the two-breath-hold subtraction technique in clinical protocols. Future studies could explore the diagnostic performance of DLR-based low-radiation dose subtraction protocols, to further aid the method's clinical acceptance.
The present study had limitations. First, the study was performed at a single center with two readers, and the sample size was small, particularly in certain subanalyses. Second, CTA was performed using a scanner from a single vendor. The generaliz-ability of the findings to scanners from other vendors requires investigation. Third, examinations with severe misregistration artifact were excluded from the subtraction analysis. This approach was used on the basis that in clinical practice subtraction images would not be used for rendering diagnoses and guiding clinical management when limited by severe misregistration artifact. Fourth, the requirement for invasive angiography as a reference standard introduced a selection bias by excluding patients with mild disease who would be less likely to proceed to angiography. Finally, the impact of the different methods on patient outcomes was not explored.
In conclusion, the combination of DLR and a two-breath-hold subtraction technique yielded optimal diagnostic performance for the detection of in-stent restenosis by coronary CTA.
Footnote
Provenance and review: Not solicited; externally peer reviewed.
Supplemental Content
File (22_27983_suppl.pdf)
- Download
- 786.62 KB
References
1.
Dangas GD, Claessen BE, Caixeta A, Sanidas EA, Mintz GS, Mehran R. Instent restenosis in the drug-eluting stent era. J Am Coll Cardiol 2010; 56:1897–1907
2.
Narula J, Chandrashekhar Y, Ahmadi A, et al. SCCT 2021 expert consensus document on coronary computed tomographic angiography: a report of the Society of Cardiovascular Computed Tomography. J Cardiovasc Comput Tomogr 2021; 15:192–217
3.
Eckert J, Renczes-Janetzko P, Schmidt M, Magedanz A, Voigtländer T, Schmermund A. Coronary CT angiography (CCTA) using third-generation dual-source CT for ruling out in-stent restenosis. Clin Res Cardiol 2019; 108:402–410
4.
Ebersberger U, Tricarico F, Schoepf UJ, et al. CT evaluation of coronary artery stents with iterative image reconstruction: improvements in image quality and potential for radiation dose reduction. Eur Radiol 2013; 23:125–132
5.
Hickethier T, Baeβler B, Kroeger JR, et al. Knowledge-based iterative reconstructions for imaging of coronary artery stents: first in-vitro experience and comparison of different radiation dose levels and kernel settings. Acta Radiol 2019; 60:160–167
6.
Andreini D, Pontone G, Mushtaq S, et al. Diagnostic accuracy of coronary CT angiography performed in 100 consecutive patients with coronary stents using a whole-organ high-definition CT scanner. Int J Cardiol 2019; 274:382–387
7.
Mark DB, Berman DS, Budoff MJ, et al.; American College of Cardiology Foundation Task Force on Expert Consensus Documents. ACCF/ACR/AHA/NASCI/SAIP/SCAI/SCCT 2010 expert consensus document on coronary computed tomographic angiography: a report of the American College of Cardiology Foundation Task Force on Expert Consensus Documents. Circulation 2010; 121:2509–2543
8.
Kalisz K, Halliburton S, Abbara S, et al. Update on cardiovascular applications of multienergy CT. RadioGraphics 2017; 37:1955–1974
9.
Tatsugami F, Higaki T, Nakamura Y, et al. Deep learning-based image restoration algorithm for coronary CT angiography. Eur Radiol 2019; 29:5322–5329
10.
Greffier J, Dabli D, Frandon J, et al. Comparison of two versions of a deep learning image reconstruction algorithm on CT image quality and dose reduction: a phantom study. Med Phys 2021; 48:5743–5755
11.
Yi Y, Xu C, Xu M, et al. Diagnostic improvements of deep learning-based image reconstruction for assessing calcification-related obstructive coronary artery disease. Front Cardiovasc Med 2021; 8:758793
12.
Viladés Medel D, Leta R, Alomar Serralach X, Carreras Costa F, Pons-Lladó G. Reliability of a new method for coronary artery calcium or metal subtraction by 320-row cardiac CT. Eur Radiol 2016; 26:3208–3214
13.
Amanuma M, Kondo T, Sano T, et al. Assessment of coronary in-stent restenosis: value of subtraction coronary computed tomography angiography. Int J Cardiovasc Imaging 2016; 32:661–670
14.
Fuchs A, Kühl JT, Chen MY, et al. Feasibility of coronary calcium and stent image subtraction using 320-detector row CT angiography. J Cardiovasc Comput Tomogr 2015; 9:393–398
15.
Fuchs A, Kühl JT, Chen MY, et al. Subtraction CT angiography improves evaluation of significant coronary artery disease in patients with severe calcifications or stents—the C-Sub 320 multicenter trial. Eur Radiol 2018; 28:4077–4085
16.
Yoshioka K, Tanaka R, Muranaka K. Subtraction coronary CT angiography for calcified lesions. Cardiol Clin 2012; 30:93–102
17.
Trattner S, Halliburton S, Thompson CM, et al. Cardiac-specific conversion factors to estimate radiation effective dose from dose-length product in computed tomography. JACC Cardiovasc Imaging 2018; 11:64–74
18.
Pflederer T, Rudofsky L, Ropers D, et al. Image quality in a low radiation exposure protocol for retrospectively ECG-gated coronary CT angiography. AJR 2009; 192:1045–1050
19.
Yang WJ, Pan ZL, Zhang H, Pang LF, Guo Y, Chen KM. Evaluation of coronary artery in-stent restenosis with prospectively ECG-triggered axial CT angiography versus retrospective technique: a phantom study. Radiol Med (Torino) 2011; 116:189–196
20.
Tatsugami F, Higaki T, Sakane H, et al. Diagnostic accuracy of in-stent restenosis using model-based iterative reconstruction at coronary CT angiography: initial experience. Br J Radiol 2018; 91:20170598
21.
Qin L, Gu S, Chen C, et al. Initial exploration of coronary stent image subtraction using dual-layer spectral CT. Eur Radiol 2019; 29:4239–4248
22.
Guo W, Tripathi P, Yang S, Qian J, Rai B, Zeng M. Modified subtraction coronary CT angiography with a two-breathhold technique: image quality and diagnostic accuracy in patients with coronary calcifications. Korean J Radiol 2019; 20:1146–1155
Information & Authors
Information
Published In
Copyright
© American Roentgen Ray Society.
History
Submitted: May 10, 2022
Revision requested: June 1, 2022
Revision received: June 28, 2022
Accepted: July 21, 2022
Version of record online: August 10, 2022
Keywords
Authors
Funding Information
Supported by grant Z210013 from the Beijing Natural Science Foundation, grant 2020-I2M-C&T-B-034 from the Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences, and grant 81873891 from the National Natural Science Foundation of China.
Metrics & Citations
Metrics
Citations
Export Citations
To download the citation to this article, select your reference manager software.