Quantifying Lumen Diameter in Coronary Artery Stents with High-Resolution Photon Counting Detector CT and Convolutional Neural Network Denoising

Abstract

Background:

Small coronary arteries containing stents pose a challenge in CT imaging due to metal induced blooming artifact. High spatial resolution imaging capability is needed as presence of highly attenuating materials limits noninvasive assessment of luminal patency.

Purpose:

The purpose of this study was to quantify the effective lumen diameter within coronary stents using a clinical photon-counting-detector (PCD) CT in concert with a convolutional neural network (CNN) denoising algorithm, compared to an energy-integrating-detector (EID) CT system.

Methods:

Seven coronary stents of different materials and inner diameters between 3.43–4.72 mm were placed in plastic tubes of diameters 3.96–4.87 mm containing 20 mg/mL of iodine solution, mimicking stented contrast-enhanced coronary arteries. Tubes were placed parallel with or perpendicular to the scanner’s z-axis in an anthropomorphic phantom emulating an average-sized patient and scanned with a clinical EID-CT and PCD-CT. EID scans were performed using our standard coronary computed tomography angiography (cCTA) protocol (120kV, 180 quality reference mAs). PCD scans were performed using the ultra-high-resolution (UHR) mode (120×0.2 mm collimation) at 120 kV with tube current adjusted so that CTDI_vol was matched to that of EID scans. EID images were reconstructed per our routine clinical protocol (Br40, 0.6 mm thickness), and with the sharpest available kernel (Br69). PCD images were reconstructed at a thickness of 0.6 mm and a dedicated sharp kernel (Br89) which is only possible with the PCD UHR mode. To address increased image noise introduced by the Br89 kernel, an image-based CNN denoising algorithm was applied to the PCD images of stents scanned parallel to the scanner’s z-axis. Stents were segmented based on full width half maximum thresholding and morphological operations, from which effective lumen diameter was calculated and compared to reference sizes measured with a caliper.

Results:

Substantial blooming artifacts were observed on EID Br40 images, resulting in larger stent struts and reduced lumen diameter (effective diameter underestimated by 41 and 47% for parallel and perpendicular orientations, respectively). Blooming artifacts were observed on EID Br69 images with 19 and 31% underestimation of lumen diameter compared to the caliper for parallel and perpendicular scans, respectively. Overall image quality was substantially improved on PCD, with higher spatial resolution and reduced blooming artifacts, resulting in clearer delineation of stent struts. Effective lumen diameters were underestimated by 9 and 19% relative to the reference for parallel and perpendicular scans respectively. CNN reduced image noise by about 50% on PCD images without impacting lumen quantification (< 0.3% difference).

Conclusion:

The PCD UHR mode improved in-stent lumen quantification for all seven stents as compared to EID images due to decreased blooming artifacts. Implementation of CNN denoising algorithms to PCD data substantially improved image quality.

Introduction

Heart disease is the leading cause of death in the world, accounting for over 600,000 deaths per year in the United States alone^1,2. Coronary artery stents are implanted percutaneously to open stenotic arteries. Accurate luminal assessment is important to identify in-stent restenosis (ISR) and disease progression^3–5. Computed Tomography (CT) has been widely used in assessment of coronary artery diseases⁶. However, due to the high attenuation of metals, small coronary artery stents and associated luminal patency are often difficult to visualize with CT due to blooming artifacts and beam hardening effects^3,7–9.

Currently, most clinical CT scanners use energy integrating detectors (EID). EIDs utilize indirect conversion technology where incoming X-ray photons pass through a solid-state scintillator and are converted to visible light⁶. The visible light then travels toward a photodiode and is converted to an electronic signal. In recent years, photon-counting detector (PCD) CT has been widely studied. PCD utilizes direct conversion technology, where X-ray photons hitting the semiconductor are directly converted to electron hole pairs¹⁰. This allows for many benefits, such as lack of septa and increased geometric efficiency of the detector, energy discrimination capabilities, elimination of electronic noise, and an ultra-high resolution (UHR) of maximum in-plane spatial resolution of 125 microns achieved by smaller detector pixel size^10–13. These features are especially advantageous in coronary artery imaging, where visualization of small arteries and discrimination of high attenuating materials such as iodine and calcium are clinically relevant. Studies using prototype PCD-CT systems (SOMATOM CounT and Count Plus, Siemens Healthineers) demonstrated improved stent visibility as compared to EID scanners^3,7,14–16. In September of 2021, the first clinical dual source PCD-CT scanner was FDA approved, which has the smallest available detector pixel size (0.2 mm) and improved spatial resolution¹⁷. Here, our study has built off previous works to use the high resolution clinical PCD-CT system and a dedicated UHR sharp reconstruction kernel of Br89, taking full advantage of the small detector pixel size such that image resolution is not limited by the kernel sharpness.

We hypothesize that PCD will outperform EID for in-stent lumen visibility in coronary artery stents. With UHR acquisition and reconstruction using a dedicated UHR sharp kernel and thin slices comes substantially increased image noise^18,19. To address this, we employed an image-based convolutional neural network (CNN) denoising algorithm to further improve image quality. Therefore, the purpose of our phantom study is to quantify lumen size (in terms of effective lumen diameter) within coronary stents using a clinical UHR PCD-CT system and compare to that of EID-CT.

Methods

Coronary artery stent phantom

Seven coronary stents made from different materials and inner diameters ranging between 3.43–4.72 mm were inflated and deployed in plastic tubes (Figure 1) of diameter 3.96–4.87 mm containing 20 mg/mL of iodine solution mimicking contrast-enhanced, stented coronary arteries.

Figure 1.

Example of the process of coronary artery balloon angioplasty (A-D) used in inflation and deployment of stents in plastic tubes. Figure created using Biorender^™.

Stent materials and manufacturers are detailed in Table 1. Stent inner diameters were measured with a caliper after inflation for a reference standard. Tubes were submerged in a 10 cm water tank and placed in a semi-anthropomorphic thorax phantom (QRM GmbH, Möhrendorf, Germany) (Figure 2) simulating an average-sized adult patient (25 cm x 35 cm) and scanned on an EID (SOMATOM Force, Siemens Healthineers) and a PCD (NAEOTOM Alpha, Siemens Healthineers) systems.

Table 1:

Manufacturer and material information of the stents used in this study

Stent Number	Manufacturer	Primary Material
1	Medtronic	Stainless Steel
2	Boston Scientific	Platinum Chromium Alloy
3	Medtronic	Cobalt-Based Alloy
4	Abbott	Stainless Steel
5	Abbott	Cobalt Chromium
6	Medtronic	Cobalt-Based Platinum Iridium Alloy
7	Medtronic	Cobalt-Based Alloy

Figure 2.

Close up image of stent on holder in axial position in the water cylinder (red arrow, A). Image of experimental setup with water cylinder in an anthropomorphic phantom and submerged stents (B). Stents scanned parallel (C) and perpendicular (D) to the scanner’s z-axis. Figure created using Biorender^™.

Data Acquisition and Image Reconstruction

Stent phantoms were placed at 0 degrees (parallel) and 90 degrees (perpendicular) to the scanner’s z-axis (Figure 2). EID scans were performed using the standard coronary CT angiography (cCTA) protocol in our practice (120kV, 196 × 0.6mm collimation, 180 quality reference mAs - a Siemens specific CAREDose4D tube current modulation parameter selected by user to determine desired image quality level²⁰). PCD scans were performed in UHR mode (120 × 0.2 mm collimation) at 120 kV, with tube current adjusted such that the CTDI_vol was matched to that of EID scans (5.5–6mGy). Both EID and PCD images were reconstructed with a field of view of 55 mm centered around the stent. EID images were reconstructed first following the clinical protocol at our institution with a 512 × 512 matrix size, a slice thickness of 0.6mm, and a medium smooth kernel (Br40) and an iterative reconstruction algorithm (ADMIRE, Siemens Healthineers) at strength 3, following the clinical protocol at our institution. In addition, EID images were reconstructed under the above parameters but with 1024 × 1024 matrix size, and the sharpest available kernel (Br69). PCD images were reconstructed with a 1024 × 1024 matrix size, a slice thickness of 0.6 mm, and a sharp kernel (Br89) that is only possible with the PCD UHR mode. Two reconstructions were performed for PCD images: one with quantum iterative reconstruction algorithm (QIR; Siemens Healthineers) at strength 3, and the other with filtered backprojection (FBP) algorithm, with all other parameters matched. The QIR images are used for stent imaging, and the difference of FBP and QIR images are used to create noise images for training of an image-based CNN denoising algorithm (to be discussed at next section).

Image-Based CNN Denoising

To address increased noise introduced by the dedicated sharp Br89 kernel, an image-based CNN denoising algorithm reported by Huber et al. was trained and applied to the PCD images of stents parallel to the scanner’s z-axis¹⁹. Specifically, all training data for the image-based CNN denoising algorithm were derived from the FBP and QIR reconstructions of the PCD-CT images (Figure 3). The difference between these inputs created a noise-only image. Afterwards, the noise-only images and QIR images from 3 adjacent slices were split into 50,000 patches (size: 64×64×3) for training. Noise-only and QIR patches were added to generate the noisy training input. The corresponding training targets were QIR patches without noise insertion. A U-Net architecture with four max pooling and upconvolutional layers, 64 filters at the first convolutional layer, exponential linear unit (ELU) activation were used for the network with a loss function combined with a VGG feature loss and mean-squared-error (MSE) loss (Figure 3)²¹. The CNN was trained with Adam optimizer of a scheduled descending learning rate (0.001 to 0.00001) within 100 epochs. After training, the CNN was applied to the reconstructed FBP images for noise reduction making the PCD + CNN image data.

Figure 3.

A schematic of CNN denoising training framework. Training data included filtered-back projections (FBP) and quantum iterative reconstruction (QIR) images.

Processing and Image Analysis

Stents were segmented using a custom Matlab script (MATLAB 2019b, The MathWorks, Inc., Natick, MA) based on full width half maximum (FWHM) thresholding and morphological operations (Figure 4). Axial and sagittal images were used for stent data acquired parallel and perpendicular to the scanner’s z-axis, respectively. Then, the script identified relevant image slices containing stents. From this, a binary image was created using FWHM Hounsfield unit (HU) thresholds based on CT numbers of the inner lumen and metal from a given slice. Using the morphological toolkit in Matlab, neighboring pixels of the stent were connected to form a circle where struts were visibly separated. The lumen of the stent was then filled, struts were subtracted, and pixels in the lumen were counted by the algorithm to calculate inner lumen area (A). The effective diameter of the inner lumen (D_E) was then estimated based on the following equation:

$$D_E=2\sqrt{\frac{\text{A}}{\text{π}}}$$ (1)

Figure 4.

Process pipeline for automated segmentation algorithm. Axial images with stents were extracted and stent location was identified. Binary images were created from stent images and connected using morphological operations. Areas inside the stents were filled for calculation of inner lumen effective diameter.

This process was repeated for EID, PCD, and PCD + CNN images. Inner lumen effective diameter was compared to reference sizes measured with a caliper. After processing, outliers in the EID, PCD, and PCD + CNN images where the algorithm included pixels far from the stent were manually excluded from the calculation. Image noise was measured as standard deviation of CT numbers in region of interest (ROI) and compared among EID, PCD, and PCD+CNN images.

Lumen measurements were calculated for each stent using the segmentation pipeline for EID, PCD, and PCD + CNN, respectively. These values were then compared to caliper measurements for each scanning orientation by calculating percent difference:

$$\text{Percent Difference}=\frac{\text{X}1–\text{ X}2}{\text{X}2}\times 100\%$$ (2)

where ${\text{X}}_{\text{1}}$ is the average lumen diameter measured from the segmentation pipeline of EID, PCD, or PCD + CNN, and ${\text{X}}_2$ is the caliper-measured diameter. Percent difference was calculated for each stent, and values were averaged across all stents and reported for EID, PCD, and PCD + CNN, respectively.

Statistical Analysis

Two-sample t-test was performed using Matlab to evaluate the difference between two variables: effective diameter of EID Br40 and Br69 versus PCD for both parallel and perpendicular orientations. Additionally, effective diameter of EID versus PCD + CNN, and effective diameter of PCD versus PCD + CNN for parallel stent scans was evaluated, respectively. A p-value < 0.05 was considered statistically significant. Standard deviation was calculated for diameter measurements for all images along the z-axis used in overall effective diameter calculation. Standard deviation values were used as error bars in plotted data.

Results

Substantial blooming artifacts were observed in both EID Br40 and Br69 images, resulting in stents appearing larger than the true size and reduced lumen sizes (Figure 5 and 6). The stent appeared as a primarily continuous piece of metal without visibility of individual struts. This was observed on axial, coronal, and volume rendered images. In addition, EID images reconstructed with Br69 kernel introduced a strong patchy noise texture and dark rings around the stent borders (Figure 5 and 6), which affected areas that could hold critical information regarding the diagnosis of in-stent restenosis and obstruct view of relevant neighboring pathology. Overall image quality was substantially improved on PCD images, with higher spatial resolution and reduced blooming artifacts compared to EID images. This resulted in clearer delineation of stent struts. From this, the lumen inside the stent appears to be more open compared to that of EID images. The original PCD images directly reconstructed from the scanner had high image noise (SD = 420 HU) due to very sharp kernel (Br89). Image noise was substantially reduced by approximately 50% (from 420 HU down to 240 HU) using the CNN model.

Figure 5.

Evaluation of EID Br40 images (left), EID Br69 (left middle), PCD (right middle) and PCD+CNN Denoising (right) for in-plane with scanner’s z-axis (top), multi-planar reformat (MPR, middle), and volume rendering technique (VRT, bottom). PCD images shown were reconstructed with QIR. WW/WL = 1000/10000 HU.

Figure 6.

Evaluation of EID Br40 images (left), EID Br69 (left middle), PCD (right middle) and PCD+CNN Denoising (right) for through-plane with scanner’s z-axis (top) and MPR (bottom). PCD images shown were reconstructed with QIR. WW/WL = 1000/10000 HU.

For quantitative measurements, effective diameter of stents was underestimated by 41 and 47% in EID Br40 images relative to caliper measurements for parallel and perpendicular scan orientations respectively (Figure 7A and B), while EID Br69 images were underestimated by 19 and 31% with respect to the caliper for parallel and perpendicular orientations. Effective lumen diameters of PCD images had an underestimation of 9 and 19%, matching the caliper reference more closely. The CNN denoising algorithm reduced noise by 50% on PCD images without significantly impacting luminal diameter measurements (<0.3% difference in comparison to PCD images, p=0.943). Implementation of CNN denoising substantially improved image quality through noise reduction (Figure 5).

Figure 7.

Results of segmented inner stent lumen effective diameter (mm) for EID Br40, EID Br69, PCD, and PCD + CNN Denoising for all seven stents parallel to the scanner’s z-axis (A) and results for EID Br40, EID Br69, and PCD perpendicular to the scanner’s z-axis (B). Error bars represent standard deviation of values used in effective diameter calculation for each stent.

Effective lumen diameter of EID Br40 images were statistically different (p<0.001) than those of PCD and PCD + CNN images for images acquired parallel to the scanner’s z-axis (Table 2A). Effective lumen diameter from EID Br69 images were smaller than those of from PCD and PCD + CNN for each of the 7 stents, but the difference was not statistically significantly (p = 0.089, p = 0.087, Table 2A). Similar trends were observed for stents scanned perpendicular to the scanner’s z-axis: effective lumen diameters measured from EID Br40 images were significantly different than those from the PCD images (p = 0.001, Table 2B); effective lumen diameters measured from the Br69 images were smaller than those from PCD but not statistically significant (p = 0.074, Table 2B). Effective lumen diameter measurements were larger with PCD for all seven stents scanned in both orientations as compared to EID.

Table 2A.

Stent lumen diameter measurements from caliper and segmentation pipeline for parallel to scanner’s z-axis images of EID Br40, EID Br69, PCD, and PCD + CNN.

Stent Number	Caliper (mm)	EID Br40 (mm)	EID Br69 (mm)	PCD (mm)	PCD + CNN (mm)
1	3.44	2.00	2.89	3.13	3.09
2	3.94	2.85	3.01	3.65	3.65
3	3.55	2.49	3.01	3.33	3.42
4	3.43	2.24	2.93	3.13	3.14
5	4.72	3.19	3.76	4.09	4.15
6	3.46	2.22	2.89	3.09	3.08
7	4.04	2.66	3.25	3.84	3.85
P-Values:
EID Br40 v. PCD		p < 0.001
EID Br69 v. PCD		p = 0.089
EID Br40 v. PCD + CNN		p < 0.001
EID Br69 v. PCD + CNN		P = 0.087
PCD v. PCD + CNN		p = 0.943

Table 2B.

Stent lumen diameter measurements from caliper and segmentation pipeline for perpendicular to scanner’s z-axis images for EID Br40, EID Br69, and PCD.

Stent Number	Caliper (mm)	EID Br40 (mm)	EID Br69 (mm)	PCD (mm)
1	3.44	1.61	2.22	2.41
2	3.94	1.85	2.71	3.24
3	3.55	2.19	2.70	3.06
4	3.43	1.46	1.89	2.76
5	4.72	2.98	3.34	3.84
6	3.46	2.13	2.63	3.01
7	4.04	1.85	2.80	3.21
P-Values:
EID Br40 v. PCD		p = 0.001
EID Br69 v. PCD		p = 0.074

Discussion

The ultra-high resolution achieved from the use of novel PCD-CT technology and sharp kernel (Br89) notably improved the visualization of the stent and quantification of inner lumen size as compared to EID images in this phantom study. The use of metallic stents is associated with up to 40% occurrence of ISR^22,23. Although introduction of drug-eluting stents (DES) has reduced rates of ISR, its prevalence remains significant at about 10% occurance²⁴. Here, our study demonstrated that use of PCD-CT scanning and dedicated sharp kernel (Br89) outperforms the clinical EID CT scanner for both Br40 and Br69 in terms of inner stent lumen quantification allowing for improved assessment of luminal patency. Reconstruction with Br69 improved effective lumen diameter measurements compared to the clinical reconstruction with Br40. However, use of Br69 for EID reconstructions introduced a strong patchy noise texture and dark ring artifacts around the stent borders (Figure 5 and 6). The affected area could hold critical information regarding the diagnosis of in-stent restenosis and obstruct view of relevant neighboring pathology. Although they provide more accurate quantitative measurements of the stent size, these images are not clinically acceptable due to these limitations. In contrary, these artifacts are not observed in PCD images as the smaller detector size supports the sharp kernel used in this study.

Stent imaging in patients using PCD-CT has been reported using an investigational, prototype spectral PCD-CT scanner (SPCCT; Philips Healthcare) with 270 um detector pixel pitch at the isocenter, demonstrating image quality improvements in stents^25–27. Sigovan et al. and Bratke et al. have exhibited improved vascular imaging in ex vivo vessel models containing stents with and without calcifications using this system^9,28. Previous work by Petritsch et al. demonstrated the benefit of using PCD over EID for stent visualization using an investigational PCD CT scanner (SOMATOM Count Plus, Siemens Healthineers) reconstructing with a non-UHR Bv60 sharp kernel¹⁶. More accurate physical dimension measurements compared to EID using a high resolution (0.25 mm detector collimation) prototype PCD system (SOMATOM CounT; Siemens Healthineers) was exhibited by Rajagopal et al¹⁵. Koons et al. evaluated use of PCD-CT UHR mode on the clinical scanner (NAOTOM Alpha, Siemens Healthineers) with phantom vessels mimicking blood and iodinated lumens with varied degree of stenosis²⁹. Mergen et al. recently demonstrated the benefits of this clinical PCD-CT system and UHR mode in patients with coronary artery calcification visualization³⁰. Here, our study quantitatively evaluated visualization of coronary artery stent inner lumen with the first clinical PCD-CT using the 0.2 mm detector pixels, a dedicated UHR sharp kernel, and an applied image-based CNN denoising algorithm. The investigated PCD-CT system has a UHR focal spot (0.4×0.5 mm²) to accommodate the small detector size. The PCD-CT system used in this study has a reported limiting spatial resolution of 125 microns in-plane with the scanner’s z-axis spatial resolution verified by the modulation transfer function (MTF) and a limiting through-plane with respect to the scanner’s z-axis of 0.3 mm measured by the section sensitivity profile (with nominal slice thickness of 0.2 mm) ¹³. This is the highest in and through-plane spatial resolution for clinically available CT scanners to date. In this study, stents were placed at 2 orientations relative to the scanner’s z-axis to test the impact of the difference of in-plane and through-plane resolution on lumen diameter assessment. This was done with the consideration that stent could be in any direction in patient scans. As expected, lumen diameter quantification performed the best when stents were oriented parallel to the scanner’s z-axis as this solely relies on the in-plane resolution which is superior to through-plane resolution. Nonetheless, PCD-CT outperformed EID-CT for both parallel and perpendicular orientations due to better resolution of PCD-CT in all directions. In addition, a dedicated UHR kernel (Br89) that is only available in the UHR mode was used in this study to take full advantage of the system resolution provided by the small detector and focal spot. As expected, image noise increased substantially with the high-resolution imaging mode and sharp reconstruction kernel, which negatively impacts image quality and may dampen the benefits of ultra-high resolution. This increase in image noise, while not significantly affecting lumen diameter measurements, may jeopardize qualitative information of intraluminal structures. Therefore, it is critical to control image noise while maintaining the ultra-high spatial resolution. In this study, we demonstrated that a CNN-based denoising algorithm can achieve approximately 50% noise reduction for UHR PCD-CT images.

There are several limitations of our study. In an animal or human model, motion artifacts may present a noteworthy challenge. Our study was restricted to use of a static model, as UHR mode was not available in electrocardiogram-gated (ECG-gated) cardiac scans at the time of this work. Future directions will implement the ECG-gated cardiac motion feature using the QRM cardiac motion phantom. Our phantom study was also limited to the seven evaluated stents; however, a range of stent materials were included in the study. Future work including more devices as well as recruitment of patients with coronary artery stents is underway.

Conclusions

Our study demonstrated that improved lumen visibility in stented coronary arteries is achieved with PCD-CT technology, and allows more accurate non-invasive assessment of ISR. The use of PCD-CT cCTA for assessment of intraluminal patency could be used to potentially avoid unnecessary interventional procedures.