Improved image quality and diagnostic performance of coronary computed tomography angiography-derived fractional flow reserve with super-resolution deep learning reconstruction

Abstract

Background

Super-resolution deep learning reconstruction (SR-DLR) algorithm has emerged as a promising image reconstruction technique for improving the image quality of coronary computed tomography angiography (CCTA) and ensuring accurate CCTA-derived fractional flow reserve (CT-FFR) assessments even in problematic scenarios (e.g., the presence of heavily calcified plaque and stent implantation). Therefore, the purposes of this study were to evaluate the image quality of CCTA obtained with SR-DLR in comparison with conventional reconstruction methods and to investigate the diagnostic performances of different reconstruction approaches based on CT-FFR.

Methods

Fifty patients who underwent CCTA and subsequent invasive coronary angiography (ICA) were retrospectively included. All images were reconstructed with hybrid iterative reconstruction (HIR), model-based iterative reconstruction (MBIR), conventional deep learning reconstruction (C-DLR), and SR-DLR algorithms. Objective parameters and subjective scores were compared. Among the patients, 22—comprising 45 lesions—had invasive FFR results as a reference, and the diagnostic performance of different reconstruction approaches based on CT-FFR were compared.

Results

SR-DLR achieved the lowest image noise, highest signal-to-noise ratio (SNR), and best edge sharpness (all P values <0.05), as well as the best subjective scores from both reviewers (all P values <0.001). With FFR serving as a reference, the specificity and positive predictive value (PPV) were improved as compared with HIR and C-DLR (72% vs. 36–44% and 73% vs. 53–58%, respectively); moreover, SR-DLR improved the sensitivity and negative predictive value (NPV) as compared to MBIR (95% vs. 70% and 95% vs. 68%, respectively; all P values <0.05). The overall diagnostic accuracy and area under the curve (AUC) for SR-DLR were significantly higher than those of the HIR, MBIR, and C-DLR algorithms (82% vs. 60–67% and 0.84 vs. 0.61–0.70, respectively; all P values <0.05).

Conclusions

SR-DLR had the best image quality for both objective and subjective evaluation. The diagnostic performances of CT-FFR were improved by SR-DLR, enabling more accurate assessment of flow-limiting lesions.

Introduction

Coronary computed tomography angiography (CCTA) is a well-established noninvasive imaging modality for the detection of obstructive coronary artery disease (CAD), with high diagnostic accuracy as compared to invasive coronary angiography (ICA) (1,2). However, the solely anatomic information provided by CCTA has poor discriminatory power for ischemia-inducing lesions (3). Fractional flow reserve (FFR) is currently the gold standard for determining the functional severity of a lesion, but it is not widely used in clinical practice due to its high cost and invasiveness (4). The advent of the CCTA-derived fractional flow reserve (CT-FFR) technique enables the comprehensive assessment of coronary anatomy and physiology with a single noninvasive exam, which outperforms CCTA and demonstrates good correlation with invasive FFR in the detection of ischemia-inducing lesions (5,6). However, severe calcification and stent evaluation remain challenging as partial averaging effects and blooming artifacts associated with limited spatial resolution can blur the vessel lumen and stent structures (7-9), which may affect lumen segmentation in the process of CT-FFR computation. Image noise, which has a tradeoff relationship with spatial resolution, can also preclude accurate assessment of CT-FFR (10).

Considerable efforts have been made to improving spatial resolution and reducing image noise, and computed tomography (CT) image reconstruction techniques have garnered substantial research interest. With the significant advances in artificial intelligence, deep learning reconstruction (DLR) algorithms based on deep convolutional neural networks (DCNNs) have been increasingly utilized. Trained with a pair of high- and low-noise images, conventional DLR (C-DLR) algorithms achieve considerable noise reduction in comparison with hybrid iterative reconstruction (HIR) and model-based iterative reconstruction (MBIR) (11,12). More recently, a novel super-resolution DLR (SR-DLR) algorithm has been developed through the use of data from ultrahigh-resolution CT (UHRCT) as the training target (13,14) and holds considerable promise in accurate CT-FFR assessments even in problematic scenarios (e.g., the presence of heavily calcified plaque and stent implantation). To our knowledge, the effect of SR-DLR on the diagnostic performance of CT-FFR remains unknown.

Therefore, the purposes of this study were to evaluate the image quality of CCTA obtained with SR-DLR in comparison with that of HIR, MBIR, and C-DLR and to assess the diagnostic performances of different reconstruction approaches based on CT-FFR. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-2075/rc).

Methods

Study participants

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by Institutional Ethics Board of Peking Union Medical College Hospital (approval No. I-23PJ1630). The requirement for individual consent was waived due to the retrospective nature of the analysis.

From April 2020 and March 2021, consecutive patients who underwent CCTA and subsequent ICA within 1 month for the assessment of known or suspected CAD were initially enrolled. The exclusion criteria were (I) poor image quality and (II) missing raw data necessary for retrospective image reconstruction.

CCTA examination and image reconstruction

All CT scans were conducted on a 320-row detector CT scanner (Aquilion ONE GENESIS Edition, Canon Medical Systems, Otawara, Japan). Each patient received nitroglycerin (1.0 mg) sublingually 1–2 minutes before examinations, and if the heart rate was higher than 75 bpm, β blockers were also administered. The contrast agent (370 mgI/mL of Iopamiro; Bracco Sine Pharma, Milan, Italy) was injected at a rate of body weight (kg) × 0.053 mL/s in 10 s (fixed), which was followed by a 30-mL injection of saline solution. The prospective one-beat mode was used for CCTA with a tube voltage of 100 kVp, and the tube current was automatically adjusted with a noise index [standard deviation (SD) =33]. A collimation of 320 mm × 0.5 mm with a z-coverage of 120–160 mm was used, and the gantry rotation time was 275 ms.

Image reconstruction was performed with HIR [adaptive iterative dose reduction 3D (AIDR 3D); FC43 kernel], MBIR [forward-projected MBIR solution (FIRST); cardiac kernel], C-DLR [advanced intelligent clear-IQ engine (AiCE); cardiac kernel], and SR-DLR [precise IQ engine (PIQE); cardiac kernel]. All reconstructions had a slice thickness of 0.5 mm with 0.25-mm intervals.

Objective image quality evaluation

A board-certified radiologist (L.M.Z., with 3 years of clinical practice experience in cardiac radiology) performed the objective image quality analysis.

Image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio assessment

All the four reconstruction datasets were simultaneously displayed on the Vitrea workstation version 7.6 (Canon Medical Systems), and regions of interest (ROIs) were placed at exactly same location via the copy and paste function. The mean CT attenuation [in Hounsfield units (HUs)] and image noise (SD of CT attenuation) were measured at the aortic root, proximal segment of left main trunk (LM), left anterior descending artery (LAD), left circumflex (LCX), and right coronary artery (RCA). Adjacent adipose tissue was also quantified for the calculation of contrast-to-noise ratio (CNR). The ROI sizes were as large as possible, and the vessel wall, calcification, and stents were carefully avoided.

The SNR and CNR were defined as follows:

$$SNR=CTattenuatio{n}_{vessel}/imagenois{e}_{vessel}$$ [1]

$$CNR=\frac{CTattenuatio{n}_{vessel}-CTattenuatio{n}_{adjacentadiposetissue}}{imagenois{e}_{vessel}}$$ [2]

Margin sharpness of the coronary artery, calcification, and stent

CT attenuation profiles of the coronary artery, calcification, and stent were generated via the setting of a linear ROI on a cross-sectional image with ImageJ software (National Institutes of Health, Bethesda, MD, USA; https://imagej.net/ij/) (15). Margin sharpness was represented by edge sharpness and the full-width at half maximum (FWHM) (Figure 1). The calculations were performed with in-house code developed in MATLAB version R2019a (MathWorks, Natick, MA, USA). Edge sharpness was examined on both sides of the profiles, and the mean values were recorded for subsequent analysis.

Figure 1.

Measurement of edge sharpness and FWHM. Profile curves for the definition of edge sharpness, the maximum slope of the profile curve, and FWHM. (A) The FWHM of the coronary artery without stent or calcification (FWHM vessel). (B) The FWHM of the calcification (FWHM calcification). (C) The FWHMs of the lumen (FWHM lumen) and stent strut (FWHM stent). The yellow lines mean the linear ROI in the cross-sectional images. CT, computed tomography; FWHM, full-width at half maximum; HU, Hounsfield unit; ROI, region of interest.

Subjective image quality evaluation

Two board-certified radiologists (L.M.Z. and C.X., with 3 and 7 years of clinical practice experience in cardiac radiology, respectively), independently assessed the overall image quality. The readers were blinded to clinical information and the reconstruction methods. The image quality of CCTA was rated on a 4-point Likert rating scale according to the severity of image noise, vessel attenuation, quality of contour delineation, and general image impression (1, nondiagnostic; 2, adequate; 3, good; 4, excellent).

CT-FFR analysis

Patients with invasive FFR results were further referred for CT-FFR measurement, and one of the abovementioned board-certified radiologists (L.M.Z.) performed the analysis with dedicated software (cFFR version 3.2.5, Siemens Healthineers, Erlangen, Germany) while being blinded to the results of FFR values. Centerline and luminal contours for coronary arteries were generated automatically and modified manually when necessary. All sites of stenosis were marked to generate a patient-specific model of the coronary artery tree, and the CT-FFR value of the target lesion was obtained 20 mm distal to the stenotic area.

ICA and FFR measurements

ICA was performed with an Allura Xper UNIQ FD10 system (Philips Healthcare, Best, the Netherlands), with images acquired through multiple projections, and at least two orthogonal projections were obtained to evaluate the target vessels. FFR was also measured at 20 mm distal to the stenotic area through use of a 0.014-in pressure wire with sensor tips (PressureWire, Radi Medical Systems, Wilmington, MA, USA) during rest and during maximal myocardial congestion induced by intravenous infusion of adenosine triphosphate (140 µg/kg/min).

Statistical analysis

Statistical analyses were performed with R software version 3.6.1 (The R Foundation for Statistical Computing; http://www.R-project.org). Quantitative parameters are expressed as the mean ± SD, and qualitative parameters are expressed as the frequency and composition ratio (%). For continuous variables with a normal distribution, one-way analysis of variance was used, while post hoc pairwise t-tests with Bonferroni correction were applied for multiple comparisons. The Friedman test was used to analyze nonnormally distributed data, and the Wilcoxon signed-rank test with Bonferroni correction was applied for subsequent multiple comparisons. For subjective evaluation, interobserver agreement was assessed via kappa coefficients (≤0.40, poor; 0.41–0.60, moderate; 0.61–0.80, good; >0.80, excellent). Mixed-effects models were used to account for within-patient correlations in the comparison of CT-FFR values. An FFR value and a CT-FFR value ≤0.8 was considered to be hemodynamically significant. With invasive FFR serving as the reference standard, the sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), diagnostic accuracy, and area under the receiver operator characteristics curve (AUC) of CT-FFR were calculated. AUCs were compared via the DeLong method, and other diagnostic parameters were compared through the bootstrapping method with Bonferroni correction (16). A P value <0.05 was considered statistically significant.

Results

Patient population

By searching our radiological datasets, we identified 1,952 patients who underwent CCTA from April 2020 and March 2021, 59 (3%) of whom underwent subsequent ICA within 1 month. Among these initial patients, 1 was excluded due to poor image quality and 8 due to CT raw data loss, leaving 50 patients (39 men and 11 women; mean age 64.5±7.8 years) for enrollment in the final analysis. The patient characteristics are specified in Table 1. Figure 2 shows the patient selection flowchart.

Table 1. Patient baseline characteristics.

Characteristic	Value (n=50)
Sex
Male	39 [78]
Female	11 [22]
Age (years)	64.5±7.8
Body mass index (kg/m2)	26.0±3.5
Risk factors for CAD
Diabetes	26 [52]
Hypertension	35 [70]
Hypercholesterolemia	34 [68]
Smoking	38 [76]
CAD family history	17 [34]

Data are presented as number [percentage] and mean ± standard deviation. CAD, coronary artery disease.

Figure 2.

Flowchart of patient inclusion and study design. CCTA, coronary computed tomography angiography; C-DLR, conventional deep learning reconstruction; CT, computed tomography; CT-FFR, coronary computed tomography angiography-derived fractional flow reserve; FFR, fractional flow reserve; HIR, hybrid iterative reconstruction; ICA, invasive coronary angiography; MBIR, model-based iterative reconstruction; SR-DLR, super-resolution deep learning reconstruction.

Objective image quality evaluation

Image noise, SNR and CNR

The image noise was lower and SNR was higher on SR-DLR than on HIR, MBIR, and C-DLR (all P values <0.05). The measured CNR of SR-DLR was higher than that of HIR (P<0.001) and comparable or superior to that of MBIR and C-DLR. Specifically, in the aortic root, the image noise was lowest, while the CNR and SNR were highest with SR-DLR (image noise: 15.40 vs. 19.87–27.21; SNR: 30.41 vs. 15.86–23.04; CNR: 37.06 vs. 21.95–32.56; all P values <0.001).

Margin sharpness of the coronary artery, calcification, and stent

The edge sharpness of the vessel, calcification, and stent were significantly higher with SR-DLR than with other reconstruction methods (all P values <0.001). The FWHM for the vessel under SR-DLR was comparable to that under MBIR (P>0.05) but smaller than that under HIR and C-DLR (all P values <0.001). The FWHM for calcification under SR-DLR was smaller than that under the other three methods (2.52 vs. 2.75–3.04 mm; all P values <0.001). As for the stent, the FWHM for the lumen calculated on SR-DLR was the largest among the four reconstruction methods (1.59 vs. 1.26–1.36 mm; all P values <0.05). FWHM for stents under SRDLR was smaller than that under the other three methods, but not statistically significant. The detailed results are provided in Table 2.

Table 2. Objective image quality assessment in CCTA reconstructed with different algorithms.

Parameter	HIR	MBIR	C-DLR	SR-DLR	P value
					All	HIR vs. SR-DLR	MBIR vs. SR-DLR	C-DLR vs. SR-DLR
CT attenuation (n=50)
Ao	419.46±58.92	483.34±87.14	454.04±71.49	458.54±72.10	<0.001	<0.001	<0.001	<0.001
LM	409.50±72.91	447.31±87.08	421.90±70.36	430.24±69.67	<0.001	<0.001	0.001	<0.001
LAD	391.63±72.60	405.89±65.19	385.35±74.53	397.86±73.73	<0.001	>0.99	0.82	<0.001
LCX	384.41±72.31	407.86±81.33	378.79±69.16	402.17±67.56	0.002	0.13	>0.99	<0.001
RCA	391.03±92.20	428.35±93.97	407.48±88.41	414.65±89.03	<0.001	<0.001	<0.001	0.001
Noise (n=50)
Ao	27.21±4.48	26.24±3.80	19.87± 1.96	15.40±2.19	<0.001	<0.001	<0.001	<0.001
LM	24.45±9.49	22.48±9.12	18.80±7.49	15.23±6.17	<0.001	<0.001	<0.001	<0.001
LAD	21.93± 8.76	22.55±9.67	19.21±7.64	16.06±7.01	<0.001	<0.001	<0.001	<0.001
LCX	19.88± 9.67	20.14±10.88	16.42±8.13	13.76±6.20	<0.001	<0.001	<0.001	0.03
RCA	23.02± 8.96	20.85±9.09	18.81±7.09	15.03±5.99	<0.001	<0.001	<0.001	<0.001
SNR (n=50)
Ao	15.86±3.61	18.64±3.52	23.04±4.19	30.41±6.82	<0.001	<0.001	<0.001	<0.001
LM	19.77±10.38	22.96±10.47	26.80±14.50	36.29±29.41	<0.001	<0.001	<0.001	<0.001
LAD	20.43±7.78	21.54±10.71	23.43±10.29	29.83±14.07	<0.001	<0.001	0.001	<0.001
LCX	23.37±10.73	24.08±9.68	29.43±17.87	34.23±14.74	<0.001	<0.001	<0.001	0.007
RCA	20.07±10.68	24.29±12.14	26.12±15.44	34.27±25.43	<0.001	<0.001	<0.001	<0.001
CNR (n=50)
Ao	21.95± 6.18	29.61±7.64	32.56±8.92	37.06±11.36	<0.001	<0.001	<0.001	<0.001
LM	23.39±9.00	30.09±12.24	34.71±14.59	36.34±15.43	<0.001	<0.001	<0.001	0.08
LAD	23.77±8.63	35.82±13.94	34.29±10.71	34.78±12.72	<0.001	<0.001	>0.99	>0.99
LCX	25.75±10.08	29.87±12.21	36.60±16.46	39.11±16.45	<0.001	<0.001	0.003	0.27
RCA	21.26±6.33	28.09±9.07	30.36±8.74	33.91±9.79	<0.001	<0.001	<0.001	<0.001
Vessel (n=50)
Edge sharpness (HU/mm)	200.59±72.84	222.10±76.37	208.32±86.58	243.04±88.39	<0.001	<0.001	<0.001	<0.001
FWHM (mm)	4.15±0.99	4.06±0.99	4.10±0.97	4.05±1.02	<0.001	<0.001	>0.99	<0.001
Calcification (n=105)
Edge sharpness (HU/mm)	444.71±195.85	566.53±262.89	531.30±247.03	590.26±249.54	<0.001	<0.001	<0.001	<0.001
FWHM (mm)	3.04±0.79	2.85±0.83	2.75±0.90	2.52±0.89	<0.001	<0.001	<0.001	<0.001
Stent (n=17)
Edge sharpness (HU/mm)	204.18±135.90	223.82±163.92	309.66±257.65	528.85±295.46	<0.001	<0.001	<0.001	<0.001
FWHM stent (mm)	0.98±0.16	0.96±0.18	0.96±0.19	0.90±0.17	0.22	0.36	0.15	0.39
FWHM lumen (mm)	1.26±0.21	1.32±0.27	1.36±0.23	1.59±0.32	<0.001	0.006	0.006	0.01

Data are presented as the mean ± standard deviation. Ao, aortic root; C-DLR, conventional deep learning reconstruction; CNR, contrast-to-noise ratio; CTCA, coronary computed tomography angiography; FWHM, full-width at half maximum; HIR, hybrid iterative reconstruction; HU, Hounsfield unit; LAD, left anterior descending artery; LCX, left circumflex artery; LM, left main trunk; MBIR, model-based iterative reconstruction; RCA, right coronary artery; SNR, signal-to-noise ratio; SR-DLR, super-resolution deep learning reconstruction.

Subjective image quality evaluation

The subjective image quality scores were significantly higher with SR-DLR than with HIR, MBIR, and C-DLR for both reviewers (reviewer 1: 3.26 vs. 2.26–2.76; reviewer 2: 3.30 vs. 2.24–2.76; all P values <0.001) (Table 3). There was excellent interobserver agreement with respect to the overall image quality (κ=0.89).

Table 3. Subjective image quality assessment in CCTA reconstructed with different algorithms.

Parameter	HIR	MBIR	C-DLR	SR-DLR	P value
					All	HIR vs. SR-DLR	MBIR vs. SR-DLR	C-DLR vs. SR-DLR
Reviewer 1	2.26±0.60	2.38±0.64	2.76±0.62	3.26±0.60	<0.001	<0.001	<0.001	0.001
Reviewer 2	2.24±0.56	2.34±0.59	2.76±0.66	3.30±0.61	<0.001	<0.001	<0.001	0.001

Data are presented as the mean ± standard deviation. CCTA, coronary computed tomography angiography; C-DLR, conventional deep learning reconstruction; HIR, hybrid iterative reconstruction; MBIR, model-based iterative reconstruction; SR-DLR, super-resolution deep learning reconstruction.

Values and diagnostic performances of different reconstruction approaches based on CT-FFR

Among the patients, 22 underwent invasive FFR, with a total of 45 lesions including 35 calcified plaques and 10 stents. The mean FFR value was 0.78±0.12, and ischemia-inducing lesions were identified in 20 lesions (44%), 15 calcified plaques, and 5 stents, respectively. The mean CT-FFR values were 0.70±0.15, 0.75±0.14, 0.72±0.13, and 0.77±0.12, for HIR, MBIR, C-DLR, and SR-DLR, respectively, representing a significant difference between the four groups (P<0.05). With invasive FFR serving as the reference standard, the diagnostic performance of SR-DLR was compared to that of HIR, MBIR and C-DLR in all patient groups and in the subgroups of calcified plaques and stents (Table 4).

Table 4. Diagnostic performances of the different reconstruction approaches based on CT-FFR.

Parameter	HIR	MBIR	C-DLR	SR-DLR	P value
					HIR vs. SR-DLR	MBIR vs. SR-DLR	C-DLR vs. SR-DLR
Calcified-related stenoses (n=35)
TP	14	11	14	14	–	–	–
TN	8	12	10	15	–	–	–
FP	12	8	10	5	–	–	–
FN	1	4	1	1	–	–	–
Sensitivity (%)	93 [79, 100]	73 [50, 94]	93 [79, 100]	93 [79, 100]	>0.99	0.03	>0.99
Specificity (%)	40 [19, 63]	60 [38, 81]	50 [28, 73]	75 [55, 94]	0.005	0.40	0.03
PPV (%)	54 [35, 73]	58 [35, 80]	58 [38, 78]	74 [53, 94]	0.03	0.10	0.08
NPV (%)	89 [63, 100]	75 [53, 94]	91 [70, 100]	94 [80, 100]	0.47	0.12	0.48
Accuracy (%)	63 [46, 77]	66 [49, 80]	69 [54, 83]	83 [69, 94]	0.01	0.09	0.04
AUC	0.67 [0.54, 0.80]	0.67 [0.51, 0.83]	0.72 [0.59, 0.85]	0.84 [0.72, 0.96]	0.004	0.07	0.04
Stents (n=10)
TP	4	3	5	5	–	–	–
TN	1	1	1	3	–	–	–
FP	4	4	4	2	–	–	–
FN	1	2	0	0	–	–	–
Sensitivity (%)	80 [33, 100]	60 [0, 100]	100 [100, 100]	100 [100, 100]	>0.99	0.03	>0.99
Specificity (%)	20 [0, 67]	20 [0, 67]	20 [0, 67]	60 [0, 100]	0.16	0.78	0.16
PPV (%)	50 [14, 86]	43 [0, 83]	56 [22, 89]	71 [33, 100]	0.27	0.35	0.29
NPV (%)	50 [0, 100]	33 [0, 100]	100 [100, 100]	100 [100, 100]	0.82	0.08	>0.99
Accuracy (%)	50 [20, 80]	40 [10, 70]	60 [30, 90]	80 [50, 100]	0.15	0.17	0.36
AUC	0.50 [0.22, 0.78]	0.40 [0.09, 0.71]	0.60 [0.40, 0.80]	0.80 [0.56, 1.00]	0.17	0.26	0.31
All (n=45)
TP	18	14	19	19	–	–	–
TN	9	13	11	18	–	–	–
FP	16	12	14	7	–	–	–
FN	2	6	1	1	–	–	–
Sensitivity (%)	90 [75, 100]	70 [48, 90]	95 [83, 100]	95 [83, 100]	>0.99	0.03	>0.99
Specificity (%)	36 [17, 55]	52 [32, 72]	44 [24, 64]	72 [53, 89]	<0.001	0.27	0.006
PPV (%)	53 [35, 69]	54 [34, 73]	58 [40, 74]	73 [55, 89]	0.008	0.06	0.02
NPV (%)	82 [56, 100]	68 [47, 89]	92 [73, 100]	95 [83, 100]	0.70	0.03	0.48
Accuracy (%)	60 [44, 73]	60 [44, 73]	67 [53, 80]	82 [71, 93]	<0.001	0.01	0.006
AUC	0.63 [0.51, 0.75]	0.61 [0.47, 0.75]	0.70 [0.58, 0.81]	0.84 [0.73, 0.94]	<0.001	0.009	0.007

The values for sensitivity, specificity, PPV, NPV, accuracy, and AUC are presented as [95% CI]. Bonferroni correction was applied for multiple comparisons, and a P value <0.05 was considered as statistically significant. AUC, area under the receiver operator characteristics curve; CI, confidence interval; CT-FFR, coronary computed tomography angiography-derived fractional flow reserve; C-DLR, conventional deep learning reconstruction; FP, false positive; FN, false negative; HIR, hybrid iterative reconstruction; MBIR, model-based iterative reconstruction; NPV, negative predictive value; PPV, positive predictive value; SR-DLR, super-resolution deep learning reconstruction; TN, true negative; TP, true positive.

Diagnostic performance in the overall cohort

Specificity and PPV were improved by SR-DLR as compared with HIR and C-DLR (specificity: 72% vs. 36–44%; PPV: 73% vs. 53–58%; all P values <0.05), whereas sensitivity and NPV were improved by SR-DLR as compared with MBIR (sensitivity: 95% vs. 70%; NPV: 95% vs. 68%; all P values <0.05). Therefore, compared with other three algorithms, SR-DLR had a significantly higher overall diagnostic accuracy (82% vs. 60–67%) and AUC (0.84 vs. 0.61–0.70) (all P values <0.05).

Diagnostic performance in the subgroup of calcified plaques

Similar results were found between the subgroup of calcified plaques and in the overall patient group, except that no significant difference was found in the PPV between C-DLR and SR-DLR (58% vs. 74%; P>0.05). When compared to MBIR, SR-DLR enabled correct reclassification from false negative to true positive in three lesions, and sensitivity was improved from 73% to 93% (P<0.05).

Diagnostic performance in subgroup of stents

Compared to HIR and C-DLR, two lesions were correctly reclassified as nonischemic by SR-DLR. However, the improvement in the diagnostic performance did not reach statistical significance (all P values >0.05). SR-DLR enabled correct reclassification from false negative to true positive in two lesions relative to MBIR, thereby improving the sensitivity (60% vs. 100%; P<0.05). The stent characteristics are summarized in Table S1, and Figures 3,4 show two example cases.

Figure 3.

Case example 1. A 62-year-old patient with severe calcification at the proximal segment of the LAD. Paired CPR and CT-FFR results on images based on HIR (A), MBIR (B), C-DLR (C), and SR-DLR (D), with CT-FFR values of 0.76, 0.77, 0.76, and 0.82, respectively. (E) ICA confirmed a stenotic degree of 50–70% and an FFR value of 0.83. C-DLR, conventional deep learning reconstruction; CPR, curved multiplanar reformation; CTA, computed tomography angiography; CT-FFR, coronary computed tomography angiography-derived fractional flow reserve; FFR, fractional flow reserve; HIR, hybrid iterative reconstruction; ICA, invasive coronary angiography; LAD, left anterior descending artery; MBIR, model-based iterative reconstruction; SR-DLR, super-resolution deep learning reconstruction.

Figure 4.

Case example 2. A 78-year-old patient with a stent placed at the proximal segment of the LAD. Paired CPR and CT-FFR results on images based on MBIR (A), and SR-DLR (B), with CT-FFR values of 0.86 and 0.67, respectively. ICA confirmed in-stent restenosis and an FFR value of 0.72. CPR, curved multiplanar reformation; CTA, computed tomography angiography; CT-FFR, coronary computed tomography angiography-derived fractional flow reserve; FFR, fractional flow reserve; ICA, invasive coronary angiography; LAD, left anterior descending artery; MBIR, model-based iterative reconstruction; SR-DLR, super-resolution deep learning reconstruction.

Discussion

SR-DLR is a novel image reconstruction technique for CCTA. In this study, the effect of SR-DLR on image quality and the diagnostic performance of CT-FFR were evaluated. The principal findings were as follows: (I) SR-DLR had the best image quality for both objective and subjective evaluation and (II) the diagnostic performance of SR-DLR was superior to that of HIR, MBIR, and C-DLR.

DLR based on DCNNs has been increasingly in clinical contexts. C-DLR is trained with a pair of high- and low-noise images to be more efficient in noise reduction than HIR and MBIR, whereas its ability to improve spatial resolution is rather limited as the training pairs share the same resolution. Using image data obtained from UHRCT and MBIR processing as the training data, SR-DLR can achieve the best possible image quality using DCNNs, with high-resolution and low noise (13,15,17-19), and thus has considerable potential in dose optimization (20). In this study, SR-DLR, in comparison to HIR, MBIR, and C-DLR, yielded significantly lower image noise (P<0.05) and increased the edge sharpness of the vessel, calcification, and stent (all P values <0.001). Moreover, compared with the other techniques, SR-DLR had the least calcium blooming (2.52 vs. 2.753.04 mm; all P values <0.001) and the largest in-stent lumen diameter (1.59 vs. 1.26–1.36 mm; all P values <0.05). The diagnostic performance of SR-DLR, in terms of AUC, was superior to that of HIR, MBIR, and C-DLR (0.84 vs. 0.61–0.70; all P values <0.05), which can be mainly attributed to the correct reclassification from false positive to true negative enabled by SR-DLR. When compared with those of HIR and C-DLR, the specificity and PPV of SR-DLR were superior (specificity: 72% vs. 36–44%; PPV: 73% vs. 53–58%; all P values <0.05), whereas the improvement in specificity and PPV compared with those of MBIR did not reach statistical significance (all P values >0.05). Image noise has been reported to be predictive of false-positive CT-FFR findings (10). Xu et al.’s study found that C-DLR had a relatively smaller number of false-positive cases as compared with HIR and MBIR, but no evidence of a difference was found regarding the diagnostic performances of different reconstruction approach-based CT-FFR techniques (12). Another important factor underlying false-positive CT-FFR findings is the limited spatial resolution of CCTA, which results in an underestimated tracing of the inner lumen and overestimated degree of stenosis (10,21) and is pronounced in the presence of severe calcification and stent implantation. We believe that the combination of reduced image noise and improved spatial resolution with SR-DLR may overcome the weaknesses of conventional CT reconstruction methods and ensure accurate CT-FFR measurements, even in challenging scenarios. In our study, SR-DLR had the smallest number of false-positive results and improved the specificity and PPV. Of note, unlike other reconstruction methods, MBIR had a relatively high false-negative rate. The reason for this may be the increased CT attenuation measured on MBIR images, an inherent characteristic of the algorithm. In this regard, the difference in CT attenuation between the vessel lumen and calcified plaque or stent was minimized, leading to an underestimated tracing of lesions. SR-DLR also allowed for the correct reclassification from false negative to true positive, thereby improving the sensitivity and NPV compared to MBIR (sensitivity: 95% vs. 70%; NPV: 95% vs. 68%; all P values <0.05).

It is worth mentioning that CT-FFR ≤0.80 and invasive FFR ≤0.80 were mismatched in eight lesions with SR-DLR, including seven false positives and one false negatives. However, six out of seven false-positive lesions were confirmed by ICA to be significantly stenotic, and the other one was caused by severe stent-related artifacts. For the false-negative case, the FFR value was 0.78—within the gray zone.

The primary limitations of our study are as follows. First, invasive FFR results were only available in 22 patients, accounting for 45 lesions and including 35 calcified plaques and 10 stents. The small sample size may account for the nonsignificant differences and limit the generalization of our findings. Nonetheless, this study represents the first of its kind to examine the impact of SR-DLR on the diagnostic performance of CT-FFR and provides a foundation for future research. Moreover, the possible effect of calcification severity and stent materials on the magnitude of blooming reduction enabled by SR-DLR remains unknown, and more granular analyses (e.g., comparison of low-moderate vs. high calcium score or of drug-eluting vs. bare-metal stents) might generate novel insights into the capability of SR-DLR. Further large-scale multicenter studies are warranted to validate and expand our initial observations. Second, in the qualitative image quality evaluation, it might have been too easy for our observers—despite being blind—to visually distinguish between the reconstruction methods due to marked differences in the image characteristics. We also acknowledge that as deep learning-based algorithms become more prevalent, subtle image alterations may occur and SR-DLR may introduce unknown or novel artifacts. Finally, large-scale prospective studies with long-term outcomes are needed to investigate the clinical impact of our findings, especially regarding the real-world effect on patient management and prognosis.

Conclusions

In conclusion, the novel SR-DLR algorithm improves both the objective and subjective image quality of CCTA in comparison with HIR, MBIR, and C-DLR. SR-DLR may overcome the weaknesses of conventional CT reconstruction methods and ensure accurate CT-FFR measurements even in the presence of heavy calcification and stent implantation, thereby improving the assessment of flow-limiting lesions.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by Institutional Ethics Board of Peking Union Medical College Hospital (approval No. I-23PJ1630). The requirement for individual consent was waived due to the retrospective nature of the analysis.

Data Sharing Statement

Available at https://qims.amegroups.com/article/view/10.21037/qims-24-2075/dss

qims-15-09-8541-dss.pdf ^{(185.2KB, pdf)}

DOI: 10.21037/qims-24-2075