Novel fast FFR derived from coronary CT angiography based on static first-pass algorithm: a comparison study

Abstract

BACKGROUND

Fractional flow reserve (FFR) is the invasive gold standard for evaluating coronary arterial stenosis. However, there have been a few non-invasive methods such as computational fluid dynamics FFR (CFD-FFR) with coronary CT angiography (CCTA) images that can perform FFR assessment. This study aims to develop a new method based on the principle of static first-pass of CT perfusion imaging technique (SF-FFR) and evaluate the efficacy in direct comparisons between CFD-FFR and the invasive FFR.

METHODS

A total of 91 patients (105 coronary artery vessels) who were admitted from January 2015 to March 2019 were enrolled in this study, retrospectively. All patients underwent CCTA and invasive FFR. 64 patients (75 coronary artery vessels) were successfully analyzed. The correlation and diagnostic performance of SF-FFR method on per-vessel basis were analyzed, using invasive FFR as the gold standard. As a comparison, we also evaluated the correlation and diagnostic performance of CFD-FFR.

RESULTS

The SF-FFR showed a good Pearson correlation (r = 0.70, P < 0.001) and intra-class correlation (r = 0.67, P < 0.001) with the gold standard. The Bland-Altman analysis showed that the average difference between the SF-FFR and invasive FFR was 0.03 (0.11–0.16); between CFD-FFR and invasive FFR was 0.04 (-0.10–0.19). Diagnostic accuracy and area under the ROC curve on a per-vessel level were 0.89, 0.94 for SF-FFR, and 0.87, 0.89 for CFD-FFR, respectively. The SF-FFR calculation time was about 2.5 s per case while CFD calculation was about 2 min on an Nvidia Tesla V100 graphic card.

CONCLUSIONS

The SF-FFR method is feasible and shows high correlation compared to the gold standard. This method could simplify the calculation procedure and save time compared to the CFD method.

In the past 20 years, fractional flow reserve (FFR) of coronary artery has gradually become a recognized functional evaluation index for coronary arterial stenosis. The treatment strategy guided by FFR has been proved to be safe, cost-friendly, and can improve the prognosis of patients.^[1] The gold standard FFR examinations are acquired from invasive coronary angiography (ICA). Non-invasive coronary CT angiography (CCTA) has shown superior performance in detecting the coronary arterial stenosis and the possibility to provide functional evaluation of coronary artery diseases with the fractional flow reserve derived from CT (CT-FFR).^{[2, 3]}

CT-FFR used in multiple previous studies has been mostly based on the computed fluid dynamics (CFD) analysis.^[4-7] These models have been shown to give good agreement with the invasive gold standard FFR. However, there are multiple parameters and boundary conditions to set in the CFD stimulation. It is time-consuming (about 30 min to 4 h per case) mainly because of solving governing equations.^[8] The discretization of the partial differential equations solving progress have hindered the application in clinical routine.^[9] A variety of software have been developed to increase the calculation speed and the accuracy of non-invasive CT-FFR. A study pointed out that based on the transluminal attenuation gradient (TAG), it was possible to further improve the diagnostic accuracy of CFD model.^[10]

Most recently, machine learning (ML) and deep learning (DL) have been applied in computational FFR applications. ML methods performed well in cardiovascular medical imaging.^[11,12] A multi-center study proved that CT-FFR based on ML improved the performance of CCTA by correctly reclassifying hemodynamically nonsignificant stenosis and performed equally well as CFD-based CT-FFR.^[13] With the development of DL, there have been a few novel DL CT-FFR approaches like DEEPVESSEL-FFR and feed-forward neural network (FFNN).^[14,15] These ML or DL approaches usually use CFD simulation results as the ground truth to train a model, therefore, the upper limit of the algorithm is the simulation results. However, there are several limitations with CFD simulations. Appropriate boundary conditions and parameters directly affected the simulation results and they required domain expertise. The whole process usually costs several minutes or longer time.

In our study, we aimed to propose a fast and simple method to reduce the CT-FFR calculation time while ensuring accuracy. This new CT-FFR method was based on the principle of static first-pass of CT perfusion (CTP) imaging technique (SF-FFR). We completed the automatic extraction of coronary artery tree based on a deep learning method to improve the SF-FFR’s performance under different image quality conditions. We evaluated the efficacy in direct comparisons with the invasive gold standard FFR. Also, we calculated the CFD-FFR using an open-source software and conducted comparisons.

METHODS

Patients

The Independent Ethics Committee of Beijing Anzhen Hospital approved this study and waived the requirement for informed patient consent. There were 91 patients with 105 vessels from a single-center cohort who had underwent CCTA for evaluation of coronary artery disease (CAD) as well as a subsequent invasive coronary angiography including invasive FFR measurements between August 1, 2014 and March 30, 2019. All patients had CCTA performed within 30 days before the invasive coronary angiography. The exclusion criteria were: (1) invasive FFR measured in stented vessel (3 vessels in 3 cases); (2) very poor image quality (2 vessels in 2 cases); (3) more than one lesion per vessel (13 vessels in 10 cases); and (4) Unable to CFD simulate (12 vessels in 12 cases). Figure 1 showed the inclusion flowchart. The demographic and clinical characteristics of the remained 64 patients with 75 vessels were shown in Table 1.

Figure 1.

Flowchart of study inclusion.

CFD: computed fluid dynamics; FFR: fractional flow reserve; SF: static first-pass.

Table 1. Demographic and clinical characteristics (n = 64).

Characteristics	Value
Data are presented as mean ± SD or n (%). BMI: body mass index; CAD: coronary artery disease.
Age, yrs	61.4 ± 8.1
Male	40 (65%)
Female	22 (35%)
BMI, kg/m2	25.6 ± 3.1
Obesity (BMI ≥ 30 kg/m2)	6 (9%)
Family history of CAD	17 (27%)
Angina pectoris	9 (15%)
Myocardial infarction	4 (6%)
Current smoking	30 (48%)
Current drinking	17 (27%)
High pressure	36 (58%)
Hyperlipidemia	33 (53%)

CCTA Imaging Acquisition

All examinations were conducted on a 256-detector row CT scanner (Revolution CT, GE Healthcare, Milwaukee, USA), following the guidelines of the Society of Cardiovascular Computed Tomography (SCCT). ^[16] Sublingual nitroglycerin (0.5 mg per dose; Nitroglycerin spray, Jingwei Pharmacy, Jinan, China) was administered 5 min before scanning in all patients. Beta-blockers were not administered to any of the patients. The CCTA data were acquired after 50-60 mL contrast agent (350 mg iodine/mL, Omnipaque, GE Healthcare, USA; or 370 mg iodine/ mL, Ultravist, Bayer Schering Pharma, Berlin, Germany) injected at a rate of 4.5-5 mL/s followed by IV injection of a saline bolus chaser of 30-35 mL at a rate of 5 mL/s. Prospectively, ECG triggered axial-mode single heartbeat acquisition was used with scanners with a wide Z-axis coverage. A bolus tracking technique was used for scan triggering the CCTA acquisition. Gantry rotation time was in the range of 0.28 s per rotation, depending on the CT scanner. Axial images were reconstructed with 0.625 mm slice thickness.

Process of the Quantification of SF-FFR from CCTA

Process of the quantification of SF-FFR included three steps: (1) data preprocessing for image quality control; (2) vessel segmentation with a deep learning method; and (3) automatic detection of stenosis site and the calculation of SF-FFR value voxel by voxel. The flow chart of the SF-FFR algorithm is shown in Figure 2, and the detailed progress of SF-FFR measurement is described as follows.

Figure 2.

Schemes of the segmentation pipeline and network architecture.

The schematic illustrates coronary artery segmentation process with datasets preparation, pre- and post-processing to optimize the segmentation result, network architecture, training progress, and test progress.

Data preprocessing for image quality control

We performed a series of data preprocessing to achieve image quality control. First of all, we used linear interpolation to resample all data to a voxel size of 0.5 mm × 0.5 mm × 0.5 mm. A Gaussian filter was added to the layers among the slice intervals. The empirical value of the aortic root area was 400 ± 50 HU, the maximum CT value within aorta area was standardized to 400 HU. In order to reduce the effect of calcified lesions with high calcium scores on the SF-FFR calculation, a Laplace filter (sharping) was added to the CCTA image to improve visualization of the coronary artery lumen.^[17] For heavily calcified coronary arteries, plaques were also extracted with threshold segmentation with a CT value over 600 HU, and then excluded from the lumen of vessels.^[18]

Vessel segmentation based on a deep learning algorithm

Segmentation of the coronary arteries and aortic root is an important step for further visualization and quantification of the vessels. In our study, the vessels, including aortic root and 3 main coronary arteries, were segmented using a pretrained deep learning method on CCTA image.^[18] We have created our models using Nvidia Corporation’s deep learning GPU training system. Both training and testing were done on a 2 Intel(R) Xeon Sliver 4110 2,1GHz, 16GB processor equipped with a Nvidia Tesla V100 graphic card, under Windows Server 2016 Standard 64 bits operating system. Figure 2 depicts a scheme of the segmentation pipeline and network architecture.

Calculation of SF-FFR

Generally, the myocardial CTP imaging can be performed with two protocols, including the static CTP imaging (acquisition of two single phases under the rest and stress conditions) and the dynamic CTP imaging (repeated acquisitions during the first-pass of the contrast media).^[19] For the static CTP, a single phase of the first-pass of contrast material through the myocardium is acquired.^[20] Hence, in our study, CCTA imaging was regarded as the static arterial first-pass imaging of coronary artery which is like the myocardial perfusion techniques.

CCTA imaging is routinely performed after the administration of iodinated contrast media through an intravenous (IV) access during the early portion of the first pass circulation of the contrast media bolus, of which a high flow rate of at least 5 mL/s is needed to optimize the strength of enhancement in the first-pass arterial phase.^[21] The contrast agent barely reaches the vein when the CCTA imaging is captured. Thus, the venous concentration of contrast agent is approximately zero. Meanwhile, the diffusion of the contrast agent to the extra-vascular extra-cellular space (EES) is started about 1min after the IV administration. The principle of coronary blood flow calculation based on CCTA image is shown in Figure 3.

Figure 3.

Principle of coronary blood flow calculation based on CCTA image.

The figure illustrates the principle of blood flow calculation based on CCTA images that was performed by measuring tracer concentration (CT values) and using a simplified formula. CCTA: coronary CT angiography; Cv(t): the venous ICM concentration; Ct(t): the concentration of ICM in the target tissue; f: the blood flow to tissue; ICM: Iodinated Contrast Media.

According to the Fick principle,^[22-26] after the administration of iodinated contrast media, the change of its concentration over time in tissues can be described by the following formula:

Note: is iodine concentration in target tissue [g/g]; f is blood flow of target tissue [mL/g/min]; (t) and (t) are iodine concentration of target tissue with arterial inflow and venous outflow, respectively [g/mL/s]; t is time [s].

The iodinate contrast agent barely reaches the vein during first pass when the CCTA imaging is captured, thus, we hypothesis that is close to zero while the iodine content within the aorta and the coronary artery reach the peak. Hence, the blood flow of target tissue can be obtained using the following calculation formula (6):

FFR refers to the ratio between the maximum achievable blood flow in a diseased coronary artery and the theoretical maximum flow in a normal coronary artery. A series of methods, such as the computational fluid dynamic approach, were proposed using the ratio of distal intracoronary pressure to aortic pressure during a maximal hyperemia across a stenosis, of which the pressure was used to replace the blood flow. ^[26] In this study, we proposed a direct method to estimate the blood flow, rather than the pressure, based on the static first-pass imaging at the diseased coronary and the normal coronary artery to calculate the ratio, thus, the SF-FFR can be obtained using the following calculation formular (7):

Note: refers to the blood flow of the reference point, which is located at the plane of aortic root at the ostium of coronary artery, since the aortic iodine concentration and coronary iodine concentration reach the peak (maximum value) at the same time. refers to the blood flow in the diseased coronary which is commonly presented as stenosis.

When substitute the in the formula (6) to the (7), the FFR value can be calculated using the following formula (8):

Since the iodinated contrast media attenuates X-rays directly proportionally to iodine content in tissue, ^[27] the can be quantified by CT number. In this way, by comparing the CT number of coronary artery stenosis with the CT number of the reference point, the formula (8) can be rewritten as:

CFD-FFR Calculation

We used SimVascular, an interactive application, to calculate the CFD-FFR. ^[28] SimVascular creates patient-specific geometric models of human vasculature from 3D medical imaging data. These models are then used as the basis for blood flow simulations using various numerical methods. Limited by computational resources, we cropped the vessels with lesions. We followed the coronary simulation guideline to set parameters and the blood flow was then calculated based on formula (7).

Invasive Measurement of FFR

Selective invasive coronary angiography and FFR were performed under the standard practice. ^[29] The pressure-wire was positioned in a vessel segment ≥ 2 mm that is 20-30 mm distal to a stenosis. Hyperemia was induced by intravenous injection of adenosine (140–180 µg/kg per minute). The name of the coronary vessel measured by invasive FFR was recorded to enable the SF-FFR analysis to be performed from the same vessel without knowing the invasive FFR results. FFR ≤ 0.80 was considered the diagnosis of lesion-specific ischemia.^[30]

Statistical Analysis

Categorical variables were presented as absolute numbers and percentages. Continuous variables were expressed as means ± SD or median (interquartile range, IQR) based on the normality of the variable. Comparisons were made between the SF-FFR, the CFD-FFR and the gold standard. We used the Pearson correlation coefficient and the intra-class correlation (ICC) to evaluate the correlations between these methods. An ICC < 0.4 indicated poor correlation, an ICC between 0.4 and 0.75 indicated fair to good correlation, and an ICC greater than 0.75 indicated excellent correlation.^[31] Thereafter, the Bland–Altman analysis was applied to assess the limits of agreement between the methods. A two-tailed P-value < 0.05 indicated statistical significance. For the performance of diagnostic, the sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were employed to evaluate the discrimination of ischemia on vessel-based and patient-based, referenced by invasive FFR. A comparison of sensitivity, specificity, and diagnostic accuracy between SF-FFR and CFD-FFR was performed by using the McNemar’s chi-squared test. The area under the curve (AUC) derived from receiver operating characteristic (ROC) analysis were employed to evaluate the discrimination of ischemia on vessel-based and patient-based, referenced by invasive FFR. And the comparison of AUC between SF-FFR and CFD-FFR was performed by using the DeLong test.^[32] Statistical analyses were performed using Python (version 3.5.6).

RESULTS

Coronary Arteries Segmentation and Calculation of SF-FFR, CFD-FFR

SF-FFR approach could be calculated successfully in the 105 vessels. Only 73% vessels could be successfully simulated using the complex CFD method. Finally, the cohort consisted of 64 patients with 75 vessels. The mean dice coefficient for the segmentation model is 0.90. An example of SF-FFR and CFD-FFR result is shown in Figure 4. The computation of SF-FFR costed only 2.5 ± 0.5 s and the whole calculation process of CFD-FFR costed about 2-3 min.

Figure 4.

An example of SF-FFR.

The gold standard invasive FFR was 0.83 in the middle part of LAD coronary artery. The SF-FFR was 0.81 and the CFD FFR was 0.82. CFD: computed fluid dynamics; FFR: fractional flow reserve; LAD: left anterior descending artery; SF: static first-pass.

Correlations between SF-FFR, CFD-FFR and Invasive FFR

On a per-vessel basis, there was a moderate correlation between SF-FFR and invasive FFR (r = 0.70 and ICC = 0.68). The correlation between CFD-FFR and invasive FFR shows a similar performance of r = 0.65 and ICC = 0.67. Figure 5 displays the regression lines for the correlation. The Bland–Altman analysis of the gold standard-invasive FFR, SF-FFR and CFD-FFR is shown in Figure 6.

Figure 5.

Regression and correlation between the CFD, SF-FFR and gold standard.

Per-vessel correlation of SF-FFR versus invasive FFR (left) shows r = 0.70 (P < 0.001). And the correlation of CFD-FFR versus invasive FFR (right) shows r = 0.67 (P < 0.001). CFD: computed fluid dynamics; FFR: fractional flow reserve; SF: static first-pass.

Figure 6.

The Bland–Altman analysis of the CFD, SF-FFR and gold standard.

Bland-Altman plots of SF-FFR versus invasive FFR (left) shows the mean difference 0.03 (−0.11−0.16). And the mean difference of CFD-FFR versus invasive FFR (right) was 0.04 (−0.10−0.19). CFD: computational fluid dynamics; FD: computed fluid dynamics; FFR: fractional flow reserve; SF: static first-pass.

Diagnostic Performance of SF-FFR and CFD-FFR

The diagnostic performance of SF-FFR and CFD-FFR on per vessel and per patient basis were illustrated in Table 2. In comparison with invasive FFR measurement, the overall accuracy of SF-FFR for discrimination of the ischemia lesion (obtained from invasive FFR ≤ 0.80) was 0.89 which was a bit higher than CFD-FFR (0.87) on the vessel level. The performance of the two methods is similar on sensitivity, specificity, PPV and NPV. The ROC curves of the two methods were displayed in Figure 7. And there was no significant difference on AUC of the two methods on per vessel (P = 0.256) and per patient basis (P = 0.337).

Table 2. Diagnostic Performance of SF-FFR versus CFD-FFR for Demonstration of Ischemia (FFR ≤ 0.80).

Method		TP	TN	FP	FN	Sensitivity	Specificity	PPV	NPV	Accuracy
FN: false negative; FP: false positive; NPV: negative predictive value; PPV: positive predictive value; TN: true negative; TP: true positive.
Patient level	SF-FFR	34	24	2	4	0.89 (0.81,0.97)	0.92 (0.82,1.00)	0.94 (0.88,1.00)	0.86 (0.75,0.96)	0.91 (0.84,0.97)
	CFD-FFR	32	23	3	6	0.84 (0.74,0.94)	0.88 (0.77,0.97)	0.91 (0.83,0.97)	0.79 (0.66,0.92)	0.86 (0.78,0.92)
	P-value					0.731	0.995	−	−	0.552
Vessel level	SF-FFR	39	28	4	4	0.91 (0.83-0.98)	0.88 (0.78-0.97)	0.91 (0.83-0.98)	0.88 (0.78-0.97)	0.89 (0.84-0.94)
	CFD-FFR	36	29	3	7	0.84 (0.74-0.93)	0.91 (0.81-0.98)	0.92 (0.85-0.98)	0.81 (0.69-0.91)	0.87 (0.80-0.93)
	P-value					0.344	0.999	−	−	0.603

Figure 7.

The ROC curves of SF-FFR (A) and CFD-FFR (B) for diagnosis of ischemia.

ROC curves show the discrimination of ischemia by SF-FFR and CFD-FFR on the per-vessel and per-patient basis, referenced by invasive FFR. AUC: area under the curve; CFD: computational fluid dynamics; FFR: fractional flow reserve; ROC: receiver operator characteristic; SF: static first-pass.

DISCUSSION

In this study, we developed a new method (SF-FFR) based on the principle of static first-pass imaging of CTP imaging technique and evaluated its efficacy in direct comparisons with CFD-FFR and the invasive gold standard FFR.

With the improvement of CFD models and algorithms, the computing speed and diagnostic accuracy of the CFD-FFR software have been continuously improved. The results of Pearson correlation analysis and Bland-Altman analysis showed that the comparison results between CFD and gold standard were similar to those of SF-FFR. The SF-FFR calculation time was about 2 s per case while CFD calculation was about 2 min on a Nvidia Tesla V100 graphic card. The overall accuracy of SF-FFR for discrimination of the ischemia was 0.89 which was a bit higher than CFD-FFR (0.87) on the vessel level. The performance of the two methods is similar on sensitivity, specificity, PPV and NPV.

Static First pass calculation of FFR (SF-FFR) by CCTA is radically different from invasive FFR, CFD-FFR and ML-based FFR in computing principles and processes compared with previous studies. The SF-FFR method has several advantages, including: (1) In SF-FFR calculation, there is no need to use the Navier-Stokes equation that is commonly used in CFD calculations and offline model training in ML algorithms.^[11,12,33] For SF-FFR, the result was calculated voxel by voxel. There are multiple parameters and boundary conditions to set in the CFD stimulation. It is time-consumed (about 30 min to 4 h per case) mainly because of solving governing equations. The ML or DL approaches usually use CFD stimulation results as ground truth to train a model, therefore, the upper limit of the algorithm is CFD. (2) In SF-FFR calculation, the selected reference point is the aortic root rather than the coronary vessels, which ensures that the reference standard is more stable and less vulnerable to influences of location, quantity, and severity of coronary artery lesions. (3) The SF-FFR is calculated directly at the site of coronary stenosis (lesion), rather than at a specific location distal to the lesion as in invasive FFR or CFD-based FFR measurement. (4) At present, the great majority existing deep learning methods are based on CFD algorithm for data training. Therefore, its accuracy and effectiveness could not exceed CFD-based methods. However, SF-FFR based on static first pass basis, which is directly compared with invasive FFR value, can somehow exceeds CFD algorithm for its high accuracy and computing speed. Most CFD-based FFR methods require manual segmentation or substantially longer computation, or a combination of those.^[1,2,6-9] Compared with the CFD method, our SF-FFR is more feasible and user-friendly. There were no complex parameters in our calculation process. The whole calculation time was shorter than the CFD stimulation. Compared with the ML or DL method, there was no training process, so the software and hardware computational cost was relatively low. As for the diagnostic performance, compared with previous studies in which accuracy ranged from 72% to 90%, our proposed SF-FFR method outperformed the previous methods under the condition of the same amount of data.^{[15, 34]}

There are still shortcomings in this study. First, the number of selected patients was from single-center retrospective data, subsequent verification of large samples and multi-centers is needed. Besides, the patient enrollment conditions were relatively strict in this study. For example, bolus injection of contrast agent and one lesion at most per vessel were required. In this way, 27 patients 30 vessels were excluded, which wasted a lot of test samples. Also, our investigation was limited by a selected population with a high prevalence of CAD and is potentially underpowered by the limited number of positive cases included. Finally, SF-FFR was derived based on static first pass basis and the value were measured with a few assumptions. Further studies need to be performed to validate the generalization and universality of the methodology. We still work on improving the generalization of this method.

In conclusion, SF-FFR method is feasible and shows moderate consistency and correlation with the gold standard. Also, it could provide equivalent results to the CFD method, which could somehow simplify the calculation process and saving time.