Impact of deep learning image reconstructions (DLIR) on coronary artery calcium quantification

Background Deep learning image reconstructions (DLIR) have been recently introduced as an alternative to filtered back projection (FBP) and iterative reconstruction (IR) algorithms for computed tomography (CT) image reconstruction. The aim of this study was to evaluate the effect of DLIR on image quality and quantification of coronary artery calcium (CAC) in comparison to FBP. Methods One hundred patients were consecutively enrolled. Image quality–associated variables (noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR)) as well as CAC-derived parameters (Agatston score, mass, and volume) were calculated from images reconstructed by using FBP and three different strengths of DLIR (low (DLIR_L), medium (DLIR_M), and high (DLIR_H)). Patients were stratified into 4 risk categories according to the Coronary Artery Calcium - Data and Reporting System (CAC-DRS) classification: 0 Agatston score (very low risk), 1–99 Agatston score (mildly increased risk), Agatston 100–299 (moderately increased risk), and ≥ 300 Agatston score (moderately-to-severely increased risk). Results In comparison to standard FBP, increasing strength of DLIR was associated with a significant and progressive decrease of image noise (p < 0.001) alongside a significant and progressive increase of both SNR and CNR (p < 0.001). The use of incremental levels of DLIR was associated with a significant decrease of Agatston CAC score and CAC volume (p < 0.001), while mass score remained unchanged when compared to FBP (p = 0.232). The underestimation of Agatston CAC led to a CAC-DRS misclassification rate of 8%. Conclusion DLIR systematically underestimates Agatston CAC score. Therefore, DLIR should be used cautiously for cardiovascular risk assessment. Key Points • In coronary artery calcium imaging, the implementation of deep learning image reconstructions improves image quality, by decreasing the level of image noise. • Deep learning image reconstructions systematically underestimate Agatston coronary artery calcium score. • Deep learning image reconstructions should be used cautiously in clinical routine to measure Agatston coronary artery calcium score for cardiovascular risk assessment.


Introduction
Coronary artery calcium (CAC) is a well-established surrogate marker of atherosclerotic plaque burden [1]. Accordingly, an increasing body of evidence has demonstrated the incremental prognostic value of CAC for hard clinical endpoints in asymptomatic patients [2,3]. Furthermore, recent reports indicate that quantification of CAC may help identify asymptomatic patients who benefit from statin therapy [4]. As such, current guidelines on primary prevention of cardiovascular disease have endorsed CAC as a risk modifier in patients with intermediate cardiovascular diseases risk [5][6][7], hence highlighting the requirement for an accurate and precise measurement of CAC.
Several technical parameters, such as image reconstruction algorithms, type of computed tomography (CT) scanner, and analysis software, have been shown to affect CAC measurements [8,9]. Deep learning image reconstructions (DLIR) based on a convolutional neural network have been recently introduced as an alternative to filtered back projection (FBP) and iterative reconstruction (IR) algorithms for coronary CT angiography (CCTA) [10,11]. Although deviations from standard FBP reconstruction settings are discouraged for CAC imaging [12], preliminary results have shown that the implementation of DLIR is associated with superior image quality [10] while maintaining a similar texture than FBP images [13]. Since data on the impact of DLIR on CAC quantification are still scarce [14], the aim of this study was to evaluate the effect of DLIR on image quality and CAC quantification in comparison to standard FBP by using non-enhanced electrocardiogram (ECG)-triggered cardiac CT.

Population
From May 2019 to November 2019, 100 patients with suspected coronary artery disease clinically referred for CCTA were consecutively enrolled in this study. Exclusion criteria comprised previous coronary revascularization by either percutaneous coronary intervention or coronary artery by-pass graft, mechanical prosthetic valves, pacemaker, or implantable cardioverter defibrillators. The study was approved by the local ethics committee (BASEC Nr. 2020-00675), and all patients included gave written informed consent for the scientific use of their data.

CT acquisition and reconstruction
Scans were performed on a 256-slice CT scanner (Revolution CT, GE Healthcare). Since a non-contrast enhanced CT scan for CAC scoring was obtained as part of the CCTA examination, patients with heart rate ≥ 65 beats/min received up to 30 mg of metoprolol intravenously prior to the scan. The CAC scan was acquired within one heartbeat using prospective ECG triggering set at 75% of the R-R interval. The scan parameters were as follows: 120 kVp, 200 mA, 256 × 0.625 mm collimation with a z-coverage of 12-16 cm [15].
Image reconstruction and analysis for CAC imaging CAC images were reconstructed with slice thickness and increment of 2.5 mm using standard FBP and three strength levels of DLIR (low (DLIR_L), medium (DLIR_M), and high (DLIR_H)). The display field of view was set to 25 cm. For each dataset, noise, signal-to-noise ratio (SNR), and contrastto-noise ratio (CNR) were measured. Noise was defined as the standard deviation (SD) of the mean attenuation measured in the aortic root at the level of the left main ostium by placing a circular region of interest (ROI) with a diameter of 20mm (corresponding to an area of 314 mm 2 ). The SNR was calculated by dividing the mean attenuation of the aortic root at the level of the left main ostium, obtained from a circular 20-mmdiameter ROI, by its SD. Finally, CNR was derived as the difference in the mean attenuation between a calcification of the proximal left anterior descending coronary artery and the adjacent perivascular adipose tissue, both obtained from a circular ROI with a diameter of 2 mm (corresponding to an area of 3.14 mm 2 ), divided by the SD of the mean attenuation of the aortic root. The CNR was evaluated only in patients with at least one calcification in the left descending coronary artery, with a size exceeding the area of the ROI.

Statistical analysis
Statistical analysis was performed using STATA (17.0, StataCorp LLC) and R (Version 4.1.1, https://www.r-project. org). Continuous variables are presented as mean ± SD or as median (interquartile range), as appropriate, whereas categorical variables are reported as frequencies and corresponding percentages. The Friedman test was applied to determine the effect of DLIR on image quality and CACderived parameters. If a significant difference was present, the Wilcoxon signed-rank post hoc test between groups was performed. Bonferroni correction for multiple comparisons was applied. In patients with any degree of CAC detected on the FBP dataset, the difference ratios between each strength of DLIR and the FBP dataset were calculated for each CAC-derived parameter by using the following formula: (CAC-derived parameter from DLIR dataset − CAC-derived parameter from FBP dataset) * 100 / CAC-derived parameter from FBP dataset. The differences between FBP and incremental strengths of DLIR for CAC-derived parameters were also assessed graphically by using Bland-Altman analysis. All statistical analyses were two-sided, and a p < 0.05 was considered statistically significant.

Population
The final population consisted of 73 men (73%) and 27 women (27%), with a mean ± SD age of 59 ± 11years and a body mass index of 26 ± 4.8 kg/m 2 . The patient characteristics are listed in Table 1.

Discussion
The main findings of our study are as follows: (1) Compared to standard FBP, the implementation of DLIR significantly decreased image noise, thus improving overall image quality; (2) Agatston CAC score decreased progressively with increasing strength of DLIR resulting erroneously in a negative CAC score in up to 6% of the patients.
Over the last few years, thanks to technological developments, IR algorithms have been introduced for cardiac CT image reconstruction as an alternative to FBP, aiming to improve image quality, and therefore, diagnostic accuracy. Several studies have reported on the impact of IR from different vendors on CAC score quantification. A reduction of Agatston CAC score up to 48% and 39% was shown when the sinogram-affirmed IR and the highest level of advanced modeled IR (ADMIRE) were used, respectively [17,18]. In line with these findings, Gebhard et al demonstrated that the reduction of Agatston CAC score progressively increased up to 22% with increasing strengths of adaptive statistical IR (ASIR) [9]. Overall, this inaccurate estimation of CAC Agatston score has discouraged the use of IR for CAC imaging.
Recently, a DLIR algorithm has been proposed to overcome the limitations related to IR. As such, similar to the results by Wang et al [14], the implementation of DLIR in our population led to a smaller reduction of Agatston CAC score as compared to IR algorithms. This may be explained by the fact that deep learning networks employed by DLIR are trained by using FBP datasets to generate images with a similar texture, hence suppressing noise without impacting anatomical and pathological structures [13]. Nevertheless, increasing DLIR strength was associated with an increasing rate of patients being erroneously classified as having zero CAC score. This may have clinical consequences since individuals with minimal CAC score (CAC 1-10) have been reported to be at 3-fold increased risk for incident cardiovascular events in comparison to those with 0 CAC score [19]. Therefore, the detection of any CAC could be used to trigger aggressive preventive therapy, especially in young adults < 40 years [20].
Several limitations are to be acknowledged. First, this is a single-center study using a single platform for CT image reconstructions. Therefore, our results are not generalizable to artificial intelligence technologies developed by other vendors. Second, the population used for our analysis does not reflect the population usually referred for cardiovascular risk stratification. Nevertheless, the aim of our study was to evaluate the impact of DLIR on CAC scoring quantification and not on patient's prognosis.

Conclusions
Although the implementation of DLIR improves image quality mainly by reducing noise, it systematically underestimates Agatston CAC score. Therefore, DLIR should be used cautiously to assess cardiovascular risk in asymptomatic patients since it could negatively impact patient management strategies. Follow-up data are warranted to assess the impact of DLIR in clinical routine.