In recent years, with the emergence of artificial intelligence in medicine, algorithms for the reconstruction of computed tomography (CT) images based on deep learning (DLR) have been developed [1]. These algorithms feature a deep neural network (DNN) [2,3] or a convolutional neural network (CNN) [4], [5], [6] to differentiate the signal from the noise and thus reduce the noise in the image without altering its texture. These new algorithms are used in clinical routine and can provide better image quality for the same dose level or even a lower dose, whilst maintaining a diagnostic image quality similar to that of iterative reconstruction algorithms [2,3,[5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17]].

A new DLR algorithm (Advanced intelligent Clear-IQ Engine [AiCE]) that features a DNN trained with high-quality model-based iterative reconstruction patient datasets has been recently developed [7,18]. A second version of this algorithm was later developed. Since the first version, new reconstruction kernels and slice thicknesses have become available [7]. In addition, its DNN has been trained with a larger patient database, resulting in reduced noise and better spatial resolution and detectability without changing the noise texture [7].

In cardiac imaging, high-resolution imaging is of great interest for detecting lesions such as calcified or non-calcified coronary plaques [19], [20], [21], [22]. Such imaging protocols require acquisitions with low nominal slice thicknesses thus increasing image noise, which can be tackled either by an increase in the X-ray dose or with image reconstruction. AiCE, reduces noise, which is especially abundant in low radiation dose images, improves signal-to-noise ratio, and maintains noise texture, which is a limitation of iterative reconstruction [1]. The Precise IQ Engine (PIQE) is a new specially trained super-resolution DLR (SR-DLR) algorithm, dedicated to cardiac image reconstruction. The neural network of PIQE features a three-dimensional deep CNN trained using cardiac image data acquired on the commercially available ultra-high resolution CT scanner (Aquilion Precision), which in clinical practice uses 0.25 mm detector elements. The raw data coming from this CT system and used to train the neural network are reconstructed with AiCE. This new SR-DLR algorithm is currently available only for cardiac cases, on the wide-area detector Aquilion ONE / PRISM Edition CT system [23]. On this CT system, PIQE can be used with two reconstructed image thicknesses of 0.5 and 1 mm combining volume acquisitions.

Recent attempts of high-resolution imaging with spiral CT are encountering difficulties such as radiation exposure and scan heart rate variability [24,25]. Several studies have highlighted AiCE's contribution to improving the quality of cardiac CT images compared to the iterative reconstruction algorithm Adaptive Iterative Dose Reduction (AIDR 3D) [10,16,26]. These studies showed that, compared to AIDR 3D, at the same dose level, AiCE improves the image quality of coronary CT angiography images and reduces the image noise [10,16,26]. Another study has shown that AiCE improves the quality of coronary CT angiography images with a dose reduction of about 40 % compared to AIDR 3D [14]. One clinical study showing the superiority of PIQE over an iterative algorithm for coronary CT angiography has been published [27]. Two phantom studies have also compared PIQE with AiCE and two other iterative reconstruction algorithms using task-based image quality assessment [28,29]. In the first study, acquisition and reconstruction parameters different from those used for routine clinical cardiac CT were used [29]. In addition, the detectability index was calculated on the iodine insert and not on simulated lesions with clinical features. In the second study, only one of the 3 levels (the level Standard) available for AiCE and PIQE were assessed [28].

The purpose of this study was to assess the impact of the new SR-DLR algorithm, PIQE on the quality of cardiac CT images compared with the DLR algorithm AiCE using all the levels available for these two algorithms and dose levels used in clinical routine.