Artificial intelligence in coronary artery calcium score: rationale, different approaches, and outcomes

Abstract

Almost 35 years after its introduction, coronary artery calcium score (CACS) not only survived technological advances but became one of the cornerstones of contemporary cardiovascular imaging. Its simplicity and quantitative nature established it as one of the most robust approaches for atherosclerotic cardiovascular disease risk stratification in primary prevention and a powerful tool to guide therapeutic choices. Groundbreaking advances in computational models and computer power translated into a surge of artificial intelligence (AI)-based approaches directly or indirectly linked to CACS analysis. This review aims to provide essential knowledge on the AI-based techniques currently applied to CACS, setting the stage for a holistic analysis of the use of these techniques in coronary artery calcium imaging. While the focus of the review will be detailing the evidence, strengths, and limitations of end-to-end CACS algorithms in electrocardiography-gated and non-gated scans, the current role of deep-learning image reconstructions, segmentation techniques, and combined applications such as simultaneous coronary artery calcium and pulmonary nodule segmentation, will also be discussed.

Keywords: Artificial intelligence; Computed tomography; Coronary artery calcium; Coronary artery calcium score; Deep-learning; Machine learning.

Fig. 1

Deep-learning and convolutional neural network algorithm architecture. A Deep learning (DL) relies on multiple hidden layers of artificial neural networks (ANN), hence the name "deep". These layers are usually defined as "hidden" because they do not belong either to the input or output layer. The number of hidden layers determines the depth of a model. Their role is to capture patterns and features, transforming input data into other data forms usable by the subsequent layer of neurons. Indeed, each neuron contained in these layers relates to the formers. Information flows from layer to layer, moving from input to output, progressively increasing its complexity and abstractedness. B Convolutional neural networks (CNN) are the most common DL architecture used in image analysis. This architecture has two major components: convolutional and pooling layers. The former is the core building of a CNN and works by applying filters to the input data, generating an activation map. Pooling layers combine the outputs of the convolutional step, reducing the number of features extracted. These steps can be repeated multiple times. Usually, the last step is allocated to layers of artificial neural networks, which in turn generates the output. ANN artificial neural network, CNN Convolutional neural network, DL deep learning

Fig. 2

Steps required to create, validate, and commercialize an AI algorithm. AI algorithm/model creation always starts by identifying a research/clinical question, which dictates the starting dataset. Similarly, the algorithm’s architecture is selected among those performing best according to the data type. Subsequently, the starting dataset is subdivided into separate datasets of different dimensions, naming training dataset, validation dataset, and test dataset. The latter does not need to be generated from the starting dataset. Indeed, it is preferable to have an external test dataset. During the training step, the model analyzes the training dataset, deriving features that are tested against the ground truth. The identical process is performed on a separate dataset, the validation dataset. This validates the performances of the algorithm and fine-tunes it. Subsequently, the algorithm is tested on an additional separate dataset (the test dataset), and its final performances are evaluated. Valuable AI models are finally commercialized. After commercialization, the algorithms learn continuously from real-world data. Also, the model can be re-trained to overcome some flaws encountered when dealing with a real-world scenario. AI artificial intelligence

Fig. 3

Coronary artery calcium segmentation. Automatic detection, segmentation, and classification of coronary artery calcium in the left main (light green) and left anterior descending artery (purple blue) in an 80-year-old man with severe coronary artery calcification (i.e., Agatston score: 2187).

Fig. 4

Computational time taken to quantify CACS according to different technical set-ups. The computational time taken to analyze the images was extremely heterogeneous between studies, varying from a few seconds (invisible cones) to approximately ten minutes. This heterogeneity was highly dependent on the computational approach used. However, most studies reported computational times lower than that taken by experts (red ring). The latter was based on the results by Eng et al. [83] CPU central processing unit, GPU graphics processing unit, min minutes, sec seconds

Fig. 5

CAC detectability according to different image reconstruction algorithms. Coronary artery calcium detectability according to different image reconstruction algorithms in a 78-year-old hypertensive male. Two small calcifications were detectable on the filtered back projection image (dotted white arrowhead) along the course of the right coronary artery. However, the same calcifications (dotted white arrowhead) on the corresponding images, reconstructed using various deep-learning strengths, were less evident. Specifically, using the highest deep-learning reconstruction strength (DLIR-H), the margins of the bigger calcification were more blurred, while the smaller classification became barely evident. The Agatston score reduced from 691 to 688, 674, and 667, with FBP, DLIR-L, DLIR-M, and DLIR-H, respectively. CAC coronary artery calcium, DLIR deep-learning image reconstruction, FBP filtered back projections