SPECT represents the most commonly used technique for assessing myocardial perfusion known or suspected CAD. The most recent hardware makes use of solid-state cadmium zinc telluride (CZT) detectors that directly detect single emitted 80–140 keV gamma rays (photons) from the injected and redistributed radiotracer (mainly ^99mTc-sestamibi or ^99mTc-tetrofosmin) in the patient’s heart. Recently, detectors are being integrated into novel collimator designs that increase photon sensitivity in the myocardial region; scanners are being reduced in size to facilitate implementation with high-throughput centers, and acquisition/processing protocols and software optimized to approach discrete PET performance (lower acquisition time, radiation exposure with higher resolution, and potential for quantification) with promising results [4].

The pooled performance of SPECT myocardial perfusion for the evaluation of myocardial ischemia resulting from anatomically obstructive CAD as in [5] is represented in Table 1.

In terms of prognosis, the most recent long-term estimates for the occurrence of major adverse cardiovascular events (MACE) have emerged from the REgistry of Fast Myocardial Perfusion Imaging with NExt generation (REFINE) registry. In this study evaluating all-cause mortality, nonfatal myocardial infarction, unstable angina, or late coronary revascularization at roughly 5 years, a significant increase in MACE with worsening of visually assessed perfusion deficits, namely from 2.0% in normal to 7.4% in abnormal scans (while for semi-quantification through total perfusion deficit (TPD) values of 1.3% and 7.8% were respectively found for a TPD of 0% and > 10%) [6].

On the other hand, PET post-dates SPECT technology and represents the modern reference standard for quantification in nuclear imaging and thus, for the absolute quantification of myocardial blood flow. PET works through detection of photon pairs which are emitted from the injected radiotracer (⁸²Rubidium, ¹³ N-ammonia, or ¹⁵O-water). A crucial difference in this detection is that by optimizing for pairs of photons and not for single photons, which is known as coincidence, less noise is recorded and more effective and organ located emissions can be detected. In conjunction, these characteristics yield lower radiation burden, higher imaging quality, and greater versatility in clinical acquisition logistics. Recently, fully digital PET/CT scanners are being deployed and are expected to boost current performance.

The pooled performance of PET myocardial perfusion for the evaluation of myocardial ischemia resulting from anatomically obstructive CAD has been introduced in [5] and depict in Table 1.

Prognostically, absolute quantitative perfusion with PET is independently associated with a range of MACE (i.e., cardiac death, acute coronary syndromes, revascularization, heart failure, stroke, and even peripheral vascular disease) with hazard ratios in the spectrum of 1.19–2.93] [7, 8].

Complementarily, an important proportion of nuclear studies is currently performed in conjunction with CT. In its simplest form, this provides a low-dose CT scan that serves as topographic base for the necessary corrections in the acquisition/processing of SPECT and PET, especially at the quantitative standard. Alternatively, a full CCTA can be performed prior to the nuclear acquisition either as gate-keeper (selective hybrid) for the nuclear study or as standard support in the same diagnostic step (standard hybrid). In low-dose images, apparent low image quality hinders diagnostic statements for obstructive CAD. However, modern approaches through machine learning (vide infra) open the possibility to extract sensitive and possibly diagnostically useful information from such coarse images. On the other hand, high-quality contrast images strongly supplement diagnostic considerations (with high sensitivity and excellent negative predictive value [5]) and also present a substrate for novel analytics [9].

In its simplest form, artificial neural networks consist of an input (structured, imaging, sound data), at least one intermediate, and an output layer (pseudoprobability of a category, regression or prediction). Additional intermediate layers are called hidden layers, and the more layers, the deeper the network becomes (as such now colloquially known as deep learning). During the last decade, the convolutional operator in the intermediate layers has become the gold standard in image segmentation and classification. Deep learning elucidates complex non-linear and high-dimensional patterns better than other machine learning strategies. Consequently, it has become the state-of-the-art for image (computer vision) and language (natural language processing) recognition tasks [10]. In essence, convolutions allow to stepwise flatten a 2- or 3-dimensional image to an array concerning its most relevant patterns. For example, before deep learning existed, convolutions were used for edge detection in image analysis. In training, neural networks are typically trained by in multiple epochs (whole runs of training data through a network) that allow the network to adjust its weights (coefficients) and select features that minimize a predefined loss-function (the penalties the network receives for incorrect predictions). This method has shown excellent performance, allowing cardiovascular researchers to deploy it in several forms using myocardial perfusion imaging data [11, 12]. Since training hyperparameters of these huge models from scratch could require massive amounts of data and computer resources, the possibility to exploit common characteristics of data (for example images) in pre-trained models for tasks that are broadly similar (picture classification vs. medical image classification) allows rapid adjustment of earlier networks to specific new tasks. The process of training a network on other similar data and then further fine-tuning it on actual data of interest is called transfer learning.

In the current standard for segmentation, a particularly successful image segmentation architecture is the “U-net.” This network exploits the information obtained during stepwise flattening (contracting path) in a stepwise up-sampling of the convolutional layers (expansive path) in order to propagate context information from the original image into the segmentation output image (https://​arxiv.​org/​abs/​1505.​04597).

In order to safeguard ensure generalizability (i.e., the capacity to apply a model to new data and achieve comparable performance, which suggests that the learned features are relevant and universal for the task at hand) in clinical practice, training on broad real-world data is essential. Presently, a variety of hardware vendors, tracers, acquisition protocols, and alignment methods are used in cardiac nuclear imaging. Yet, most datasets available for training have originated from trails or study protocols with a high degree of homogeneity. These datasets therefore might not reflect the full data variety in imaging and may prove biased towards images of certain quality standards. A timely argument around this underlying risk comes from the recent work by DeGrave et al. [23]. They showed that many algorithms trained to classify COVID-19 from chest X-rays did not in fact learn disease patterns within the image. Instead, they learned to recognize the difference in acquisition position between the COVID-19 and non-COVID cases in the training data as chest films were mostly made in a seated position in the COVID-19 cases at the emergency room instead of in conventional acquisition positions.

Careful organization of training-dataset to reflect variations in clinical practice is therefore essential. Besides this database curation, image pre-processing techniques can be used to simulate variation in image-acquisition (such as adding noise or generating variations in pixel-value distributions of the image data) [24].

In myocardial perfusion evaluation, analysts consciously and unconsciously perform quality control on both the acquired data and analysis output. Quality of the images, scatter, and alignment between the tracer signal and myocardial segmentation are all scrutinized. In automated analysis using machine learning-based AI, similar quality control (QC) is paramount. In consequence, attention has been recently focused on implementing QC steps into AI analysis frameworks. Image quality is a topic affecting all areas of medical imaging. With respect to cardiac imaging, image QC steps have been implemented for cardiac MRI analysis [25, 26] as well as echocardiography [27] and brain PET imaging [28]. For PET/CT specifically, registration steps proposed for alignment of PET and CT images [29] could also be exploited to provide a QC step for (mis)alignment.

Output data can also be scrutinized against known clinical priors. For example, ventricular shape/morphology [30, 31] and myocardial and ventricular volume dynamics over the cardiac cycle can be assessed for physiological conceivability [26]. Such steps aid to flag potential weaknesses or errors in the analysis for clinician review. Moreover, they could also be used to inform further training of the algorithm using newly analyzed data to boost its performance [32].

The features that machine learning-based AI uses in decision-making are self-taught and can be made up out of a complex set of neural activations (in deep learning) that are not readily translated to concepts recognizable to humans. However, understanding the motivation/decision process in algorithms may provide better insight in their clinical validity. Interpretability could additionally support discovery of previously unknown associations and concepts within the data that determine outcome. Although it is impossible to fully comprehend the algorithm’s decision process, several methods have been developed to gain insight in an algorithm’s decisions.

Attention/activation maps are illustrations that demonstrate what part of the image contributed mostly to the decision. In myocardial perfusion imaging through SPECT, such maps can be used to correlate the classification of perfusion defects to expected coronary anatomy [16••]. Another approach is to embed known clinical concepts into deep learning models. In previous study, we have proposed a method that ensures clinical concepts (in this case ejection fraction or septal flash) are encoded into the latent space of a variational auto-encoder, a dimensionality reduction algorithm frequently used for classification [33]. This approach allows to solve the classification process while simultaneously interrogating the importance of known clinical features in the decision process.

In its traditional approach, the weight of a deep learning model is assigned during training as point estimates (single values are assigned for each weight). Predictions made by the model are therefore fixed maximum likelihood predictions, and do not include information about the certainty or uncertainty of the output. Several methods have been proposed to incorporate confidence estimates into DL algorithms. Monte Carlo drop-out is an approach in which random weights of a conventional neural network are temporarily disabled at test-time. This results in variations in the model predictions that can be used to classify the confidence (i.e., robustness) of these predictions. Additionally, novel Bayesian neural networks bring the possibility to quantify the uncertainty of the network allowing more comprehensive output interpretation by physicians for considering or not model’s prediction in their clinical judgment [34]. Such networks use distributions that allow the models to generalize better over new datasets. As randomness is added to the weights during training, overfitting of the model to the training-data is reduced. This makes them suitable for applications in cardiac imaging, where variations in image-data (vendors and institutional differences in acquisition protocols) are common.