AI refers to any computer algorithm which performs a task normally characteristic of human intelligence.⁴ ML algorithms are an example of AI which utilize input data (typically tabular variables) to perform a task, without being explicitly programmed to understand the importance of the different features.⁵ ML algorithms are well suited to integrate the vast amount of information from nuclear cardiology studies while accounting for non-linearity, collinearity, and interactions.⁶ DL is a subset of ML, which refers to algorithms that use a multi-layered learning approach. Convolutional neural networks (CNN) are one common example in which nodes from one layer are only connected to nearby nodes in subsequent layers to maintain the spatial relationships of input images.^7-9 The model architecture (e.g., the number of layers, the number of nodes per layer, and the connectivity between nodes) influences the functionality of DL models. Another common approach is a generative adversarial network (GAN) in which a generator model (such as CNN) is trained in parallel with a discriminator network until the simulated/predicted images are indistinguishable from actual images.

All AI algorithms require data to learn patterns, which can then be applied to predict future outcomes. The process of determining the optimal method for combining parameters to predict the outcome of interest is referred to as training. Hyperparameters are variables that control the learning process, for example, the learning rate or number of iterations. Methods used to optimize the hyperparameters of the model are termed “tuning.” Testing (sometimes called validation) is the process in which an AI model is applied to a new data set to evaluate its performance. Two commonly used procedures are k-fold cross-validation and external testing. In k-fold cross-validation, the population is randomly divided into k-folds (or sets of data), with each fold sequentially held out for testing.¹⁰ Repeated testing decreases the variance in estimating model accuracy and is preferred to testing in a single random sample.¹¹ In external testing, the patients selected for training are from a different population than was used for training and tuning.¹² External testing provides an estimate of the model performance when it is applied in a new, unrelated population and may also be repeated.¹² Most importantly, the data used for training and tuning the model should not be used in any way during model testing.¹³

AI algorithms are increasingly being utilized to enhance nuclear cardiology image quality, which can then be used to reduce radiation exposure, shorten acquisition times, or both. For instance, Sun et al. employed a conditional generative adversarial network (cGAN) to enhance denoising performance and facilitate ultra-low dose SPECT imaging.¹⁴ They demonstrated that a cGAN could generate images more similarly to noise-free images, even at dose levels as low as 1/20^th, compared to conventional post-reconstruction filtering from simulated half-dose images.¹⁴ Similarly, Du et al. applied a 3D integrated attention-guided multi-frequency cGAN, which improved denoising efficiency and performance compared to conventional reconstruction across multiple noise levels.¹⁵ Image denoising is frequently applied as a post-processing step, which means that images can still be reconstructed using camera or vendor-specific algorithms.

While many studies apply denoising as a post-processing step, it may be preferable to integrate denoising into image reconstruction. Yousefzadeh et al. evaluated a GAN (with CNN generator) for denoising images during each iteration of an ordered subset expectation maximization algorithm.¹⁶ In a dataset of 247 SPECT-MPI patients, they demonstrated higher contrast-to-noise ratios for the GAN model (improved by 54.5%) compared to traditional Gaussian filtering.¹⁶ They also demonstrated that denoising after each iteration outperformed post-reconstruction denoising concerning the contrast-to-noise ratio.¹⁶ Wang et al. trained a two-component DL model to first reconstruct PET images from a sinogram and then denoise the same images.¹⁷ This dual approach had a higher structural similarity index compared to U-Net-based denoising in complex images.¹⁷ However, it is important to note that assessing improvements in task-specific outcomes, like defect identification, is essential, as denoising may boost technical quality without translating into better clinical performance.¹⁸

Nuclear cardiology studies are influenced by cardiac, respiratory, and gross patient motion. While these can be addressed with traditional techniques, there is also a potential role for AI. For instance, Zhang et al. assessed the impact of a DL network on cardiac motion correction in SPECT MPI.¹⁹ Quantitative analysis of perfusion from DL motion-corrected images showed the highest prediction performance for abnormal perfusion.¹⁹ Chang et al. developed and used an artificial neural network (ANN) to provide motion correction directly from sinograms which demonstrated a higher contrast-to-noise ratio compared to traditional methods.²⁰ DL can also estimate respiratory signals from SPECT projection data to provide respiratory motion correction.²¹ Ko et al. developed a CNN-based motion prediction model that accurately predicts respiratory motion amplitude from static MPI.²² Their model enabled the identification of images with significant motion and predicted potential artifacts affecting total perfection deficit (TPD) results.²² Builoff et al. have implemented an automatic algorithm for automated motion correction of stress and rest dynamic PET MPI acquisitions.²³ Compared to manual image correction, which took 10 minutes per case, the automatic algorithm completed motion correction in <17 seconds while achieving similar diagnostic accuracy.²³ Motion during PET scans can cause mismatches between frames, often resulting in altered images with lowered quality.²⁴ Guo et al. also developed a promising model for inter-frame motion correction in dynamic cardiac PET imaging.²⁵ Any of these methods could be used to significantly reduce the technical time required for motion correction during reconstruction without sacrificing image quality or accuracy.

Nuclear cardiology studies are increasingly acquired on hybrid camera systems, which include computed tomography (CT) for AC.²⁶ While AC imaging enhances the diagnostic accuracy of SPECT, it requires precise image registration to avoid inducing artifacts.^27,28 AI can help minimize image misregistration. For example, Ko et al. trained a CNN to predict the offset between images using 402 cases and demonstrated an average residual misalignment of 2.38±2.00mm during testing on 100 cases.²⁹ Another method is to generate synthetic AC maps (which should be perfectly co-registered with the emission images used to generate them). Canalejo et al. found that DL-based synthetic AC maps for SPECT MPI achieved high mean structural similarity (MSSIM = 0.97) and low normalized mean absolute error (NMAE = 3.08%) compared to traditional CT-based AC maps.³⁰ Chen et al. proposed a Dual-Channel-Squeeze-Fusion-Excitation co-attention module for the registration of cardiac SPECT and CT-derived attenuation maps, demonstrating lower residual misregistration compared to conventional approaches.³¹ Synthetic AC maps (generated using a GAN-based approach) have also been used to enhance the registration of coronary CT angiography with sodium fluoride PET images.³²

CT-based AC can account for soft-tissue attenuation on SPECT and PET, but it requires additional radiation exposure and imaging time.²⁶ AI could address this limitation by generating synthetic AC imaging. Yu et al. developed a DL-based method for generating synthetic AC SPECT images.³³ The synthetic AC images were visually similar to actual AC images in cases without misregistration but more similar to the reference standard in cases with misregistration.³³ Chen et al. developed a DL network (incorporating clinical information with images) to predict AC images that better matched actual images compared to a U-Net model, with a normalized mean square error of 2.01% versus 2.23%.³⁴ Nguyen et al. utilized a GAN to simulate AC images from non-AC data, achieving a structural similarity index of 0.946 compared to true AC images, outperforming U-Net alone.³⁵ Hagio et al. developed a CNN that generated synthetic AC polar maps from non-AC polar map images,³⁶ with improvement in specificity ranging from 26% (in higher quality studies) to 8% (when considering lower quality studies as well). Shanbhag et al. devised a cGAN strategy for generating synthetic short-axis AC SPECT images from non-attenuation corrected images using a training set of 5000 cases,³⁷ with representative images in Figure 1. In external testing, the area under the receiver operating characteristic curve (AUC) for obstructive CAD of synthetic AC images (0.79) was higher than non-AC images (0.72, p<0.001) and similar to actual AC imaging (0.81, p=0.196).³⁷ Synthetic AC images can also be generated using images from stand-alone SPECT camera systems, with AUC improving from 0.717 to 0.752.³⁸ This last study suggests that synthetic AC imaging could be applied broadly to improve diagnostic accuracy without additional cost or radiation exposure.

Figure 1 NC: non-attenuation corrected; AC: attenuation correction; DeepAC: synthetic attenuation corrected; sTPD: Stress total perfusion deficit

ML can incorporate a multitude of features into predictive models to improve patient selection or disease diagnosis. For instance, ML models have been developed to predict abnormal myocardial perfusion from pre-test features in patients referred for SPECT⁵⁸ or PET⁵⁹ imaging. ML has also been used to automatically identify abnormal perfusion from polar map images.⁶⁰ ML models have also been developed to identify patients with a low likelihood of obstructive CAD who could safely have their rest scan canceled.⁶¹ ML can also predict ischemia from stress-only SPECT, suggesting that it may be possible to utilize stress-only imaging in all patients.⁶² ML has also been used to identify patients with obstructive CAD from PET MPI variables, with similar diagnostic accuracy compared to expert image interpretation (89% vs 88%).⁶³ Lastly, ML can be applied to provide therapeutic decision support. For example, ML may be able to help identify cardiac resynchronization therapy super-responders from electrocardiography and gated SPECT MPI.⁶⁴

ML models typically need variables to be coded or quantified, whereas DL algorithms can process images directly to generate predictions. One such DL model utilizes polar maps as the input to predict the likelihood of obstructive CAD, demonstrating significant improvements in prediction performance compared to quantitative perfusion assessment (AUC 0.81 vs 0.78).⁸ DL can also be utilized to automate the detection of transthyretin cardiac amyloidosis.⁶⁵ In addition to directly predicting the likelihood of disease, DL models may help identify important abnormalities on SPECT or PET images. Kusumoto et al. used SPECT MPI images to develop and validate a DL-based support system to identify perfusion defects with an AUC of 0.91.⁶⁶ DL has also been utilized to identify wall motion abnormalities from SPECT MPI.⁶⁷

MPI has a critical role in estimating cardiovascular risk, an area in which AI may be particularly valuable as the number of features increases. Hu et al. incorporated clinical, stress, and imaging features to identify patients with low risk of major adverse cardiovascular events.⁶⁸ The addition of DL-based CAC scores can further improve this risk stratification.⁶⁹ Feher et al. incorporated clinical, stress test, and imaging parameters to create an AI model to predict heart failure hospitalizations in patients undergoing SPECT MPI.⁷⁰ The combined ML model displayed superior prediction performance in external validation (AUC 0.80) compared to both stress left ventricular ejection fraction (AUC 0.72) and a clinical model (AUC 0.72).⁷⁰

DL approaches have also demonstrated significant potential in predicting risks in MPI. Singh et al. developed an explainable DL model that used PET-derived relative perfusion, myocardial blood flow, and myocardial flow reserve polar maps to predict mortality. In simulated prospective testing with 1,135 patients, the DL model achieved a higher prediction accuracy (AUC 0.82) compared to myocardial flow reserve (AUC 0.70) and a comprehensive regression model (AUC 0.75).⁷¹ A similar methodology has been applied to predict death or myocardial infarction using SPECT polar maps, where the DL model outperformed both logistic regression (AUC 0.70) and stress TPD (AUC 0.65) in external testing (n = 9,019).⁷²