Artificial Intelligence (AI) has been increasingly applied to nuclear cardiology to improve image segmentation,1 disease diagnosis, and risk prediction.2,3 When applied to image segmentation, AI can be used to automate processes which normally require tedious manual adjustments.1 In risk prediction, the major advantage of AI is the ability to objectively integrate multiple potential parameters4,5,6 or directly predict the outcome of interest from images while explaining predictions.7 Additionally, AI has been used to facilitate automated structured clinical reporting.8

AI techniques have also been applied extensively to improve image reconstruction, resulting in better image quality, reduced radiation exposure, or both.2, 3 Convolutional autoencoders, a type of deep-learning algorithm, are structured with hidden inner layers with fewer dimensions compared to the input or output layers. Convolutional autoencoders are frequently tasked with copying an image from the input to output layer, but the reduced inner dimensions force the algorithm to identify only the most critical aspect of images to carry forward. This can be applied to provide image noise reduction as a specialized post-processing technique. Ramon et al demonstrated the feasibility of denoising single photon emission computed tomography (SPECT) using stacked autoencoders designed to predict full-dose images from low-dose image reconstructions.9 In simulations, images with as little as 1/16th of the clinical dose de-noised with stacked autoencoders achieved similar quality to images with 1/8th of the dose reconstructed with standard methods.9 Other approaches have also been applied. For example, Shiri et al integrated a deep residual neural network to predict full time and projection acquisitions from either half-time acquisitions or half of the projections.10 Using 10-fold cross-validation in 363 patients, the residual network was able to generate images resulting in similar automatic quantitation of perfusion (stress total perfusion deficit) and function (volume, eccentricity, and shape index) compared to full acquisition studies.10

Another important application of AI to image reconstruction is to provide automated attenuation correction (AC). Nguyen et al developed a generative adversarial network (GAN) to simulate AC images from non-AC data with data from 491 patients for training and 112 for testing.11 The generator network, based on the UNet architecture which is a specialized convolutional autoencoder, simulated AC images while the discriminator network was tasked with differentiating the simulated images from actual AC images.11 The goal of such architecture is to make the images indistinguishable. The UNet-GAN achieved structural similarity index of 0.946 compared to true AC images and outperformed UNet alone.11 In this issue, Chen et al present a novel method to incorporate multiple imaging parameters within the reconstruction process.12 The proposed dual squeeze-and-excitation residual dense network (DuRDN) incorporates gender, body mass index (BMI), and images from 3 scatter windows together with non-AC images in order to predict AC images.12 The authors demonstrate that the predicted AC images from DuRDN more closely matched the actual AC images compared to UNet, with normalized mean square error of 2.01% vs 2.23%.12 Additionally, the images generated by DuRDN, on a segmental basis, were quantitatively similar compared to actual AC images while those generated using UNet were not.12 Lastly, the authors show that the addition of each element (incorporating gender/BMI, incorporating scatter window images, and the addition of the dual-squeeze block) incrementally improved model performance.12 Establishing the value of the components is important since each increases the computational complexity which may be a barrier to clinical implementation.

While the study by Chen et al represents a step forward in the application of AI to image reconstruction, there are a few important considerations. The study was performed in a relatively small patient population, imaged with a single camera system, with final testing performed using only 42 datasets. A larger study will be needed to evaluate this approach more robustly. Critically, while demonstrating that DuRDN-generated images are quantitatively similar to AC images is an important step, most physicians would have more interest in the diagnostic and prognostic accuracy of the approach. To date there is no evidence that simulated AC images can offer equivalent improvements in diagnostic performance of risk stratification as the actual AC images.13

Another consideration is that the anatomic information available from computed tomographic (CT) AC imaging carries significant independent diagnostic and prognostic information.14,15,16,17,18 Therefore, centers with existing SPECT/CT scanners may decide that the added diagnostic and prognostic information more than outweighs the minimal increase in radiation exposure. While synthetic CT datasets have been generated using convolutional neural networks;19 it is unlikely that information regarding coronary calcification could be obtained with this technique. Therefore, the simulated AC imaging, if indeed proven viable, is more likely to be implemented in centers without SPECT/CT systems (which is the majority of centers). Nevertheless, in spite of these considerations, the work presented by Chen et al represents a potentially promising future application of AI, although further investigations are needed to more conclusively demonstrate the feasibility of this approach.