SPECT remains a cornerstone of nuclear medicine imaging that provides functional insights into a range of medical conditions, including cardiovascular diseases,1, 2 neurodegenerative disorders,3, 4 oncologic disease,5, 6, 7 and hepatobiliary diseases.8, 9 The current landscape of nuclear medicine is marked by a revival of SPECT, propelled by advancements in both hardware 10 and software,11 and further accelerated by the essential role that SPECT and SPECT-based dosimetry are poised to play in the rapidly evolving field of radiopharmaceutical therapy (RPT).12, 13 SPECT is widely applied in clinical practice, with well-established roles in multiple domains. The largest proportion of clinical SPECT scans are for myocardial perfusion imaging (MPI), which can detect myocardial ischemia and infarction by measuring regional blood flow and is widely used for detecting coronary artery disease (CAD), microvascular dysfunction, and transplant vasculopathy. In neurological imaging, SPECT is commonly used for dopamine transporter (DaT) imaging, aiding in the diagnosis of Parkinson’s disease by quantifying striatal DaT availability.14, 15 In oncology, prostate-specific membrane antigen (PSMA) SPECT is increasingly used for the precise quantification of the absorbed dose to guide 177Lu-PSMA therapy. Additionally, bone SPECT plays a crucial role in detecting skeletal metastases.16

Despite its broad clinical utility, however, SPECT has inherent limitations arising from practical and physical constraints. Limited spatial resolution is a continuing challenge for SPECT systems, especially when compared to positron emission tomography (PET) 17, 18. Noise levels in SPECT images are typically high as a result of the need to shorten acquisition times to reduce patient inconvenience and the need to limit tracer dose to minimize patient radiation exposure.19 Photon attenuation can degrade SPECT image quality and quantitation, and yet attenuation correction in SPECT remains a continuing challenge since the vast majority of clinical SPECT scanners are single-modality systems and cannot generate CT derived attenuation maps.20 Alongside attenuation, scatter continues to pose a significant challenge to the image quality and quantitative measurements for modern SPECT systems, especially when imaging with radioisotopes that have complex emission spectra.21 Susceptibility to motion artifacts could further impact the accuracy of functional assessment in SPECT imaging.22 Larger voxel sizes lead to partial volume effects, where tracer signals from different tissues mix, such as spill-over between the myocardium and blood pool in cardiac imaging. These challenges highlight the need for improved image generation and correction techniques for SPECT.

There are also many open research problems in the domains of SPECT image analysis and interpretation. A key step in the SPECT image analysis pipeline is image segmentation, which is required to precisely delineate anatomical and pathological structures for quantifying regional tracer uptake and detecting disease. Given that manual segmentation can be slow and is subject to inter-observer variability, there is a need for automated segmentation tools.23, 24 Notably, many partial volume correction (PVC) methods rely on anatomical priors and on accurate image segmentation. There is also an increasing interest in multimodal image fusion tools, which could strengthen SPECT’s role in integrative diagnostics by combining functional and/or structural information from PET, CT, and MRI. Examples of multimodal fusion tasks include incorporation of cross-modality information in image reconstruction, multimodality registration, and leveraging multimodal information for diagnostics.

Recent advances in artificial intelligence (AI) have introduced innovative solutions to overcome these challenges in SPECT imaging, particularly in image reconstruction, quantitative corrections, post-reconstruction image-enhancement, and automated image analysis. Early studies demonstrate the effectiveness of deep learning models, including convolutional neural networks (CNNs) and generative adversarial networks (GANs), in improving both SPECT image generation and analysis. Although many of these ideas were applied to PET earlier, the SPECT field has been rapidly catching up. Previous reviews on AI in SPECT imaging have focused on AI applications in specific areas like image enhancement,11, 25 attenuation correction,20 and specific diseases.26 However, given the current rapid pace and scale of AI-driven advancements in SPECT, a broader perspective on this topic is necessary. This review aims to revisit the role of AI in SPECT imaging using a broader lens to elucidate its full clinical scope while focusing mainly on innovations from the last two years.

Figure 1 presents a Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart, outlining the systematic process used for the selection of articles in this review, ensuring a comprehensive and unbiased inclusion of relevant studies. Furthermore, Figure 2 provides an overview of the distribution of selected articles in different AI applications in SPECT imaging. The key focus areas of this review are image reconstruction and enhancement, attenuation and other quantitative correction, reorientation and segmentation, predictive analytics, and multimodal fusion. Each subsequent section of this article provides an in-depth discussion of current methodologies, highlighting both the opportunities and challenges associated with AI implementation in these domains. The review concludes with a discussion on potential future directions, emphasizing the need for standardized datasets, model interpretability, and clinical validation to facilitate the seamless integration of AI into routine nuclear imaging practice.