Artificial intelligence in PET/CT is evolving from narrowly optimized systems toward reusable pretrained representations and multimodal models. PET/CT is a demanding test case because clinically meaningful interpretation must preserve quantitative molecular information while integrating anatomic CT, acquisition metadata, longitudinal examinations, and free-text imaging reports. This focused narrative review synthesizes literature reported from January 2021 through July 2026, spanning PET reconstruction and enhancement, attenuation correction, segmentation, outcome modeling, pretrained representations, vision-language systems, and agentic workflows. We distinguish task-specific models, reusable pretrained representations, promptable models, vision-language systems, generalist foundation-model candidates, and agentic systems according to demonstrated capabilities, while separately considering evaluation benchmarks. Task-specific deep learning has demonstrated promising technical performance in low-count imaging and whole-body tumor analysis, but most evidence remains retrospective and restricted by radiotracer, disease, scanner, institution, or endpoint. Emerging PET-native studies demonstrate label-efficient segmentation, image-text grounding, report generation, and modular tool orchestration; however, they do not yet establish broad cross-tracer generalization or prospective clinical benefit. We therefore propose MOSAIC-PET, an author-developed heuristic six-domain evaluation framework for PET/CT foundation models encompassing Molecular fidelity, Organ and anatomic grounding, Semantic faithfulness, Adaptability, Integration and clinical impact, and Continuous governance. MOSAIC-PET has not undergone Delphi consensus or formal validation and is intended to organize PET-specific evidence requirements rather than function as a validated clinical-readiness tool. The near-term translational priority is the development of modular, auditable systems in which validated quantitative tools and evidence-linked language outputs remain under specialist oversight, with progression from external testing to silent evaluation, assisted-use studies, clinical-impact assessment, and continuous lifecycle monitoring.
Ahmad Alenezi, Masoud Garashi, Layla Ghadenfer, Satish Panchadar, Gautam Biswas, Kawther Ali
Rubidium-82 (⁸ 2 Rb) PET is typically interpreted using perfusion category and left ventricular ejection fraction (LVEF), but absolute myocardial blood flow (MBF) and flow reserve (MFR) may add information. We investigated whether absolute stress MBF differs between patients with and without prior cardiac disease when studies appear normal, in an Arabian Gulf population with high rates of diabetes and obesity. A secondary exploratory aim compared quantitative parameters between predefined PET categories. This retrospective single-centre study included 292 source-verified patients (mean age 63.1 ± 12.1 years; 56.8% male) who underwent adenosine-stress ⁸ 2 Rb PET/CT, classified by the PET C-category system. The primary analysis compared, within normal-appearing studies (C1 perfusion with resting LVEF ≥ 50%; n = 84), those with prior cardiac disease (n = 38) and those without (n = 46). A secondary exploratory analysis compared these same 46 patients (Normal) with the remaining 246 (Abnormal). Analyses used Mann–Whitney U tests with Cohen's d and 95% confidence intervals, with Bonferroni correction. Among normal-appearing studies, patients with prior cardiac disease had lower global stress MBF (2.25 vs. 2.99 mL/min/g; d = −0.78, 95% CI −1.22 to −0.33) and lower RCA stress MBF (2.25 vs. 3.29; d = −1.02, −1.47 to −0.56); rest MBF was also lower (d = −0.47, −0.91 to −0.04), whereas MFR (d = −0.37, −0.80 to 0.07), LVEF and LV volumes did not differ. In the secondary comparison, stress MBF and MFR separated the predefined categories comparably (DeLong Z = 0.89, p = 0.372) and both more strongly than LVEF (Z = 3.54, p < 0.001); because the categories are defined in part by these parameters, this represents relative separation within the classification scheme and not diagnostic accuracy. In this Arabian Gulf cohort, patients with prior cardiac disease had lower absolute stress MBF despite normal perfusion category and ejection fraction. These exploratory findings support further prospective evaluation of absolute stress MBF as an adjunct to conventional PET interpretation. Because rest and stress flow fell together with a preserved ratio, and because haemodynamic, medication and infarct-related explanations could not be excluded, no mechanism is inferred.
Quantitative bone SPECT/CT enables objective assessment of skeletal radiotracer uptake using standardised uptake values (SUVs) and related parameters. Advances in reconstruction, correction methods, calibration and detector technology have improved measurement reliability and expanded interest in oncological and benign skeletal applications. This review evaluates current clinical applications, technical limitations and future directions of quantitative bone SPECT/CT. PubMed was searched for English-language full-text human studies published from January 2015 to March 2025 using a reproducible Boolean strategy focused on quantitative or semi-quantitative bone SPECT/CT. Screening followed PRISMA 2020 principles. Thirteen primary studies were included in the qualitative synthesis; reviews, case reports, guidelines, position papers and other supporting publications were used only for narrative context. Quantitative bone SPECT/CT showed potential value for differentiating benign from malignant bone lesions, comparison with 18F-NaF PET/CT, longitudinal response assessment and prognostic evaluation before bone-targeted radionuclide therapy. Potential applications were also identified in condylar hyperplasia, fibrous dysplasia, vertebral compression fractures and degenerative joint disease. However, most studies were retrospective, single-centre and technically heterogeneous. Reported SUV thresholds were influenced by tumour type, lesion phenotype, uptake time, acquisition/reconstruction parameters, calibration, segmentation and recovery-coefficient methods, limiting transferability. Quantitative bone SPECT/CT is a promising technique for objective assessment of bone metabolism, but current evidence supports cautious use as an adjunct to morphology and clinical context. Prospective multicentre studies with harmonised protocols, risk-of-bias control and patient-relevant endpoints are required. Artificial intelligence, radiomics, advanced detectors and theranostics should currently be considered research opportunities rather than established applications.
Mallareddy Banala, Shabbir Ezuddin, Mark Foley, Russ Kuker
Artificial intelligence (AI) has been studied across radiopharmaceutical therapy (RPT); however, evidence for treatment-changing use remains limited. We conducted a structured narrative review with descriptive mapping of learned models for patient stratification, segmentation, tumor-burden quantification, quantitative preprocessing, dosimetry, toxicity prediction, response assessment, and radiation-safety/logistics support. The mapped set contained 73 direct full journal reports and eight direct conference abstracts, with one additional meeting abstract retained as contextual evidence. Among the full reports, 17 were PSMA-related, 16 SSTR/PRRT, 15 radioiodine, 14 ⁹⁰Y radioembolization, and 11 cross-platform or emerging-target reports; 16 addressed segmentation or quantification, 32 selection, response, prognosis, toxicity, or safety/logistics, and 25 registration, preprocessing, or dosimetry. Fifty-one reports were retrospective, one was an explicitly prospective clinical/technical evaluation, nine were technical, synthetic, or phantom evaluations, and temporal design was unclear or conflicting in 12. Seven reported an external-type held-out evaluation, including four that clearly held out an institution; one reported explicit calibration, one propagated task-level uncertainty, and none evaluated a prospective AI-guided treatment policy. The mapped evidence most directly supports human-reviewed measurement and workflow assistance. An eight-domain RPT evidence-to-decision synthesis organizes quantitative fidelity, reference standards, validation, endpoints, biological linkage, oversight, uncertainty and quality assurance, and decision impact; it is not a validated score or adoption standard.
Anton Adelöw, Hamidreza Rashidykanan, Alessandro Guazzo, Massimiliano Colarieti-Tosti
Positron Emission Tomography (PET) is widely used for assessing metabolic activity and diagnosing cancer. Due to the inherently high noise levels in PET data, advanced reconstruction algorithms are essential for accurate imaging. Convolutional Neural Networks (CNNs) within the Learned Primal-Dual (LPD) framework have shown good performance for PET reconstruction; however, the locality assumption imposed by CNNs can limit their ability to capture long-range contextual dependencies in sinogram data. Motivated by our previous work on transformer-based architecture for low-dose and low-count sinogram denoising, we comprehensively investigate the integration of attention mechanisms, both self-attention and cross-attention, within the LPD framework. We propose and evaluate three novel transformer-based LPD architectures: the Dual-Domain Stacked Transformer-based LPD, which employs sinusoidal patches to model long-range dependencies in the dual domain (sinogram); the Dual-Domain Restormer-based LPD, a hybrid design that combines CNNs for fine-grained local feature extraction with transformers for global information exchange; and the Dual-Domain UNet-based LPD incorporating Cross-Attention mechanism, which augments the U-Net LPD with cross-attention to capture complementary information across image and sinogram domains. To enable robust transfer learning, we further introduce a system-aligned synthetic data generation process that replicates the MiniPET-3 system, our in-house preclinical PET scanner dedicated to small-animal imaging, incorporating realistic acquisition geometry, resolution blurring, positron range effects, and Poisson noise. Evaluations on synthetic data demonstrate that both the Stacked Transformer-based LPD and Dual-Domain Restormer-based LPD outperform the UNet-based LPD baseline in MSE and PSNR across all noise levels, with the Restormer additionally achieving higher SSIM at low noise. The UNet-based LPD incorporating a cross-attention mechanism also improves MSE and PSNR over the baseline, while exhibiting a slight reduction in SSIM, indicating a trade-off between pixel-wise accuracy and structural similarity. Trained exclusively on synthetic data, the proposed models generalize well to experimental measurements, producing high-quality reconstructions. Overall, this work highlights the promise of transformer-integrated LPD frameworks for enhancing PET reconstruction quality and robustness in low-count imaging settings.
